NLOGIT

Version 6
Reference Guide

by

William H. Greene
Econometric Software, Inc.
© 1986 - 2016 Econometric Software, Inc. All rights reserved.

This software product, including both the program code and the accompanying
documentation, is copyrighted by, and all rights are reserved by Econometric Software, Inc. No part
of this product, either the software or the documentation, may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means without prior written permission of Econometric
Software, Inc.

LIMDEP® and NLOGIT® are registered trademarks of Econometric Software, Inc. All other
brand and product names are trademarks or registered trademarks of their respective companies.

Econometric Software, Inc.

15 Gloria Place
Plainview, NY 11803
USA
Tel: +1 516-938-5254
Fax: +1 516-938-2441
Email: [email protected]
Websites: www.limdep.com and www.nlogit.com.

Econometric Software, Australia

6/16 Carr Street
Waverton NSW 2060
Australia
Tel: +61 (0) 418-433-057
Email: [email protected]
 End-User License Agreement
This is a contract between you and Econometric Software, Inc. The software product refers
to the computer software and documentation as well as any upgrades, modified versions, copies or
supplements supplied by Econometric Software. By installing, downloading, accessing or otherwise
using the software product, you agree to be bound by the terms and conditions of this Agreement.
Subject to the terms and conditions of this Agreement, Econometric Software, Inc. grants
you a non-assignable, non-transferable license, without the right to sublicense, to use the licensed
software and documentation in object-code form only, solely for your internal business, research, or
educational purposes.

Copyright, Trademark, and Intellectual Property

This software product is copyrighted by, and all rights are reserved by Econometric Software,
Inc. No part of this software product, either the software or the documentation, may be reproduced,
distributed, downloaded, stored in a retrieval system, transmitted in any form or by any means, sold or
transferred without prior written permission of Econometric Software. You may not, or permit any
person, to: (i) modify, adapt, translate, or change the software product; (ii) reverse engineer, decompile,
disassemble, or otherwise attempt to discover the source code of the software product; (iii) sublicense,
resell, rent, lease, distribute, commercialize, or otherwise transfer rights or usage to the software
product; (iv) remove, modify, or obscure any copyright, registered trademark, or other proprietary
notices; (v) embed the software product in any third-party applications; or (vi) make the software
product, either the software or the documentation, available on any website.
LIMDEP® and NLOGIT® are registered trademarks of Econometric Software, Inc. The
software product is licensed, not sold. Your possession, installation and use of the software product
does not transfer to you any title and intellectual property rights, nor does this license grant you any
rights in connection with software product registered trademarks.

Use of the Software Product

You have only the non-exclusive right to use this software product. A single user license is
registered to one specific individual as the sole authorized user, and is not for multiple users on one
machine or for installation on a network, in a computer laboratory or on a public access computer.
For a single user license only, the registered user may install the software on a primary stand alone
computer and one home or portable secondary computer for his or her exclusive use. However, the
software may not be used on the primary computer by another person while the secondary computer
is in use. For a multi-user site license, the specific terms of the site license agreement apply for scope
of use and installation.
Limited Warranty

Econometric Software warrants that the software product will perform substantially in
accordance with the documentation for a period of ninety (90) days from the date of the original
purchase. To make a warranty claim, you must notify Econometric Software in writing within ninety
(90) days from the date of the original purchase and return the defective software to Econometric
Software. If the software does not perform substantially in accordance with the documentation, the
entire liability and your exclusive remedy shall be limited to, at Econometric Software’s option, the
replacement of the software product or refund of the license fee paid to Econometric Software for the
software product. Proof of purchase from an authorized source is required. This limited warranty is
void if failure of the software product has resulted from accident, abuse, or misapplication. Some states
and jurisdictions do not allow limitations on the duration of an implied warranty, so the above
limitation may not apply to you. To the extent permissible, any implied warranties on the software
product are limited to ninety (90) days.
Econometric Software does not warrant the performance or results you may obtain by using
the software product. To the maximum extent permitted by applicable law, Econometric Software
disclaims all other warranties and conditions, either expressed or implied, including, but not limited
to, implied warranties of merchantability, fitness for a particular purpose, title, and non-infringement
with respect to the software product. This limited warranty gives you specific legal rights. You may
have others, which vary from state to state and jurisdiction to jurisdiction.

Limitation of Liability

Under no circumstances will Econometric Software be liable to you or any other person for
any indirect, special, incidental, or consequential damages whatsoever (including, without limitation,
damages for loss of business profits, business interruption, computer failure or malfunction, loss of
business information, or any other pecuniary loss) arising out of the use or inability to use the
software product, even if Econometric Software has been advised of the possibility of such damages.
In any case, Econometric Software’s entire liability under any provision of this agreement shall not
exceed the amount paid to Econometric Software for the software product. Some states or
jurisdictions do not allow the exclusion or limitation of liability for incidental or consequential
damages, so the above limitation may not apply to you.
 Preface
NLOGIT is used for estimation of discrete multinomial choice models. The program is a
superset of LIMDEP Version 11; NLOGIT 6 is LIMDEP 11 plus the NLOGIT command (and
numerous variants) which invokes the multinomial choice estimators. The centerpiece of the
estimation and analysis package is the multinomial logit model. The many variations include
multinomial probit, latent class logit, and, most importantly, the many forms of the mixed (or,
random parameters) logit model, including the most recently developed formulation, the generalized
mixed logit model. No other program supports as wide a variety of multinomial choice model
estimators and post estimation analysis tools.
NLOGIT has pioneered many of the methods described in this manual, such as several forms
of the mixed logit model, the attribute nonattendance model and estimation of mixed models in WTP
space. This version continues the ongoing collaboration of William Greene (Econometric Software,
Inc.) and David Hensher (Econometric Software, Australia.) Recent developments, especially the
random parameters and generalized mixed logit in its cross section and panel data variants have also
benefited from the enthusiastic collaboration of John Rose. We note, the recent practitioner’s guide,
Applied Choice Analysis, 2nd Edition (Hensher, D., Rose, J. and Greene, W., Cambridge University
Press, 2015). This is a wide ranging introduction to discrete choice modeling that contains numerous
applications developed with NLOGIT. This book should provide a useful companion to the
documentation for NLOGIT.

Econometric Software, Inc.

Plainview, NY 11803
2016
 NLOGIT 6 Table of Contents vi

Table of Contents
Table of Contents....................................................................................................................vi

What’s New in Version 6? ................................................................................................... N-1

WN1 New Multinomial Choice Models ..................................................................................... N-1
WN1.1 Fixed Effects in Multinomial Logit Models ..................................................... N-1
WN1.2 Random Effects Multinomial Logit Models ..................................................... N-2
WN1.3 Random Regret Logit Model ............................................................................ N-2
WN1.4 Sequential Logit Model..................................................................................... N-2
WN1.5 Best/Worst Outcome Data ................................................................................ N-3
WN1.6 Berry, Levinsohn and Pakes Random Parameters Logit Model ....................... N-3
WN.2 Model Extensions............................................................................................................. N-4
WN2.1 Willingness to Pay ............................................................................................ N-4
WN2.2 Model Simulation ............................................................................................. N-4
WN2.3 Estimated Elasticities and Partial Effects.......................................................... N-5
WN2.4 Robust Covariance Matrix ................................................................................ N-5
WN2.5 Random Number Generator .............................................................................. N-5
WN2.6 Posterior Estimates from Latent Class Models ................................................. N-6
WN2.7 Coefficients in RP Models ................................................................................ N-6
WN2.8 Simplified WALD Command ........................................................................... N-7

N1: Introduction to NLOGIT Version 6................................................................................ N-8

N1.1 Introduction ....................................................................................................................... N-8
N1.2 The NLOGIT Program ....................................................................................................... N-8
N1.3 NLOGIT and LIMDEP Integration and Documentation .................................................... N-8
N1.4 Discrete Choice Modeling with NLOGIT .......................................................................... N-9
N1.5 Types of Discrete Choice Models in NLOGIT ................................................................ N-10
N1.5.1 Random Regret Logit Model ........................................................................... N-12
N1.5.2 Scaled Multinomial Logit Model..................................................................... N-13
N1.5.3 Latent Class and Random Parameters LC Models .......................................... N-13
N1.5.4 Heteroscedastic Extreme Value Model............................................................ N-13
N1.5.5 Multinomial Probit Model ............................................................................... N-13
N1.5.6 Nested Logit Models ....................................................................................... N-14
N1.5.7 Random Parameters and Nonlinear RP Logit Model ...................................... N-14
N1.5.8 Error Components Logit and Fixed Effects Models ........................................ N-15
N1.5.9 Generalized Mixed Logit Model ..................................................................... N-15
N1.6 Functions of NLOGIT ...................................................................................................... N-16

N2: Discrete Choice Models .............................................................................................. N-17

N2.1 Introduction ..................................................................................................................... N-17
N2.2 Random Utility Models ................................................................................................... N-17
N2.3 Binary Choice Models ..................................................................................................... N-18
N2.4 Bivariate and Multivariate Binary Choice Models .......................................................... N-20
N2.5 Ordered Choice Models ................................................................................................... N-21
N2.6 Multinomial Logit Model ................................................................................................ N-24
 NLOGIT 6 Table of Contents vii

N2.6.1 Random Effects and Common (True) Random Effects ................................... N-25
N2.6.2 A Dynamic Multinomial Logit Model ............................................................. N-26
N2.7 Conditional Logit Model ................................................................................................. N-26
N2.7.1 Fixed Effects ................................................................................................... N-27
N2.7.2 Random Regret Logit and Hybrid Utility Models ........................................... N-27
N2.7.3 Scaled MNL Model ......................................................................................... N-28
N2.8 Error Components Logit Model....................................................................................... N-29
N2.9 Heteroscedastic Extreme Value Model............................................................................ N-30
N2.10 Nested and Generalized Nested Logit Models .............................................................. N-30
N2.10.1 Alternative Normalizations of the Nested Logit Model ................................ N-32
N2.10.2 A Model of Covariance Heterogeneity .......................................................... N-34
N2.10.3 Generalized Nested Logit Model ................................................................... N-34
N2.10.4 Box-Cox Nested Logit ................................................................................... N-35
N2.11 Random Parameters Logit Models ................................................................................ N-35
N2.11.1 Nonlinear Utility RP Model........................................................................... N-37
N2.11.2 Generalized Mixed Logit Model ................................................................... N-37
N2.12 Latent Class Logit Models ............................................................................................. N-38
N2.12.1 2K Latent Class Model for Attribute Nonattendance ..................................... N-39
N2.12.2 Latent Class – Random Parameters Model .................................................... N-39
N2.13 Multinomial Probit Model ............................................................................................. N-39

N3: Model and Command Summary for Discrete Choice Models .................................. N-41
N3.1 Introduction ..................................................................................................................... N-41
N3.2 Model Dimensions ........................................................................................................... N-41
N3.3 Basic Discrete Choice Models ......................................................................................... N-42
N3.3.1 Binary Choice Models ..................................................................................... N-42
N3.3.2 Bivariate Binary Choices ................................................................................. N-42
N3.3.3 Multivariate Binary Choice Models ................................................................ N-42
N3.3.4 Ordered Choice Models ................................................................................... N-43
N3.4 Multinomial Logit Models............................................................................................... N-43
N3.4.1 Multinomial Logit............................................................................................ N-43
N3.4.2 Conditional Logit ............................................................................................. N-44
N3.5 NLOGIT Extensions of Conditional Logit ...................................................................... N-44
N3.5.1 Random Regret Logit ...................................................................................... N-44
N3.5.2 Scaled Multinomial Logit ................................................................................ N-45
N3.5.3 Heteroscedastic Extreme Value ....................................................................... N-45
N3.5.4 Error Components Logit and Fixed Effects ..................................................... N-45
N3.5.5 Nested and Generalized Nested Logit ............................................................. N-46
N3.5.6 Random Parameters Logit ............................................................................... N-46
N3.5.7 Generalized Mixed Logit ................................................................................. N-47
N3.5.8 Nonlinear Random Parameters Logit .............................................................. N-48
N3.5.9 Latent Class Logit ............................................................................................ N-48
N3.5.10 2K Latent Class Logit ..................................................................................... N-49
N3.5.11 Latent Class Random Parameters .................................................................. N-49
N3.5.12 Multinomial Probit......................................................................................... N-49
N3.6 Command Summary ........................................................................................................ N-50
N3.7 Subcommand Summary ................................................................................................... N-51
 NLOGIT 6 Table of Contents viii

N4: Data for Binary and Ordered Choice Models............................................................. N-55

N4.1 Introduction ..................................................................................................................... N-55
N4.2 Grouped and Individual Data for Discrete Choice Models ............................................. N-55
N4.3 Data Used in Estimation of Binary Choice Models ......................................................... N-56
N4.3.1 The Dependent Variable .................................................................................. N-56
N4.3.2 Problems with the Independent Variables ....................................................... N-56
N4.3.3 Dummy Variables with Empty Cells ............................................................... N-59
N4.3.4 Missing Values ................................................................................................ N-62
N4.4 Bivariate Binary Choice .................................................................................................. N-63
N4.5 Ordered Choice Model Structure and Data...................................................................... N-63
N4.5.1 Empty Cells ..................................................................................................... N-63
N4.5.2 Coding the Dependent Variable....................................................................... N-64
N4.6 Constant Terms ................................................................................................................ N-64

N5: Models for Binary Choice ........................................................................................... N-65

N5.1 Introduction ..................................................................................................................... N-65
N5.2 Modeling Binary Choices ................................................................................................ N-65
N5.2.1 Underlying Processes....................................................................................... N-65
N5.2.2 Modeling Approaches...................................................................................... N-66
N5.2.3 The Linear Probability Model ......................................................................... N-67
N5.3 Grouped and Individual Data for Binary Choice Models .................................................. N-68
N5.4 Variance Normalization ................................................................................................... N-68
N5.5 The Constant Term in Index Function Models ................................................................ N-69

N6: Probit and Logit Models: Estimation ......................................................................... N-70

N6.1 Introduction ..................................................................................................................... N-70
N6.2 Probit and Logit Models for Binary Choice .................................................................... N-70
N6.3 Commands ....................................................................................................................... N-70
N6.4 Output .............................................................................................................................. N-71
N6.4.1 Reported Estimates .......................................................................................... N-71
N6.4.2 Fit Measures .................................................................................................... N-73
N6.4.3 Covariance Matrix ........................................................................................... N-75
N6.4.4 Retained Results and Generalized Residuals ................................................... N-76
N6.5 Robust Covariance Matrix Estimation............................................................................. N-77
N6.5.1 The Sandwich Estimator .................................................................................. N-77
N6.5.2 Clustering......................................................................................................... N-77
N6.5.3 Stratification and Clustering ............................................................................ N-80
N6.6 Analysis of Partial Effects ............................................................................................... N-81
N6.6.1 The Krinsky and Robb Method ....................................................................... N-82
N6.7 Simulation and Analysis of a Binary Choice Model ....................................................... N-87
N6.8 Using Weights and Choice Based Sampling ................................................................... N-89
N6.9 Heteroscedasticity in Probit and Logit Models................................................................ N-91

N7: Tests and Restrictions in Models for Binary Choice ................................................ N-97
N7.1 Introduction ..................................................................................................................... N-97
N7.2 Testing Hypotheses.......................................................................................................... N-97
N7.2.1 Wald Tests ....................................................................................................... N-97
 NLOGIT 6 Table of Contents ix

N7.2.2 Likelihood Ratio Tests..................................................................................... N-99

N7.2.3 Lagrange Multiplier Tests.............................................................................. N-102
N7.3 Two Specification Tests ................................................................................................ N-104
N7.3.1 A Test for Nonnested Probit Models ............................................................. N-104
N7.3.2 A Test for Normality in the Probit Model ..................................................... N-105
N7.4 The WALD Command .................................................................................................. N-106
N7.5 Imposing Linear Restrictions ......................................................................................... N-107

N8: Extended Binary Choice Models .............................................................................. N-109

N8.1 Introduction ................................................................................................................... N-109
N8.2 Sample Selection in Probit and Logit Models ............................................................... N-109
N8.3 Endogenous Variable in a Probit Model ........................................................................ N-110

N9: Fixed and Random Effects Models for Binary Choice ............................................ N-113
N9.1 Introduction ................................................................................................................... N-113
N9.2 Commands ..................................................................................................................... N-114
N9.3 Clustering, Stratification and Robust Covariance Matrices........................................... N-115
N9.4 One and Two Way Fixed Effects Models...................................................................... N-117
N9.5 Conditional MLE of the Fixed Effects Logit Model ..................................................... N-123
N9.5.1 Command....................................................................................................... N-124
N9.5.2 Application .................................................................................................... N-125
N9.5.3 Estimating the Individual Constant Terms .................................................... N-127
N9.5.4 A Hausman Test for Fixed Effects in the Logit Model ................................. N-128
N9.6 Random Effects Models for Binary Choice................................................................... N-129

N10: Random Parameter Models for Binary Choice ...................................................... N-136

N10.1 Introduction ................................................................................................................. N-136
N10.2 Probit and Logit Models with Random Parameters ..................................................... N-137
N10.2.1 Command for the Random Parameters Models ........................................... N-137
N10.2.2 Results from the Estimator and Applications .............................................. N-139
N10.2.3 Controlling the Simulation .......................................................................... N-146
N10.2.4 The Parameter Vector and Starting Values.................................................. N-147
N10.2.5 A Dynamic Probit Model............................................................................. N-148
N10.3 Latent Class Models for Binary Choice....................................................................... N-150
N10.3.1 Application .................................................................................................. N-151

N11: Semiparametric and Nonparametric Models for Binary Choice........................... N-158

N11.1 Introduction ................................................................................................................. N-158
N11.2 Maximum Score Estimation - MSCORE..................................................................... N-159
N11.2.1 Command for MSCORE.............................................................................. N-160
N11.2.2 Options Specific to the Maximum Score Estimator .................................... N-160
N11.2.3 General Options for MSCORE .................................................................... N-162
N11.2.4 Output from MSCORE ................................................................................ N-163
N11.3 Klein and Spady’s Semiparametric Binary Choice Model ............................................ N-164
N11.3.1 Command..................................................................................................... N-165
N11.3.2 Output .......................................................................................................... N-165
N11.3.3 Application .................................................................................................. N-166
 NLOGIT 6 Table of Contents x

N11.4 Nonparametric Binary Choice Model .......................................................................... N-168

N11.4.1 Output from NPREG ................................................................................... N-170
N11.4.2 Application .................................................................................................. N-170

N12: Bivariate and Multivariate Probit and Partial Observability Models .................... N-173
N12.1 Introduction ................................................................................................................. N-173
N12.2 Estimating the Bivariate Probit Model ........................................................................ N-174
N12.2.1 Options for the Bivariate Probit Model ....................................................... N-174
N12.2.2 Proportions Data .......................................................................................... N-176
N12.2.3 Heteroscedasticity ........................................................................................ N-177
N12.2.4 Specification Tests ....................................................................................... N-177
N12.2.5 Model Results for the Bivariate Probit Model ............................................. N-179
N12.2.6 Partial Effects .............................................................................................. N-180
N12.3 Tetrachoric Correlation ................................................................................................ N-186
N12.4 Bivariate Probit Model with Sample Selection............................................................ N-188
N12.5 Simultaneity in the Binary Variables ........................................................................... N-188
N12.6 Recursive Bivariate Probit Model................................................................................ N-189
N12.7 Panel Data Bivariate Probit Models............................................................................. N-191
N12.8 Simulation and Partial Effects ..................................................................................... N-197
N12.9 Multivariate Probit Model ........................................................................................... N-199
N12.9.1 Retrievable Results ...................................................................................... N-200
N12.9.2 Partial Effects .............................................................................................. N-200
N12.9.3 Sample Selection Model .............................................................................. N-201

N13: Ordered Choice Models .......................................................................................... N-203

N13.1 Introduction ................................................................................................................. N-203
N13.2 Command for Ordered Probability Models ................................................................. N-204
N13.3 Data Problems.............................................................................................................. N-205
N13.4 Output from the Ordered Probability Estimators ......................................................... N-205
N13.4.1 Robust Covariance Matrix Estimation......................................................... N-208
N13.4.2 Saved Results ............................................................................................... N-209
N13.5 Partial Effects and Simulations .................................................................................... N-210

N14: Extended Ordered Choice Models ......................................................................... N-215

N14.1 Introduction ................................................................................................................. N-215
N14.2 Weighting and Heteroscedasticity ............................................................................... N-215
N14.3 Multiplicative Heteroscedasticity ................................................................................ N-216
N14.3.1 Testing for Heteroscedasticity ..................................................................... N-217
N14.3.2 Partial Effects in the Heteroscedasticity Model ........................................... N-221
N14.4 Sample Selection and Treatment Effects ..................................................................... N-223
N14.4.1 Command..................................................................................................... N-224
N14.4.2 Saved Results ............................................................................................... N-224
N14.4.3 Applications ................................................................................................. N-225
N14.5 Hierarchical Ordered Probit Models ............................................................................ N-229
N14.6 Zero Inflated Ordered Probit (ZIOP, ZIHOP) Models ................................................ N-232
N14.7 Bivariate Ordered Probit and Polychoric Correlation.................................................. N-234
 NLOGIT 6 Table of Contents xi

N15: Panel Data Models for Ordered Choice ................................................................. N-239

N15.1 Introduction ................................................................................................................. N-239
N15.2 Fixed Effects Ordered Choice Models......................................................................... N-240
N15.3 Random Effects Ordered Choice Models .................................................................... N-243
N15.3.1 Commands ................................................................................................... N-244
N15.3.2 Output and Results....................................................................................... N-245
N15.3.3 Application .................................................................................................. N-246
N15.4 Random Parameters and Random Thresholds Ordered Choice Models ...................... N-248
N15.4.1 Model Commands........................................................................................ N-249
N15.4.2 Results ......................................................................................................... N-252
N15.4.3 Application .................................................................................................. N-252
N15.4.4 Random Parameters HOPIT Model ............................................................. N-256
N15.5 Latent Class Ordered Choice Models .......................................................................... N-262
N15.5.1 Command..................................................................................................... N-262
N15.5.2 Results ......................................................................................................... N-263

N16: The Multinomial Logit Model .................................................................................. N-272

N16.1 Introduction ................................................................................................................. N-272
N16.2 The Multinomial Logit Model – MLOGIT.................................................................. N-273
N16.3 Model Command for the Multinomial Logit Model .................................................... N-274
N16.3.1 Imposing Constraints on Parameters ........................................................... N-274
N16.3.2 Starting Values ............................................................................................ N-275
N16.4 Robust Covariance Matrix ........................................................................................... N-275
N16.5 Cluster Correction........................................................................................................ N-276
N16.6 Choice Based Sampling ............................................................................................... N-277
N16.7 Output for the Logit Models ........................................................................................ N-278
N16.8 Partial Effects............................................................................................................... N-281
N16.8.1 Computation of Partial Effects with the Model ........................................... N-282
N16.8.2 Partial Effects Using the PARTIALS EFFECTS Command ....................... N-285
N16.9 Predicted Probabilities ................................................................................................. N-286
N16.10 Generalized Maximum Entropy (GME) Estimation .................................................. N-287
N16.11 Technical Details on Optimization ............................................................................ N-289
N16.12 Sequential Logit Model ............................................................................................. N-290
N16.13 Panel Data Multinomial Logit Models ...................................................................... N-292
N16.13.1 Random Effects and Common (True) Random Effects ............................. N-292
N16.13.2 Dynamic Multinomial Logit Model........................................................... N-298

N17: Conditional Logit Model.......................................................................................... N-300

N17.1 Introduction ................................................................................................................. N-300
N17.2 The Conditional Logit Model – CLOGIT.................................................................... N-301
N17.3 Clogit Data for the Applications .................................................................................. N-302
N17.3.1 Setting Up the Data...................................................................................... N-304
N17.4 Command for the Discrete Choice Model ................................................................... N-305
N17.5 Results for the Conditional Logit Model ..................................................................... N-307
N17.5.1 Robust Standard Errors ................................................................................ N-310
N17.5.2 Descriptive Statistics ................................................................................... N-311
N17.6 Estimating and Fixing Coefficients ............................................................................. N-313
 NLOGIT 6 Table of Contents xii

N17.7 Generalized Maximum Entropy Estimator .................................................................. N-315

N17.8 MLOGIT and CLOGIT ............................................................................................... N-317

N18: Data Setup for NLOGIT ........................................................................................... N-319

N18.1 Introduction ................................................................................................................. N-319
N18.2 Basic Data Setup for NLOGIT ..................................................................................... N-319
N18.3 Types of Data on the Choice Variable ......................................................................... N-320
N18.3.1 Unlabeled Choice Sets ................................................................................. N-322
N18.3.2 Simulated Choice Data ................................................................................ N-323
N18.3.3 Checking Data Validity ............................................................................... N-323
N18.4 Weighting .................................................................................................................... N-324
N18.5 Choice Based Sampling ............................................................................................... N-325
N18.6 Entering Data on a Single Line .................................................................................... N-327
N18.7 Converting One Line Data Sets for NLOGIT .............................................................. N-329
N18.7.1 Converting the Data Set to Multiple Line Format ....................................... N-330
N18.7.2 Writing a Multiple Line Data File for NLOGIT .......................................... N-333
N18.8 Merging Invariant Variables into a Panel .................................................................... N-333
N18.9 Modeling Choice Strategy ........................................................................................... N-335
N18.10 Scaling the Data ......................................................................................................... N-336
N18.11 Data for the Applications ........................................................................................... N-337
N18.12 Merging Revealed Preference (RP) and Stated Preference (SP) Data Sets ............... N-339

N19: NLOGIT Commands and Results ........................................................................... N-340

N19.1 Introduction ................................................................................................................. N-340
N19.2 NLOGIT Commands .................................................................................................... N-340
N19.3 Other Optional Specifications on NLOGIT Commands .............................................. N-344
N19.4 Estimation Results ....................................................................................................... N-345
N19.4.1 Descriptive Headers for NLOGIT Models ................................................... N-345
N19.4.2 Standard Model Results ............................................................................... N-346
N19.4.3 Retained Results .......................................................................................... N-349
N19.4.4 Descriptive Statistics for Alternatives ......................................................... N-350
N19.5 Calibrating a Model ..................................................................................................... N-352

N20: Choice Sets and Utility Functions.......................................................................... N-354

N20.1 Introduction ................................................................................................................. N-354
N20.2 Choice Sets .................................................................................................................. N-354
N20.2.1 Fixed and Variable Numbers of Choices ..................................................... N-356
N20.2.2 Restricting the Choice Set ........................................................................... N-358
N20.2.3 A Shorthand for Choice Sets ....................................................................... N-360
N20.2.4 Large Choice Sets – A Panel Data Equivalence .......................................... N-360
N20.2.5 An Alternative Data Arrangement ............................................................... N-362
N20.3 Specifying the Utility Functions with Rhs and Rh2 .................................................... N-363
N20.3.1 Utility Functions .......................................................................................... N-364
N20.3.2 Generic Coefficients .................................................................................... N-365
N20.3.3 Alternative Specific Constants and Interactions with Constants ................. N-365
N20.3.4 Command Builders ...................................................................................... N-367
N20.4 Building the Utility Functions ..................................................................................... N-368
 NLOGIT 6 Table of Contents xiii

N20.4.1 Notations for Sets of Utility Functions ........................................................ N-370

N20.4.2 Alternative Specific Constants and Interactions .......................................... N-371
N20.4.3 Logs and the Box Cox Transformation ....................................................... N-373
N20.4.4 Equality Constraints..................................................................................... N-374
N20.5 Starting and Fixed Values for Parameters ................................................................... N-375
N20.5.1 Fixed Values ................................................................................................ N-376
N20.5.2 Starting Values and Fixed Values from a Previous Model .......................... N-376

N21: Post Estimation Results for Conditional Logit Models......................................... N-377

N21.1 Introduction ................................................................................................................. N-377
N21.2 Partial Effects and Elasticities ..................................................................................... N-377
N21.2.1 Elasticities.................................................................................................... N-379
N21.2.2 Standard Errors for Estimated Partial Effects .............................................. N-380
N21.2.3 Influential Observations and Probability Weights ....................................... N-381
N21.2.4 Saving Elasticities in the Data Set ............................................................... N-382
N21.2.5 Computing Partial Effects at Data Means.................................................... N-383
N21.2.6 Exporting Results in a Spreadsheet ............................................................. N-385
N21.3 Predicted Probabilities and Logsums (Inclusive Values) ............................................... N-387
N21.3.1 Fitted Probabilities ....................................................................................... N-387
N21.3.2 Computing and Listing Model Probabilities................................................ N-388
N21.3.3 Utilities and Inclusive Values ...................................................................... N-390
N21.3.4 Fitted Values of the Choice Variable........................................................... N-390
N21.4 Specification Tests of IIA and Hypothesis .................................................................. N-391
N21.4.1 Hausman-McFadden Test of the IIA Assumption ....................................... N-391
N21.4.2 Small-Hsiao Likelihood Ratio Test of IIA .................................................. N-394
N21.4.3 Lagrange Multiplier, Wald, and Likelihood Ratio Tests ............................. N-397

N22: Simulating Probabilities in Discrete Choice Models............................................. N-399

N22.1 Introduction ................................................................................................................. N-399
N22.2 Essential Subcommands .............................................................................................. N-400
N22.3 Multiple Attribute Specifications and Scenarios ......................................................... N-401
N22.4 Simulation Commands................................................................................................. N-402
N22.4.1 Observations Used for the Simulations ....................................................... N-402
N22.4.2 Variables Used for the Simulations ............................................................. N-402
N22.4.3 Choices Simulated ....................................................................................... N-402
N22.4.4 Other NLOGIT Options ............................................................................... N-403
N22.4.5 Observations Used for the Simulations ....................................................... N-403
N22.5 Arc Elasticities ............................................................................................................. N-403
N22.6 Plotting Simulated Choice Probabilities ...................................................................... N-404
N22.7 Applications ................................................................................................................. N-405
N22.8 A Case Study ............................................................................................................... N-411
N22.8.1 Base Model – Multinomial Logit (MNL) .................................................... N-412
N22.8.2 Scenarios...................................................................................................... N-414

N23: The Multinomial Logit and Random Regret Models.............................................. N-422

N23.1 Introduction ................................................................................................................. N-422
N23.2 Command for the Multinomial Logit Model ............................................................... N-423
 NLOGIT 6 Table of Contents xiv

N23.3 Results for the Multinomial Logit Model .................................................................... N-425

N23.4 Application .................................................................................................................. N-425
N23.5 Partial Effects............................................................................................................... N-430
N23.6 Technical Details on Maximum Likelihood Estimation ................................................ N-432
N23.7 Random Regret Model................................................................................................. N-434
N23.7.1 Commands for Random Regret ................................................................... N-434
N23.7.2 Application .................................................................................................. N-435
N23.7.3 Technical Details: Random Regret Elasticities ........................................... N-437
N23.8 Fixed Effects Multinomial Logit Model ...................................................................... N-439
N23.8.1 Estimation of the Fixed Effects Multinomial Logit Model ......................... N-439
N23.8.2 Application .................................................................................................. N-439
N23.8.3 Technical Background for the FE MNL Estimator ..................................... N-441

N24: The Scaled Multinomial Logit Model ...................................................................... N-444

N24.1 Introduction ................................................................................................................. N-444
N24.2 Command for the Scaled MNL Model ........................................................................ N-445
N24.3 Application .................................................................................................................. N-445
N24.4 Technical Details ......................................................................................................... N-448

N25: Latent Class and 2K Multinomial Logit Model........................................................ N-449

N25.1 Introduction ................................................................................................................. N-449
N25.2 Model Command ......................................................................................................... N-450
N25.3 Individual Specific Results .......................................................................................... N-451
N25.4 Constraining the Model Parameters............................................................................. N-452
N25.5 Accessing Class Specific Information ......................................................................... N-455
N25.6 An Application............................................................................................................. N-456
N25.7 The 2K Model ............................................................................................................... N-459
N25.8 Individual Results ........................................................................................................ N-461
N25.8.1 Parameters ................................................................................................... N-462
N25.8.2 Willingness to Pay ....................................................................................... N-462
N25.8.3 Elasticities.................................................................................................... N-464
N25.8.4 Accessing Individual Coefficients After Estimation ................................... N-465
N25.9 Scaled Latent Class Model .......................................................................................... N-467
N25.10 Random Regret Latent Class Model .......................................................................... N-468
N25.11 Technical Details ....................................................................................................... N-468

N26: Heteroscedastic Extreme Value Model .................................................................. N-470

N26.1 Introduction ................................................................................................................. N-470
N26.2 Command for the HEV Model..................................................................................... N-471
N26.3 Application .................................................................................................................. N-473
N26.4 Constraining the Precision Parameters ........................................................................ N-475
N26.5 Individual Heterogeneity in the Variances .................................................................. N-480
N26.6 Technical Details ......................................................................................................... N-482

N27: Multinomial Probit Model ........................................................................................ N-484

N27.1 Introduction ................................................................................................................. N-484
N27.2 Model Command ......................................................................................................... N-485
 NLOGIT 6 Table of Contents xv

N27.3 An Application............................................................................................................. N-486

N27.4 Modifying the Covariance Structure............................................................................ N-488
N27.4.1 Specifying the Standard Deviations............................................................. N-489
N27.4.2 Specifying the Correlation Matrix ............................................................... N-491
N27.5 Testing IIA with a Multinomial Probit Model ............................................................. N-494
N27.6 A Model of Covariance Heterogeneity ........................................................................ N-495
N27.7 Panel Data – The Multinomial Multiperiod Probit Model ............................................ N-495
N27.8 Technical Details ......................................................................................................... N-496
N27.9 Multivariate Normal Probabilities ............................................................................... N-497

N28: Nested Logit and Covariance Heterogeneity Models ........................................... N-499

N28.1 Introduction ................................................................................................................. N-499
N28.2 Mathematical Specification of the Model.................................................................... N-500
N28.3 Commands for FIML Estimation................................................................................. N-502
N28.3.1 Data Setup.................................................................................................... N-502
N28.3.2 Tree Definition ............................................................................................ N-502
N28.3.3 Utility Functions .......................................................................................... N-504
N28.3.4 Setting and Constraining Inclusive Value Parameters ................................. N-505
N28.3.5 Starting Values ............................................................................................ N-506
N28.3.6 Command Builder........................................................................................ N-508
N28.4 Partial Effects and Elasticities ..................................................................................... N-510
N28.5 Inclusive Values, Utilities, and Probabilities ............................................................... N-512
N28.6 Application of a Nested Logit Model .......................................................................... N-513
N28.7 Alternative Normalizations .......................................................................................... N-517
N28.7.1 Nondegenerate Cases ................................................................................... N-520
N28.7.2 Degenerate Cases ......................................................................................... N-523
N28.8 Technical Details ......................................................................................................... N-525
N28.9 Sequential (Two Step) Estimation of Nested Logit Models ........................................ N-527
N28.10 Combining Data Sets and Scaling in Discrete Choice Models .................................. N-530
N28.10.1 Joint Estimation ......................................................................................... N-531
N28.10.2 Sequential Estimation ................................................................................ N-533
N28.11 A Model of Covariance Heterogeneity ...................................................................... N-534
N28.12 The Generalized Nested Logit Model........................................................................ N-536
N28.13 Box-Cox Nested Logit Model ................................................................................... N-539

N29: Random Parameters Logit Model........................................................................... N-542

N29.1 Introduction ................................................................................................................. N-542
N29.2 Random Parameters (Mixed) Logit Models ................................................................ N-543
N29.3 Command for the Random Parameters Logit Models ................................................. N-547
N29.3.1 Distributions of Random Parameters in the Model ..................................... N-548
N29.3.2 Spreads, Scaling Parameters and Standard Deviations ................................ N-551
N29.3.3 Alternative Specific Constants .................................................................... N-555
N29.3.4 Heterogeneity in the Means of the Random Parameters.............................. N-556
N29.3.5 Fixed Coefficients........................................................................................ N-557
N29.3.6 Correlated Parameters.................................................................................. N-557
N29.3.7 Restricted Standard Deviations and Hierarchical Logit Models.................. N-560
N29.3.8 Special Forms of Random Parameter Specifications................................... N-562
 NLOGIT 6 Table of Contents xvi

N29.3.9 Other Optional Specifications...................................................................... N-567

N29.4 Heteroscedasticity and Heterogeneity in the Variances ................................................ N-567
N29.5 Random Effects and Error Components ...................................................................... N-568
N29.6 Controlling the Simulations ......................................................................................... N-573
N29.6.1 Number and Initiation of the Random Draws.............................................. N-573
N29.6.2 Halton Draws and Random Draws for Simulations..................................... N-574
N29.7 Model Estimates .......................................................................................................... N-574
N29.8 Individual Specific Estimates ...................................................................................... N-578
N29.8.1 Computing the Individual Specific Parameter Estimates ............................ N-579
N29.8.2 Examining the Distribution of the Parameters............................................. N-585
N29.8.3 Conditional Confidence Intervals for Parameters ........................................ N-590
N29.8.4 Willingness to Pay Estimates....................................................................... N-591
N29.9 Applications ................................................................................................................. N-593
N29.10 Panel Data .................................................................................................................. N-595
N29.10.1 Random Effects Model .............................................................................. N-596
N29.10.2 Error Components Model .......................................................................... N-598
N29.10.3 Autoregression Model ............................................................................... N-600
N29.10.4 Berry, Levinsohn and Pakes RP Logit Model ........................................... N-601
N29.11 Technical Details on RP Estimation .......................................................................... N-604
N29.11.1 The Simulated Log Likelihood .................................................................. N-604
N29.11.2 Random Draws for the Simulations ........................................................... N-606
N29.11.3 Halton Draws for the Simulations ............................................................. N-607
N29.11.4 Functions and Gradients ............................................................................ N-609
N29.11.5 Hessians ..................................................................................................... N-611
N29.11.6 Panel Data and Autocorrelation ................................................................. N-612

N30: Error Components Multinomial Logit Model ......................................................... N-614

N30.1 Introduction ................................................................................................................. N-614
N30.2 Command for the Error Components MNL Model ..................................................... N-614
N30.3 Heteroscedastic Error Components ............................................................................. N-616
N30.4 General Form of the Error Components Model ........................................................... N-617
N30.5 Results for the Error Components MNL Model .......................................................... N-618
N30.6 Application .................................................................................................................. N-621
N30.7 Technical Details on Maximum Likelihood Estimation ................................................ N-622

N31: Nonlinear Random Parameters Logit Model ......................................................... N-624

N31.1 Introduction ................................................................................................................. N-624
N31.2 Model Command for Nonlinear RP Models ................................................................ N-624
N31.2.1 Parameter Definition.................................................................................... N-625
N31.2.2 Nonlinear Components ................................................................................ N-625
N31.2.3 Utility Functions .......................................................................................... N-626
N31.2.4 The Error Components Model ..................................................................... N-626
N31.2.5 Scaling function, σi – The Scaled Nonlinear RP Model .............................. N-626
N31.2.6 Panel Data .................................................................................................... N-627
N31.2.7 Ignored Attributes ........................................................................................ N-627
N31.3 Results ......................................................................................................................... N-627
N31.3.1 Individual Specific Parameters .................................................................... N-628
 NLOGIT 6 Table of Contents xvii

N31.3.2 Willingness to Pay ....................................................................................... N-628

N31.4 Application .................................................................................................................. N-629
N31.4.1 Elasticities and Partial Effects ..................................................................... N-634
N31.4.2 Variables Saved in the Data Set................................................................... N-635
N31.5 Technical Details ......................................................................................................... N-636

N32: Latent Class Random Parameters Model .............................................................. N-638

N32.1 Introduction ................................................................................................................. N-638
N32.2 Command..................................................................................................................... N-638
N32.2.1 Output Options ............................................................................................ N-639
N32.2.2 Post Estimation ............................................................................................ N-639
N32.3 Applications ................................................................................................................. N-640
N32.4 Technical Details ......................................................................................................... N-648

N33: Generalized Mixed Logit Model .............................................................................. N-651

N33.1 Introduction ................................................................................................................. N-651
N33.2 Commands ................................................................................................................... N-652
N33.2.1 Controlling the GMXLOGIT Parameters .................................................... N-653
N33.2.2 The Scaled MNL Model .............................................................................. N-654
N33.2.3 Alternative Specific Constants .................................................................... N-654
N33.2.4 Heteroscedasticity ........................................................................................ N-654
N33.3 Estimation in Willingness to Pay Space ...................................................................... N-655
N33.4 Results ......................................................................................................................... N-657

N34: Diagnostics and Error Messages ........................................................................... N-660

N34.1 Introduction ................................................................................................................. N-660
N34.2 Discrete Choice (CLOGIT) and NLOGIT ................................................................... N-661

NLOGIT 6 References ...................................................................................................... N-669

NLOGIT 6 Index ................................................................................................................ N-673

What’s New in Version 6? N-1

What’s New in Version 6?

NLOGIT 6 takes advantage of all the new features developed in LIMDEP 11. Several new
models have been added in NLOGIT 6. We have also continued to add enhancements to give you
greater flexibility in analyzing data and organizing results. The following will summarize the
important new developments.

WN1 New Multinomial Choice Models

We have added several major model classes to the package. Some of these are extensions of
the random parameters models that are at the forefront of current practice. We have also extended the
latent class model. The following are new model classes (commands) in Version 6:

RPLOGIT = random parameters logit

ECLOGIT = error components logit
RRLOGIT = random regret logit
LCRRLOGIT = latent class random regret logit
RPRRLOGIT = random parameters random regret logit
BWLOGIT = best/worst logit
BWMNL = best/worst multinomial logit
BWRPL = best/worst random parameters logit
SEQLOGIT = sequential logit
REMLOGIT = random effects multinomial logit
FEMLOGIT = fixed effects multinomial logit
BLPLOGIT = Berry, Levinsohn and Pakes logit

WN1.1 Fixed Effects in Multinomial Logit Models

The fixed and random effects forms of the multinomial logit model will be useful for data
from stated choice experiments in which the same individual is observed in a sequence of choice
situations. The common effects model will appear

U ijt = b′xijt + e ijt + α ij , j= 1,..., J , t= 1,..., T , i= 1,..., n;

exp(b′xijt + α ij )
Prob(choiceijt ) = .
Σ exp(b′ximt + α im )
J
m =1

In the fixed effects form, it is assumed that the effects carry the unobserved attributes that might be
correlated with observed attributes. An example might be a model for brand choice in which the
observed data are on attributes, but price data are unavailable. This model builds off the
Chamberlain/Rasch formulation of the binary logit model for panel data. Rather than use maximum
likelihood estimation for the full system, which will take hours or days for even moderately sized
problems, NLOGIT uses minimum distance estimation which will take seconds for comparably sized
problems.
What’s New in Version 6? N-2

WN1.2 Random Effects Multinomial Logit Models

The random effects form of the model above adds an assumption whereby the common
(fixed) effects are uncorrelated with the observed attributes. In this instance, the effects become
random constant terms in the utility functions of the multinomial logit model. There are two forms
of this model. If there is no variation across the observed utilities so that the utilities are determined
only by the characteristics of the individual (age, gender, income, etc.) and the random term, then the
model that applies is a ‘multinomial logit,’ estimated using the MLOGIT ; Fcn = remnl command.
In the stated preference case, where there are attributes of the choices in the utility functions, then the
model takes the form of a mixed logit model with random alternative specific constants, and is
estimated using the new REMLOGIT command.

WN1.3 Random Regret Logit Model

The random regret model was introduced in Version 5 as an alternative to maximum random
utility. The base case form substitutes the minimum random regret formulation for the familiar
maximum random utility for all the attributes in the utility function. Three extensions of the random
regret model have been added in NLOGIT 6:

• The MNL may be specified with a mixture of attributes valued by minimum regret and
attributes valued by maximum utility.
• The mixed (random parameters) model may be specified as minimum regret or maximum
utility.
• The latent class logit model may be specified as minimum regret or maximum utility.

WN1.4 Sequential Logit Model

The sequential logit model treats the outcome of a multinomial choice as a sequence of
transitions. The outcomes are ordered by some construction (which might suggest an ordered choice
model as a natural formulation.) Consider the outcome level of schooling, less than high school
(y = 0), high school (y =1), college (y = 2) and advanced degree (y = 3). The sequential logit model
treats the level as a sequence of transitions starting with transition from level 0, which everyone
enters so P0 = 1. The probability of stopping at level 0 is the probability of entering and not passing
level 1, which is P0(1 – P1). The probability of stopping at level 1 (high school) is the probability of
leaving level 1 and not leaving level 2, P0P1(1 – P2). For level 2, the probability is, P0P1P2(1 – P3)
and finally for level 3, P0P1P2P3 This is treated as a sequence of conditional binary logit models;

Prob(observed level = j) = Probability(exit levels j = 1,…,j) × Probability(no exit at level j+1).

By this formulation, for the outcomes listed above,

Prob(observed level = j) = P0,1,…,j(1 – PJ+1)

What’s New in Version 6? N-3

WN1.5 Best/Worst Outcome Data

Data for multinomial choice typically come in the form of the indicator of the most preferred
outcome. In some survey situations, one might also have data on the least preferred outcome.
NLOGIT 6 supports one form of this setting, in which data consist of a set of ranks for all of the
alternatives. The new best/worst setting is a subset. Data might come in one of three forms. In the
typical case, the individual simultaneously indicates his/her most and least preferred outcomes. We
label this simply the best/worst case. Alternatively, the individual might indicate his/her most
preferred alternative, then their least preferred choice among those that remain. We label this ‘best
then worst.’ Probably less common would be a case in which the least favored alternative is chosen,
then the most favored alternative is chosen among those that remain. Intuition suggests the three
choice mechanisms would produce the same outcomes, but it is at least possible that some difference
could occur.

WN1.6 Berry, Levinsohn and Pakes Random Parameters Logit

Model
The Berry, Levinsohn and Pakes model is a random parameters logit model for market share
data. The core of the specification is a multinomial logit model for the market shares, with fixed
brand effects. The estimation alternates between a GMM estimator for β and an intermediate step to
reconcile the fixed effects with the observed market share data. The multinomial logit model
describes the utility of consumer i’s choice of brand j in market period t;

=
U ijt x′it β i + α jt + eijt ,
U i0t = ei0t for the outside good.

The assumptions of the model produce the conditional (on βi) probability,

exp(x′jt b i + a jt )
Prob(consumer i chooses brand j =
in market t ) s=
j ( Xt , a t , b i )
1 + Σ mJ =1exp(x′mt b i + a mt )

This is a mixed logit model at this point, though it is based on market share data. Estimation of the
model parameters is complicated by two factors:

1. Some attributes are endogenous due to omitted factors. In the BLP application, price is
included in the model but features of the models that consumers respond to and which affect
the price are not included.

2. The fixed brand effects must be estimated. The estimation procedure alternates between two
steps, GMM conditioned on the fixed effects and the method of moments equating market
shares to theoretical market shares to calibrate fixed effects.

Further technical details appear in Chapter N29.

What’s New in Version 6? N-4

WN.2 Model Extensions

A variety of improvements and extensions have been built into the existing models and tools
for manipulating results post estimation.

WN2.1 Willingness to Pay

Willingness to pay (WTP) measures for models with variable coefficients such as mixed and
generalized logit models are generally saved as the rows of a matrix wtp_i post estimation. They
may now be saved directly as variables in the data area.

WN2.2 Model Simulation

After estimation, the simulator can be used to examine scenarios such as changes in attribute
prices of values on the market shares predicted by the model. You may also plot the probabilities
that result from your scenario. In the figure below, mode choice probabilities from a choice model
are plotted for values of the cost of air ranging from 50 to 100.
Simulations can be based on a variety of different configurations, including setting an
attribute at a specific value, setting several attributes, or changing one or more attributes by specific
values. In the scenario below, the cost of air is fixed at 75 to examine the impact of this setting
compared to the observed values in the sample:

Figure WN.1 Predicted Choice Probabilities

What’s New in Version 6? N-5

-------------------------------------------------------------------------
Specification of scenario 1 is:
Attribute Alternatives affected Change type Value
--------- ------------------------------- ------------------- ---------
GC AIR Fix base at new vlu 75.000
-------------------------------------------------------------------------
The simulator located 210 observations for this scenario.
Simulated Probabilities (shares) for this scenario:
+----------+--------------+--------------+------------------+
|Choice | Base | Scenario | Scenario - Base |
| |%Share Number |%Share Number |ChgShare ChgNumber|
+----------+--------------+--------------+------------------+
|AIR | 27.619 58 | 11.820 25 |-15.799% -33 |
|TRAIN | 30.000 63 | 34.773 73 | 4.773% 10 |
|BUS | 14.286 30 | 18.114 38 | 3.828% 8 |
|CAR | 28.095 59 | 35.294 74 | 7.199% 15 |
|Total |100.000 210 |100.000 210 | .000% 0 |
+----------+--------------+--------------+------------------+

WN2.3 Estimated Elasticities and Partial Effects

There are numerous options for computing and displaying partial effects and elasticities in
estimated choice models. New features added allow the user to export cross elasticity matrices to
Excel and text editors. We have developed standard errors and confidence intervals for elasticities
based on the method of Krinsky and Robb.

WN2.4 Robust Covariance Matrix

Robust covariance matrices may be computed for all multinomial choice model estimators.
The ‘cluster robust’ covariance matrix may also be computed for stated choice data (though a formal
estimator such as RPLOGIT is likely to be a preferable approach).

WN2.5 Random Number Generator

Simulation based calculations in NLOGIT use the Mersenne Twister. L’Ecuyer’s generator
used in earlier versions remains available. You may also modify the sampling procedure by
combining the generator with shuffling, Modified Latin Hypercube sampling or antithetic draws.
What’s New in Version 6? N-6

WN2.6 Posterior Estimates from Latent Class Models

The latent class models create results for separate classes as well as conditional estimates for
each individual in the sample. New CREATE and MATRIX commands provide access to the
specific terms for simulations and other post estimation analyses. For example, the following fits a
three class LC model to a stated choice experiment, then obtains the predicted probabilities for each
choice situation based on the conditional (posterior) estimates:

NAMELIST ; x = gc,invc,invt,ttme $
LCLOGIT ; Lhs = mode ; Choices = air,train,bus,car
; Rhs = x ; Pts = 3
; Pds = 3 ; Parameters $
CREATE ; i =Trn(12,0) $
CREATE ; utility = Mbx(beta_i,i,x) $
CREATE ; mnlprob = mnl_probs(utility,Set=4) $

The same capability is provided for the random parameters models. (The new functions can be used
for all of the choice models, but it would usually be needed only for these two.)

For latent class models, the conditional (posterior) class probabilities may also be saved in
the data set in addition to the matrix classp_i as in previous versions.
Elasticities and other partial effects and simulations for latent class models are based on
average coefficient vectors. Some applications involve analysis of the classes (based on the class
specific parameter values). A convenient method of setting up the simulation for the class specific
components of the LC model is provided.

WN2.7 Coefficients in RP Models

It can be inconvenient to specify random ASCs in an RP model as they take the name partly
from the names given to the alternatives. To simplify the specification, the constants may be
assigned the name of the alternative defined in the utility function. For example,

RPLOGIT ; … ; Choices = air,train,bus,car

; … ; Rh2 = one
; Fcn = air(n),train(n),bus(n)

specifies what amounts to a random effects model. (As usual, one of the constants must be set to
zero.) The lognormal distribution is used in RP models to constrain the sign of a coefficient. The
specification will appear as in ; Fcn = price(l). However, as stated here, this forces the coefficient to
be positive. To force it to be negative the well known ‘trick’ is to multiply the price variable by
minus one, then force the coefficient to be positive as usual. Alternatively, the sign can be built into
the coefficient by using ; Fcn = -price(l).
What’s New in Version 6? N-7

WN2.8 Simplified WALD Command

A simplified syntax is provided for the WALD command when it is based on the previously
estimated model. In this case, the parameter values (; Parameters = name) and covariance matrix
(; Covariance = name) may be omitted. The defaults b and varb will then be assumed. The labels
are optional as well. If they are omitted, default labels, b1, b2, …, bK are used, where K is the
number of elements in B. Labels may be provided if desired. For an example, the following two
WALD commands are identical:

CLOGIT ; Lhs = mode ; Choices = air,train,bus,car ; Rhs = gc,invc,invt,ttme $

WALD ; Fn1 = wtp = -b3 / b1 $
WALD ; Labels = bcost, bicost, btime, bttme ; Fn1 = wtp = -btime/bcost $
N1: Introduction to NLOGIT Version 6 N-8

N1: Introduction to NLOGIT Version 6

N1.1 Introduction
NLOGIT is a package of programs for analyzing data on multinomial choice. The program,
itself, consists of a special set of estimation and analysis routines, specifically for this class of
models and style of analysis. LIMDEP provides the foundation for NLOGIT, including the full set of
tools used for setting up the data, such as importing data files, transforming variables (e.g.,
CREATE), and so on. NLOGIT is created by adding a set of capabilities to LIMDEP. The notes
below describe this connection in a bit more detail.

N1.2 The NLOGIT Program

NLOGIT adds one (very powerful) command to LIMDEP,

NLOGIT ; … specification of choice variable

; … specification of choice model behavioral equations
; … definition of choice modeling framework (e.g., nested logit)
; … other required and optional features $

The NLOGIT command is the gateway to the large set of features that are described in this NLOGIT
Reference Guide. All other features and commands in LIMDEP are provided in the NLOGIT package
as well.
The estimation results produced by NLOGIT look essentially the same as by LIMDEP, but at
various points, there are differences that are characteristic of this type of modeling. For example, the
standard data configuration for NLOGIT looks like a panel data set analyzed elsewhere in LIMDEP.
This has implications for the way, for example, model predictions are handled. These differences are
noted specifically in the descriptions to follow. But, at the same time, the estimation and post
estimation tools provided for LIMDEP, such as matrix algebra and the hypothesis testing procedures,
are all unchanged. That is, NLOGIT is LIMDEP with an additional special command.

N1.3 NLOGIT and LIMDEP Integration and Documentation

NLOGIT 6 is a suite of programs for estimating discrete choice models that are built around
the logit and multinomial logit form. This is a superset of LIMDEP’s models – NLOGIT 6 is all of
LIMDEP 11 plus the set of tools and estimators described in this guide. LIMDEP 11 contains the
CLOGIT command and the estimator for the ‘conditional logit’ (or multinomial logit) model.
CLOGIT is the same as the most basic form of the NLOGIT command described in Chapter N19.
The full set of features of LIMDEP 11 is part of this package. We assume that you will use
the other parts of LIMDEP as part of your analysis. To use NLOGIT, you will need to be familiar
with the LIMDEP platform. At various points in your operation of the program, you will encounter
LIMDEP, rather than NLOGIT as the program name, for example in certain menus, dialog boxes,
window headers, diagnostics, and so on. Once again, these result from the fact that in obtaining
NLOGIT, you have installed LIMDEP plus some additional capabilities. If you are uncertain which
program is actually installed on your computer, go to the About box in the main menu. It will
clearly indicate which program you are operating.
N1: Introduction to NLOGIT Version 6 N-9

This NLOGIT Reference Guide provides documentation for some aspects of discrete choice
models in general but is primarily focused on the specialized tools and estimators in NLOGIT 6 that
extend the multinomial logit model. These include, for example, extensions of the multinomial logit
model such as the nested logit, random parameters logit, generalized mixed logit and multinomial
probit models. This guide is primarily oriented to the commands added to LIMDEP that request the
set of discrete choice estimators. However, in order to provide a more complete and useful package,
Chapters N4-N17 in the NLOGIT Reference Guide describe common features of LIMDEP 11 and
NLOGIT 6 that will be integral tools in your analysis of discrete choice data, as shown, for example,
in many of the examples and applications in this manual.
Users will find the LIMDEP documentation, the LIMDEP Reference Guide and the LIMDEP
Econometric Modeling Guide, essential for effective use of this program. It is assumed throughout
that you are already a user of LIMDEP. The NLOGIT Reference Guide, by itself, will not be
sufficient documentation for you to use NLOGIT unless you are already familiar with the program
platform, LIMDEP, on which NLOGIT is placed.
The LIMDEP and NLOGIT documentation use the following format: The LIMDEP
Reference Guide chapter numbers are preceded by the letter ‘R.’ The LIMDEP Econometric
Modeling Guide chapter numbers are preceded by ‘E,’ and the NLOGIT Reference Guide chapter
numbers are preceded by ‘N.’

N1.4 Discrete Choice Modeling with NLOGIT

NLOGIT is a set of tools for building models of discrete choice among multiple alternatives.
The essential building block that underlies the set of programs is the random utility model of choice,

U(choice 1) = f1 (attributes of choice 1, characteristics of the chooser, e1,v,w)

...
U(choice J) = fJ (attributes of choice J, characteristics of the chooser, eJ,v,w)

where the functions on the right hand side describe the utility to an individual decision maker of J
possible choices, as functions of the attributes of the choices, the characteristics of the chooser,
random choice specific elements of preferences, ej, that may be known to the chooser but are
unobserved by the analyst, and random elements v and w, that will capture the unobservable
heterogeneity across individuals. Finally, a crucial element of the underlying theory is the
assumption of utility maximization,

The choice made is alternative j such that U(choice j) > U(choice q) ∀ q ≠ j.

The tools provided by NLOGIT are a complete suite of estimators beginning with the simplest binary
logit model for choice between two alternatives and progressing through the most recently developed
models for multiple choices, including random parameters, mixed logit models with individual
specific random effects for repeated observation choice settings and the multinomial probit model.
N1: Introduction to NLOGIT Version 6 N-10

Background theory and applications for the programs described here can be found in many
sources. For a primer that develops the theory for multinomial choice modeling in detail and
presents many examples and applications, all using NLOGIT, we suggest

Hensher, D., Rose, J. and Greene, W., Applied Choice Analysis, 2nd Edition, Cambridge
University Press, 2015.

A general reference for ordered choice models, also based on NLOGIT is

Greene, W. and Hensher, D., Modeling Ordered Choices, Cambridge University Press, 2010.

It is not possible (nor desirable) to present all of the necessary econometric methodology in a manual of
this sort. The econometric background needed for Applied Choice Analysis as well as for use of the
tools to be described here can be found in many graduate econometrics books. One popular choice is

Greene, W., Econometric Analysis, 7th Edition, Prentice Hall, 2012.

N1.5 Types of Discrete Choice Models in NLOGIT

The order and organization of presentations in this manual are partly oriented to the types of
models you will analyze and partly toward the types of data you will use. Chapters N2 and N3
describe discrete choice models including NLOGIT model and command summaries.
In Chapters N4-N15, we develop basic choice models that have occupied a large part of the
econometrics literature for several decades. The situations are essentially those in which the
characteristics of decision makers and the choices that they make form the observational base for the
model building. The fundamental building block for all of these, as well as for the more elaborate
models, is the binary choice model: The structural equations for a model of consumer choice based
on a single alternative – either to choose an outcome or not to choose it – are

U(choice) = b′x + e,
Prob(choice) = Prob(U > 0)
= F(b′x),
Prob(not choice) = 1 - F(b′x),

where x is a vector of characteristics of the consumer such as age, sex, education, income, and other
sociodemographic variables, b is a vector of parameters and F(.) is a suitable function that describes the
model. The choice of vote for a political candidate or party is a natural application. Models for binary
choice are developed at length in Chapters E26-E32 in the LIMDEP Econometric Modeling Guide.
They will be briefly summarized in Chapters N4-N7 to provide the departure point for the models that
follow. Useful extensions of the binary choice model presented in Chapters N8-N12 include models
for more than one simultaneous binary choice (of the same type), including bivariate binary choice
models and simultaneous binary choice models and a model for multivariate binary choices (up to 20).
N1: Introduction to NLOGIT Version 6 N-11

The ordered choice model described in Chapters N13-N15 describe a censoring of the
underlying utility in which consumers are able to provide more information about their preferences.
In the binary choice model, decision makers reveal through their decisions that the utility from
making the choice being modeled is greater than the utility of not making that choice. In the ordered
choice case, consumers can reveal more about their preferences – we obtain a discretized version of
their underlying utility. Thus, in survey data, voters might reveal their strength of preferences for a
candidate or a food or drink product, from zero (strongly disapprove), one (somewhat disapprove) to,
say, four (strongly approve).
The appropriate model might be

Prob(strongly dislike) = Prob(U < 0),

Prob(dislike) = Prob(0 < U < µ1),
Prob(indifferent) = Prob(µ1 < U < µ2),
and so on.

We can also build extensions of the ordered choice model, such as a bivariate ordered choice model
for two simultaneous choices and a sample selection model for nonrandomly selected samples.
The multinomial logit (MNL) model described in Chapters N16 and N17 is the original
formulation of this model for the situations in which, as in the binary choice and ordered choice
models already considered, we observe characteristics of the individual and the choices that they
make. The classic applications are the Nerlove and Press (1973) and Schmidt and Strauss (1975)
studies of labor markets and occupational choice. The model structure appears as follows:

exp ( β′j xi )
Prob[yi = j] = .
∑ exp ( β′q xi )
Ji
q=1

Note the signature feature, that the determinants of the outcome probability are the individual
characteristics. This model represents a straightforward special case of the more general forms of
the multinomial choice model described in Chapters N16 and N17 and in the extensions that follow
in Chapters N23-N33.
Chapters N18-N22 document general aspects of operating NLOGIT. Chapter N18 describes
the way that your data will be arranged for estimation of multinomial discrete choice models.
Chapter N19 presents an overview of the command structure for NLOGIT models. The commands
differ somewhat from one model to another, but there are many common elements that are needed to
set up the essential modeling framework. Chapter N20 describes choice sets and utility functions.
Chapter N21 describes results that are computed for the multinomial choice models beyond the
coefficients and standard errors. Finally, Chapter N22 describes the model simulator. You will use
this tool after fitting a model to analyze the effects of changes in the attributes of choices on the
aggregate choices made by individuals in the sample.
N1: Introduction to NLOGIT Version 6 N-12

The models developed in Chapters N23-N33 extend the binary choice case to situations in
which decision makers choose among multiple alternatives. These settings involve richer data sets in
which the attributes of the alternatives are also part of the observation, and more elaborate models of
behavior. The broad modeling framework is the multinomial logit model. With a particular
specification of the utility functions and distributions of the unobservable random components, we
obtain the canonical form of the logit model,

exp ( β′xij )
Prob[yi = j] = ,
∑ exp ( β′xiq )
Ji
q=1

where yi is the index of the choice made. This is the basic, core model of the set of estimators in
NLOGIT. (This is the model described in Chapters N16 and N17.)
The basic setup for this model consists of observations on N individuals, each of whom
makes a single choice among Ji choices, or alternatives. There is a subscript on J because we do not
restrict the choice sets to have the same number of choices for every individual. The data will
typically consist of the choices and observations on K ‘attributes’ for each choice. The attributes that
describe each choice, i.e., the arguments that enter the utility functions, may be the same for all
choices, or may be defined differently for each utility function. It is also possible to incorporate
characteristics of the individual which do not vary across choices in the utility functions. The
estimators described in this manual allow a large number of variations of this basic model.
In the discrete choice framework, the observed ‘dependent variable’ usually consists of an
indicator of which among Ji alternatives was most preferred by the respondent. All that is known
about the others is that they were judged inferior to the one chosen. But, there are cases in which
information is more complete and consists of a subjective ranking of all Ji alternatives by the
individual. NLOGIT allows specification of the model for estimation with ‘ranks data.’ The ranking
might be incomplete – ranks data can include ties. An interesting extension of this possibility is
‘Best/Worst’ data in which the chooser indicates both their most and least favored alternatives. In
addition, in some settings, the sample data might consist of aggregates for the choices, such as
proportions (market shares) or frequency counts. NLOGIT will accommodate these cases as well.
The multinomial model has provided a mainstay of empirical research in this literature for
decades. But, it does have limitations, notably the assumption of independence from irrelevant
alternatives, which limit its generality. Recent research has produced many new, different
formulations that have broadened the model. NLOGIT contains most of these, all of which remove
the crucial IIA assumption of the multinomial logit (MNL) model. Chapters N23-N33 describe these
frontier extensions of the multinomial logit model. In brief, these are as follows:

N1.5.1 Random Regret Logit Model

The random regret logit model is a variant of the basic conditional logit model. The form of
the utility functions involves more direct comparisons of the attributes of the alternatives. Whereas
in the essential MNL model, the utility functions enter the probability linearly in terms of the
attributes, so the coefficients are marginal utilities, in the random regret model, the attributes enter
the probabilities through the regret functions,

Rij (m) = log[1 + exp(βm ( x jm − xim ))]

which compare attribute m in alternative j to that attribute in alternative i.

N1: Introduction to NLOGIT Version 6 N-13

N1.5.2 Scaled Multinomial Logit Model

The scaled multinomial logit model accommodates individual heterogeneity in choice
structures through the scaling of the marginal utilities rather than in the location parameters. The
coefficients in the scaled MNL model take the form

bi = σib
where σi = σ × exp(δ′zi + τvi).

This is a type of random parameters model; the scale parameter can vary systematically with the
observables, zi and randomly across individuals with vi.

N1.5.3 Latent Class and Random Parameters LC Models

The latent class model is a semiparametric approximation to the random parameters
multinomial logit model. It embodies many of the features of the RPL model. But, the parameters
are modeled as having a discrete distribution with a small number of support points. An alternative
interpretation is that individuals are intrinsically sorted into a small number of classes, and
information about class membership is extracted from the sample along with class specific parameter
vectors. The RP variant, which is new with this version of NLOGIT, provides a random parameters
logit model (see Section N1.5.7) in each class.

N1.5.4 Heteroscedastic Extreme Value Model

In the base case, multinomial logit model, the assumption of equal variances produces great
simplicity in the mathematical results, but at considerable cost in the generality of the model. In
particular, if the assumption of equal variances is inappropriate, then the different scaling that is
present in the variances will, instead, be forced on the coefficients in the utility functions, in ways
that may distort the predictions of the model. The heteroscedastic extreme value model relaxes this
assumption by allowing the disturbances in the utility functions each to have their own variance. An
extension of this model allows these unequal variances to be dependent on characteristics of the
individual as well. Thus, the heteroscedasticity assumption allows us to relax the assumption of
equal variances across choices and to incorporate individual heterogeneity in the scaling as well as
the ‘locations’ of the utility functions.

N1.5.5 Multinomial Probit Model

This model is much more general than the multinomial logit model, but until recently was
largely impractical because of the multinormal integrals required for estimation. We include an
implementation based on the GHK simulation method. The multinomial probit (MNP) model
relaxes the assumptions of the MNL model by assuming joint normality for the random terms in the
utility functions and by allowing (subject to some identification restrictions) the random terms to
have different variances and unrestricted correlations.
N1: Introduction to NLOGIT Version 6 N-14

N1.5.6 Nested Logit Models

The choice among alternatives could be viewed as taking place at more than one level. For
instance, in an application developed in the chapters to follow, we consider transportation mode
choice among four alternatives, car, train, bus, and air. One might view the choice among these
four as first between public (bus, train) and private (air, car) transportation and then, within each of
the two branches of the ‘tree,’ a second choice of specific mode. This sort of hierarchical choice is
handled in the setting of ‘nested logit models.’ NLOGIT allows tree structures to have up to four
levels. There are also several specific forms of the nested logit model that enforce the implications
of utility maximization on the model parameters.
The nested logit (NL) model described in the previous paragraph is appropriately viewed as
a relaxation of the strong IID structure of the multinomial logit model that implies the IIA
assumption. In particular, the nested logit model allows for different variances for the groups of
alternatives in the branches and for (equal) correlation across the alternatives within a branch. (The
earlier interpretation of a decision structure is only superimposed on the nested logit model; it is not
the statistical basis of the NL model. The ‘decision’ part of the model rests at the lowest level,
among the alternatives.) The covariance heterogeneity model extends this model a bit further by
allowing the variances to depend on variables in the model. The covariance heterogeneity model is a
model of heteroscedasticity.
One of the weaker parts of the nested logit specification is the narrow specific assumption of
which alternative appears in each branch of the tree. This is often not known with certainty. The
generalized nested logit model allows alternatives to appear in more than one branch, in a
probabilistic fashion.

N1.5.7 Random Parameters and Nonlinear RP Logit Model

This is the most general model contained in NLOGIT. As argued by McFadden and Train
(2000), it may be the most flexible form of discrete choice model available generally, as they argue
that any behavior pattern can be captured by this model form. The random parameters logit (RPL)
model extends the MNL model by allowing its parameters to be random across individuals. The
random parameters may have their own data dependent means, their own variances, and may be
correlated. By this device, we obtain an extremely general, flexible model. The assumptions about
the covariance matrix of the random parameters are transmitted to the random terms in the utility
functions so that both the uncorrelatedness and equal variance assumptions are relaxed in the
process. This model also allows a panel data treatment, with either random effects or an
autoregressive pattern in the random terms. The error components logit model provides a method of
incorporating a rich structure of individual specific random effects in the conditional logit and
random parameters models. The nonlinear RP variant allows the utility functions in the probability
model to be arbitrary nonlinear functions of the data and parameters.
N1: Introduction to NLOGIT Version 6 N-15

N1.5.8 Error Components Logit and Fixed Effects Models

The error components logit model is essentially a random effects model for the MNL
framework. The basic model structure for a repeated choice (panel data) setting would be

exp ( β′x ji + Σ sM=1d js vis )

Prob[yit = j| vi1,...,viM) = ,
∑ exp ( β′xiq + Σ Ms =1d qs vis )
Ji
q =1

where vi1,...,viM are M individual effects that appear in the Ji utility functions and djs are binary
variables that place specific effects in the different alternatives. Different sets of effects, or only
particular ones, appear in each utility function, which allows a nested type of arrangement. A fixed
effects version of the multinomial logit model would appear as

exp ( α ij + β′x ji )
Prob[yit = j| vi1,...,viM) = ,
∑ exp ( αiq + β′xiq )
Ji
q =1

The fixed effects model is estimated using the conditional estimation method proposed in
Chamberlain (1984) with a minimum distance estimator to extend the results to the multinomial
outcome case.

N1.5.9 Generalized Mixed Logit Model

The generalized mixed logit model is an encompassing model for many of the specifications
already noted, and a variety of new specifications as well. The model follows the random
parameters model of Section N1.5.7, but adds several layers to the specification of the random
parameters. Specifically,

bi = σib + [γ + (1 - γ)σi]Γvi,

where σi is the heterogeneous scale factor noted in Section N1.5.2, γ is a distribution parameter that
moves emphasis to or away from the random part of the model, Γ is (essentially) the correlation
matrix among the random parameters. As noted, several earlier specifications are special cases.
This form of the RP model allows a number of useful extensions, including estimation of the
model in willingness to pay (WTP) space, rather than utility space.
N1: Introduction to NLOGIT Version 6 N-16

N1.6 Functions of NLOGIT

The chapters to follow will describe the different features of NLOGIT and the various
models it will estimate. The functionality of the program consists of these major features:

• Estimation programs. These are full information maximum likelihood estimators for the
collection of models.

• Description and analysis. Model results are used to compute elasticities, marginal effects,
and other descriptive measures.

• Hypothesis testing, including the IIA assumption and tests of model specification.

• Computation of probabilities, utility functions, and inclusive values for individuals in the
sample.
Simulation of the model to predict the effects of changes in the values of attributes in the aggregate
behavior of the individuals in the sample. For example, if x% of the sampled individuals choose a
particular alternative, how would x change if a certain price in the model were assumed to be p%
higher for all individuals.
N2: Discrete Choice Models N-17

N2: Discrete Choice Models

N2.1 Introduction
This chapter will provide a short, thumbnail sketch of the discrete choice models discussed
in this manual. NLOGIT supports a large array of models for both discrete and continuous variables,
including regression models, survival models, models for counts and, of relevance to this setting,
models for discrete outcomes. The group of models described in this manual are those that arise
naturally from a random utility framework, that is, those that arise from an individual choice setting
in which the model is of an individual’s selection among two or more alternatives. This includes
several of the models described in the LIMDEP manual, such as the binary logit and probit models,
but also excludes some others, including the models for count data and censored and truncated
regression models, and some of the loglinear models such as the geometric regression model.
Two groups of models are considered. The first set are the binary, ordered and multivariate
choice models that are documented at length in Chapters E26-E35 in the LIMDEP Econometric
Modeling Guide. These form the basic building blocks for the NLOGIT extensions that are the main
focus of this part of the program. Since they are developed in detail elsewhere, we will only provide
the basic forms and only the essential documentation here. The second group of estimators are the
multinomial logit models and extensions of them that form the group of tools specific to NLOGIT.

N2.2 Random Utility Models

The random utility framework starts with a structural model,

U(choice 1) = f1 (attributes of choice 1, characteristics of the consumer, e1,v,w),

...
U(choice J) = fJ (attributes of choice J, characteristics of the consumer, eJ,v,w),

where e1,...,eJ denote the random elements of the random utility functions and in our later treatments,
v and w will represent the unobserved individual heterogeneity built into models such as the error
components and random parameters (mixed logit) models. The assumption that the choice made is
alternative j such that

U(choice j) > U(choice q) ∀ q ≠ j.

The observed outcome variable is then

y = the index of the observed choice.

The econometric model that describes the determination of y is then built around the assumptions
about the random elements in the utility functions that endow the model with its stochastic
characteristics. Thus, where Y is the random variable that will be the observed discrete outcome,

Prob(Y = j) = Prob(U(choice j) > U(choice q) ∀ q ≠ j).

N2: Discrete Choice Models N-18

The objects of estimation will be the parameters that are built into the utility functions including
possibly those of the distributions of the random components and, with estimates of the parameters
in hand, useful characteristics of consumer behavior that can be derived from the model, such as
partial effects and measures of aggregate behavior.
To consider the simplest example, that will provide the starting point for our development,
consider a consumer’s random utility derived over a single choice situation, say whether to make a
purchase. The two outcomes are ‘make the purchase’ and ‘do not make the purchase.’ The random
utility model is simply

U(not purchase) = b 0′x0 + e0,

U(purchase) = b 1′x1 + e1.

Assuming that e0 and e1 are random, the probability that the analyst will observe a purchase is

Prob(purchase) = Prob(U(purchase) > U(not purchase))

= Prob(b 1′x1 + e1 > b 0′x0 + e0)
= Prob(e1 - e0 < b 1′x1 - b0′x0)
= F(b 1′x1 - b 0′x0),

where F(z) is the CDF of the random variable e1 - e0. The model is completed and an estimator,
generally maximum likelihood, is implied by an assumption about this probability distribution. For
example, if e0 and e1 are assumed to be normally distributed, then the difference is also, and the
familiar probit model emerges. (The probit model is developed in Chapters E26 and E27.)
The sections to follow will outline the models described in this manual in the context of this
random utility model. The different models derive from different assumptions about the utility
functions and the distributions of their random components.

N2.3 Binary Choice Models

Continuing the example in the previous section, the choice of alternative 1 (purchase)
reveals that U1 > U0, or that

e0 - e1 < b 1′x1 - b 0′x0.

Let e = e1 - e0 and b′x represent the difference on the right hand side of the inequality – x is the union
of the two sets of covariates, and b is constructed from the two parameter vectors with zeros in the
appropriate locations if necessary. Then, a binary choice model applies to the probability that e ≤
b′x, which is the familiar sort of model developed in Chapter E26. Two of the parametric model
formulations in NLOGIT for binary choice models are the probit model based on the normal
distribution:
β 'x i exπ(−t 2 / 2)
F = ∫−∞ 2π
dt = Φ(b′xi),
N2: Discrete Choice Models N-19

and the logit model based on the logistic distribution

exp(β′xi )
F = = Λ(b′xi).
1 + exp(β′xi )

Numerous variations on the model can be obtained. A model with multiplicative

heteroscedasticity is obtained with the additional assumption

ei ~ normal or logistic with variance ∝ [exp(γ′zi)]2,

where zi is a set of observed characteristics of the individual. A model of sample selection can be
extended to the probit and logit binary choice models. In both cases, we depart from

Prob(yi = 1 |xi) = F(b′xi),

where F(t) = Φ(t) for the probit model and Λ(t) for the logit model,
di* = α′zi + ui, ui ~ N[0,1], di = 1(di* > 0),
yi, xi observed only when di = 1.

where zi is a set of observed characteristics of the individual. In both cases, as stated, there is no
obvious way that the selection mechanism impacts the binary choice model of interest. We modify
the models as follows: For the probit model,

yi* = b′xi + ei, ei ~ N[0,1], yi = 1(yi* > 0),

which is the structure underlying the probit model in any event, and

ui, ei ~ N2[(0,0),(1,ρ,1)].

(We use NP to denote the P-variate normal distribution, with the mean vector followed by the
definition of the covariance matrix in the succeeding brackets.) For the logit model, a similar
approach does not produce a convenient bivariate model. The probability is changed to

exp(β′xi + σei )
Prob(yi = 1 | xi,ei) = .
1 + exp(β′xi + σei )

With the selection model for zi as stated above, the bivariate probability for yi and zi is a mixture of a
logit and a probit model. The log likelihood can be obtained, but it is not in closed form, and must
be computed by approximation. We do so with simulation. The model and the background results
are presented in Chapter E27.
N2: Discrete Choice Models N-20

There are several formulations for extensions of the binary choice models to panel data
setting. These include

• Fixed effects: Prob(yit = 1) = F(b′xit + αi),

αi correlated with xit.

• Random effects: Prob(yit = 1) = Prob(b′xit + eit + ui > 0),

ui uncorrelated with xit.

• Random parameters: Prob(yit = 1) = F(b i′xit),

bi | i ~ h(b|i) with mean vector b and covariance matrix Σ.

• Latent class: Prob(yit = 1|class j) = F(b j′xit),

Prob(class = j) = Gj(q,zi).

where zi is a set of observed characteristics of the individual. Other variations include simultaneous
equations models and semiparametric formulations.

N2.4 Bivariate and Multivariate Binary Choice Models

The bivariate probit model is a natural extension of the model above in which two decisions
are made jointly;
yi1* = b 1′xi1 + ei1, yi1 = 1 if yi1* > 0, yi1 = 0 otherwise,
yi2* = b 2′xi2 + ei2, yi2 = 1 if yi2* > 0, yi2 = 0 otherwise,
[ei1,ei2] ~ N2[0,0,1,1,ρ], -1 < ρ < 1,
individual observations on y1 and y2 are available for all i.

This model extends the binary choice model to two different, but related outcomes. One might, for
example, model y1 = home ownership (vs. renting) and y2 = automobile purchase (vs. leasing). The
two decisions are obviously correlated (and possibly even jointly determined).
A special case of the bivariate probit model is useful for formulating the correlation between
two binary variables. The tetrachoric correlation coefficient is equivalent to the correlation
coefficient in the following bivariate probit model:

yi1* = µ + ei1, yi1 = 1(yi1* > 0),

yi2* = µ + ei2, yi2 = 1(yi2* > 0),
(ei1,ei2) ~ N2[(0,0),(1,1,ρ)].

The bivariate probit model has been extended to the random parameters form of the panel data
models. For example, a true random effects model for a bivariate probit outcome can be formulated
as follows: Each equation has its own random effect, and the two are correlated.
N2: Discrete Choice Models N-21

The model structure is

yit1* = b 1′xit1 + eit1 + ui1, yit1 = 1 if yit1* > 0, yit1 = 0 otherwise,

yit2* = b 2′xit2 + eit2 + ui2, yit2 = 1 if yit2* > 0, yit2 = 0 otherwise,
[eit1,eit2] ~ N2[0,0,1,1,ρ], -1 < ρ < 1,
[ui1 , ui2] ~ N2[0,0,1,1,q], -1 < q < 1.

Individual observations on yi1 and yi2 are available for all i. Note, in the structure, the idiosyncratic
eitj creates the bivariate probit model, whereas the time invariant common effects, uij create the
random effects (random constants) model. Thus, there are two sources of correlation across the
equations, the correlation between the unique disturbances, ρ, and the correlation between the time
invariant disturbances, q.
The multivariate probit model is the extension to M equations of the bivariate probit model

yim* = b m′xim+ eim, m = 1,…,M

yim = 1 if yim* > 0, and 0 otherwise,
eim, m = 1,...,M ~ NM[0,R],

where R is the correlation matrix. Each individual equation is a standard probit model. This
generalizes the bivariate probit model for up to M = 20 equations.

N2.5 Ordered Choice Models

The basic ordered choice model can be cast in an analog to our random utility specification.
We suppose that preferences over a given outcome are reflected as earlier, in the random utility
function:
yi* = b′xi + ei,
ei ~ F(ei |q), θ = a vector of parameters,
E[ei|xi] = 0,
Var[ei|xi] = 1.

The consumers are asked to reveal the strength of their preferences over the outcome, but are given
only a discrete, ordinal scale, 0,1,...,J. The observed response represents a complete censoring of the
latent utility as follows:

yi = 0 if yi* ≤ µ0,
= 1 if µ0 < yi* ≤ µ1,
= 2 if µ1 < yi* ≤ µ2,
...
= J if yi* > µJ-1.
N2: Discrete Choice Models N-22

The latent ‘preference’ variable, yi* is not observed. The observed counterpart to yi* is yi. (The
model as stated does embody the strong assumption that the threshold values are the same for all
individuals. We will relax that assumption below.) The ordered probit model based on the normal
distribution was developed by Zavoina and McElvey (1975). It applies in applications such as
surveys, in which the respondent expresses a preference with the above sort of ordinal ranking. The
ordered logit model arises if ei is assumed to have a logistic distribution rather than a normal. The
variance of ei is assumed to be the standard, one for the probit model and π2/6 for the logit model,
since as long as yi*, b, and ei are all unobserved, no scaling of the underlying model can be deduced
from the observed data. (The assumption of homoscedasticity is arguably a strong one. We will also
relax that assumption.) Since the µs are free parameters, there is no significance to the unit distance
between the set of observed values of yi. They merely provide the coding. Estimates are obtained by
maximum likelihood. The probabilities which enter the log likelihood function are

Prob(yi = j) = Prob(yi* is in the jth range).

The model may be estimated either with individual data, with yi = 0, 1, 2, ... or with grouped data, in
which case each observation consists of a full set of J + 1 proportions, pi0,...,piJ.
There are many variants of the ordered probit model. A model with multiplicative
heteroscedasticity of the same form as in the binary choice models is

Var[ei] = [exp(γ′zi)]2.

The following describes an ordered probit counterpart to the standard sample selection model. (This
is only available for the ordered probit specification.) The structural equations are, first, the main
equation, the ordered choice model that was given above and, second, a selection equation, a
univariate probit model,
di* = α′zi + ui,
di = 1 if di* > 0 and 0 otherwise.

The observation mechanism is

[yi,xi] is observed if and only if di = 1,
ei,ui ~ N2[0,0,1,1,ρ]; there is ‘selectivity’ if ρ is not equal to zero.
The general set of panel data formulations is also available for the ordered probit and logit models.

• Fixed effects: Prob(yit = j) = F[µj -(b′xit + αi)] - F[µj-1-(b′xit + αi)],

αi correlated with xit.

• Random effects: Prob(yit = j) = F[µj -(b′xit + ui)] - F[µj-1-(b′xit + ui)],

ui uncorrelated with xit.

• Random parameters: Prob(yit = j) = F(µj -b i′xit) - F(-µj-1b i′xit),

bi | i ~ h(b|i) with mean vector b and covariance matrix Σ.

• Latent class: Prob(yit = j|class c) = F(µj,c -b c′xit) - F(µj-1,c -bc′xit),

Prob(class = c) = Gc(q,zi).
N2: Discrete Choice Models N-23

The hierarchical ordered probit model, or generalized ordered probit model, relaxes the
assumption that the threshold parameters are the same for all individuals. Two forms of the model
are provided.
Form 1: µij = exp(qj + δ′zi),
Form 2: µij = exp(qj + δj′zi).

Note that in Form 1, each µj has a different constant term, but the same coefficient vector, while in
Form 2, each threshold parameter has its own parameter vector.
Harris and Zhao (2004, 2007) have developed a zero inflated ordered probit (ZIOP)
counterpart to the zero inflated Poisson model. The ZIOP formulation would appear

di* = α′zi + ui, di = 1 (di* > 0),

yi* = b′xi + ei, yi = 0 if yi* < 0 or di = 0,
1 if 0 < yi* < µ1 and di = 1,
2 if µ1 < yi* < µ2 and di = 1,
and so on.
The first equation is assumed to be a probit model (based on the normal distribution) – this estimator
does not support a logit formulation. The correlation between ui and ei is ρ, which by default equals
zero, but may be estimated instead. The latent class nature of the formulation has the effect of
inflating the number of observed zeros, even if u and e are uncorrelated. The model with correlation
between ui and ei is an optional specification that analysts might want to test. The zero inflation
model may also be combined with the hierarchical (generalized) model given above.
The bivariate ordered probit model is analogous to the seemingly unrelated regressions
model for the ordered probit case:
yij* = b j′xji + eij,
yij = 0 if yij* < 0,
1 if 0 < yij* < µ1,
2, ... and so on, j = 1,2,

for a pair of ordered probit models that are linked by Cor(ei1,ei2) = ρ. The model can be estimated
one equation at a time using the results described earlier. Full efficiency in estimation and an
estimate of ρ are achieved by full information maximum likelihood estimation. Either variable (but
not both) may be binary. (If both are binary, the bivariate probit model should be used.) The
polychoric correlation coefficient is used to quantify the correlation between discrete variables that
are qualitative measures. The standard interpretation is that the discrete variables are discretized
counterparts to underlying quantitative measures. We typically use ordered probit models to analyze
such data. The polychoric correlation measures the correlation between y1 = 0,1,...,J1 and y2 = 0,1,...,J2.
(Note, J1 need not equal J2.) One of the two variables may be binary as well. (If both variables are
binary, we use the tetrachoric correlation coefficient described in Section E33.3.) For the case noted,
the polychoric correlation is the correlation in the bivariate ordered probit model, so it can be
estimated just by specifying a bivariate ordered choice model in which both right hand sides contain
only a constant term.
N2: Discrete Choice Models N-24

N2.6 Multinomial Logit Model

The canonical random utility model suggested by the structure of Section N2.2 is as follows:

U(alternative 0) = b 0′xi0 + e i0,

U(alternative 1) = b 1′xi1 + e i1,
...
U(alternative J) = b J ′xiJ + eiJ,
Observed yi = choice j if Ui (alternative j) > Ui (alternative q) ∀ q ≠ j.

The ‘disturbances’ in this framework (individual heterogeneity terms) are assumed to be

independently and identically distributed with identical type 1extreme value distribution; the CDF is

F(ej) = exp(-exp(-ej)).

Based on this specification, the choice probabilities are

Prob(choice j) = Prob(Uj > Uq), ∀ q ≠ j

exp(β′j xij )
= , j = 0,...,J.
∑ exp(β′q xiq )
J
q= 0

At this point we make a purely semantic distinction between two cases of the model. When the
observed data consist of individual choices and (only) data on the characteristics of the individual,
identification of the model parameters will require that the parameter vectors differ across the utility
functions, as they do above. The study on labor market decisions by Schmidt and Strauss (1975) is a
classic example. For the moment, we will call this the multinomial logit model. When the data also
include attributes of the choices that differ across the alternatives, then the forms of the utility
functions can change slightly – and the coefficients can be generic, that is the same across
alternatives. Again, only for the present, we will call this the conditional logit model. (It will
emerge that the multinomial logit is a special case of the conditional logit model, though the reverse
is not true.) The conditional logit model is defined in Section N2.7.
The general form of the multinomial logit model is

exp(β′j xi )
Prob(choice j) = , j = 0,...,J.
∑ exp(β′q xi )
J
q= 0

A possible J + 1 unordered outcomes can occur. In order to identify the parameters of the model, we
impose the normalization b0 = 0. This model is typically employed for individual or grouped data in
which the ‘x’ variables are characteristics of the observed individual(s), not the choices. The data
will appear as follows:

• Individual data: yi coded 0, 1, ..., J,

• Grouped data: yi0, yi1,...,yiJ give proportions or shares,
N2: Discrete Choice Models N-25

The sequential logit model (command SEQLOGIT) is an alternative interpretation of the

multinomial outcome of the choice process. In this specification, the values 0, 1, …, J represent
levels, such as schooling, in which the outcome represents the highest level reached.

N2.6.1 Random Effects and Common (True) Random Effects

The structural equations of the multinomial logit model are

Uijt = b j′xit + eijt, t = 1,...,Ti, j = 0,1,...,J,i=1,...,N,

where Uijt gives the utility of choice j by person i in period t – we assume a panel data application
with t = 1,...,Ti. The model about to be described can be applied to cross sections, where Ti = 1.
Note also that as usual, we assume that panels may be unbalanced. We also assume that eijt has a
type 1 extreme value distribution and that the J random terms are independent. Finally, we assume
that the individual makes the choice with maximum utility. Under these (IIA inducing) assumptions,
the probability that individual i makes choice j in period t is

exp(β′j xit )
Pijt = .
∑ exp(β′q xit )
J
q= 0

We now suppose that individual i has latent, unobserved, time invariant heterogeneity that enters the
utility functions in the form of a random effect, so that

Uijt = b j′xit + uij + eijt, t = 1,...,Ti, j = 0,1,...,J,i=1,...,N.

The resulting choice probabilities, conditioned on the random effects, are

exp(β′j xit + uij )

Pijt | αi1,...,αiJ = .
∑ exp(β′q xit + uiq )
J
q =0

To complete the model, we assume that the heterogeneity is normally distributed with zero means
and (J+1)×(J+1) covariance matrix, Σ. For identification purposes, one of the coefficient vectors,
bq, must be normalized to zero and one of the uiqs is set to zero. We normalize the first element –
subscript 0 – to zero. For convenience, this normalization is left implicit in what follows. It is
automatically imposed by the software. To allow the remaining random effects to be freely
correlated, we write the J×1 vector of nonzero us as

ui = Γvi

where Γ is a lower triangular matrix to be estimated and vi is a standard normally distributed (mean
vector 0, covariance matrix, I) vector.
N2: Discrete Choice Models N-26

N2.6.2 A Dynamic Multinomial Logit Model

The preceding random effects model can be modified to produce the dynamic multinomial
logit model proposed in Gong, van Soest and Villagomez (2000). The choice probabilities are

exp(β′j xit + γ ′j z it + uij )

Pijt | ui1,...,uiJ = t = 1,...,Ti, j = 0,1,...,J,i=1,...,N,
∑ exp(β′q xit + γ ′q z it + uiq )
J
q =1

where zit contains lagged values of the dependent variables (these are binary choice indicators for the
choice made in period t) and possibly interactions with other variables. The zit variables are now
endogenous, and conventional maximum likelihood estimation is inconsistent. The authors argue
that Heckman’s treatment of initial conditions is sufficient to produce a consistent estimator. The
core of the treatment is to treat the first period as an equilibrium, with no lagged effects,

exp(δ′j xi 0 + qij )
Pij0 | qi1,...,qiJ = , t = 0, j = 0,1,...,J,i=1,...,N,
∑ exp(δ′q xi 0 + qiq )
J
q=1

where the vector of effects, q, is built from the same primitives as u in the later choice probabilities.
Thus, ui = Γvi and qi = Φ vi, for the same vi, but different lower triangular scaling matrices. (This
treatment slightly less than doubles the size of the model – it amounts to a separate treatment for the
first period.) Full information maximum likelihood estimates of the model parameters,
(b 1,...,b J,γ1,...,γJ,δ1,...,δJ,Γ,Φ) are obtained by maximum simulated likelihood, by modifying the
random effects model. The likelihood function for individual i consists of the period 0 probability as
shown above times the product of the period 1,2,...,Ti probabilities defined earlier.

N2.7 Conditional Logit Model

If the utility functions are conditioned on observed individual, choice invariant
characteristics, zi, as well as the attributes of the choices, xij, then we write

U(choice j for individual i) = Uij = b′xij + γj′zi + eij, j = 1,...,Ji.

(For this model, which uses a different part of NLOGIT, we number the alternatives 1,...,Ji rather
than 0,...,Ji. There is no substantive significance to this – it is purely for convenience in the context
of the model development for the program commands.) The random, individual specific terms,
(ei1,ei2,...,eiJ) are once again assumed to be independently distributed across the utilities, each with
the same type 1 extreme value distribution

F(eij) = exp(-exp(-eij)).

Under these assumptions, the probability that individual t chooses alternative j is

Prob(Uij > Uiq) for all q ≠ j.

N2: Discrete Choice Models N-27

It has been shown that for independent type 1 extreme value distributions, as above, this probability is

exp ( β′xij + γ ′j z i )
Prob(yi = j) =
∑ exp ( β′xiq + γ ′q z i )
Ji
q=1

where yi is the index of the choice made. We note at the outset that the IID assumptions made about
ej are quite stringent, and induce the ‘Independence from Irrelevant Alternatives’ or IIA features that
characterize the model. This is functionally identical to the multinomial logit model of Section N2.6.
Indeed, the earlier model emerges by the simple restriction γj = 0. We have distinguished it in this
fashion because the nature of the data suggests a different arrangement than for the multinomial logit
model and, second, the models in the section to follow are formulated as extensions of this one. Data
for the choice variable in this context may come in several different forms that require explicit
treatment in the estimation and analysis of the model results:

• Individual data: yi coded 0 (not chosen) or 1 (for the one chosen alternative)
• Market shares: yi coded as 0 < sij < 1 for J alternatives such that Σjsij = 1
• Frequencies: yi coded as Fij > 0 for J alternatives
• Ranks: yi coded 1,2,…, J (with possible ties for last place)
• Best/worst: yi coded 0 (not best or worst), 1 (best) or 2 (worst)

N2.7.1 Fixed Effects

There are several forms of the multinomial choice models that are suitable for choice
experiments, which are essentially panel data sets. The error components logit model (ECLOGIT
command) noted in Section N2.8 is an example that encompasses random effects specifications. The
fixed effects multinomial logit model (FEMLOGIT command) is

Prob(yit = j) = exp ( α jt + β′xitj ) .

∑ q=1 exp ( α qt + β′xitq )
Ji

The estimator for this model builds on Chamberlain’s conditional logit estimator for binary choice,
but uses a far faster (extremely so) algorithm based on the minimum distance estimator.

N2.7.2 Random Regret Logit and Hybrid Utility Models

We consider two direct extensions of the conditional logit model, one related to the forms of
the utility functions and a second related to the treatment of heterogeneity.
The random utility form of the model is based on linear utility functions of the alternatives,

Uijt = b′xit + eijt, t = 1,...,Ti, j = 0,1,...,J,i=1,...,N.

N2: Discrete Choice Models N-28

The random regret form bases the choices at least partly on attribute level regret functions,

Rij(k) = log[1+exp(bk(xjk – xik))]

where k denotes the specific attribute and i and j denote association with alternatives i and j,
respectively. (See Chorus (2010) and Chorus, Greene and Hensher (2013).) The systematic regret
of choice i can then be written

∑j ∑
J K
=Ri = 1=k 1
log[1 + exp(βk ( x jk − xik ))] .

The random regret form of the choice model is then

exp(− R j )
Pj =
∑
J
j =1
exp(− R j )

This model does not impose the IIA assumptions. The model may also be specified with only a
subset of the attributes treated in the random regret format. This hybrid model is

exp(− R j + β′xij )
Pj =
∑ exp(− R j + β′xij )
J
j =1

The random regret model is also extended to the latent class framework.

N2.7.3 Scaled MNL Model

The scaled multinomial logit model allows the model to accommodate broad heterogeneity
across individuals, for example when two or more data sets from different groups are combined.
This is a special case of the generalized mixed logit model described in Section N2.11.2. The
general form of the scaled MNL model is

exp ( σi β′xij )
Prob(yi = j) =
∑ exp ( σi β′xiq )
Ji
q=1

where σi = exp(δ′zi + τvi)

The scaling factor, σi differs across individuals, but not across choices. It has a deterministic
component, exp(δ′zi), and a random component, exp(τvi). Either (or both) may equal 1.0, that is, either
or both restrictions δ = 0 or τ = 0. For example, a simple nonstochastic scaling differential between
two groups would result if τ = 0 and if zi were simply a dummy variable that identifies the two groups.
Other forms of scaling heterogeneity can be produced by different variables in zi. The scaling may also
be random through the term τvi. In this instance, vi is a random term (usually, but not necessarily
normally distributed). With δ = 0 and τ ≠ 0, we obtain a randomly scaled multinomial logit model.
N2: Discrete Choice Models N-29

N2.8 Error Components Logit Model

When the sample consists of a ‘panel’ of data, that is, when individuals are observed in more
than one choice situation, the conditional logit model can be augmented with individual effects,
similar to the use of common effects models in regression and other single equation cases. A ‘panel
data’ form of this model that is a counterpart to the random effects model is what we label the ‘error
components model.’ (This has been called the ‘kernel logit model’ in some treatments in the
literature.) The model arises by introducing M up to maxi Ji alternative and individual specific
random terms in the utility functions as in

U(choice j for individual i in choice setting t)

= Uijt
= b′xij + γj′zi + eij + Σ m =1d jm σ m uim , j = 1,...,Ji, t = 1,...,Ti.
M

where djm = 1 if effect m appears in utility function j, 0 if not,

σm = the standard deviation of effect m (to be estimated),
vim = effect m for individual i.

The M random individual specifics are σmuim. They are distributed as normal with zero means and
variances σm2. The constants djm equal one if random effect m appears in the utility function for
alternative j, and zero otherwise. The error components account for unobserved, alternative specific
variation. With this device, the sets of random effects in different utility functions can overlap, so as
to accommodate correlation in the unobservables across choices. The random effects may also be
heteroscedastic, with

σm,i2 = σm2 exp(qm′zi).

The probabilities attached to the choices are now

exp ( β′xij + γ ′j z i + Σ mM=1d jm σ m uim )

Prob(yi = j) = .
∑ exp ( β′xiq + γ ′q z i Σ mM=1d qm σ m uim )
Ji
q =1

This is precisely an analog to the random effects model for single equation models. Given the
patterns of djm, this can provide a nesting structure as well. Examples in Chapter N30 will
demonstrate.
N2: Discrete Choice Models N-30

N2.9 Heteroscedastic Extreme Value Model

In the conditional logit model,

U(choice j for individual i) = Uij = b′xij + γj′zi + eij, j = 1,...,Ji,

exp ( β′xij + γ ′j z i )
Prob(yi = j) = ,
∑ m=1 exp (β′xim + γ ′m z i )
Ji

an implicit assumption is that the variances of eji are the same. With the type 1 extreme value
distribution assumption, this common value is π2/6. This assumption is a strong one, and it is not
necessary for identification or estimation. The heteroscedastic extreme value model relaxes this
assumption. We assume, instead, that

F(eij) = exp(-exp(-qjeij)],

Var[eij] = σj2 (π2/6) where σj2 = 1/qj2,

with one of the variance parameters normalized to one for identification. (Technical details for this
model including a statement of the probabilities appears in Chapter N26.) A further extension of this
model allows the variance parameters to be heterogeneous, in the standard fashion,

σij2 = σj2 exp(γ′zi).

N2.10 Nested and Generalized Nested Logit Models

The nested logit model is an extension of the conditional logit model. The models supported
by NLOGIT are based on variations of a four level tree structure such as the following:

ROOT root
│
┌───────────────┴────────────────┐
│ │
TRUNKS trunk1 trunk2
│ │
┌───────┴───────┐ ┌────────┴──────┐
│ │ │ │
LIMBS limb1 limb2 limb3 limb4
│ │ │ │
┌───┴───┐ ┌───┴───┐ ┌───┴───┐ ┌───┴───┐
│ │ │ │ │ │ │ │
BRANCHES branch1 branch2 branch3 branch4 branch5 branch6 branch7 branch8
│ │ │ │ │ │ │ │
┌─┴─┐ ┌─┴─┐ ┌─┴─┐ ┌─┴─┐ ┌─┴─┐ ┌─┴─┐ ┌─┴─┐ ┌─┴─┐
│ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │
ALTS a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 a13 a14 a15 a16
N2: Discrete Choice Models N-31

The choice probability under the assumption of the nested logit model is defined to be the
conditional probability of alternative j in branch b, limb l, and trunk r, j|b,l,r:

exp(b′x j|b ,l , r ) exp(b′x j|b ,l ,r )

P(j|b,l,r) = = ,
∑ q|b,l ,r exp(b′xq|b,l ,r ) exp( J b|l , r )

where Jb|l,r is the inclusive value for branch b in limb l, trunk r, Jb|l,r = log Σq|b,l,rexp(b′xq|b,l,r). At the
next level up the tree, we define the conditional probability of choosing a particular branch in limb l,
trunk r,
exp(α′y b|l , r + τb|l , r J b|l , r ) exp(α′y b|l , r + τb|l , r J b|l , r )
P(b|l,r) = = ,
∑ s|l ,r exp(α′y s|l ,r + τs|l ,r J s|l ,r ) exp( I l |r )

where Il|r is the inclusive value for limb l in trunk r, Il|r = log Σs|l,rexp(α′ys|l,r + τs|l,rJs|l,r). The
probability of choosing limb l in trunk r is

exp(δ′z l |r + sl |r I l |r ) exp(δ′z l |r + sl |r I l |r )
P(l|r) = = ,
∑ s|r exp(δ′z q|r + ss|r I s|r ) exp( H r )

where Hr is the inclusive value for trunk r, Hr = log Σs|lexp(δ′zs|r + σs|rIs|r).

Finally, the probability of choosing a particular limb is

exp(θ′h r + φr H r )
P(r) = .
∑ s exp(θ′h s + φs H s )
By the laws of probability, the unconditional probability of the observed choice made by an
individual is
P(j,b,l,r) = P(j|b,l,r) × P(b|l,r) × P(l|r) × P(r).

This is the contribution of an individual observation to the likelihood function for the sample.
The ‘nested logit’ aspect of the model arises when any of the τb|l,r or σl|r or φr differ from 1.0.
If all of these deep parameters are set equal to 1.0, the unconditional probability reduces to

exp(b′x j|b ,l , r + α′y b|l , r + δ′z l |r + θ′h r )

P(j,b,l,r) = ,
∑ ∑ r l ∑ b ∑ j exp(b′x j ,b,l ,r + α′y b,l ,r + δ′z l ,r + θ′h r )
which is the probability for a one level conditional (multinomial) logit model.
N2: Discrete Choice Models N-32

N2.10.1 Alternative Normalizations of the Nested Logit Model

The formulation of the nested logit model imposes no restrictions on the inclusive value
parameters. However, the assumption of utility maximization and the stochastic underpinnings of
the model do imply certain restrictions. For the former, in principle, the inclusive value parameters
must be between zero and one. For the latter, the restrictions are implied by the way that the random
terms in the utility functions are constructed. In particular, the nesting aspect of the model is
obtained by writing
ej|b,l,r = uj|b,l,r + vb|l,r.

That is, within a branch, the random terms are viewed as the sum of a unique component, uj|b,l,r, and a
common component, vb|l,r. This has certain implications for the structure of the scale parameters in
the model. NLOGIT provides a method of imposing the restrictions implied by the underlying
theory.
There are three possible normalizations of the inclusive value parameters which will produce
the desired results. These are provided in this estimator for two and three level models only. This
includes most of the received applications. We will detail the first two of these forms here and
describe how to estimate all of them in Chapter N28. For convenience, we label these random utility
formulations RU1, RU2 and RU3. (RU3 is just a variant of RU2.)

RU1

The first form is

exp(b′x j|b ,l ) exp(b′x j|b ,l )
P(j|b,l) = = ,
∑ q|b,l exp(b′xq| j ,l ) exp( J b|l )

where Jb|l is the inclusive value for branch b in limb l,

Jb|l = log Σq|b,l exp(b′xq|b,l).

At the next level up the tree, we define the conditional probability of choosing a particular branch in
limb l,
exp l b|l (α′y b|l + J b|l )  exp l b|l (α′y b|l + J b|l ) 
P(b|l) = = ,
∑ s|l exp l s|l (α′y s|l + J s|l )  exp( I l )

where Il is the inclusive value for limb l,

Il = log Σs|l exp[ls|l (α′ys|l + Js|l)].

The probability of choosing limb l is

exp[ γ l (δ′z l | + I l )] exp[ γ l (δ′z l + I l )]

P(l) = = .
∑ s exp [ γ s (δ′z s + I s )] exp( H )
N2: Discrete Choice Models N-33

Note that this the same as the familiar normalization used earlier; this form just makes the scaling
explicit at each level.

RU2

The second form moves the scaling down to the twig level, rather than at the branch level.
Here it is made explicit that within a branch, the scaling must be the same for alternatives.

exp µb|l (b′x j|b ,l )  exp µb|l (b′x j|b ,l ) 

P(j|b,l) = = .
∑ q|b ,l
exp µb|l (b′x q|b ,l )  exp( J b|l )

Note in the summation in the inclusive value that the scaling parameter is not varying with the
summation index. It is the same for all twigs in the branch. Now, Jb|l is the inclusive value for
branch j in limb l,

Jb|l = log Σq|b,l exp[µb|l (b′xq|b,l)].

At the next level up the tree, we define the conditional probability of choosing a particular branch in
limb l,

exp  γ l ( α′y b|l + (1/ µb|l ) J b|l )  exp  γ l ( α′y b|l + (1/ µb|l ) J b|l ) 
P(b|l) = = ,
∑ s exp  γ s ( α′y s|l + (1/ µ s|l ) J s|l ) exp( I l )

where Il is the inclusive value for limb l,

Il = log ∑ s|l exp  g l ( α ' y s|l + (1/ µ s|l ) J s|l )  .

Finally, the probability of choosing limb l is

exp [ δ′z l + (1/ γ l ) I l ] exp [ δ′z l + (1/ γ l ) I l ]

P(l) = = ,
∑ s exp [δ′z s + (1/ γ s ) I s ] exp( H )

where the log sum for the full model is

H = log ∑ s exp [ δ ' z s + (1/ g s ) I s ] .

N2: Discrete Choice Models N-34

N2.10.2 A Model of Covariance Heterogeneity

This is a modification of the two level nested logit model. The base case for the model is

exp(b′x j|b )
P ( j | b) = .
∑ exp(b′x q|b )
J |b
q =1

Denote the logsum, the log of the denominator, as Jb = inclusive value for branch b = IV(b). Then,

exp(α′y b + τb J b )
P (b) = .
∑ exp(α′y s + τs J s )
B
s =1

The covariance heterogeneity model allows the τb inclusive value parameters to be functions of a set
of attributes, vb , in the form

τb* = τb × exp(δ′vb),

where δ is a new vector of parameters to be estimated. Since the inclusive parameter is a scaling
parameter for a common random component in the alternatives within a branch, this is equivalent to
a model of heteroscedasticity.

N2.10.3 Generalized Nested Logit Model

The generalized nested logit model is an extension of the nested logit model in which
alternatives may appear in more than one branch. Alternatives that appear in more than one branch
are allocated across branches probabilistically. The model estimated includes the usual nested logit
framework (only two levels are supported in this framework), as well as the matrix of allocation
parameters. The only difference between this and the more basic nested logit model is the
specification of the tree. For the allocations of choices to branches, a multinomial logit form is used,

πj,b = Prob(alternative j is in branch b) = exp(qj,b) / Σs exp(qj,s),

where the parameters q are estimated by the program. Note the denominator summation is over
branches that the alternative appears in. The probabilities sum to one. The identification rule that
one of the qs for each alternative modeled equals one is imposed. These allocations may depend on
an individual characteristic (not a choice attribute), such as income. In this instance, the multinomial
logit probabilities become functions of this variable,

πj,b = Prob(alternative j is in branch b) = exp(qj,b + γj,bzi ) / Σs exp(qj,s+ γj,szi).

Now, to achieve identification, one of the qs is set equal to zero and one of the γs is set equal to zero.
It is convenient to form the matrix Π = [πj,b]. This is a J×B matrix of allocation parameters. The
rows sum to one, and note that some values in the matrix are zero. But, no rows have all zeros –
every alternative appears in at least one branch, and no columns have all zeros – every branch
contains at least one alternative.
N2: Discrete Choice Models N-35

The probabilities for the observed choices are formed as

Prob(alternative, branch) = P(j,b)

= P(j|b) × P(b)

[π j ,bU j ]sb
where P ( j | b) =
∑ [π j , sU s ]ss
B
s =1

(the denominator summation is over the alternatives in that branch)

1/ σb
 ∑ [π j ,bU j ]σb 
P (b) =  
j |b
and 1/ σb
.
∑ b=1 ∑ j|b [π j ,bU j ] 
 σb 
B

N2.10.4 Box-Cox Nested Logit

The Box-Cox form of the nested logit model automates a model specification that was
already in NLOGIT 4. This form can replace the function transformation BCX(variable) in the utility
functions.

N2.11 Random Parameters Logit Models

In its most general form, we write the multinomial logit probability as

exp(α ji + θ′j z i + fβ
′j f ji + ′ji x ji )
P( j | vi ) = ,
∑ exp(α qi + θ′q z i + fβ
′ ′
q f qi + qi x qi )
J
q =1

where ′j z i + ′j f ji + ′ji x ji , j = 1,...,Ji alternatives in individual i’s choice set

U(j,i) = α ji + θfβ

αji is an alternative specific constant which may be fixed or random, αJi = 0,

qj is a vector of nonrandom (fixed) coefficients, qJi = 0,
φj is a vector of nonrandom (fixed) coefficients,
bji is a coefficient vector that is randomly distributed across individuals;
vi enters b ji,
zi is a set of choice invariant individual characteristics such as age or income,
fji is a vector of M individual and choice varying attributes of choices,
multiplied by φj,
xji is a vector of L individual and choice varying attributes of choices,
multiplied by b ji.
N2: Discrete Choice Models N-36

The term ‘mixed logit’ is often used in the literature (e.g., Revelt and Train (1998)) for this model. The
choice specific constants, αji and the elements of bji are distributed randomly across individuals such
that for each random coefficient, ρki = any (not necessarily all of) αji or bjki, the coefficient on attribute
xjik, k = 1,...,K,
ρjki = αji or bjki = ρjk + δk′wi + σkvki,
or ρjki = αji or bjki = exp(ρjk + δjk′wi + σjkvjki).

The vector wi (which does not include one) is a set of choice invariant characteristics that produce
individual heterogeneity in the means of the randomly distributed coefficients; ρjk is the constant
term and δjk is a vector of ‘deep’ coefficients which produce an individual specific mean. The
random term, vjki is normally distributed (or distributed with some other distribution) with mean 0
and standard deviation 1, so σjk is the standard deviation of the marginal distribution of ρjki. The vjkis
are individual and choice specific, unobserved random disturbances – the source of the
heterogeneity. Thus, as stated above, in the population

αji or bjki ~ Normal or Lognormal [ρjk + δjk′wi, σjk2].

(Other distributions may be specified.) For the full vector of K random coefficients in the model, we
may write
ρi = ρ + Dwi + Γvi

where Γ is a diagonal matrix which contains σk on its diagonal. A nondiagonal Γ allows the random
parameters to be correlated. Then, the full covariance matrix of the random coefficients is Σ = ΓΓ′.
The standard case of uncorrelated coefficients has Γ = diag(σ1,σ2 ,…,σk). If the coefficients are
freely correlated, Γ is a full, unrestricted, lower triangular matrix and Σ will have nonzero off
diagonal elements. An additional level of flexibility is obtained by allowing the distributions of the
random parameters to be heteroscedastic,

σijk2 = σjk2 × exp(γjk′hi).

This is now built into the model by specifying

ρi = ρ + Dwi + Γ Ωi vi
where Ωi = diag[σijk2]

and now, Γ is a lower triangular matrix of constants with ones on the diagonal. Finally,
autocorrelation can also be incorporated by allowing the random components of the random
parameters to obey an autoregressive process,

vki,t = τki vki,t-1 + cki,t

where cki,t is now the random element driving the random parameter.
N2: Discrete Choice Models N-37

This produces, then, the full random parameters logit model

exp(α ji + β′i x ji )
P( j | vi ) = ,
∑ exp(α mi + β′i x mi )
J
m =1

b i = b + Dzi + Γ Ωi vi
vi ~ with mean vector 0 and covariance matrix I.

The specific distributions may vary from one parameter to the next. We also allow the parameters to
be lognormally distributed so that the preceding specification applies to the logarithm of the specific
parameter.

N2.11.1 Nonlinear Utility RP Model

In the NLRP model, the model parameters may be specified as in the model above. But, the
utility functions need not be linear in the attributes and characteristics. This more general model is

exp[U j (β′i , x ji )]
P( j | vi ) = ,
∑ exp[U j (β′i , x ji )]
J
m =1

where b i = b + Dzi + Γ Ωi vi
vi ~ with mean vector 0 and covariance matrix I.
and U j (β′i , x ji ) is any nonlinear function of the data and parameters.

N2.11.2 Generalized Mixed Logit Model

The second major extension of the random parameters model is the generalized mixed logit
model developed by Fiebig, Keane, Louviere and Wasi (2010). The extension of the random
parameters model is
bi = σib + γΓvi + (1 - γ)σiΓvi

The generalized mixed logit model embodies several different forms of heterogeneity in the random
parameters and random scaling, as well as the distribution parameter, γ, which allocates the influence
of the parameter heterogeneity and the scaling heterogeneity. Several interesting model forms are
produced by different restrictions on the parameters. For example, if γ = 0 and Γ = 0, we obtain the
scaled MNL model in Section N2.7.3. A variety of other special cases are also provided. One
nonlinear normalization in particular allows the model to be transformed from a specification in
‘utility space’ as above to ‘willingness to pay space’ by analyzing an implicit ratio of coefficients.
N2: Discrete Choice Models N-38

N2.12 Latent Class Logit Models

In the latent class formulation, parameter heterogeneity across individuals is modeled with a
discrete distribution, or set of ‘classes.’ The situation can be viewed as one in which the individual
resides in a ‘latent’ class, c, which is not revealed to the analyst. There are a fixed number of
classes, C. Estimates consist of the class specific parameters and for each person, a set of
probabilities defined over the classes. Individual i’s choice among J alternatives at choice situation t
given that individual i is in class c is the one with maximum utility, where the utility functions are

Ujit|c = bc′xjit + ejit

where Ujit = utility of alternative j to individual i in choice situation t
xjit = union of all attributes that appear in all utility functions. For
some alternatives, xjit,k may be zero by construction for some
attribute k which does not enter their utility function for
alternative j.
ejit = unobserved heterogeneity for individual i and alternative j in
choice situation t.
bc = class specific parameter vector.

Within the class, choice probabilities are assumed to be generated by the multinomial logit model

exp ( β′c x jit )

Prob[yit = j | class = c] = .
∑ exp ( β′c x jit )
Ji
j =1

As noted, the class is not observed. Class probabilities are specified by the multinomial logit form,

exp ( θ′c z i )
Prob[class = c] = Qic = , qC = 0.
∑ exp ( θ′c z i )
C
c=1

where zi is an optional set of person, situation invariant characteristics. The class specific
probabilities may be a set of fixed constants if no such characteristics are observed. In this case, the
class probabilities are simply functions of C parameters, qc, the last of which is fixed at zero. This
model does not impose the IIA property on the observed probabilities.
For a given individual, the model’s estimate of the probability of a specific choice is the
expected value (over classes) of the class specific probabilities. Thus,

 exp ( β′ x ) 
Prob(yit = j) = Ec  Ji 
c jit

 ∑ exp ( β′c x jit ) 

 j =1 

 exp ( b′ x ) 
∑ c =1  .
C
=
c jit
= Prob(class c )
 ∑ Ji exp ( b′c x jit ) 
 j =1 
N2: Discrete Choice Models N-39

N2.12.1 2K Latent Class Model for Attribute Nonattendance

NLOGIT accommodates attribute ‘nonattendance’ by the ‘-888’ feature described in Chapter
N18. In particular, in some choice analyses, some, but not all individuals indicate that they did not
pay attention to certain attributes. The appropriate model building strategy is to impose zero
restrictions on the utility parameters, b, for these specific individuals. NLOGIT provides this
capability throughout the estimation suite – all models are fit with this capability. (This feature is
unique to NLOGIT.) This feature accommodates cases in which individuals explicitly reveal the
form of their utility functions. The model noted here is usable when the sorting of individuals in this
way is latent – there is no observed indicator. Consider a model with four attributes, x1, x2, x3, x4.
All individuals attend to x1 and x2. Some ignore x3, some ignore x4, and some ignore both x3 and
x4 (and some attend both). Thus, in terms of the possible utility functions, there are four types of
individuals in the population, distinguished by the type of utility function that is appropriate:

(x3 and x4) Uij = b1x1 + b2x2 + b3x3 + b4x4 +ε

(x3 only) Uij = b1x1 + b2x2 + b3x3 +ε
(x4 only) Uij = b1x1 + b2x2 + b4x4 +ε
(Neither) Uij = b1x1 + b2x2 +ε

The difference that is built into this model form is that the analyst does not know which individual is
in which group. This can be treated as a latent class model. The number of classes is 2K where K is
the number of attributes that treated by the latent class specification.

N2.12.2 Latent Class – Random Parameters Model

The LCRP model is a combination of the latent class model described above and the random
parameters model in Section N2.11. This is a latent class model in which a random parameters
model applies within each class.

N2.13 Multinomial Probit Model

In this model, the individual’s choice among J alternatives is the one with maximum utility,
where the utility functions are

Uji = b′xji + eji

where Uji = utility of alternative j to individual i
xjit = union of all attributes that appear in all utility functions. For
some alternatives, xjit,k may be zero by construction for some
attribute k which does not enter their utility function for
alternative j.
N2: Discrete Choice Models N-40

The multinomial logit model specifies that eji are draws from independent extreme value
distributions (which induces the IIA condition). In the multinomial probit model, we assume that eji
are normally distributed with standard deviations Sdv[eji] = σj and correlations Cor[eji, eqi] = ρjq (the
same for all individuals). Observations are independent, so Cor[eji,eqs ] = 0 if i is not equal to s, for
all j and q. A variation of the model allows the standard deviations and covariances to be scaled by a
function of the data, which allows some heteroscedasticity across individuals.
The correlations ρjq are restricted to -1 < ρjq < 1, but they are otherwise unrestricted save for
a necessary normalization. The correlations in the last row of the correlation matrix must be fixed at
zero. The standard deviations are unrestricted with the exception of a normalization – two standard
deviations are fixed at 1.0 – NLOGIT fixes the last two.
This model may also be fit with panel data. In this case, the utility function is modified as
follows:
Uji,t = b′xji,t + eji,t + vji,t

where ‘t’ indexes the periods or replications. There are two formulations for vji,t,

Random effects vji,t = vji,t (the same in all periods)

First order autoregressive vji,t = αj vji,t-1 + aji,t.

It is assumed that you have a total of Ti observations (choice situations) for person i. Two situations
might lend themselves to this treatment. If the individual is faced with a set of choice situations that
are similar and occur close together in time, then the random effects formulation is likely to be
appropriate. However, if the choice situations are fairly far apart in time, or if habits or knowledge
accumulation are likely to influence the latter choices, then the autoregressive model might be the
better one.
You can also add a form of individual heterogeneity to the disturbance covariance matrix.
The model extension is

Var[ei] = exp[γ′hi] × Σ

where Σ is the matrix defined earlier (the same for all individuals), and hi is an individual (not
alternative) specific set of variables not including a constant.
N3: Model and Command Summary for Discrete Choice Models N-41

N3: Model and Command Summary for

Discrete Choice Models
N3.1 Introduction
The chapters to follow will provide details on the various discrete choice models you can
estimate with NLOGIT and on the model commands you will use to request the estimates. This
chapter will provide a brief summary listing of the models and model commands. The variety of
logit models now use a set of specific names, rather than qualifiers to more general model classes as
in earlier versions. For example, the model name OLOGIT can be used instead of ORDERED ;
Logit. The earlier formats remain available, but the newer ones may prove more convenient. The
full listing of these commands is also given below. The commands below specify the essential parts
needed to fit the model. The numerous options and different forms are discussed in the chapters to
follow (and, were noted in the LIMDEP Econometric Modeling Guide as well).

N3.2 Model Dimensions

The descriptions below present the different discrete choice models that are the main feature
of NLOGIT. NLOGIT contains all of LIMDEP, so all of the models documented in the LIMDEP
Econometric Modeling Guide, including the regression models, limited dependent variable models,
generalized linear models, sample selection models, and so on are supported in NLOGIT, as well as
the ancillary tools including MATRIX, etc.
There are various built in limits in the estimators. These are noted at the specific points
below where necessary. The following lists the most important internal constraints on the
estimators:

• Multinomial choice model estimators in NLOGIT: maximum numbers of:

° Alternatives 500
° Attributes 300
° Branches in nested logit models 25
° Limbs in nested logit models 10
° Random error components 10
• Maximum number of choices in the MLOGIT form of the model 25
• Heteroscedasticity models, maximum number of variables 75
• Ordered choice models: maximum number of outcomes 25
• Unconditional fixed effects models, number of individuals 100,000
• Random parameters models, maximum number of RPs 25
• Latent class models, maximum number of classes 30
N3: Model and Command Summary for Discrete Choice Models N-42

N3.3 Basic Discrete Choice Models

The binomial probit and logit models and the ordered probit and logit models are the primary
model frameworks for single equation, single decision, discrete choice models. The ordered choice
and the bivariate and multivariate probit models are multivariate extensions of the simple probit model.

N3.3.1 Binary Choice Models

There are six binary choice models, probit, logit, complementary log log, Gompertz, Burr,
and arctangent documented in Chapter E27. The ones that interest us here are the binary probit and
logit models. The probit model is requested with
PROBIT ; Lhs = dependent variable
; Rhs = independent variables $
The binary logit model may be invoked with
BLOGIT ; Lhs = dependent variable
; Rhs = independent variables $
In earlier versions, you would use the LOGIT command, which is still useable. LOGIT is the same
as BLOGIT when the data on the dependent variable are either binary (zeros and ones) or
proportions (strictly between zero and one). Chapters E26-E29 document numerous extensions of
these models. Chapters E30-E32 consider semiparametric and nonparametric approaches and
extensions of the binary choice models for panel data.

N3.3.2 Bivariate Binary Choices

The command for the bivariate probit model is

BVPROBIT ; Lhs = variable 1, variable 2

; Rh1 = independent variables for equation 1
; Rh2 = independent variables for equation 2 $
In this form, the Lhs specifies two binary dependent variables. You may use proportions data
instead, in which case, you will provide four proportions variables, in order, p00, p01, p10, p11.
This command is the same as BIVARIATE PROBIT in earlier versions. (You may still use
BIVARIATE PROBIT.)

N3.3.3 Multivariate Binary Choice Models

The multivariate probit model is specified with
MVPROBIT ; Lhs = y1, y2, ..., yM
; Eq1 = Rhs variables for equation 1
; Eq2 = Rhs variables for equation 2
...
; EqM = Rhs variables for equation M $
Data for this model must be individual. The Lhs specifies a set of binary dependent variables. This
command is the same as MPROBIT (which may still be used) in earlier versions.
N3: Model and Command Summary for Discrete Choice Models N-43

N3.3.4 Ordered Choice Models

Chapter E34 describes five forms for the ordered choice model, probit, logit, complementary
log log, Gompertz and arctangent. The first two interest us here. The ordered probit model is
requested with

OPROBIT ; Lhs = dependent variable

; Rhs = independent variables $

This is the same as the ORDERED PROBIT command, which may still be used. In this model, the
dependent variable is integer valued, taking the values 0, 1, ..., J. All J+1 values must appear in the
data set, including zero. You may supply a set of J+1 proportions variables instead. Proportions will
sum to 1.0 for every observation. Chapter E35 documents a bivariate version of the ordered probit
model for two joint ordered outcomes, and a sample selection model.
The ordered logit model is requested with

OLOGIT ; Lhs = dependent variable

; Rhs = independent variables $

The same arrangement for the dependent variables as for the ordered probit model is assumed. This
command is the same as ORDERED ; Logit in earlier versions.

N3.4 Multinomial Logit Models

The ‘multinomial logit model’ is a special case of the conditional logit model, which, itself,
is the gateway model to the main model extensions described in Chapter N2.

N3.4.1 Multinomial Logit

The multinomial logit model described in Section N2.6 and Chapter E37 is invoked with

MLOGIT ; Lhs = dependent variable

; Rhs = independent variables $

Data for the MLOGIT model consist of an integer valued variable taking the values 0, 1, ..., J. This
model may also be fit with proportions data. In that case, you will provide the names of J+1 Lhs
variables that will be strictly between zero and one, and will sum to one at every observation. The
MLOGIT command is the same as LOGIT. The program inspects the command (Lhs) and the data,
and determines internally whether BLOGIT or MLOGIT is appropriate. Note, on proportions data,
if you want to fit a binary logit model with proportions data, you will supply a single proportions
variable, not two. (What would be the second one is just one minus the first.) If you want to fit a
multinomial logit model with proportions data with three or more outcomes, you must provide the
full set of proportions. Thus, you would never supply two Lhs variables in a LOGIT, BLOGIT or
MLOGIT command. Three other forms of this canonical model are the sequential logit model,
SEQLOGIT, and two forms for panel data, REMLOGIT for random effects and FEMLOGIT for
fixed effects.
N3: Model and Command Summary for Discrete Choice Models N-44

N3.4.2 Conditional Logit

The command for the conditional model, and the commands in the sections to follow, are
variants of the NLOGIT command. This is a full class of estimators based on the conditional logit
form. There are several forms of the essential command for fitting the conditional logit model with
NLOGIT. The simpler one is

CLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $

As discussed in Chapter N20 and in Section E38.3, the data for this estimator consist of a set of J
observations, one for each alternative. (The observation resembles a group in a panel data set.) The
command just given assumes that every individual in the sample chooses from the same size choice
set, J. The choice sets may have different numbers of choices, in which case, the command is
changed to
; Lhs = dependent variable, choice set size variable

The second Lhs variable is structured exactly the same as a ; Pds variable for a panel data estimator.
In the second form of the model command, the utility functions are specified directly, symbolically.
The ; Rhs and ; Rh2 specifications can be replaced with

; Model: ... specification of the utility functions

This is discussed in Chapter N21 and Chapter E39.

The CLOGIT command is the same as DISCRETE CHOICE. It is also the same as
NLOGIT when the only information given in the command is that specified above, that is when
none of the specifications that invoke the model extensions that are described in the sections to
follow are provided.

N3.5 NLOGIT Extensions of Conditional Logit

The conditional logit model provides the basic framework for a very large number of
extensions that are provided by NLOGIT. The following lists the basic commands for most of these.
Each model form is developed in greater detail in one of the chapters that follow. Each model may
be specified with a variety of options and different specifications for numerous variants. The
following shows the essential command for the most basic form of the model.

N3.5.1 Random Regret Logit

The random regret form of the model is specified with
RRLOGIT ; Lhs = dependent variable
; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $
N3: Model and Command Summary for Discrete Choice Models N-45

The command is otherwise the same as CLOGIT, with the same formats for variable choice set sizes,
etc. The utility functions must be specified as above, not using ; Model: …, owing to the particular
form of the utility functions in the random regret format. The random regret logit model can be framed
in several forms, including latent class, random parameters and best/worst multinomial logit.

N3.5.2 Scaled Multinomial Logit

The scaled multinomial logit model is a randomly scaled MNL, with b i = σib, where σi is a
heterogeneous scalar. The model command is

SMNLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $

N3.5.3 Heteroscedastic Extreme Value

The heteroscedastic extreme value model is requested with the command
HLOGIT ; Lhs = dependent variable
; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $
The command is otherwise the same as the CLOGIT command, with the same formats for variable
choice set sizes and utility function specifications. The HLOGIT command is the same as

NLOGIT ; Heteroscedasticity
; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $
as used in earlier versions of NLOGIT. (This may still be used if desired.)

N3.5.4 Error Components Logit and Fixed Effects

The error components model is described in Section N2.8 and in Chapter N30. The model
command is

ECLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; ECM = specification of the tree structure for the error components $
This command is the same as NLOGIT ; ECM = specification ... $ The error components model
may also be specified as a part of the random parameters model. Thus, your RPLOGIT command
may also contain the ; ECM = specification. The error components logit model is a random effects
model if each alternative (save for one) has a separate choice situation invariant random term. The
default specification is consistent with a random effects model.
N3: Model and Command Summary for Discrete Choice Models N-46

A ‘fixed effects’ specification,

Uijt = αij + xijt'β + εijt

is also provided as the FEMLOGIT command.

N3.5.5 Nested and Generalized Nested Logit

The nested logit model is the default form of the NLOGIT command. Request the nested
logit model with

NLOGIT ; Tree = specification of the tree structure

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $

The generalized nested logit model command is

GNLOGIT ; Tree = specification of the tree structure

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $

The GNLOGIT command in place of the NLOGIT command tells NLOGIT that the tree structure
may have overlapping branch specifications. (You may also use NLOGIT ; GNL.) If you specify
that alternatives appear in more than one branch in the NLOGIT command, this will produce an
error message. The option is available only for the GNLOGIT command. The specification of
variable choice set sizes and utility functions is the same as for the CLOGIT command.

N3.5.6 Random Parameters Logit

The random parameters logit model (mixed logit model) is requested by specifying a
conditional logit model, and adding the specification of the random parameters. The model command is

RPLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Fcn = the specifications of the random parameters
; ... other specifications for the random parameters model $

Once again, variable choice set sizes and utility function specifications are specified as in the
CLOGIT command. This command is the same as

NLOGIT ; RPL
; ... the rest of the command $
N3: Model and Command Summary for Discrete Choice Models N-47

There is one modification that might be necessary. If you are providing variables that affect the
means of the random parameters, you would generally use

NLOGIT ; RPL = the list of variables

; ... the rest of the command $

The RPL specification may still be used this way. The command can be NLOGIT as above, or

RPLOGIT ; RPL = the list of variables

; ... the rest of the command $

These are identical.

The random parameters model may also include an error components specification defined
in the next section. The command will be

RPLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Fcn = the specifications of the random parameters
; ... other specifications for the random parameters model
; ECM = specification $

There are many variants of the random parameters logit model supported in NLOGIT. The most
flexible is the generalized mixed logit model described in the next section. Three somewhat
narrower, specific forms are

Best/worst multinomial logit: BWRPLOGIT ; …

Random regret multinomial logit: RPRRLOGIT
Berry, Levinsohn, Pakes (BLP) logit: BLPLOGIT

The logit model of Berry, Levinsohn and Pakes is used for aggregate, market share data. NLOGIT
uses a new algorithm (see Lee and Seo (2015)) for estimation of the BLP model that promises to be
much faster than the conventional contraction mapping approach.

N3.5.7 Generalized Mixed Logit

The generalized mixed logit model is an extension of the random parameters model. The
command has several parts that produce the various model types. The essential command is

GMXLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Fcn = specification of the random parameters $
N3: Model and Command Summary for Discrete Choice Models N-48

N3.5.8 Nonlinear Random Parameters Logit

This command extends the random parameters model by allowing the utility functions to be
any nonlinear that you specify. There are numerous variants of this model. The essential command
is
NLRPLOGIT ; Lhs = dependent variable
; Choices = the names of the J alternatives
; Labels = the labels used for the model parameters
; Start = starting values for iterations
; Fn1 = specification of a nonlinear function
; … up to 50 nonlinear function specifications
; Model: U(name…) = one of the nonlinear functions defined /
U(name…) = another one of the functions, etc.
; Fcn = specifications of the random parameters $

The model is set up by defining the choice variable and a set of nonlinear functions that will be
combined to make the utility functions. The functions may be arbitrarily complex

N3.5.9 Latent Class Logit

The essential form of the command for the latent class model is

LCLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Pts = the number of classes $

Like the RPLOGIT command, you need to modify this command if you are providing variables that
affect the class probabilities. You would generally use

NLOGIT ; LCM = the list of variables

; ... the rest of the command $

The LCM specification may still be used this way. The command can be NLOGIT as above, or
identically,

LCLOGIT ; LCM = the list of variables

; ... the rest of the command $

The default framework for the latent class model is random utility. The model may be changed to
random regret with model command LCRRLOGIT. Another variant of the latent class model may
have a mix of random regret based classes and maximum random utility classes.
N3: Model and Command Summary for Discrete Choice Models N-49

N3.5.10 2K Latent Class Logit

The 2K model is a particular latent class model in which there are simple constraints across
the classes, but only one parameter vector used for the whole model. The model is set up as a latent
class model with an additional specification:

LCLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Pts = the number of classes $

In this form of the model, the number of points is specified as 102, 103, or 104, corresponding to
whether the first 2, 3, or 4 variables in the RHS list are given the special treatment that defines the
model.

N3.5.11 Latent Class Random Parameters

The latent class random parameters model extends the latent class model. The essential
command is

LCRPLOGIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics
; Fcn = definition of the random parameters part
; Pts = the number of classes $

N3.5.12 Multinomial Probit

The multinomial probit model is described in Chapter N27 and Section N2.13. The essential
command is

MNPROBIT ; Lhs = dependent variable

; Choices = the names of the J alternatives
; Rhs = list of choice specific attributes
; Rh2 = list of choice invariant individual characteristics $

Variable choice set sizes and utility function specifications are specified as in the CLOGIT
command. This command is the same as

NLOGIT ; MNP
; ... the rest of the command $
N3: Model and Command Summary for Discrete Choice Models N-50

N3.6 Command Summary

The following lists the current and where applicable, alternative forms of the discrete choice
model commands. The two sets of commands are identical, and for each model, in NLOGIT 6, either
command may be used for that model.
Models Command Alternative Command Form
Binary Choice Models
Binary Probit PROBIT PROBIT
Binary Logit BLOGIT LOGIT
Bivariate Probit BVPROBIT BIVARIATE PROBIT
Multivariate Probit MVPROBIT MPROBIT
Ordered Choice Models
Ordered Probit OPROBIT ORDERED PROBIT
Ordered Logit OLOGIT ORDERED ; Logit
Multinomial Logit Models
Multinomial Logit MLOGIT LOGIT
Conditional Logit CLOGIT DISCRETE CHOICE
Conditional Logit Extensions
Conditional Logit CLOGIT CLOGIT
Multinomial Logit NLOGIT NLOGIT (same as CLOGIT)
Scaled Multinomial Logit SMNLOGIT GMXLOGIT ; SMNL
Best/Worst Model BWLOGIT
Best/Worst Multinomial Logit BWMNLOGIT
Best/Worst Random Parameters BWRPLOGIT
Sequential Logit SEQLOGIT
Random Effects Multinomial Logit REMLOGIT
Fixed Effects Multinomial Logit FEMLOGIT
Random Regret Multinomial Logit RRLOGIT
Latent Class Random Regret Model LCRRLOGIT
Error Components Logit ECLOGIT NLOGIT ; ECM = ...
Heteroscedastic Extreme Value HLOGIT NLOGIT ; Het
Nested Logit NLOGIT ; Tree = ... NLOGIT ; Tree = ...
Generalized Nested Logit GNLOGIT ; Tree = ... NLOGIT ; GNL ; Tree = ...
Random Parameters Logit RPLOGIT NLOGIT ; RPL
Berry, Levinsohn, Pakes RP Logit BLPLOGIT
Latent Class Random Parameters RPRRLOGIT
Generalized Mixed Logit GMXLOGIT
Nonlinear Random Parameters NLRPLOGIT
Latent Class Logit LCLOGIT NLOGIT ; LCM
2K Latent Class LCLOGIT
Random Parameters Latent Class LCRPLOGIT
Multinomial Probit MNPROBIT NLOGIT ; MNP
N3: Model and Command Summary for Discrete Choice Models N-51

NLOGIT contains an additional command that is used for a specific purpose:

NLCONVERT ; Lhs = ... ; Rhs = ... ; Other parameters $

This command is used to reconfigure a data set from a one line format to a multiple line format that is
more convenient in NLOGIT. NLCONVERT is described in Chapter N18.

N3.7 Subcommand Summary

The following subcommands are used in NLOGIT model commands. The BLOGIT,
BPROBIT, BVPROBIT, MVPROBIT, OLOGIT and OPROBIT commands have additional
specifications that are documented in the LIMDEP Econometric Modeling Guide for these specific
models. The specifications below are those that may appear in the NLOGIT command or the
conditional logit extensions described above.

General Model Specification and Data Setup

Data on Dependent Variable
; Ranks indicates that data are in the form of ranks, possibly ties at last place.
; Shares indicates that data are in the form of proportions or shares.
; Frequencies indicates that data are in the form of frequencies or counts.
; Checkdata checks validity of the data before estimation.
; Wts = name specifies a weighting variable. (Noscale is not used here.)
; Scale (list of variables) = values for scaling loop specifies scaling of certain variables
during iterations.
; Pds = spec indicates multiple choice situations for individuals. Used by RPL, LCM, ECM,
MNP and by binary choice models to indicate a panel data set.
Specification of the Dependent Variable
; Lhs = names specifies model dependent variable(s).
Second Lhs variable indicates variable choice set size.
Third Lhs variable indicates specific choices in a universal choice set.
First Lhs variable is a set of utilities if ; MCS is used.
; MCS requests data generated by Monte Carlo simulation.
; Choices = list lists names for alternatives.
Specification of Utility Functions
; Rhs = names lists choice varying attribute variables.
; Rh2 = names lists choice invariant characteristic variables.
; Model: alternative way to specify utility functions, followed by definitions of
utility functions.
; Fix = list lists names of and values for coefficients that are to be fixed.
; Uset (list of alternatives) = list of values or [list of values] alternative method of
specifying starting values or fixed coefficients.
; Lambda = value specifies coefficient to use for Box-Cox transformation.
; Attr = list lists names for attributes used in one line entry format.
N3: Model and Command Summary for Discrete Choice Models N-52

Output Control

List and Retain Variables and Results

; Prob = name keeps predicted probabilities from estimated model as variable.

; Keep = name keeps predicted values from estimated model as variable. Used by
PROBIT and BLOGIT only.
; Utility = name keeps predicted utilities as variable.
; List lists predicted probabilities and predicted outcomes with model results.
; Parameters retains additional parameters as matrices. With RPL and LCM, keeps
matrices of individual specific parameter means.
; WTP = list lists specifications to retain computations of willingness to pay.

Covariance Matrices

; Covariance Matrix displays estimated asymptotic covariance matrix (normally not shown),
; Robust computes robust sandwich estimator for asymptotic covariance matrix.
; Cluster = spec computes robust cluster corrected asymptotic covariance matrix.

Display of Estimation Results

; Show displays model specification and tree structure.

; Describe lists descriptive statistics for attributes by alternative.
; Odds includes odds ratios in estimation results. Used only by BLOGIT.
; Crosstab includes cross tabulation of predicted and actual outcomes.
; Table = name adds model results to stored tables.

Marginal Effects

; Effects: spec displays estimated marginal effects. Used by NLOGIT.

; Partial Effects displays marginal effects, same as ; Marginal Effects. Used by PROBIT,
BLOGIT, BVPROBIT, MVPROBIT, OLOGIT, OPROBIT.
; Means computes marginal effects using data means. Uses average partial effects
if this is not specified.
; Pwt uses probability weights to compute average partial effects.

Hypothesis Testing

; Test: spec defines a Wald test of linear restrictions.

; Wald: spec defines a Wald test of linear restrictions, same as ; Test: spec.
; IAS = list lists choices used with CLOGIT to test IIA assumption.
N3: Model and Command Summary for Discrete Choice Models N-53

Optimization

Iterations Controls

; Alg = name specifies optimization method.

; Maxit = n sets the maximum iterations.
; Tlg [ = value] sets the convergence value for convergence on the gradient.
; Tlf [ = value] sets the convergence value for function convergence.
; Tlb[ = value] sets the convergence value for convergence on change in parameters.
; Set keeps current setting of optimization parameters as permanent.
; Output = n requests technical output during iterations; the level ‘n’ is 1, 2, 3 or 4.

Starting Values

; Start = list provides starting values for all model parameters.

; PR0 = list provides starting values for free parameters only. (Generally not used.)

Constrained Estimation

; CML: spec defines a constrained maximum likelihood estimator.

; Rst = list imposes fixed value and equality constraints.
; Calibrate fixes parameters at previously estimated values.
; ASC initially fit model with just ASCs.

Criterion Function for CLOGIT

; GME [ = number of support points] generalized maximum entropy. Used by MLOGIT

and CLOGIT.
; Sequential sequential two step estimator for nested logit. (Generally not used.)
; Conditional conditional estimator for two step nested logit. (Generally not used.)

Simulation Based Estimation

; Pts = number sets number of replications for simulation estimator. Used by ECM and
MNP. (Also used by LCM to specify number of latent classes.)
; Shuffled uses shuffled uniform draws to compute draws for simulations.
; Halton uses Halton sequences for simulation based estimators.

Simulation Processor (BINARY CHOICE Command for PROBIT and BLOGIT)

; Simulation [ = list of choices] simulates effect of changes in attributes on aggregate outcomes.

; Scenarios specifies changes in attributes for simulations.
; Arc computes arc elasticities during simulations.
; Merge merges revealed and stated preference data during simulations.
N3: Model and Command Summary for Discrete Choice Models N-54

Specific NLOGIT Model Commands

; LCM [ = list of variables] specifies latent class model. Optionally, specifies variables that
enter the class probabilities. (Command is also LCLOGIT.) Also used
by PROBIT and BLOGIT.
; ECM = list of specifications specifies error components logit model. (Command is also
ECLOGIT.)
; HEV specifies heteroscedastic extreme value model. (Command is also
HCLOGIT.)
Heteroscedastic Models
; Het specifies a heteroscedastic model. Used by RPL, ECL and HEV.
; Hfr = names specifies heteroscedastic function in RPL, HEV and covariance
heterogeneity form of nested logit model.
; Hfe = names specifies heteroscedasticity for ECM.
Nested Logit Model
; Tree = spec specifies tree structure in nested logit model.
; GNL specifies generalized nested logit model. (Command is also GNLOGIT.)
; RU1 specifies parameterization of second and third levels of the tree.
; RU2 specifies parameterization of second and third levels of the tree.
; RU3 specifies parameterization of second and third levels of the tree.
; IVSET: spec imposes constraints on inclusive value parameters.
; IVB = name keeps branch level inclusive values as a variable.
; IVL = name keeps limb level inclusive values as a variable.
; IVT = name keeps trunk level inclusive values as a variable.
; Prb = name keeps branch level probabilities as a variable.
; Cprob = name keeps conditional probabilities for alternatives.
Random Parameters Logit Model
; RPL [ = list of variables] requests mixed logit model. Optionally specifies variables to
enter means of random parameters.
; AR1 AR(1) structure for random terms in random parameters.
; Fcn: defines names and types of random parameters.
; Correlation specifies that random parameters are correlated.
; Hfr = names defines variables in heteroscedasticity. Also used by HEV and covariance
heterogeneity.
Multinomial Probit
; MNP specifies multinomial probit model. (Command is also MNPROBIT.)
; EQC = list specifies a set of choices whose pairwise correlations are all equal.
; RCR = list specifies configurations for correlations for multinomial probit model.
Also used by RPL.
; SDV = list specifies diagonal elements of covariance matrix. Also used by RPL
and HEV.
; REM specifies random effects form of the model.
N4: Data for Binary and Ordered Choice Models N-55

N4: Data for Binary and Ordered Choice

Models
N4.1 Introduction
The data arrangement needed for discrete choice modeling depends on the model you are
estimating. For the models described in Chapters N4-N15, you are fitting either cross section or
panel models, and the observations are arranged accordingly. This is needed because in this part of
the environment, you are fitting models for a single choice, and you need only a single observation to
record that choice. For the models in Chapters N16 and N17 and N23-N33, the basic format of your
data set will resemble a panel, even though it will usually be a cross section. This is because you are
fitting models for choice sets with multiple alternatives, with one ‘observation’ (data record) for each
alternative. For ‘panel’ models in the discrete choice environment, your data will consist of sets of
groups of observations. This is developed in detail in Chapter N20.

N4.2 Grouped and Individual Data for Discrete Choice

Models
There are two types of data which may be analyzed. We say that the data are individual if
the measurement of the dependent variable is physically discrete, consisting of individual responses.
The familiar case of the probit model with measured 0/1 responses is an example. The data are
grouped if the underlying model is discrete but the observed dependent variable is a proportion. In
the probit setting, this arises commonly in bioassay. A number of respondents have the same values
of the independent variables, and the observed dependent variable is the proportion of them with
individual responses equal to one. Voting proportions are a common application from political
science.
With only two exceptions, all of the discrete response models estimated by LIMDEP and
NLOGIT can be estimated with either individual or grouped data. The two exceptions are

• the multivariate probit model described in Chapter N12 (and E33)

• the multinomial probit model described in Chapter N27

You do not have to inform the program which type you are using. If necessary, the data are inspected
to determine which applies. The differences in estimation arise only in the way starting values are
computed and, occasionally, in the way the output should be interpreted. Cases sometimes arise in
which grouped data contain cells which are empty (proportion is zero) or full (proportion is one).
This does not affect maximum likelihood estimation and is handled internally in obtaining the
starting values. No special attention has to be paid to these cells in assembling the data set. We do
note, zero and unit ‘proportions’ data are sometimes indicative of a flawed data set, and can distort
your results.
N4: Data for Binary and Ordered Choice Models N-56

N4.3 Data Used in Estimation of Binary Choice Models

The following lists the specific features of the data needed to enable estimation of binary
choice models. Certain features of the data that are inconsequential or irrelevant in linear regression
modeling can impede estimation of a discrete choice model.

N4.3.1 The Dependent Variable

Data on the dependent variable for binary choice models may be individual or grouped. The
estimation program will check internally, and adjust accordingly where necessary. The log
likelihood function computed takes the same form for either case. The only special consideration
concerns the computation of the starting values for the iterations. If you do not provide your own
starting values, they are determined for the individual data case by simple least squares. The OLS
estimator is not useful in itself, but it does help to adjust the scale of the coefficient vector for the
first iteration. For the grouped data case, however, the initial values are determined by the minimum
chi squared, weighted least squares computation. Since this will generally involve logarithms or
other transformations which become noncomputable at zero or one, they are not computed for
individual data.

N4.3.2 Problems with the Independent Variables

There is a special consideration for the independent variables in a binary choice model. If a
variable xk is such that the range of xk can be divided into two parts and within the two parts, the
value of the dependent variable is always the same, then this variable becomes a perfect predictor for
the model. The estimator will break down, sometimes by iterating endlessly as the coefficient vector
drifts to extreme values. The following program illustrates the effect: The variable z is positive
when y equals one and negative when it equals zero. Notice, first, it spun for 100 iterations, which
is almost certainly problematic. A probit model should take less than 10 iterations. Second, note that
the log likelihood function is essentially zero, indicative of a perfect fit. Finally, note that the
coefficients are nonsensical, and the standard errors are essentially infinite. All are indicators of a
bad data set and/or model. The extreme (perfect) values for the fit measures on the next page
underscore the point. Finally, note the prediction table shows that the model predicts the dependent
variable perfectly.

SAMPLE ; 1-100 $
CALC ; Ran(12345) $
CREATE ; x = Rnn(0,1) ; d = Rnu(0,1) > .5 $
CREATE ; y = (-.5 + x + d + Rnn(0,1)) > 0 $
CREATE ; If(y = 1)z = Rnu(0,1)
; If(y = 0)z = -Rnu(0,1) $
PROBIT ; Lhs = y
; Rhs = one,x,z
; Output = 4 $
N4: Data for Binary and Ordered Choice Models N-57

Maximum of 100 iterations. Exit iterations with status=1.

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable Y
Log likelihood function .00000
Restricted log likelihood -69.13461
Chi squared [ 2 d.f.] 138.26922
Significance level .00000
McFadden Pseudo R-squared 1.0000000
Estimation based on N = 100, K = 3
Inf.Cr.AIC = 6.0 AIC/N = .060
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| -.98505 148462.2 .00 1.0000 *********** ***********
X| .14766 120032.6 .00 1.0000 *********** ***********
Z| 144.424 345728.4 .00 .9997 -677470.698 677759.546
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

+----------------------------------------+
| Fit Measures for Binomial Choice Model |
| Probit model for variable Y |
+----------------------------------------+
| Y=0 Y=1 Total|
| Proportions .53000 .47000 1.00000|
| Sample Size 53 47 100|
+----------------------------------------+
| Log Likelihood Functions for BC Model |
| P=0.50 P=N1/N P=Model|
| LogL = -69.31 -69.13 .00|
+----------------------------------------+
| Fit Measures based on Log Likelihood |
| McFadden = 1-(L/L0) = 1.00000|
| Estrella = 1-(L/L0)^(-2L0/n) = 1.00000|
| R-squared (ML) = .74910|
| Akaike Information Crit. = .06000|
| Schwartz Information Crit. = .13816|
+----------------------------------------+
| Fit Measures Based on Model Predictions|
| Efron = 1.00000|
| Ben Akiva and Lerman = 1.00000|
| Veall and Zimmerman = 1.00000|
| Cramer = 1.00000|
+----------------------------------------+
N4: Data for Binary and Ordered Choice Models N-58

+---------------------------------------------------------+
|Predictions for Binary Choice Model. Predicted value is |
|1 when probability is greater than .500000, 0 otherwise.|
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Value | |
|Value | 0 1 | Total Actual |
+------+----------------+----------------+----------------+
| 0 | 53 ( 53.0%)| 0 ( .0%)| 53 ( 53.0%)|
| 1 | 0 ( .0%)| 47 ( 47.0%)| 47 ( 47.0%)|
+------+----------------+----------------+----------------+
|Total | 53 ( 53.0%)| 47 ( 47.0%)| 100 (100.0%)|
+------+----------------+----------------+----------------+
+---------------------------------------------------------+
|Crosstab for Binary Choice Model. Predicted probability |
|vs. actual outcome. Entry = Sum[Y(i,j)*Prob(i,m)] 0,1. |
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Probability | |
|Value | Prob(y=0) Prob(y=1) | Total Actual |
+------+----------------+----------------+----------------+
| y=0 | 52 ( 52.0%)| 0 ( .0%)| 53 ( 52.0%)|
| y=1 | 0 ( .0%)| 46 ( 46.0%)| 47 ( 46.0%)|
+------+----------------+----------------+----------------+
|Total | 53 ( 52.0%)| 46 ( 46.0%)| 100 ( 98.0%)|
+------+----------------+----------------+----------------+
-----------------------------------------------------------------------
Analysis of Binary Choice Model Predictions Based on Threshold = .5000
-----------------------------------------------------------------------
Prediction Success
-----------------------------------------------------------------------
Sensitivity = actual 1s correctly predicted 97.872%
Specificity = actual 0s correctly predicted 98.113%
Positive predictive value = predicted 1s that were actual 1s 100.000%
Negative predictive value = predicted 0s that were actual 0s 98.113%
Correct prediction = actual 1s and 0s correctly predicted 98.000%
-----------------------------------------------------------------------
Prediction Failure
-----------------------------------------------------------------------
False pos. for true neg. = actual 0s predicted as 1s .000%
False neg. for true pos. = actual 1s predicted as 0s .000%
False pos. for predicted pos. = predicted 1s actual 0s .000%
False neg. for predicted neg. = predicted 0s actual 1s .000%
False predictions = actual 1s and 0s incorrectly predicted .000%
-----------------------------------------------------------------------

In general, for every Rhs variable, x, the minimum x for which y is one must be less than the
maximum x for which y is zero, and the minimum x for which y is zero must be less than the maximum
x for which y is one. If either condition fails, the estimator will break down. This is a more subtle, and
sometimes less obvious failure of the estimator. Unfortunately, it does not lead to a singularity and the
eventual appearance of collinearity in the Hessian. You might observe what appears to be convergence
of the estimator on a set of parameter estimates and standard errors which might look reasonable. The
main indication of this condition would be an excessive number of iterations – the probit model will
usually reach convergence in only a handful of iterations – and a suspiciously large standard error is
reported for the coefficient on the offending variable, as in the preceding example.
N4: Data for Binary and Ordered Choice Models N-59

You can check for this condition with the command:

CALC ; Chk (names of independent variables to check,

name of dependent variable) $

The offending variable in the previous example would be tagged by this check;

CALC ; Chk(z,y) $
Error 462: 0/1 choice model is inestimable. Bad variable = Z
Error 463: Its values predict 1[Y = 1] perfectly.
Error 116: CALC - Unable to compute result. Check earlier message.

This computation will issue warnings when the condition is found in any of the variables listed.
(Some computer programs will check for this condition automatically, and drop the offending
variable from the model. In keeping with LIMDEP’s general approach to modeling, this program
does not automatically make functional form decisions. The software does not accept the job of
determining the appropriate set of variables to include in the equation. This is up to the analyst.)

N4.3.3 Dummy Variables with Empty Cells

A problem similar to the one noted above arises when your model includes a dummy
variable that has no observations equal to one in one of the two cells of the dependent variable, or
vice versa. An example appears in Greene (1993, p. 673) in which the Lhs variable is always zero
when the variable ‘Southwest’ is zero. Professor Terry Seaks has used this example to examine a
number of econometrics programs. He found that no program which did not specifically check for
the failure – only one did – could detect the failure in some other way. All iterated to apparent
convergence, though with very different estimates of this coefficient and differing numbers of
iterations because of their use of different convergence rules. This form of incomplete matching of
values likewise prevents estimation, though the effect is likely to be more subtle. In this case, a
likely outcome is that the iterations will fail to converge, though the parameter estimates will not
necessarily become extreme.
Here is an example of this effect at work. The probit model looks excellent in the full
sample. In the restricted sample, d never equals zero when y equals zero. The estimator appears to
have converged, the derivatives are zero, but the standard errors are huge:

SAMPLE ; 1-100 $
CALC ; Ran(12345) $
CREATE ; x = Rnn(0,1)
; d = Rnu(0,1) > .5 $
CREATE ; y = (-.5 + x + d + Rnn(0,1)) > 0 $
PROBIT ; Lhs = y
; Rhs = one,x,d $

In this subset of data, d is always one when y equals zero.

REJECT ;y=0&d=0$
PROBIT ; Lhs = y
; Rhs = one,x,d $
N4: Data for Binary and Ordered Choice Models N-60

Nonlinear Estimation of Model Parameters

Method=NEWTON; Maximum iterations=100
1st derivs. .35811D+02 -.19962D+02 .12369D+01
Itr 1 F= .6981D+02 gtHg= .6608D+01 chg.F= .6981D+02 max|db|= .9613D+01
1st derivs. .49044D+01 -.74989D+01 -.29693D+00
Itr 2 F= .4521D+02 gtHg= .2003D+01 chg.F= .2460D+02 max|db|= .5302D+00
...
Itr 5 F= .4282D+02 gtHg= .1305D-03 chg.F= .4625D-03 max|db|= .2534D-04
1st derivs. -.10201D-08 -.76739D-08 -.32583D-08
Itr 6 F= .4282D+02 gtHg= .2445D-08 chg.F= .8516D-08 max|db|= .4705D-09
* Converged
Normal exit from iterations. Exit status=0.
Function= .69808104286D+02, at entry, .42822158396D+02 at exit

+---------------------------------------------+
| Binomial Probit Model |
| Dependent variable Y |
| Number of observations 100 |
| Iterations completed 6 |
| Log likelihood function -42.82216 |
+---------------------------------------------+
+--------+--------------+----------------+--------+--------+----------+
|Variable| Coefficient | Standard Error |b/St.Er.|P[|Z|>z]| Mean of X|
+--------+--------------+----------------+--------+--------+----------+
---------+Index function for probability
Constant| -.93917517 .23373657 -4.018 .0001
X | 1.17177061 .24254318 4.831 .0000 .10291147
D | 1.53191876 .35304007 4.339 .0000 .45000000

The second model required 24 iterations to converge, and produced these results: The apparent
convergence is deceptive, as evidenced by the standard errors.

Nonlinear Estimation of Model Parameters

Method=NEWTON; Maximum iterations=100
Itr 21 F= .1660D+02 gtHg= .3006D-04 chg.F= .1614D-08 max|db|= .2668D-01
1st derivs. -.19854D-08 .10979D-08 -.28588D-14
Parameters: .70037D+01 .14126D+01 -.63569D+01
Itr 22 F= .1660D+02 gtHg= .1787D-04 chg.F= .5692D-09 max|db|= .2530D-01
1st derivs. -.72119D-09 .39979D-09 .11824D-13
Parameters: .71645D+01 .14126D+01 -.65178D+01
Itr 23 F= .1660D+02 gtHg= .1064D-04 chg.F= .2012D-09 max|db|= .2406D-01
1st derivs. -.26221D-09 .14554D-09 -.35527D-14
Parameters: .73213D+01 .14126D+01 -.66746D+01
Itr 24 F= .1660D+02 gtHg= .6336D-05 chg.F= .7126D-10 max|db|= .2294D-01
* Converged
Normal exit: 24 iterations. Status=0, F= 16.60262
Function= .26413087151D+02, at entry, .16602624379D+02 at exit
N4: Data for Binary and Ordered Choice Models N-61

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable Y
Log likelihood function -16.60262
Restricted log likelihood -32.85957
Chi squared [ 2 d.f.] 32.51388
Significance level .00000
McFadden Pseudo R-squared .4947400
Estimation based on N = 61, K = 3
Inf.Cr.AIC = 39.2 AIC/N = .643
Hosmer-Lemeshow chi-squared = 4.91910
P-value= .08547 with deg.fr. = 2
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 7.32134 24162.78 .00 .9998 *********** 47365.49187
X| 1.41264*** .39338 3.59 .0003 .64163 2.18365
D| -6.67459 24162.78 .00 .9998 *********** 47351.49594
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

You can check for this condition if you suspect it is present by using a crosstab. The command is

CROSSTAB ; Lhs = dependent variable

; Rhs = independent dummy variable $

The 2×2 table produced should contain four nonempty cells. If any cells contain zeros, as in the
table below, then the model will be inestimable.

+-----------------------------------------------------------------+
|Cross Tabulation |
|Row variable is Y (Out of range 0-49: 0) |
|Number of Rows = 2 (Y = 0 to 1) |
|Col variable is D (Out of range 0-49: 0) |
|Number of Cols = 2 (D = 0 to 1) |
|Chi-squared independence tests: |
|Chi-squared[ 1] = 6.46052 Prob value = .01103 |
|G-squared [ 1] = 9.92032 Prob value = .00163 |
+-----------------------------------------------------------------+
| D |
+--------+--------------+------+ |
| Y| 0 1| Total| |
+--------+--------------+------+ |
| 0| 0 14| 14| |
| 1| 16 31| 47| |
+--------+--------------+------+ |
| Total| 16 45| 61| |
+-----------------------------------------------------------------+
N4: Data for Binary and Ordered Choice Models N-62

N4.3.4 Missing Values

Missing values in the current sample will always impede estimation. In the case of the
binary choice models, if your sample contains missing observations for the dependent variable, you
will receive a warning about improper coding of the values of the Lhs variable. This message will be
given whenever values of the dependent variable appear to be neither binary (0/1) or a proportion,
strictly between 0 and 1.

Probit: Data on Y are badly coded. (<0,1> and <=0 or >= 1).

Missing values for the independent variables will also badly distort the estimates. Since the
program assumes you will be deciding what observations to use for estimation, and -999 (the missing
value code) is a valid value, missing values on the right hand side of your model are not flagged as
an error. You will generally be able to see their presence in the model results. The sample means
for variables which contain missing values will usually look peculiar. In the small example below,
x2 is a dummy variable. Both coefficients are one, which should be apparent in a sample of 1,000.
The results, which otherwise look quite normal, suggest that missing values are being used as data in
the estimation. With SKIP, the results, based on the complete data, look much more reasonable.

CALC ; Ran(12345) $
SAMPLE ; 1-1000 $
CREATE ; x1 = Rnn(0,1)
; x2 = (Rnu(0,1) > .5) $
CREATE ; y = (-.5 + x1 +x2+rnn(0,1)) > 0 $
CREATE ; If(_obsno > 900)x2 = -999 $
PROBIT ; Lhs = y
; Rhs = one,x1,x2 $
SKIP $
PROBIT ; Lhs = y
; Rhs = one,x1,x2 $

Normal exit: 5 iterations. Status=0, F= 549.5785

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable Y
Log likelihood function -549.57851
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| -.08623* .04601 -1.87 .0609 -.17640 .00394
X1| .81668*** .05541 14.74 .0000 .70807 .92529
X2| .00029* .00015 1.95 .0517 .00000 .00058
--------+--------------------------------------------------------------------
N4: Data for Binary and Ordered Choice Models N-63

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable Y
Log likelihood function -441.38989
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| -.57123*** .07004 -8.16 .0000 -.70850 -.43396
X1| .97268*** .06611 14.71 .0000 .84310 1.10225
X2| .98082*** .10134 9.68 .0000 .78219 1.17945
--------+--------------------------------------------------------------------

You should use either SKIP or REJECT to remove the missing data from the sample. (See
Chapter R7 for details on skipping observations with missing values.)

N4.4 Bivariate Binary Choice

The bivariate probit model can be fit with either grouped data (you provide four proportions
variables) or individual data (you provide two binary variables). In either case, the data must contain
observations in both off diagonal cells. If your binary data are such that either the (y1=0,y2=1) or the
(y1=1,y2=0) have no observations, then the correlation coefficient cannot be estimated, and the
estimator will iterate endlessly, eventually ‘converging’ to a value of -1 or +1 for ρ. Note that this
does not apply to the bivariate probit with selection, but that is a different model. For the grouped
data case, if one of the proportions variables is always zero, the same problem will arise.

N4.5 Ordered Choice Model Structure and Data

Data for the ordered choice models must obey essentially the same rules as those for binary
choice models. Data may be grouped or individual. (Survey data might logically come in grouped
form.) If you provide individual data, the dependent variable is coded 0, 1, 2, ..., J. There must be at
least three values. Otherwise, the binary probit model applies. If the data are grouped, a full set of
proportions, p0, p1, ..., pJ, which sum to one at every observation must be provided. In the individual
data case, the data are examined to determine the value of J, which will be the largest observed value
of y that appears in the sample. In the grouped data case, J is one less than the number of Lhs
variables you provide. There are two additional considerations for ordered choice modeling.

N4.5.1 Empty Cells

If you are using individual data, the Lhs variable must be coded 0,1,...,J. All the values must
be present in the data. NLOGIT will look for empty cells. If there are any, the estimation is halted.
(If the value ‘j’ is not represented in the data, then the threshold parameter, µj cannot be estimated.
In this case, you will receive a diagnostic such as
ORDE, Panel, BIVA PROBIT: A cell has (almost) no observations.
Empty cell: Y never takes the value 2.

This diagnostic means exactly what it says. The ordered probability model cannot be estimated
unless all cells are represented in the data
N4: Data for Binary and Ordered Choice Models N-64

N4.5.2 Coding the Dependent Variable

Users frequently overlook the coding requirement, y = 0,1,... If you have a dependent
variable that is coded 1, 2,..., you will see the following diagnostic

Models - Insufficient variation in dependent variable

The reason this particular diagnostic shows up is that NLOGIT creates a new variable from your
dependent variable, say y, which equals zero when y equals zero and one when y is greater than zero.
It then tries to obtain starting values for the model by fitting a regression model to this new variable.
If you have miscoded the Lhs variable, the transformed variable always equals one, which explains
the diagnostic. In fact, there is no variation in the transformed dependent variable. If this is the case,
you can simply use CREATE to subtract 1.0 from your dependent variable to use this estimator.

N4.6 Constant Terms

In general, discrete choice models should contain constant terms. Omitting the constant term
is analogous to leaving the constant term out of a linear regression. This imposes a restriction that
rarely makes sense.
The ordered probit model must include a constant term, one, as the first Rhs variable. Since
the equation does include a constant term, one of the µs is not identified. We normalize µ0 to zero.
(Consider the special case of the binary probit model with something other than zero as its threshold
value. If it contains a constant, this cannot be estimated.) Other programs sometimes use different
normalizations of the model. For example, if the constant term is forced to equal zero, then one will
instead, have a nonzero threshold parameter, µ0, which equals zero in the presence of a nonzero
constant term.
In the more general multinomial choice models, when choices are unlabeled, there may be
no case for including alternative specific constants (ASCs) in the model, since they are not actually
associated with a particular choice. On the other hand, ASCs in a model with unlabeled choices
might simply imply that after controlling for the effects of the attributes, the indicated alternative is
chosen more or less frequently than the base alternative. It is possible that this might occur because
the alternative is close to the reference alternative or that culturally, those undertaking the
experiment might tend to read left to right. Failure to include ASCs in the model would in this case
correlate the alternative order effect into the other estimated parameters, possibly distorting the
model results.
N5: Models for Binary Choice N-65

N5: Models for Binary Choice

N5.1 Introduction
We define models in which the response variable being described is inherently discrete as
qualitative response (QR) models. This and the next several chapters will describe NLOGIT’s
qualitative dependent variable model estimators. The simplest of these are the binomial choice
models, which are the subject of this chapter and Chapters E27-E29. This will be followed by the
progressively more intricate formulations such as bivariate and multivariate probit, multinomial logit
and ordered choice models. NLOGIT supports a large variety of models and extensions for the
analysis of binary choice. The parametric model formulations, probit, logit, extreme value
(complementary log log) etc. are treated in detail in Chapter E27. We will focus on the first two of
these here.

N5.2 Modeling Binary Choices

A binomial response may be the outcome of a decision or the response to a question in a
survey. Consider, for example, survey data which indicate political party choice, mode of
transportation, occupation, or choice of location. We model these in terms of probability
distributions defined over the set of outcomes. There are a number of interpretations of an
underlying data generating process that produce the binary choice models we consider here. All of
them are consistent with the models that NLOGIT estimates, but the exact interpretation is a function
of the modeling framework.

N5.2.1 Underlying Processes

Consider a process with two possible outcomes indicated by a dependent variable, y, labeled
for convenience, y = 0 and y = 1. We assume, as well, that there is a set of measurable covariates, x,
which will be used to help explain the occurrence of one outcome or the other. Most models of
binary choice set up in this fashion will be based upon an index function, b′x, where b is a vector of
parameters to be estimated. The modeling of discrete, binary choice in these terms, is typically done
in one of the following frameworks:

Random Utility Approach

The respondent derives utility

U0 = b 0′x + e0 from choice 0, and U1 = b 1′x + e1 from choice 1,

in which e0 and e1 are the individual specific, random components of the individual’s utility that are
unaccounted for by the measured covariates, x. The choice of alternative 1 reveals that U1 > U0, or that

e0 - e1 < b 0′x - b 1′x.

N5: Models for Binary Choice N-66

Let e = e0 - e1 and let b′x represent the difference on the right hand side of the inequality – x is the
union of the two sets of covariates, and b is constructed from the two parameter vectors with zeros in
the appropriate locations if necessary. Then, the binary choice model applies to the probability that e
≤ b′x, which is the familiar sort of model shown in the next paragraph. This is a convenient way to
view migration behavior and survey responses to questions about economic issues.

Latent Regression Approach

A latent regression is specified as

y* = b′x + e.
The observed counterpart to y* is
y = 1 if and only if y* > 0.
This is the basis for most of the binary choice models in econometrics, and is described in further
detail below. It is the same model as the reduced form in the previous paragraph. Threshold models,
such as labor supply and reservation wages lend themselves to this approach.

Conditional Mean Function Approach

We assume that y is a binary variable, taking values 0 and 1, and formulate a priori that
Prob[y=1] = F(b′x), where F is any function of the index that satisfies the axioms of probability,

0 < F(b′x) < 1

F ′ (b′x) > 0,

limz↓-∞ F(z) = 0, limz↑+∞ F(z) = 1.

It follows that,
F(b′x) = 0 × Prob[y = 0 | x] + 1 × Prob[y = 1 | x]

is the conditional mean function for the observed binary y. This may be treated as a nonlinear
regression or as a binary choice model amenable to maximum likelihood estimation. This is a useful
departure point for less parametric approaches to binary choice modeling.

N5.2.2 Modeling Approaches

NLOGIT provides estimators for three approaches to formulating the binary choice models
described above:

Parametric Models – Probit, Logit, Extreme etc.

Most of the material below (and the received literature) focuses on models in which the full
functional form, including the probability distribution, are defined a priori. Thus, the probit model
which forms the basis of most of the results in econometrics, is based on a latent regression model in
which the disturbances are assumed to have a normal distribution. The logit model, in contrast, can be
construed as a random utility model in which it is assumed that the random parts of the utility functions
are distributed as independent extreme value.
N5: Models for Binary Choice N-67

Semiparametric Models – Maximum Score, Semiparametric Analysis

A semiparametric approach to modeling the binary choice steps back one level from the
previous model in that the specific distributional assumption is dropped, while the covariation (index
function) nature of the model is retained. Thus, the semiparametric approach analyzes the common
characteristics of the observed data which would arise regardless of the specific distribution
assumed. Thus, the semiparametric approach is essentially the conditional mean framework without
the specific distribution assumed. For the models that are supported in NLOGIT, MSCORE and
Klein and Spady’s framework, it is assumed only that F(b′x) exists and is a smooth continuous
function of its argument which satisfies the axioms of probability. The semiparametric approach is
more general (and more robust) than the parametric approach, but it provides the analyst far less
flexibility in terms of the types of analysis of the data that may be performed. In a general sense, the
gain to formulating the parametric model is the additional precision with which statements about the
data generating process may be made. Hypothesis tests, model extensions, and analysis of, e.g.,
interactions such as marginal effects, are difficult or impossible in semiparametric settings.

Nonparametric Analysis – NPREG

The nonparametric approach, as its name suggests, drops the formal modeling framework. It
is largely a bivariate modeling approach in which little more is assumed than that the probability that
y equals one depends on some x. (It can be extended to a latent regression, but this requires prior
specification and estimation, at least up to scale, of a parameter vector.) The nonparametric
approach to analysis of discrete choice is done in NLOGIT with a kernel density (largely based on
the computation of histograms) and with graphs of the implied relationship. Nonparametric analysis
is, by construction, the most general and robust of the techniques we consider, but, as a consequence,
the least precise. The statements that can be made about the underlying DGP in the nonparametric
framework are, of necessity, very broad, and usually provide little more than a crude overall
characterization of the relationship between a y and an x.

N5.2.3 The Linear Probability Model

One approach to modeling binary choice has been to ignore the special nature of the
dependent variable, and use conventional least squares. The resulting model,

Prob[yi = 1] = b′xi + ei

has been called the linear probability model (LPM). The LPM is known to have several problems,
most importantly that the model cannot be made to satisfy the axioms of probably independently of
the particular data set in use. Some authors have documented approaches to forcing the LPM on the
data, e.g., Fomby, et al., (1984), Long (1997) and Angrist and Pischke (2009). These computations
can easily be done with the other parts of NLOGIT, but will not be pursued here.
N5: Models for Binary Choice N-68

N5.3 Grouped and Individual Data for Binary Choice Models

There are two types of data which may be analyzed. We say that the data are individual if the
measurement of the dependent variable is physically discrete, consisting of individual responses. The
familiar case of the probit model with measured 0/1 responses is an example. The data are grouped if
the underlying model is discrete but the observed dependent variable is a proportion. In the probit
setting, this arises commonly in bioassay. A number of respondents have the same values of the
independent variables, and the observed dependent variable is the proportion of them with individual
responses equal to one. Voting proportions are a common application from political science.
All of the qualitative response models estimated by NLOGIT can be estimated with either
individual or grouped data. You do not have to inform the program which type you are using; if
necessary, the data are inspected to determine which applies. The differences arise only in the way
starting values are computed and, occasionally, in the way the output should be interpreted. Cases
sometimes arise in which grouped data contain cells which are empty (proportion is zero) or full
(proportion is one). This does not affect maximum likelihood estimation and is handled internally in
obtaining the starting values. No special attention has to be paid to these cells in assembling the data set.

N5.4 Variance Normalization

In the latent regression formulation of the model, the observed data are generated by the
underlying process
y = 1 if and only if b′x + e > 0.

The random variable, e, is assumed to have a zero mean (which is a simple normalization if the
model contains a constant term). The variance is left unspecified. The data contain no information
about the variance of e. Let σ denote the standard deviation of e. The same model and data arise if
the model is written as

y = 1 if and only if (b/σ)′x + e/σ > 0.

which is equivalent to
y = 1 if and only if γ′x + w > 0.

where the variance of w equals one. Since only the sign of y is observed, no information about
overall scaling is contained in the data. Therefore, the parameter σ is not estimable; it is assumed
with no loss of generality to equal one. (In some treatments (Horowitz (1993)), the constant term in
b is assumed to equal one, instead, in which case, the ‘constant’ in the model is an estimator of 1/σ.
This is simply an alternative normalization of the parameter vector, not a substantive change in the
model.)
N5: Models for Binary Choice N-69

N5.5 The Constant Term in Index Function Models

A question that sometimes arises is whether the binary choice model should contain a
constant term. The answer is yes, unless the underlying structure of your model specifically dictates
that none be included. There are a number of useful features of the parametric models that will be
subverted if you do not include a constant term in your model:

• Familiar fit measures will be distorted. Indeed, omitting the constant term can seriously
degrade the fit of a model, and will never improve it.

• Certain useful test statistics, such as the overall test for the joint significance of the
coefficients, may be rendered noncomputable if you omit the constant term.

• Some properties of the binary choice models, such as their ability to reproduce the average
outcome (sample proportion) will be lost.

Forcing the constant term to be zero is a linear restriction on the coefficient vector. Like any other
linear restriction, if imposed improperly, it will induce biases in the remaining coefficients.
(Orthogonality with the other independent variables is not a salvation here. Thus, putting variables
in mean deviation form does not remove the constant term from the model as it would in the linear
regression case.)
N6: Probit and Logit Models: Estimation N-70

N6: Probit and Logit Models: Estimation

N6.1 Introduction
We define models in which the response variable being described is inherently discrete as
qualitative response (QR) models. This and the next several chapters will describe two of NLOGIT’s
qualitative dependent variable model estimators, the probit and logit models. More extensive
treatment and technical background are given in Chapters E27-29. Several model extensions such as
models with endogenous variables, and sample selection, are treated in Chapter E29. Panel data
models for binary choice appear in Chapters E30 and E31. Semi- and nonparametric models are
documented in Chapter E32.

N6.2 Probit and Logit Models for Binary Choice

These parametric model formulations are provided as internal procedures in NLOGIT for
binary choice models. The probabilities and density functions are as follows:

Probit

β 'x i exπ(−t 2 / 2)
F= ∫−∞ 2π
dt = Φ(b′xi), f = φ(b′xi)

Logit

exp(β′xi )
F= = Λ(b′xi), f = Λ(b′xi)[1 - Λ(b′xi)]
1 + exp(β′xi )

N6.3 Commands
The basic model commands for the two binary choice models of interest here are:

PROBIT ; Lhs = dependent variable

or BLOGIT ; Rhs = regressors $

Data on the dependent variable may be either individual or proportions for both cases. When the
dependent variable is binary, 0 or 1, the model command may be LOGIT – the program will inspect
the data and make the appropriate adjustments for estimation of the model.
N6: Probit and Logit Models: Estimation N-71

N6.4 Output
The binary choice models can produce a very large amount of optional output.
Computation begins with some type of least squares estimation in order to obtain starting values.
With ungrouped data, we simply use OLS of the binary variable on the regressors. If requested, the
usual regression results are given, including diagnostic statistics, e.g., sum of squared residuals, and
the coefficient ‘estimates.’ The OLS estimates based on individual data are known to be inconsistent.
They will be visibly different from the final maximum likelihood estimates. For the grouped data
case, the estimates are GLS, minimum chi squared estimates, which are consistent and efficient. Full
GLS results will be shown for this case.

NOTE: The OLS results will not normally be displayed in the output. To request the display, use
; OLS in any of the model commands.

N6.4.1 Reported Estimates

Final estimates include:

• logL = the log likelihood function at the maximum,

• logL0 = the log likelihood function assuming all slopes are zero. If your Rhs variables do
not include one, this statistic will be meaningless. It is computed as

logL0 = n[PlogP + (1-P)log(1-P)]

where P is the sample proportion of ones.

• McFadden’s pseudo R2 - 1 - logL/logL0.

• The chi squared statistic for testing H0: b = 0 (not including the constant) and the
significance level = probability that χ2 exceeds test value. The statistic is

χ2 = 2(logL - logL0).

• Akaike’s information criterion, -2(logL - K) and the normalized AIC, = -2(logL - K)/n.

• The sample and model sizes, n and K.

• Hosmer and Lemeshow’s fit statistic and associated chi squared and p value. (The Hosmer
and Lemeshow statistic is documented in Section E27.8.)

The standard statistical results, including coefficient estimates, standard errors, t ratios, p
values and confidence intervals appear next. A complete listing is given below with an example.
After the coefficient estimates are given, two additional sets of results can be requested, an analysis
of the model fit and an analysis of the model predictions.
N6: Probit and Logit Models: Estimation N-72

We will illustrate with binary logit and probit estimates of a model for visits to the doctor
using the German health care data described in Chapter E2. The first model command is

LOGIT ; Lhs = doctor

; Rhs = one,age,hhninc,hhkids,educ,married
; OLS ; Summary
; Output = IC $ (Display all variants of information criteria)
Note that the command requests the optional listing of the OLS starting values and the additional fit
and diagnostic results. The results for this command are as follows. With the exception of the table
noted below, the same results (with different values, of course) will appear for all five parametric
models. Some additional optional computations and results will be discussed later.
-----------------------------------------------------------------------------
Binomial Logit Model for Binary Choice
There are 2 outcomes for LHS variable DOCTOR
These are the OLS estimates based on the
binary variables for each outcome Y(i)=j.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .63280*** .05584 11.33 .0000 .52335 .74224
AGE| .00387*** .00082 4.73 .0000 .00226 .00547
HHNINC| -.08338** .03967 -2.10 .0356 -.16114 -.00563
HHKIDS| -.08456*** .01943 -4.35 .0000 -.12264 -.04647
EDUC| -.00804** .00355 -2.27 .0234 -.01500 -.00109
MARRIED| .03209 .02131 1.51 .1321 -.00968 .07387
--------+--------------------------------------------------------------------
Binary Logit Model for Binary Choice
Dependent variable DOCTOR
Log likelihood function -2121.43961
Restricted log likelihood -2169.26982
Chi squared [ 5 d.f.] 95.66041
Significance level .00000
McFadden Pseudo R-squared .0220490
Estimation based on N = 3377, K = 6
Inf.Cr.AIC = 4254.879 AIC/N = 1.260
FinSmplAIC = 4254.904 FIC/N = 1.260
Bayes IC = 4291.628 BIC/N = 1.271
HannanQuinn = 4268.018 HIC/N = 1.264
Hosmer-Lemeshow chi-squared = 17.65094
P-value= .02400 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .52240** .24887 2.10 .0358 .03463 1.01018
AGE| .01834*** .00378 4.85 .0000 .01092 .02575
HHNINC| -.38750** .17760 -2.18 .0291 -.73559 -.03941
HHKIDS| -.38161*** .08735 -4.37 .0000 -.55282 -.21040
EDUC| -.03581** .01576 -2.27 .0230 -.06669 -.00493
MARRIED| .14709 .09727 1.51 .1305 -.04357 .33774
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
N6: Probit and Logit Models: Estimation N-73

N6.4.2 Fit Measures

The model results are followed by a cross tabulation of the correct and incorrect predictions
of the model using the rule
∧ ∧
y = 1 if F( β ‘xi ) > .5, and 0 otherwise.

For the models with symmetric distributions, probit and logit, the average predicted probability will
equal the sample proportion. If you have a quite unbalanced sample – high or low proportion of ones
– the rule above is likely to result in only one value, zero or one, being predicted for the Lhs variable.
You can choose a threshold different from .5 by using

; Limit = the value you wish

in your command. There is no direct counterpart to an R2 in regression. Authors very commonly

report the
log L(model)
Pseudo – R2 = 1 − .
log L(constants only)

We emphasize, this is not a proportion of variation explained. Moreover, as a fit measure, it has some
peculiar features. Note, for our example above, it is 1 - (-17673.10)/(-18019.55) = 0.01923, yet with
the standard prediction rule, the estimated model predicts almost 63% of the outcomes correctly.

+----------------------------------------+
| Fit Measures for Binomial Choice Model |
| Logit model for variable DOCTOR |
+----------------------------------------+
| Y=0 Y=1 Total|
| Proportions .34202 .65798 1.00000|
| Sample Size 1155 2222 3377|
+----------------------------------------+
| Log Likelihood Functions for BC Model |
| P=0.50 P=N1/N P=Model|
| LogL = -2340.76 -2169.27 -2121.44|
+----------------------------------------+
| Fit Measures based on Log Likelihood |
| McFadden = 1-(L/L0) = .02205|
| Estrella = 1-(L/L0)^(-2L0/n) = .02824|
| R-squared (ML) = .02793|
| Akaike Information Crit. = 1.25996|
| Schwartz Information Crit. = 1.27084|
+----------------------------------------+
| Fit Measures Based on Model Predictions|
| Efron = .02693|
| Ben Akiva and Lerman = .56223|
| Veall and Zimmerman = .04899|
| Cramer = .02735|
+----------------------------------------+
N6: Probit and Logit Models: Estimation N-74

The next set of results examines the success of the prediction rule

Predict yi = 1 if Pi > P* and 0 otherwise

where P* is a defined threshold probability. The default value of P* is 0.5, which makes the
prediction rule equivalent to ‘Predict yi = 1 if the model says the predicted event yi = 1 | xi is more
likely than the complement, yi = 0 | xi.’ You can change the threshold from 0.5 to some other value
with
; Limit = your P*

+---------------------------------------------------------+
|Predictions for Binary Choice Model. Predicted value is |
|1 when probability is greater than .500000, 0 otherwise.|
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Value | |
|Value | 0 1 | Total Actual |
+------+----------------+----------------+----------------+
| 0 | 21 ( .6%)| 1134 ( 33.6%)| 1155 ( 34.2%)|
| 1 | 12 ( .4%)| 2210 ( 65.4%)| 2222 ( 65.8%)|
+------+----------------+----------------+----------------+
|Total | 33 ( 1.0%)| 3344 ( 99.0%)| 3377 (100.0%)|
+------+----------------+----------------+----------------+
+---------------------------------------------------------+
|Crosstab for Binary Choice Model. Predicted probability |
|vs. actual outcome. Entry = Sum[Y(i,j)*Prob(i,m)] 0,1. |
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Probability | |
|Value | Prob(y=0) Prob(y=1) | Total Actual |
+------+----------------+----------------+----------------+
| y=0 | 415 ( 12.3%)| 739 ( 21.9%)| 1155 ( 34.2%)|
| y=1 | 739 ( 21.9%)| 1482 ( 43.9%)| 2222 ( 65.8%)|
+------+----------------+----------------+----------------+
|Total | 1155 ( 34.2%)| 2221 ( 65.8%)| 3377 ( 99.9%)|
+------+----------------+----------------+----------------+

This table computes a variety of conditional and marginal proportions based on the results using the
defined prediction rule. For examples, the 66.697% equals (1482/2222)100% while the 66.727% is
(1482/2221)100%.

-----------------------------------------------------------------------
Analysis of Binary Choice Model Predictions Based on Threshold = .5000
-----------------------------------------------------------------------
Prediction Success
-----------------------------------------------------------------------
Sensitivity = actual 1s correctly predicted 66.697%
Specificity = actual 0s correctly predicted 35.931%
Positive predictive value = predicted 1s that were actual 1s 66.727%
Negative predictive value = predicted 0s that were actual 0s 35.931%
Correct prediction = actual 1s and 0s correctly predicted 56.174%
-----------------------------------------------------------------------
N6: Probit and Logit Models: Estimation N-75

-----------------------------------------------------------------------
Prediction Failure
-----------------------------------------------------------------------
False pos. for true neg. = actual 0s predicted as 1s 63.983%
False neg. for true pos. = actual 1s predicted as 0s 33.258%
False pos. for predicted pos. = predicted 1s actual 0s 33.273%
False neg. for predicted neg. = predicted 0s actual 1s 63.983%
False predictions = actual 1s and 0s incorrectly predicted 43.767%
-----------------------------------------------------------------------

N6.4.3 Covariance Matrix

The estimated asymptotic covariance matrix of the coefficient estimator is not automatically
displayed – it might be huge. You can request a display with

; Covariance

If the matrix is not larger than 5×5, it will be displayed in full. If it is larger, the covariance matrix
will be placed in the matrix area in your project window with the name COV.[B^]. By double
clicking the name, you can display the matrix in a window. An example appears in Figure N6.1
below.

Figure N6.1 Covariance Matrix

N6: Probit and Logit Models: Estimation N-76

N6.4.4 Retained Results and Generalized Residuals

The results saved by the binary choice models are:

Matrices: b = estimate of b (also contains γ for the Burr model)

varb = asymptotic covariance matrix

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Variables: logl_obs = individual contribution to log likelihood

Last Model: b_variables

Last Function: Prob(y = 1 | x) = F(b′x). This varies with the model specification.

Models that are estimated using maximum likelihood automatically create a variable named
logl_obs, that contains the contribution of each individual observation to the log likelihood for the
sample. Since the log likelihood is the sum of these terms, you could, in principle, recover the
overall log likelihood after estimation with

CALC ; List ; Sum(logl_obs) $

The variable can be used for certain hypothesis tests, such as the Vuong test for nonnested models.
The following is an example (albeit, one that appears to have no real power) that applies the Vuong
test to discern whether the logit or probit is a preferable model for a set of data:

LOGIT ;…$
CREATE ; lilogit = logl_obs $
PROBIT ;…$
CREATE ; liprobit = logl_obs ; di = liprobit - lilogit $
CALC ; List ; vtest = Sqr(n) * Xbr(di) / Sdv(di) $

The ‘generalized residuals’ in a parametric binary choice model are the derivatives of the log
likelihood with respect to the constant term in the model. These are sometimes used to check the
specification of the model (see Chesher and Irish (1987)). These are easy to compute for the models
listed above – in each case, the generalized residual is the derivative of the log of the probability with
respect to b′x. This is computed internally as part of the iterations, and kept automatically in your
data area in a variable named score_fn. The formulas for the generalized residuals are provided in
Section E27.12 with the technical details for the models. For example, you can verify the
convergence of the estimator to a maximum of the log likelihood with the instruction

CALC ; List ; Sum(score_fn) $

N6: Probit and Logit Models: Estimation N-77

N6.5 Robust Covariance Matrix Estimation

The preceding describes a covariance estimator that accounts for a specific, observed aspect
of the data. The concept of the ‘robust’ covariance matrix is that it is meant to account for
hypothetical, unobserved failures of the model assumptions. The intent is to produce an asymptotic
covariance matrix that is appropriate even if some of the assumptions of the model are not met. (It is
an important, but infrequently discussed issue whether the estimator, itself, remains consistent in the
presence of these model failures – that is, whether the so called robust covariance matrix estimator is
being computed for an inconsistent estimator.) (Chapter R10 provides general discussion of robust
covariance matrix estimation.)

N6.5.1 The Sandwich Estimator

A robust covariance matrix estimator adjusts the estimated asymptotic covariance matrix for
possible misspecification in the model which leaves the MLE consistent but the estimated asymptotic
covariance matrix incorrectly computed. One example would be a binary choice model with
unspecified latent heterogeneity. A frequent adjustment for this case is the ‘sandwich estimator,’
which is the choice based sampling estimator suggested above with weights equal to one. (This
suggests how it could be computed.) The desired matrix is

−1 −1
 n  ∂ 2 log Fi    n  ∂ log Fi  ∂ log Fi    n  ∂ 2 log Fi 
βˆ  =  ∑ 
ˆ ∂ ˆ ′    ∑ i 1  ∂ ˆ =
  '  ∑ i 1  ˆ ˆ 
Est.Asy.Var  
 ∂ ˆ ′     ∂ ∂ ′
= i 1=
  ∂ββββββ    

Three ways to obtain this matrix are

; Wts = one ; Choice based sampling

or ; Robust
or ; Cluster = 1

The computation is identical in all cases. (As noted below, the last of them will be slightly larger, as
it will be multiplied by n/(n-1).)

N6.5.2 Clustering
A related calculation is used when observations occur in groups which may be correlated.
This is rather like a panel; one might use this approach in a random effects kind of setting in which
observations have a common latent heterogeneity. The parameter estimator is unchanged in this
case, but an adjustment is made to the estimated asymptotic covariance matrix. The calculation is
done as follows: Suppose the n observations are assembled in G clusters of observations, in which
the number of observations in the ith cluster is ni. Thus,

∑
G
i =1
ni = n.
N6: Probit and Logit Models: Estimation N-78

Let the observation specific gradients and Hessians be

∂ log Lij
gij =
∂β

∂ 2 log Lij
Hij = .
∂ββ
∂ '

The uncorrected estimator of the asymptotic covariance matrix based on the Hessian is

( −∑ )
−1
∑
G ni
VH = -H-1 = =i 1 =j 1
H ij

Estimators for some models such as the Burr model will use the BHHH estimator, instead. In
general,
(∑ )
−1
∑ g ij g′ij
G ni
VB = =i 1 =j 1

Let V be the estimator chosen. Then, the corrected asymptotic covariance matrix is

Est.Asy.Var βˆ  = V
G  G
∑ i 1 =
G −1 
= ∑
ni
j 1 ∑
g ij =
ni
j 1 (
′
g ij  V

)( )
Note that if there is exactly one observation per cluster, then this is G/(G-1) times the sandwich
estimator discussed above. Also, if you have fewer clusters than parameters, then this matrix is
singular – it has rank equal to the minimum of G and K, the number of parameters.
This procedure is described in greater detail in Section E27.5.3. To request the estimator,
your command must include

; Cluster = specification

where the specification is either the fixed value if all the clusters are the same size, or the name of an
identifying variable if the clusters vary in size. Note, this is not the same as the variable in the Pds
function that is used to specify a panel. The cluster specification must be an identifying code that is
specific to the cluster. For example, our health care data used in our examples is an unbalanced
panel. The first variable is a family id, which we will use as follows

; Cluster = id

The results below demonstrate the effect of this estimator. Three sets of estimates are given. The
first are the original logit estimates that ignore the cross observation correlations. The second use the
correction for clustering. The third is a panel data estimator – the random effects estimator described
in Chapter E30 – that explicitly accounts for the correlation across observations. It is clear that the
different treatments change the results noticeably.
N6: Probit and Logit Models: Estimation N-79

Uncorrected covariance matrix:

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -.20205** .09397 -2.15 .0315 -.38622 -.01787
AGE| .01935*** .00130 14.90 .0000 .01681 .02190
EDUC| -.02477*** .00578 -4.28 .0000 -.03611 -.01344
MARRIED| .12023*** .03376 3.56 .0004 .05405 .18640
HHNINC| -.21388*** .07580 -2.82 .0048 -.36245 -.06532
HHKIDS| -.24879*** .02983 -8.34 .0000 -.30726 -.19032
FEMALE| .58305*** .02620 22.26 .0000 .53171 .63439
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Cluster corrected covariance matrix:

+---------------------------------------------------------------------+
| Covariance matrix for the model is adjusted for data clustering. |
| Sample of 27326 observations contained 7293 clusters defined by |
| variable ID which identifies by a value a cluster ID. |
+---------------------------------------------------------------------+
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -.20205 .12997 -1.55 .1200 -.45678 .05269
AGE| .01935*** .00176 11.00 .0000 .01590 .02280
EDUC| -.02477*** .00811 -3.05 .0023 -.04067 -.00888
MARRIED| .12023*** .04556 2.64 .0083 .03093 .20953
HHNINC| -.21388** .09276 -2.31 .0211 -.39568 -.03209
HHKIDS| -.24879*** .03842 -6.48 .0000 -.32409 -.17349
FEMALE| .58305*** .03744 15.57 .0000 .50967 .65644
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Random effects estimates:

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -.70495*** .18028 -3.91 .0001 -1.05830 -.35160
AGE| .03656*** .00241 15.18 .0000 .03184 .04128
EDUC| -.03703*** .01132 -3.27 .0011 -.05923 -.01484
MARRIED| .05481 .05570 .98 .3251 -.05435 .16397
HHNINC| .00772 .11698 .07 .9474 -.22156 .23700
HHKIDS| -.23497*** .04727 -4.97 .0000 -.32763 -.14232
FEMALE| .77202*** .05357 14.41 .0000 .66702 .87702
Rho| .39909*** .00586 68.07 .0000 .38760 .41058
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N6: Probit and Logit Models: Estimation N-80

N6.5.3 Stratification and Clustering

The clustering estimator is extended to include stratum level grouping, where a stratum
includes one or more clusters, and weighting to allow finite population correction. We suppose that
there are a total of S strata in the sample. Each stratum, ‘s,’ contains Cs clusters. The number of
observations in a cluster is Ncs. Neglecting the weights for the moment,

Variance estimator = VGV

V = the inverse of conventional estimator of the Hessian
G = ∑ s =1 ws G s
S

Gs = (∑ Cs
c =1 )
g cs g′cs - C1s g s g′s

g s = ∑ c =1
sC
g cs
g cs = ∑ i=1cs wics g ics
N

where gics is the derivative of the contribution to the log likelihood of individual i in cluster c in
stratum s. The remaining detail in the preceding is the weighting factor, ws. The stratum weight is
computed as
ws = fs × hs × d
where fs = 1 or a finite population correction, 1 - Cs/Cs* where Cs* is the true
number of clusters in stratum s, where Cs* > Cs.
hs = 1 or Cs/(Cs - 1)
d = 1 or (N-1)/(N-K) where N is the total number of observations in the
entire sample and K is the number of parameters (rows in V).
Use
; Cluster = the number of observations in a cluster (fixed) or the name of a
stratification variable which gives the cluster an identification. This
is the setup that is described above.
; Stratum = the number of observations in a stratum (fixed) or the name of a
stratification variable which gives the stratum an identification
; Wts = the name of the usual weighting variable for model estimation if
weights are desired. This defines wics.
; FPC = the name of a variable which gives the number of clusters in the
stratum. This number will be the same for all observations in a
stratum – repeated for all clusters in the stratum. If this number is
the same for all strata, then just give the number.
; Huber Use this switch to request hs. If omitted, hs = 1 is used.
; DFC Use this switch to request the use of d given above. If omitted,
d = 1 is used.

Further details on this estimator may be found in Section E30.3 and Section R10.3.
N6: Probit and Logit Models: Estimation N-81

N6.6 Analysis of Partial Effects

Partial effects in a binary choice model are
′x) dF ( ′x)
∂E[ y | x] ∂F (ββ
= = β = F′(b′x)b = f(b′x)b
∂x ∂x d (β′x)

That is, the vector of marginal effects is a scalar multiple of the coefficient vector. The scale factor,
f(b′x), is the density function, which is a function of x. This function can be computed at any data
vector desired. Average partial effects are computed by averaging the function over the sample
observations. The elasticity of the probability is
∂ log E[ y | x] xk ∂E[ y | x] xk
= = × marginal effect
∂ log xk E[ y | x] ∂xk E[ y | x ]

When the variable in x that is changing in the computation is a dummy variable, the
derivative approach to estimating the marginal effect is not appropriate. An alternative which is
closer to the desired computation for a dummy variable, that we denote z, is
DFz = Prob[y = 1 | z = 1] - Prob[y = 1 | z = 0]
= F(b′x + αz | z = 1) - F(b′x + αz | z = 0)
= F(b′x + α) - F(b′x).
NLOGIT examines the variables in the model and makes this adjustment automatically.
There are two programs in NLOGIT for obtaining partial effects for the binary choice (and
most other) models, the built in computation provided by the model command and the PARTIAL
EFFECTS command. Examples of both are shown below.
The LOGIT, PROBIT, etc. commands provide a built in, basic computation for partial
effects. You can request the computation to be done automatically by adding
; Partial Effects (or ; Marginal Effects)
to your command. The results below are produced for logit model in the earlier example. The
standard errors for the partial effects are computed using the delta method. See Section E27.12 for
technical details on the computation. The results reported are the average partial effects.
-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00402*** .26013 4.92 .0000 .00242 .00562
HHNINC| -.08666** -.05857 -2.22 .0267 -.16331 -.01001
HHKIDS| -.08524*** -.05021 -4.33 .0000 -.12382 -.04667 #
EDUC| -.00779** -.13620 -2.24 .0252 -.01461 -.00097
MARRIED| .03279 .03534 1.52 .1288 -.00952 .07510 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
N6: Probit and Logit Models: Estimation N-82

The equivalent PARTIAL EFFECTS command, which would immediately follow the LOGIT
command, would be

PARTIAL EFFECTS ; Effects: age / hhninc / hhkids / educ / married

; Summary $

---------------------------------------------------------------------
Partial Effects for Probit Probability Function
Partial Effects Averaged Over Observations
* ==> Partial Effect for a Binary Variable
---------------------------------------------------------------------
Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
AGE .00402 .00082 4.92 .00242 .00562
HHNINC -.08666 .03911 2.22 -.16331 -.01001
* HHKIDS -.08524 .01968 4.33 -.12382 -.04667
EDUC -.00779 .00348 2.24 -.01461 -.00097
* MARRIED .03279 .02159 1.52 -.00952 .07510
---------------------------------------------------------------------

The second method provides a variety of options for computing partial effects under various
scenarios, plotting the effects, etc. See Chapter R11 for further details.

NOTE: If your model contains nonlinear terms in the variables, such as age^2 or interaction terms
such as age*female, then you must use the PARTIAL EFFECTS command to obtain partial effects.
The built in routine in the command, ; Partial Effects, will not give the correct answers for variables
that appear in nonlinear terms.

N6.6.1 The Krinsky and Robb Method

An alternative to the delta method described above that is sometimes advocated is the
Krinsky and Robb method. By this device, we have our estimate of the model coefficients, b, and
the estimated asymptotic covariance matrix, V. The marginal effects are computed as a function of b
and the vector of means of the sample data, x , say gk(b, x ) for the kth variable. The Krinsky and
Robb technique involves sampling R draws from the asymptotic normal distribution of the estimator,
computing the function with these R draws, then computing the empirical variance. This is not done
automatically by the binary choice estimator, but you can easily do the computation using the
WALD command. For an example, we will use this method to compute the marginal effects for two
variables in the logit model estimated earlier. The program would be

NAMELIST ; x = one,age,hhninc,hhkids,educ,married $
LOGIT ; Lhs = doctor ; Rhs = x ; Partial Effects $
MATRIX ; xbar = Mean(x) $
CALC ; kx = Col(x) ; Ran(12345) $
WALD ; Start = b ; Var = varb ; Labels = kx_b
; Fn1 = b2 * Lgd(b1'xbar)
; Fn2 = b3 * Lgd(b1'xbar)
; K&R ; Pts = 2000 $
N6: Probit and Logit Models: Estimation N-83

-----------------------------------------------------------------------------
WALD procedure. Estimates and standard errors
for nonlinear functions and joint test of
nonlinear restrictions.
Wald Statistic = 27.72506
Prob. from Chi-squared[ 2] = .00000
Krinsky-Robb method used with 2000 draws
Functions are computed at means of variables
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
WaldFcns| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Fncn(1)| .00409*** .00084 4.85 .0000 .00244 .00575
Fncn(2)| -.08694** .03913 -2.22 .0263 -.16363 -.01025
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
---------------------------------------------------------------------
Partial Effects for Probit Probability Function
Partial Effects Averaged Over Observations
* ==> Partial Effect for a Binary Variable
---------------------------------------------------------------------
Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
AGE .00402 .00082 4.92 .00242 .00562
HHNINC -.08666 .03911 2.22 -.16331 -.01001
---------------------------------------------------------------------

There is a second source of difference between the Krinsky and Robb estimates and the delta method
results that follow: The Krinsky and Robb procedure is based on the means of the data while the
delta method averages the partial effects over the observations. It is possible to perform the K&R
iteration at every observation to reproduce the APE calculations by adding ; Average to the WALD
command. The results below illustrate.

--------+--------------------------------------------------------------------
Fncn(1)| .00407*** .00085 4.80 .0000 .00241 .00573
Fncn(2)| -.08673** .03929 -2.21 .0273 -.16373 -.00973
--------+--------------------------------------------------------------------

We do not recommend this as a general procedure, however. It is enormously time consuming and
does not produce a more accurate result.

Estimating Marginal Effects by Strata

Marginal effects may be calculated for indicated subsets of the data by using

; Margin = variable
where ‘variable’ is the name of a variable coded 0,1,... which designates up to 10 subgroups of the
data set, in addition to the full data set. For example, a common application would be
; Margin = sex
in which the variable sex is coded 0 for men and 1 for women (or vice versa). The variable used in
this computation need not appear in the model; it may be any variable in the data set.
N6: Probit and Logit Models: Estimation N-84

For example, using our logit model above, we now compute marginal effects separately for
men and women:

LOGIT ; Lhs = doctor

; Rhs = one,age,hhninc,hhkids,educ,married
; Margin = female $

-----------------------------------------------------------------------------
Binary Logit Model for Binary Choice
Dependent variable DOCTOR
Log likelihood function -2121.43961
Restricted log likelihood -2169.26982
Chi squared [ 5 d.f.] 95.66041
Significance level .00000
McFadden Pseudo R-squared .0220490
Estimation based on N = 3377, K = 6
Inf.Cr.AIC = 4254.879 AIC/N = 1.260
Hosmer-Lemeshow chi-squared = 17.65094
P-value= .02400 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .52240** .24887 2.10 .0358 .03463 1.01018
AGE| .01834*** .00378 4.85 .0000 .01092 .02575
HHNINC| -.38750** .17760 -2.18 .0291 -.73559 -.03941
HHKIDS| -.38161*** .08735 -4.37 .0000 -.55282 -.21040
EDUC| -.03581** .01576 -2.27 .0230 -.06669 -.00493
MARRIED| .14709 .09727 1.51 .1305 -.04357 .33774
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Observations used are FEMALE=0
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00414*** .26343 4.84 .0000 .00247 .00582
HHNINC| -.08756** -.06038 -2.18 .0291 -.16619 -.00893
HHKIDS| -.08714*** -.05161 -4.34 .0000 -.12645 -.04783 #
EDUC| -.00809** -.14612 -2.27 .0234 -.01509 -.00109
MARRIED| .03351 .03549 1.50 .1334 -.01025 .07728 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N6: Probit and Logit Models: Estimation N-85

-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Observations used are FEMALE=1
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00404*** .26337 4.88 .0000 .00242 .00567
HHNINC| -.08545** -.05555 -2.18 .0290 -.16217 -.00873
HHKIDS| -.08519*** -.04911 -4.33 .0000 -.12379 -.04659 #
EDUC| -.00790** -.13086 -2.28 .0225 -.01468 -.00111
MARRIED| .03279 .03550 1.50 .1345 -.01015 .07573 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Observations used are All Obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00410*** .26352 4.86 .0000 .00244 .00575
HHNINC| -.08660** -.05811 -2.18 .0291 -.16436 -.00884
HHKIDS| -.08626*** -.05044 -4.34 .0000 -.12524 -.04727 #
EDUC| -.00800** -.13893 -2.27 .0230 -.01490 -.00110
MARRIED| .03318 .03551 1.50 .1339 -.01021 .07658 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
+-------------------------------------------+
| Marginal Effects for Logit |
+----------+----------+----------+----------+
| Variable | FEMALE=0 | FEMALE=1 | All Obs. |
+----------+----------+----------+----------+
| AGE | .00414 | .00404 | .00410 |
| HHNINC | -.08756 | -.08545 | -.08660 |
| HHKIDS | -.08714 | -.08519 | -.08626 |
| EDUC | -.00809 | -.00790 | -.00800 |
| MARRIED | .03351 | .03279 | .03318 |
+----------+----------+----------+----------+

The computation using the built in estimator is done at the strata means of the data. The
computation can be done by averaging across observations using the PARTIAL EFFECTS (or just
PARTIALS) command. For example, the corresponding results for the income variable are
obtained with

PARTIAL EFFECTS ; Effects: hhninc @ female=0,1 $

N6: Probit and Logit Models: Estimation N-86

---------------------------------------------------------------------
Partial Effects Analysis for Logit Probability Function
---------------------------------------------------------------------
Effects on function with respect to HHNINC
Results are computed by average over sample observations
Partial effects for continuous HHNINC computed by differentiation
Effect is computed as derivative = df(.)/dx
---------------------------------------------------------------------
df/dHHNINC Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
Subsample for this iteration is FEMALE = 0 Observations: 1812
APE. Function -.08585 .03925 2.19 -.16278 -.00892
---------------------------------------------------------------------
Subsample for this iteration is FEMALE = 1 Observations: 1565
APE. Function -.08355 .03820 2.19 -.15841 -.00868

Examining the Effect of a Variable Over a Range of Values

Another useful device is a plot of the probability (conditional mean) over the range of a
variable of interest either holding other variables at their means, or averaging over the sample values.
The figure below does this for the income variable in the logit model for doctor visits. The figure is
plotted for hhkids = 1 and hhkids = 0 to show the two effects. We see that the probability falls with
increased income, and also for individuals in households in which there are children.

SIMULATE ; Scenario: & hhninc = 0(.05).5 | hhkids=0,1 ; Plot $

Figure N6.2 Probabilities Varying with Income

N6: Probit and Logit Models: Estimation N-87

N6.7 Simulation and Analysis of a Binary Choice Model

This section describes a procedure that is used with all of the parametric models described
above. It is used for two specific analyses. This procedure allows you to analyze the predictions
made by a binary choice when the variables in the model are changed. The analysis is provided in
two parts:

• Change specific variables in the model by a prescribed amount, and examine the changes in
the model predictions.

• Vary a particular variable over a range of values and examine the predicted probabilities
when other variables are held fixed at their means.

This program is available for the six parametric binary choice models: probit, logit, Gompertz,
complementary log log, arctangent and Burr. The probit and logit models may also be
heteroscedastic. The routine is accessed as follows. First fit the model as usual. Then, use the
identical model specification as shown below with the specifications indicated:

(MODEL) ; Lhs = ... ; Rhs = ... $

Then

BINARY CHOICE ; Lhs = (the same) ; Rhs = (the same) ; ... (also the same)
; Model = Probit, Logit, Gompertz, Comploglog or Burr
; Start = B (from the preceding model)

(optional, the value to use for predicting Lhs = 1, default = .5)

; Threshold = P*

(optional) ; Scenario: variable operation = value /

(variable operation = value) / ... (may be repeated)

(optional) ; Plot: variable (lower limit, upper limit) $

In the ; Plot specification, the limits part may be omitted, in which case the range of the variable is
used. This will replicate for the one variable the computation of the program in the preceding section.
The ; Scenario section computes all predicted probabilities for the model using the sample
data and the estimated parameters. Then, it recomputes the probabilities after changing the variables
in the way specified in the scenarios. (The actual data are not changed – the modification is done
while the probabilities are computed.) The scenarios are of the form

variable operation = value

such as hhkids + = 1 (effect of additional kids in the home)

or hhninc * = 1.1 (effect of a 10% increase in income)
N6: Probit and Logit Models: Estimation N-88

You may provide multiple scenarios. They are evaluated one at a time. This is an extension of the
computation of marginal effects.
In the example below, we extend the analysis of marginal effect in the logit model used
above. The scenario examined is the impact of every individual having one more child in the
household then having a 50% increase in income. (Since hhkids is actually a dummy variable for the
presence of kids in the home, increasing it by one is actually an ambiguous experiment. We retain it
for the sake of a simple numerical example.) The plot shows the effect of income on the probability
of visiting the doctor, according to the model.

NAMELIST ; x = one,age,educ,married,hhninc,hhkids $
LOGIT ; Lhs = doctor ; Rhs = x $
BINARY ; Lhs = doctor ; Rhs = x
; Model = Logit ; Start = b
; Scenario: hhkids + = 1 / hhninc * = 1.5 $

The model output is omitted for brevity.

+-------------------------------------------------------------+
|Scenario 1. Effect on aggregate proportions. Logit Model |
|Threshold T* for computing Fit = 1[Prob > T*] is .50000 |
|Variable changing = HHKIDS , Operation = +, value = 1.000 |
+-------------------------------------------------------------+
|Outcome Base case Under Scenario Change |
| 0 33 = .98% 831 = 24.61% 798 |
| 1 3344 = 99.02% 2546 = 75.39% -798 |
| Total 3377 = 100.00% 3377 = 100.00% 0 |
+-------------------------------------------------------------+
+-------------------------------------------------------------+
|Scenario 2. Effect on aggregate proportions. Logit Model |
|Threshold T* for computing Fit = 1[Prob > T*] is .50000 |
|Variable changing = HHNINC , Operation = *, value = 1.500 |
+-------------------------------------------------------------+
|Outcome Base case Under Scenario Change |
| 0 33 = .98% 106 = 3.14% 73 |
| 1 3344 = 99.02% 3271 = 96.86% -73 |
| Total 3377 = 100.00% 3377 = 100.00% 0 |
+-------------------------------------------------------------+

The SIMULATE command used in the example provides a greater range of scenarios that
one can examine to see the effects of changes in a variable on the overall prediction of the binary
choice model. The advantage of the BINARY command used here is that for straightforward
scenarios, it can be used to provide useful tables such as the ones shown above.
N6: Probit and Logit Models: Estimation N-89

N6.8 Using Weights and Choice Based Sampling

The ; Wts option can always be used in the usual fashion for the probit and logit models.
However, in the grouped data case, a somewhat different treatment may be desired. The
observations may consist of pi, xi and ni, where ni is the number of replications used to obtain pi. The
usual treatment assumes that pi is a sample of one from a distribution with variance pi(1-pi). But pi is
more precise than this. Its unconditional variance is pi(1-pi)/ni. Thus, the efficiency of the estimator
of b is underestimated. There is also an inherent heteroscedasticity which must be accounted for.
The heteroscedasticity due to pi is built into the likelihood function. But if your proportions are
based on different numbers of observations, the variances will differ correspondingly. This can be
accounted for by including ni as a weighting variable. Since the weighting procedure automatically
scales the weights so that they sum to the sample size, which would be inappropriate here, it is
necessary to modify the specification. Use

; Wts = variable, Noscale

or just ; Wts = variable, N

to prevent the automatic scaling. This produces a replication of the observations, which is what is
needed for grouped data.
This usage often has the surprising side effect of producing implausibly small standard
errors. Consider, for example, using unscaled weights for statewide observations on election
outcomes. The implication of the Noscale parameter is that each proportion represents millions of
observations. Once again, this is an issue that must be considered on a case by case basis.

Choice Based Sampling

In some individual data cases, the data are deliberately sampled so that one or the other
outcome is overrepresented in the sample. For example, suppose that in a binary response setting,
the true proportion of ones in the population is .05 and the true proportion of zeros is .95. One might
over sample the ones in order to learn more about the decision process. However, some account
must be taken of this fact in the estimation since it obviously will impart some biases. The following
assumes that these population proportions are known, which must be true to apply the technique.
We use the assumed values to demonstrate the technique; other values would be substituted in the
analogous manner.
The general principle involved is as follows: Suppose that the sample is deliberately drawn
so that it contains 50% ones and 50% zeros while it is known that the true proportions in the
population are .05 and .95. Then, the ones are overrepresented by a factor of .50/.05 = 10 while the
zeros are underrepresented by a factor of .50/.95 = .5263. To obtain the right ‘mix’ in the sample, it
is necessary to scale down the ones by a factor of .05/.50 = .1 and scale up the zeros by a factor of
.95/.50 = 1.9. This can be handled simply by using a weighting variable during estimation to
reweight the observations. The precise method of doing so is discussed below. (See, also, Manski
and McFadden (1981).)
N6: Probit and Logit Models: Estimation N-90

An additional change must be made in order to obtain the correct asymptotic covariance
matrix for the estimates. Let H be the Hessian of the (weighted) log likelihood, i.e., the usual
estimator for the variance matrix of the estimates, and let G′G be the summed outer products of the
first derivatives of the (weighted) log likelihood. (This is the inverse of the BHHH estimator.)
Manski and McFadden (1981) show that the appropriate covariance matrix for the estimates is

V = (-H)-1 G′G (-H)-1.

The computation of the weighted estimator and the corrected asymptotic covariance is handled
automatically in NLOGIT by the following estimation programs:

• univariate probit, logit, extreme value and Gompertz model,

• bivariate probit model with and without sample selection,
• binomial and multinomial logit models,
• discrete choice (conditional logit).

With the exception of the last of these, you request the estimator with

; Wts = name of weighting variable

; Choice Based

The weighting variable can usually be created with a single command. For example, the weighting
variable suggested in the example used above would be specified as follows:

CREATE ; wt = (.95/.50)*(y = 0) + (.05/.50)*(y = 1) $

For models that do not appear in the list above, there is a general way to do this kind of
computation. How the weights are obtained will be specific to your application if you wish to do
this. To compute the counterpart to V above, you can do the following:

CREATE ; wt = the desired weighting variable $

Model name ; ... specification of the model
; Wts = the weighting variable
; Cluster = 1 $

Since the ‘cluster’ estimator computes a sandwich estimator, we need only ‘trick’ the program by
specifying that each cluster contains one observation. The observations in the parts will be weighted
by the variable given, so this is exactly what is needed.
N6: Probit and Logit Models: Estimation N-91

N6.9 Heteroscedasticity in Probit and Logit Models

The univariate choice model with multiplicative heteroscedasticity is

yi* = b′xi + ei,

yi = 1 if yi* > 0 and yi = 0 if yi* ≤ 0,
2
ei ~ Normal or Logistic with mean 0, and variance ∝ [exp(γ′wi)]

(In the logistic case, the true variance is scaled by π2/3.)

NOTE: These heteroscedasticity models require individual data.

Request the model with heteroscedasticity with

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = regressors in x
; Rh2 = list of variables in w
; Heteroscedasticity (or just ; Het) $

Other options and specifications for this model are the same as the basic model. Two general
options that are likely to be useful are

; Keep = name to retain predicted values

; Prob = name to retain fitted probabilities

and the controls of the iterations and the amount of output.

NOTE: Do not include one in the Rh2 list. A constant in γ is not identified.

This model differs from the basic model only in the presence of the variance term. The
output for this model is also the same, with the addition of the coefficients for the variance term. The
initial OLS results are computed without any consideration of the heteroscedasticity, however.
Since the log likelihood for this model, unlike the basic model, is not globally concave, the
default algorithm is BFGS, not Newton’s method.
For purposes of hypothesis testing and imposing restrictions, the parameter vector is

q = [b1,...,bK,γ1,...,γL].

If you provide your own starting values, give the right number of values in exactly this order.
You can also use WALD and ; Test: to test hypotheses about the coefficient vector. Finally,
you can impose restrictions with

; Rst = ....
or ; CML: restrictions...
N6: Probit and Logit Models: Estimation N-92

NOTE: In principle, you can impose equality restrictions across the elements of b and γ with
; Rst = ..., (i.e., force an element in b to equal one in γ), but the results are unlikely to be satisfactory.
Implicitly, the variables involved are of different scales, and this will place a rather stringent
restriction on the model.

Use
; Robust
or ; Cluster = id variable or group size

to request the sandwich style robust covariance matrix estimator or the cluster correction.

NOTE: There is no ‘robust’ covariance matrix for the logit or probit model that is robust to
heteroscedasticity, in the form of the White estimator for the linear model. In order to accommodate
heteroscedasticity in a binary choice model, you must model it explicitly.

NOTE: ; Maxit = 0 provides an easy way to test for heteroscedasticity with an LM test.

To test the hypothesis of homoscedasticity against the specification of this more general
model, the following template can be used: (The model may be LOGIT if desired.)

NAMELIST ; x = ... the Rhs of the probit model

; w = ... the Rh2 of the heteroscedasticity model $
CALC ; m = Col(w) $
PROBIT ; Lhs = ...
; Rhs = x $
PROBIT ; Lhs = ...
; Rhs = x
; Rh2 = w ; Het
; Start = b, m_0
; Maxit = 0 $

This produces an LM statistic and (superfluously) reproduces the restricted model.

The results that are saved automatically are the same as for the basic model, that is, b, varb,
and the scalars. In this case, b will contain the full set of estimates, with the slopes followed by the
variance parameters, i.e., [b,c]. The Last Model labels for the WALD command are [b_variable,
c_variable].
We note, this model may be rather weakly identified by the observed data, unless they are
plentiful and the model is sharply consistent with the data. In fact, identification is not a problem,
and the model is straightforward to estimate. But, one could argue that the specification problem
addressed by this model is one of functional form rather than heteroscedasticity. That is, the model
specification is arguably indistinguishable from one with a peculiar kind of conditional mean
function, which, in turn, could be standing in for some other, perhaps reasonable, albeit nonlinear
model. In addition, it is common for the estimated standard errors that are computed for this model
to be quite large, as a result of a kind of multicollinearity – the high correlation of the derivatives of
the log likelihood.
N6: Probit and Logit Models: Estimation N-93

Application

To illustrate the model, we have refit the specification of the previous section with a
variance term of the form Var[e] = [exp(γ1female + γ2working )]2. Since both of these are binary
variables, this is equivalent to a groupwise heteroscedasticity model. The variances are 1.0, exp(2γ1),
exp(2γ2) and exp(2γ1+2γ2) for the four groups. We have fit the original model without
heteroscedasticity first. The second LOGIT command carries out the LM test of heteroscedasticity.
The third command fits the full heteroscedasticity model.

INCLUDE ; New ; year = 1994 $

NAMELIST ; x = one,age,educ,married,hhninc,hhkids,female $
LOGIT ; Lhs = doctor ; Rhs = x
; Partial Effects $
NAMELIST ; w = female,working $
CALC ; m = Col(w) $
LOGIT ; Lhs = doctor ; Rhs = x
; Heteroscedasticity ; Rh2 = w
; Start = b,m_0
; Maxit = 0 $
LOGIT ; Lhs = doctor ; Rhs = x
; Heteroscedasticity ; Rh2 = w
; Partial Effects $
PARTIALS ; Effects: female $

The model results have been rearranged in the listing below to highlight the differences in the
models. Also, for convenience, some of the results have been omitted.

Binary Logit Model for Binary Choice

Dependent variable DOCTOR
Log likelihood function -2085.33796

The LM statistic is included in the initial diagnostic statistics for the second model estimated.

LM Stat. at start values 3.11867

LM statistic kept as scalar LMSTAT

These are the results for the model with homoscedastic disturbances.

Inf.Cr.AIC = 4184.676 AIC/N = 1.239

Restricted log likelihood -2169.26982
McFadden Pseudo R-squared .0386913

These are the coefficient estimates for the two models.

N6: Probit and Logit Models: Estimation N-94

Homoscedastic disturbances
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .14726 .25460 .58 .5630 -.35173 .64626
AGE| .01643*** .00384 4.28 .0000 .00891 .02395
EDUC| -.01965 .01608 -1.22 .2219 -.05117 .01188
MARRIED| .15536 .09904 1.57 .1167 -.03875 .34947
HHNINC| -.39474** .17993 -2.19 .0282 -.74739 -.04208
HHKIDS| -.41534*** .08866 -4.68 .0000 -.58911 -.24157
FEMALE| .64274*** .07643 8.41 .0000 .49295 .79253
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Heteroscedastic disturbances
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .12927 .30739 .42 .6741 -.47320 .73174
AGE| .02036*** .00501 4.06 .0000 .01053 .03018
EDUC| -.02913 .01984 -1.47 .1421 -.06803 .00976
MARRIED| .19969 .12639 1.58 .1141 -.04803 .44742
HHNINC| -.36965* .22169 -1.67 .0954 -.80414 .06485
HHKIDS| -.53029*** .12783 -4.15 .0000 -.78083 -.27974
FEMALE| 1.24685*** .45754 2.73 .0064 .35009 2.14361
|Disturbance Variance Terms
FEMALE| .44128* .25946 1.70 .0890 -.06725 .94982
WORKING| .08459 .10082 .84 .4014 -.11300 .28219
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

These are the marginal effects for the two models. Note that the effects are also computed for the
terms in the variance function. The explanatory text indicates the treatment of variables that appear
in both the linear part and the exponential part of the probability.

+-------------------------------------------+
| Partial derivatives of probabilities with |
| respect to the vector of characteristics. |
| They are computed at the means of the Xs. |
| Effects are the sum of the mean and var- |
| iance term for variables which appear in |
| both parts of the function. |
+-------------------------------------------+
+--------+--------------+----------------+--------+--------+----------+
|Variable| Coefficient | Standard Error |b/St.Er.|P[|Z|>z]|Elasticity|
+--------+--------------+----------------+--------+--------+----------+
N6: Probit and Logit Models: Estimation N-95

Homoscedastic disturbances
-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00352*** -.00205 4.29 .0000 .00191 .00512
EDUC| -.00421 .00058 -1.22 .2218 -.01096 .00254
MARRIED| .03357 -.00031 1.56 .1194 -.00868 .07582 #
HHNINC| -.08452** .00044 -2.20 .0282 -.16000 -.00905
HHKIDS| -.09058*** .00027 -4.65 .0000 -.12876 -.05240 #
FEMALE| .13842*** -.00119 8.60 .0000 .10687 .16997 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Heteroscedastic disturbances
-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effects are the sum of the mean and var-
iance term for variables which appear in
both parts of the function.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
AGE| .00337*** .20980 3.84 .0001 .00165 .00509
EDUC| -.00482 -.08104 -1.47 .1404 -.01123 .00159
MARRIED| .03306 .03424 1.59 .1119 -.00769 .07380
HHNINC| -.06119 -.03975 -1.63 .1038 -.13492 .01254
HHKIDS| -.08778*** -.04969 -4.45 .0000 -.12640 -.04916
FEMALE| .20639*** .13969 5.09 .0000 .12687 .28592
|Disturbance Variance Terms
FEMALE| -.07388 -.05000 -1.08 .2784 -.20747 .05972
WORKING| -.01416 -.01493 -.71 .4801 -.05347 .02514
|Sum of terms for variables in both parts
FEMALE| .13252*** .08969 3.52 .0004 .05875 .20629
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The partial effects for the heteroscedasticity model are computed at the means of the
variables. It is possible to obtain average partial effects by using the PARTIAL EFFECTS program
rather than the built in marginal effects routine. The following shows the results for female, which
appears in both parts of the model.

PARTIAL EFFECTS ; Effects: female $

N6: Probit and Logit Models: Estimation N-96

---------------------------------------------------------------------
Partial Effects Analysis for Heteros. Logit Prob.Function
---------------------------------------------------------------------
Effects on function with respect to FEMALE
Results are computed by average over sample observations
Partial effects for binary var FEMALE computed by first difference
---------------------------------------------------------------------
df/dFEMALE Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
APE. Function .13430 .01653 8.12 .10190 .16669

These are the summaries of the predictions of the two estimated models. The performance of the
two models in terms of the simple count of correct predictions is almost identical – the
heteroscedasticity model correctly predicts three observations more than the homoscedasticity
model. The mix of correct predictions is very different, however.

Homoscedastic disturbances
+---------------------------------------------------------+
|Predictions for Binary Choice Model. Predicted value is |
|1 when probability is greater than .500000, 0 otherwise.|
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Value | |
|Value | 0 1 | Total Actual |
+------+----------------+----------------+----------------+
| 0 | 82 ( 2.4%)| 1073 ( 31.8%)| 1155 ( 34.2%)|
| 1 | 85 ( 2.5%)| 2137 ( 63.3%)| 2222 ( 65.8%)|
+------+----------------+----------------+----------------+
|Total | 167 ( 4.9%)| 3210 ( 95.1%)| 3377 (100.0%)|
+------+----------------+----------------+----------------+

Heteroscedastic disturbances
+---------------------------------------------------------+
|Predictions for Binary Choice Model. Predicted value is |
|1 when probability is greater than .500000, 0 otherwise.|
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Value | |
|Value | 0 1 | Total Actual |
+------+----------------+----------------+----------------+
| 0 | 131 ( 3.9%)| 1024 ( 30.3%)| 1155 ( 34.2%)|
| 1 | 139 ( 4.1%)| 2083 ( 61.7%)| 2222 ( 65.8%)|
+------+----------------+----------------+----------------+
|Total | 270 ( 8.0%)| 3107 ( 92.0%)| 3377 (100.0%)|
+------+----------------+----------------+----------------+
N7: Tests and Restrictions in Models for Binary Choice N-97

N7: Tests and Restrictions in Models for Binary

Choice
N7.1 Introduction
We define models in which the response variable being described is inherently discrete as
qualitative response (QR) models. Chapter N6 presented the model formulation and estimation and
analysis tools. This chapter will detail some aspects of hypothesis testing. Most of these results are
generic, and will apply in other models as well.

N7.2 Testing Hypotheses

The full set of options is available for testing hypotheses and imposing restrictions on the
binary choice models. In using these, the set of parameters is

b1, ..., bK plus γ for the Burr model

In the parametric models, hypotheses can be done with the standard trinity of tests: Wald,
likelihood ratio and Lagrange Multiplier. All three are particularly straightforward for the binary
choice models.

N7.2.1 Wald Tests

Wald tests are carried out in two ways, with the ; Test: specification in the model command
and by using the WALD command after fitting the model. The former is used for linear restrictions.
The WALD command is more general and allows for tests of nonlinear restrictions on parameters.
The Wald statistic is computed using the estimates of an unrestricted model. The hypothesis
implies a set of restrictions

H0: c(b) = 0.

(This may involve linear distance from a constant, such as 2b3 - 1.2 = 0. The preceding formulation
is used to achieve the full generality that NLOGIT allows.) The Wald statistic is computed by the
formula
( ) ( ){ ( )} ( ) ()
−1
ˆ ' G ˆ Est. Asy.Var ˆ G ˆ ' c ˆ
W = c βββββ
 

()
∂c βˆ
where ( )
G β̂ =
∂βˆ '

and β̂ is the vector of estimated parameters.

N7: Tests and Restrictions in Models for Binary Choice N-98

You can request Wald tests of simple restrictions by including the request in the model
command. For example:

PROBIT ; Lhs = doctor

; Rhs = one,age,educ,married,hhninc,hhkids
; Test: age + educ = 0,
married = 0 ,
hhninc + 2*hhkids = -.3 $

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -17670.94233
Restricted log likelihood -18019.55173
Chi squared [ 5 d.f.] 697.21881
Significance level .00000
McFadden Pseudo R-squared .0193462
Estimation based on N = 27326, K = 6
Inf.Cr.AIC =35353.885 AIC/N = 1.294
Hosmer-Lemeshow chi-squared = 105.22799
P-value= .00000 with deg.fr. = 8
Wald test of 3 linear restrictions
Chi-squared = 26.06, P value = .00001
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| .15500*** .05652 2.74 .0061 .04423 .26577
AGE| .01283*** .00079 16.24 .0000 .01129 .01438
EDUC| -.02812*** .00350 -8.03 .0000 -.03498 -.02125
MARRIED| .05226** .02046 2.55 .0106 .01216 .09237
HHNINC| -.11643** .04633 -2.51 .0120 -.20723 -.02563
HHKIDS| -.14118*** .01822 -7.75 .0000 -.17689 -.10548
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Note that the results reported are for the unrestricted model, and the results of the Wald test are
reported with the initial header information. To fit the model subject to the restriction, we change
; Test: in the command to ; CML: with the following results:

PROBIT ; Lhs = doctor

; Rhs = one,age,educ,married,hhninc,hhkids
; CML: age + educ = 0,
married = 0 ,
hhninc + 2*hhkids = -.3 $
N7: Tests and Restrictions in Models for Binary Choice N-99

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -2125.57999
Restricted log likelihood -2169.26982
Chi squared [ 2 d.f.] 87.37966
Significance level .00000
McFadden Pseudo R-squared .0201403
Estimation based on N = 3377, K = 3
Inf.Cr.AIC = 4257.160 AIC/N = 1.261
Linear constraints imposed 3
Hosmer-Lemeshow chi-squared = 20.93392
P-value= .00733 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| .04583 .06144 .75 .4557 -.07458 .16624
AGE| .01427*** .00192 7.44 .0000 .01052 .01803
EDUC| -.01427*** .00192 -7.44 .0000 -.01803 -.01052
MARRIED| 0.0 .....(Fixed Parameter).....
HHNINC| -.06304 .07079 -.89 .3731 -.20178 .07569
HHKIDS| -.11848*** .03539 -3.35 .0008 -.18785 -.04911
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

When the restrictions are built into the estimator with CML, the information reported is only that the
restrictions were imposed. The results of the Wald or LR test cannot be reported because the
unrestricted model is not computed.

N7.2.2 Likelihood Ratio Tests

Use the log likelihood functions from both restricted and unrestricted models. Log
likelihood functions are saved automatically by the estimators. Do keep in mind that these are
overwritten each time – the scalar logl gets replaced by each model command. Your general strategy
for carrying out a likelihood ratio test would be

Model name ; ... - specifies the unrestricted model

CALC ; lu = logl $ Capture log likelihood function
Model name ; ... - specifies the restricted model
CALC ; lr = logl
; List ; chisq = 2*(lu - lr )
; 1 - Chi(chisq, degrees of freedom) $

You must supply the degrees of freedom. If the result of the last line is less than your significance
level – usually 0.05 – then, the null hypothesis of the restriction would be rejected. Here are two
examples: We continue to examine the German health care data. For purposes of these tests, just for
the illustrations, we will switch to a probit model.
N7: Tests and Restrictions in Models for Binary Choice N-100

Simple Linear Restriction

The following tests the pair of linear restrictions suggested above. Looking at the unrestricted
results from earlier, the restrictions don’t look like they are going to pass. The results bear this out.

SAMPLE ; All $
NAMELIST ; x = one,age,educ,married,hhninc,hhkids $
LOGIT ; Lhs = doctor ; Rhs = x $
CALC ; lu = logl $
LOGIT ; Lhs = doctor ; Rhs = x
; Rst = b0, b1, b1, 0, b2, b3 $
CALC ; lr = logl
; List ; chisq = 2*(lu - lr) ; 1 - Chi(chisq,2) $
[CALC] CHISQ = 158.9035080
[CALC] *Result*= .0000000
Calculator: Computed 3 scalar results

Homogeneity Test
We are frequently asked about this. The sample can be partitioned into a number of
subgroups. The question is whether it is valid to pool the subgroups. Here is a general strategy that
is the maximum likelihood counterpart to the Chow test for linear models: Define a variable, say,
group, that takes values 1,2,...,G, that partitions the sample. This is a stratification variable. The test
statistic for homogeneity is

χ2 = 2[(Σgroups log likelihood for the group) - log likelihood for the pooled sample]
The degrees of freedom is G-1 times the number of coefficients in the model.
Create the group variable.

SAMPLE ; Pooled sample ... however defined ... $

Model name ; ... ; Quiet $ Specify the appropriate model. Suppress the output.
CALC ; chisq = -2*logl ; df = -kreg $
Automate the model fitting estimation, and accumulate the statistic.
PROC $
INCLUDE ; New ; Group = i $
Model name ; ... ; Quiet $ Specify the same model. Suppress the output.
CALC ; chisq = chisq + 2*logl ; df = df + kreg $
ENDPROC $
Determine the number of groups.

CALC ; g = Max(group) $
Estimate the model once for each group.
EXEC ; i = 1,g $
CALC ; List ; chisq ; df ; 1 - Chi(chisq,df) $
N7: Tests and Restrictions in Models for Binary Choice N-101

This procedure produces only the output of the last CALC command, which will display the test
statistic, the degrees of freedom and the p value for the test.
To illustrate, we’ll test the hypothesis that the same probit model for doctor visits applies to
both men and women. This command suppresses all output save for the actual test of the hypothesis.

NAMELIST ; x = one,age,educ,married,hhninc,hhkids $
PROBIT ; If [ female = 0] ; Lhs = doctor ; Rhs = x ; Quiet $
CALC ; l0 = logl $
PROBIT ; If [ female = 1] ; Lhs = doctor ; Rhs = x ; Quiet $
CALC ; l1 = logl $
PROBIT ; Lhs = doctor ; Rhs = x ; Quiet $
CALC ; l01 = logl ; List
; chisq = -2*(l01 - l0 - l1)
; df = 2*kreg ; pvalue = 1 - Chi(chisq,df) $

The results of the chi squared test strongly reject the homogeneity restriction.

[CALC] CHISQ = 549.2873320

[CALC] DF = 12.0000000
[CALC] PVALUE = .0000000
Calculator: Computed 4 scalar results

The homogeneity test shown above can be automated in the probit command. For the
preceding, we would use

PROBIT ; For [(test) female] ; Lhs = doctor ; Rhs = x ; Quiet $

This will produce the following:

----------------------------------------------------
Setting up an iteration over the values of FEMALE
The model command will be executed for 2 values
of this variable. In the current sample of 27326
observations, the following counts were found:
Subsample -Observations Subsample -Observations
FEMALE = 0 14243 FEMALE = 1 13083
FEMALE = **** 27326
Actual subsamples may be smaller if missing values
are being bypassed. Subsamples with 0 observations
will be bypassed.
----------------------------------------------------
Subsample analyzed for this command is FEMALE = 0
Subsample analyzed for this command is FEMALE = 1
Full pooled sample is used for this iteration.
-----------------------------------------------------------------------
Homogeneity Test for Estimated Model
-----------------------------------------------------------------------
The model was estimated for 2 subsamples and the full sample
The likelihood ratio statistic is 2[Sum(g=1...G)logL(g) - logL(pooled)]
Chi squared = 549.2873 Estimated degrees of freedom = 6
Estimated P value for this test is .0000
-----------------------------------------------------------------------
N7: Tests and Restrictions in Models for Binary Choice N-102

N7.2.3 Lagrange Multiplier Tests

The third procedure available for testing hypotheses is the Lagrange Multiplier, or LM
approach. The Lagrange Multiplier statistic is computed as a Wald statistic for testing the hypothesis
that the derivatives of the log likelihood are zero when evaluated at the restricted maximum
likelihood estimator;
( ) { ( )} ( )
−1
LM ˆ '  Est. Asy.Var g ˆ
= g βββ  g ˆ
R
 R
 R

where βˆ R = MLE of the parameters of the model, with restrictions imposed

( )
g βˆ R = derivatives of log likelihood of full model, evaluated at βˆ R

The estimated asymptotic covariance matrix of the gradient is any of the usual estimators of the
asymptotic covariance matrix of the coefficient estimator, negative inverse of the actual or expected
Hessian, or the BHHH estimator based on the first derivatives only.
Your strategy for carrying out LM tests with NLOGIT is as follows:

Step 1. Obtain the restricted parameter vector. This may involve an unrestricted parameter vector in
some restricted model, padded with some zeros, or a similar arrangement.

Step 2. Set up the full, unrestricted model as if it were to be estimated, but include in the command

; Start = restricted parameter vector

; Maxit = 0

The rest of the procedure is automated for you. The ; Maxit = 0 specification takes on a particular
meaning when you also provide a set of starting values. It implies that you wish to carry out an LM
test using the starting values.
To demonstrate, we will carry out the test of the hypothesis

b_age + b_educ = 0
b_married = 0
b_hhninc + b_hhkids = - .3

that we tested earlier with a Wald statistic, now with the LM test. The commands would be as follows:

PROBIT ; Lhs = doctor

; Rhs = one,age,educ,married,hhninc,hhkids
; CML: age+educ = 0, married = 0 , hhninc + 2*hhkids = -.3 $
PROBIT ; Lhs = doctor
; Rhs = one,age,educ,married,hhninc,hhkids
; Maxit = 0 ; Start = b $

The results of the second model command provide the Lagrange multiplier statistic. The value of
26.06032 is the same as the Wald statistic computed earlier, 26.06.
N7: Tests and Restrictions in Models for Binary Choice N-103

Maximum of 0 iterations. Exit iterations with status=1.

Maxit = 0. Computing LM statistic at starting values.
No iterations computed and no parameter update done.
-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
LM Stat. at start values 26.06032
LM statistic kept as scalar LMSTAT
Log likelihood function -17683.96508
Restricted log likelihood -18019.55173
Chi squared [ 5 d.f.] 671.17331
Significance level .00000
McFadden Pseudo R-squared .0186235
Estimation based on N = 27326, K = 6
Inf.Cr.AIC =35379.930 AIC/N = 1.295
Hosmer-Lemeshow chi-squared = 132.57086
P-value= .00000 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| -.06593 .05655 -1.17 .2437 -.17678 .04491
AGE| .01484*** .00079 18.76 .0000 .01329 .01639
EDUC| -.01484*** .00351 -4.23 .0000 -.02171 -.00796
MARRIED| 0.0 .02049 .00 1.0000 -.40156D-01 .40156D-01
HHNINC| -.09655** .04636 -2.08 .0373 -.18741 -.00568
HHKIDS| -.10173*** .01821 -5.59 .0000 -.13742 -.06603
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

To complete the trinity of tests, we can carry out the likelihood ratio test, which we could do
as follows:

PROBIT ; Quiet ; Lhs = doctor

; Rhs = one,age,educ,married,hhninc,hhkids
; CML: b(2) + b(3) = 0, b(4) = 0, b(5) + b(6) = -.3 $
CALC ; lr = logl $
PROBIT ; Quiet ; Lhs = doctor
; Rhs = one,age,educ,married,hhninc,hhkids $
CALC ; lu = logl ; List
; lrstat = 2*(lu – lr) $

The result of the computation (which displays only the last statistic) is

[CALC] LRSTAT = 26.0455042

Calculator: Computed 2 scalar results

The value of 26.0455 differs only trivially from the other values. This is actually not surprising,
since they should all converge to the same statistic, and the sample in use here is very large.
N7: Tests and Restrictions in Models for Binary Choice N-104

N7.3 Two Specification Tests

The following are two specialized tests for the probit model, one for testing which of two
competing models appears to be appropriate, and one test against the hypothesis of normality that
underlies the probit model.

N7.3.1 A Test for Nonnested Probit Models

Davidson and MacKinnon (1993) present a test of the nonnested hypothesis that an
alternative set of variables, zi, is the appropriate one for the structural equation of the probit model.

Testing y* = x′b + e vs. y* = z′γ + u

NAMELIST ; x = the independent variables

; z = the competing list of independent variables $
CREATE ; y = the dependent variable $
PROBIT ; Quiet ; Lhs = y ; Rhs = x $
CREATE ; xbeta = x’b ; fx = N01(xbeta) ; px = Phi(xbeta)
; v = Sqr(px*(1-px)) ; dev = (y - px) / v
; xv = fx*xbeta / v $
PROBIT ; Quiet ; Lhs = y ; Rhs = z $
CREATE ; pz = Phi(z’b) ; test = (px - pz) / v $
REGRESS ; Lhs = dev ; Rhs = xv,test $

The test is carried out by referring the t ratio on test to the t table. A value larger than the critical
value argues in favor of z as the correct specification. For example, the following tests for which of
two specifications of the right hand side of the probit model is preferred.

NAMELIST ; x = one,age,educ,married,hhninc,hhkids,self
; z = one,age,educ,married,hhninc,female,working $
CREATE ; y = doctor $

The remaining commands are identical.

The essential regression results are as follows. We also reversed the roles of x and z.
Unfortunately, as often happens in specifications, the results are contradictory.

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DEV| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
XV| .04569** .01985 2.30 .0214 .00678 .08459
TEST| -.79517*** .03995 -19.90 .0000 -.87348 -.71687
--------+--------------------------------------------------------------------
XV| .04668** .02033 2.30 .0217 .00684 .08652
TEST| -.26126*** .04273 -6.11 .0000 -.34500 -.17751

The t ratio of -19.9 in the first regression argues in favor of z as the appropriate specification. But,
the also significant t ratio of -6.11 in the second argues in favor of x.
N7: Tests and Restrictions in Models for Binary Choice N-105

N7.3.2 A Test for Normality in the Probit Model

The second test is a Lagrange multiplier test against the null hypothesis of normality in the
probit model. (The test was developed in Bera, Jarque and Lee (1984).) As usual in normality tests,
the statistic is computed by comparing the third and fourth moments of an underlying variable to
their expected value under normality. The computations are as follows, where i indicates the ith
observation:
ai = xi′b
φi = φ(ai)
Φi = Φ(ai)
di = φi (yi - Φi) / [Φi(1 - Φi)]
ci = φi2 / [Φi(1 - Φi)]
m3i = -1/2(ai2 – 1)
m4i = 1/4 (ai (ai2 + 3))
zi = (xi′, m3i, m4i)′

( ′
)( )( )
−1
∑ ∑ ′ ∑ i 1 di z i
N N N
Then, LM =
=i 1 =i 1
d i z i ci z=i z i

The commands below will carry out the test. The chi squared reported by the last line has two
degrees of freedom.

NAMELIST ; x = one,... $
CREATE ; y = the dependent variable $
PROBIT ; Lhs = y ; Rhs = x $
CREATE ; ai = b'x ; fi = Phi(ai) ; dfi = N01(ai)
; di = (y-fi) * dfi /(fi*(1-fi)) ; ci = dfi^2 /(fi*(1-fi))
; m3i = -1/2*(ai^2-1) ; m4i = 1/4*(ai*(ai^2+3)) $
NAMELIST ; z = x,m3i,m4i $
MATRIX ; List ; LM = di’z * <z'[ci]z> * z'di $

We executed the routine for our probit model estimated earlier, with

NAMELIST ; x = one,age,educ,married,hhninc,hhkids,self $
CREATE ; y = doctor $

The result of 93.12115 would lead to rejection of the hypothesis of normality; the 5% critical value
for the chi squared variable with two degrees of freedom is 5.99.

LM| 1
--------+--------------
1| 93.1211
N7: Tests and Restrictions in Models for Binary Choice N-106

N7.4 The WALD Command

The WALD command may be used for linear and nonlinear restrictions. The model
commands produce a set of names that can be used in WALD commands after estimation. For the
binary choice commands, these are b_variable. The WALD command can be used with these names
in specified restrictions, with no other information needed. For example:

PROBIT ; Lhs = doctor

; Rhs = one,age,educ,married,hhninc,hhkids $
WALD ; Fn1 = b_age + b_educ - 0
; Fn2 = b_married - 0
; Fn3 = b_hhninc + b_hhkids + .3 $

(The latter restriction doesn’t make much sense, but we can test it anyway.) The results of this pair
of commands are shown below. (The PROBIT command was shown earlier.)

-----------------------------------------------------------------------------
WALD procedure. Estimates and standard errors
for nonlinear functions and joint test of
nonlinear restrictions.
Wald Statistic = 24.95162
Prob. from Chi-squared[ 3] = .00002
Functions are computed at means of variables
-----------------------------------------------------------------------------

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
WaldFcns| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Fncn(1)| -.01528*** .00369 -4.14 .0000 -.02252 -.00805
Fncn(2)| .05226** .02046 2.55 .0106 .01216 .09237
Fncn(3)| .04239 .05065 .84 .4027 -.05689 .14166
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

You may follow a model command with as many WALD commands as you wish.
You can use WALD to obtain standard errors for linear or nonlinear functions of parameters.
Just ignore the test statistics. Also, WALD produces some useful output in addition to the displayed
results. The new matrix varwald will contain the estimated asymptotic covariance matrix for the set of
functions. The new vector waldfns will contain the values of the specified functions. A third matrix,
jacobian, will equal the derivative matrix, ∂c(b)/∂b′. For the computations above, the three matrices are
N7: Tests and Restrictions in Models for Binary Choice N-107

Figure N7.1 Matrix Results for the WALD Command

Thus, the command

MATRIX ; w = waldfns’ <varwald> waldfns $

would recompute the Wald statistic.

Matrix W has 1 rows and 1 columns.
1
+--------------
1| 24.95162

N7.5 Imposing Linear Restrictions

Fixed Value and Equality Restrictions
Fixed value and equality restrictions are imposed with

; Rst = the list of settings symbols for free parameters,

values for specific values
For example,

NAMELIST ; x = one,age,educ,married,hhninc,hhkids $
LOGIT ; Lhs = doctor ; Rhs = x
; Rst = b0, b1, b1, 0, b2, b3 $

will force the second and third coefficients to be equal and the fourth to equal zero.
N7: Tests and Restrictions in Models for Binary Choice N-108

Linear Restrictions

These are imposed with

; CML: the set of linear restrictions

(See Section R13.6.3.) This is a bit more general than the Rst function, but similar. For example, to
force the restriction that the coefficient on age plus that on educ equal twice that on hhninc, use

; CML: age + educ - 2*hhninc = 0

N8: Extended Binary Choice Models N-109

N8: Extended Binary Choice Models

N8.1 Introduction
NLOGIT supports a large variety of models and extensions for the analysis of binary choice.
This chapter documents sample selection models, models with endogenous right hand side variables
and two step estimation of models that build on probit and logit models.

N8.2 Sample Selection in Probit and Logit Models

The model of sample selection can be extended to the probit and logit binary choice models.
In both cases, we depart from

Prob[yi = 1 |xi] = F(b′xi)

where F(t) = Φ(t) for the probit model and Λ(t) for the logit model,
zi* = α′wi + ui, ui ~ N[0,1], zi = 1(zi* > 0)
yi, xi observed only when zi = 1.

In both cases, as stated, there is no obvious way that the selection mechanism impacts the binary
choice model of interest. We modify the models as follows:
For the probit model,

yi* = b′xi + ei, ei ~ N[0,1], yi = 1(yi* > 0)

which is the structure underlying the probit model in any event, and

ui, ei ~ BVN[(0,0),(1,ρ,1)].

This is precisely the structure underlying the bivariate probit model. Thus, the probit model with
selection is treated as a bivariate probit model. Some modification of the model is required to
accommodate the selection mechanism. The command is simply

BIVARIATE ; Lhs = y,z

; Rh1 = variables in x
; Rh2 = variables in w
; Selection $

For the logit model, a similar approach does not produce a convenient bivariate model. The
probability is changed to
exp(β′xi + σei )
Prob(yi = 1 | xi,ei) = .
1 + exp(β′xi + σei )
N8: Extended Binary Choice Models N-110

With the selection model for zi as stated above, the bivariate probability for yi and zi is a mixture of a
logit and a probit model. The log likelihood can be obtained, but it is not in closed form, and must
be computed by approximation. We do so with simulation. The commands for the model are

PROBIT ; Lhs = z ; Rhs = variables in w ; Hold $

LOGIT ; Lhs = y ; Rhs = variables in x ; Selection $

The motivation for a probit selection mechanism into a logit model does seem ambiguous.

N8.3 Endogenous Variable in a Probit Model

This estimator is for what is essentially a simultaneous equations model. The model
equations are
y1=* β′x + αy2 + ε, y=
1 1[ y1 * > 0],
y=2 γ ′z + u ,
 0   1 ρσ  
( ε, u ) ~ N    ,  2 
.
 0   ρσ σ  

Probit estimation based on y1 and (x1,y2) will not consistently estimate (b,α) because of the
correlation between y2 and e induced by the correlation between u and e. Several methods have been
proposed for estimation. One possibility is to use the partial reduced form obtained by inserting the
second equation in the first. This will produce consistent estimates of b/(1+α2σ2+2ασρ)1/2 and
αγ/(1+α2σ2+2ασρ)1/2. Linear regression of y2 on z produces estimates of γ and σ2, but there is no
method of moments estimator of ρ produced by this procedure, so this estimator is incomplete.
Newey (1987) suggested a ‘minimum chi squared’ estimator that does estimate all parameters. A
more direct, and actually simpler approach is full information maximum likelihood. Details on the
estimation procedure appear in Section E29.4.
To estimate this model, use the command

PROBIT ; Lhs = y1, y2

; Rh1 = independent variables in probit equation
; Rh2 = independent variables in regression equation $

(Note, the probit must be the first equation.) Other optional features relating to fitted values,
marginal effects, etc. are the same as for the univariate probit command. We note, marginal effects
are computed using the univariate probit probabilities,

Prob[y1 = 1] ~ Φ[b′x + αy2]

These will approximate the marginal effects obtained from the conditional model (which contain u).
When averaged over the sample values, the effect of u will become asymptotically negligible.
Predictions, etc. are kept with ; Keep = name, and so on. Likewise, options for the optimization,
such as maximum iterations, etc. are also the same as for the univariate probit model.
N8: Extended Binary Choice Models N-111

Retained Results
The results saved by this binary choice estimator are:

Matrices: b = estimate of (b,α,γ). Using ; Par adds σ and ρ to b.

varb = asymptotic covariance matrix.

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Last Model: b_variable (includes α) and, c_variables.

Last Function: Φ(b′x + ay2) = Prob(y1 = 1 | x,y2).

The Last Model names are used with WALD to simplify hypothesis tests. The last function is the
conditional mean function. The extra complication of the estimator has been used to obtain a
consistent estimator of b,α. With that in hand, the interesting function is E[y1| x,y2].

NAMELIST ; xdoctor = one,age,hsat,public,hhninc $

NAMELIST ; xincome = one,age,age*age,educ,female,hhkids $
PROBIT ; Lhs = doctor,hhninc
; Rh1 = xdoctor ; Rh2 = xincome $
-----------------------------------------------------------------------------
Probit Regression Start Values for DOCTOR
Dependent variable DOCTOR
Log likelihood function -16634.88715
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33279.774 AIC/N = 1.218
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Constant| 1.05627*** .05508 19.18 .0000 .94831 1.16423
AGE| .00895*** .00073 12.24 .0000 .00752 .01038
HSAT| -.17520*** .00395 -44.31 .0000 -.18295 -.16745
PUBLIC| .12985*** .02515 5.16 .0000 .08056 .17914
HHNINC| -.01332 .04581 -.29 .7712 -.10310 .07645
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Ordinary least squares regression ............
LHS=HHNINC Mean = .35208
Standard deviation = .17691
No. of observations = 27326 Degrees of freedom
Regression Sum of Squares = 88.9621 5
Residual Sum of Squares = 766.216 27320
Total Sum of Squares = 855.178 27325
Standard error of e = .16747
Fit R-squared = .10403 R-bar squared = .10386
Model test F[ 5, 27320] = 634.40260 Prob F > F* = .00000
Diagnostic Log likelihood = 10059.42844 Akaike I.C. = -3.57369
Restricted (b=0) = 8558.60603 Bayes I.C. = -3.57189
Chi squared [ 5] = 3001.64483 Prob C2 > C2* = .00000
N8: Extended Binary Choice Models N-112

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HHNINC| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Constant| -.40365*** .01704 -23.68 .0000 -.43705 -.37024
AGE| .02555*** .00079 32.43 .0000 .02400 .02709
AGE*AGE| -.00029*** .9008D-05 -31.68 .0000 -.00030 -.00027
EDUC| .01989*** .00045 44.22 .0000 .01901 .02077
FEMALE| .00122 .00207 .59 .5538 -.00283 .00527
HHKIDS| -.01146*** .00231 -4.96 .0000 -.01599 -.00693
--------+--------------------------------------------------------------------
Note: nnnnn.D-xx or D+xx => multiply by 10 to -xx or +xx.
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Initial iterations cannot improve function.Status=3
Error 805: Initial iterations cannot improve function.Status=3
Function= .61428384629D+04, at entry, .61358027527D+04 at exit
-----------------------------------------------------------------------------
Probit with Endogenous RHS Variable
Dependent variable DOCTOR
Log likelihood function -6135.80156
Restricted log likelihood -16599.60800
Chi squared [ 11 d.f.] 20927.61288
Significance level .00000
McFadden Pseudo R-squared .6303647
Estimation based on N = 27326, K = 13
Inf.Cr.AIC =12297.603 AIC/N = .450
--------+--------------------------------------------------------------------
DOCTOR| Standard Prob. 95% Confidence
HHNINC| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Coefficients in Probit Equation for DOCTOR
Constant| 1.05627*** .07626 13.85 .0000 .90681 1.20574
AGE| .00895*** .00074 12.03 .0000 .00749 .01041
HSAT| -.17520*** .00392 -44.72 .0000 -.18288 -.16752
PUBLIC| .12985*** .02626 4.94 .0000 .07838 .18131
HHNINC| -.01332 .14728 -.09 .9279 -.30200 .27535
|Coefficients in Linear Regression for HHNINC
Constant| -.40301*** .01712 -23.55 .0000 -.43656 -.36946
AGE| .02551*** .00081 31.37 .0000 .02391 .02710
AGE*AGE| -.00028*** .9377D-05 -30.39 .0000 -.00030 -.00027
EDUC| .01986*** .00040 50.26 .0000 .01908 .02063
FEMALE| .00122 .00207 .59 .5552 -.00284 .00528
HHKIDS| -.01144*** .00226 -5.06 .0000 -.01587 -.00701
|Standard Deviation of Regression Disturbances
Sigma(w)| .16720*** .00026 639.64 .0000 .16669 .16772
|Correlation Between Probit and Regression Disturbances
Rho(e,w)| .02412 .02550 .95 .3442 -.02586 .07409
--------+--------------------------------------------------------------------
Note: nnnnn.D-xx or D+xx => multiply by 10 to -xx or +xx.
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-113

N9: Fixed and Random Effects Models for

Binary Choice
N9.1 Introduction
The parametric models discussed in Chapters N5-N6 are extended to panel data formats.
Four specific parametric model formulations are provided as internal procedures in NLOGIT for
these binary choice models. These are the same ones described earlier, less the Burr distribution
which is not included in this set. Four classes of models are supported:

• Fixed effects: Prob[yit = 1] = F(b′xit + αi),

αi may be correlated with xit.

• Random effects: Prob[yit = 1] = Prob[b′xit + eit + ui > 0],

ui is uncorrelated with xit.

• Random parameters: Prob[yit = 1] = F(b i′xit),

bi | i ~ h(b|i) with mean vector b and covariance matrix Σ.

• Latent class: Prob[yit = 1|class j] = F(b j′xit),

Prob[class = j] = Fj(q).

The last two models provide various extensions of the basic form shown above.

NOTE: None of these panel data models require balanced panels. The group sizes may always
vary.

NOTE: None of these panel data models are provided for the Burr (scobit) model.

All formulations are treated the same for the five models, probit, logit, extreme value, Gompertz and
arctangent.

NOTE: The random effects estimator requires individual data. The fixed effects estimator allows
grouped data.

The third and fourth arise naturally in a panel data setting, but in fact, can be used in cross section
frameworks as well. The fixed and random effects estimators require panel data. The fixed and
random effects models are described in this chapter. Random parameters and latent class models are
documented in Chapter N10.
N9: Fixed and Random Effects Models for Binary Choice N-114

The applications in this chapter are based on the German health care data used throughout
the documentation. The data are an unbalanced panel of observations on health care utilization by
7,293 individuals. The group sizes in the panel number as follows: Ti: 1=1525, 2=2158, 3=825,
4=926, 5=1051, 6=1000, 7=987. There are altogether 27,326 observations. The variables in the file
that are used here are

doctor = 1 if number of doctor visits > 0, 0 otherwise,

hhninc = household nominal monthly net income in German marks / 10000,
hhkids = 1 if children under age 16 in the household, 0 otherwise,
educ = years of schooling,
married = marital status,
female = 1 for female, 0 for male,
docvis = number of visits to the doctor,
hospvis = number of visits to the hospital,
newhsat = self assessed health satisfaction, coded 0,1,...,10.

The data on health satisfaction in the raw data file, in variable hsat, contained some obvious coding
errors. Our corrected data are in newhsat.

N9.2 Commands
The essential model command for the models described in this chapter are

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = independent variables - not including one
; Panel
; ... specification of the panel data model $

As always, panels may be balanced or unbalanced. The panel is indicated with

SETPANEL ; Group = group identifier

; Pds = count variable to be created $

The ; Pds = variable is optional in the SETPANEL command. The default name for the created
variable is ti. You may change this with ; Pds. Thereafter,

; Panel

in the model command is sufficient to specify the panel setting. In circumstances where you have set
up the count variable yourself, you may also use the explicit declaration in the command:

; Pds = the fixed number of periods if the panel is balanced

; Pds = a variable which, within a group, repeats the number
of observations in the group

One or the other of these two specifications is required for the fixed and random effects estimators.
N9: Fixed and Random Effects Models for Binary Choice N-115

NOTE: For these estimators, you should not attempt to manage missing data. Just leave
observations with missing values in the sample. NLOGIT will automatically bypass the missing
values. Do not use SKIP, as it will undermine the setting of ; Pds = specification.

The estimator produces and saves the coefficient estimator, b and covariance matrix, varb, as usual.
Unless requested, the estimated fixed effects coefficients are not retained. (They are not reported
regardless.) To save the vector of fixed effects estimates, α in a matrix named alphafe, add

; Parameters

to the command. The fixed effects estimators allow up to 100,000 groups. However, only up to
50,000 estimated constant terms may be saved in alphafe.

N9.3 Clustering, Stratification and Robust Covariance

Matrices
The robust estimator based on sample clustering and stratification is available for the
parametric binary choice models. Full details appear in Chapter R10 for the general case and
Section E27.5.2 for the parametric binary choice models of interest here. The option for clustering is
offered in the command builders for most of the nonlinear model and binary choice routines in the
Model Estimates submenu. This will differ a bit from model to model. The one for the probit
model is shown below in Figure N9.1. The Model Estimates dialog box is selected at the bottom of
the Output page, then the clustering is specified in the next dialog box.

Figure N9.1 Command Builder for a Probit Model

N9: Fixed and Random Effects Models for Binary Choice N-116

This sampling setup may be used with any of the binary choice estimators. Do note, however, you
should not use it with panel data models. The so called ‘clustering’ corrections are already built into
the panel data estimators. (This is unlike the linear regression case, in which some authors argue that
the correction should be used even when fixed or random effects models are estimated.)
To illustrate, the following shows the setup for the panel data set described in the preceding
section. We have also artificially reduced the sample to 1,015 observations, 29 groups of 35
individuals, all of whom were observed seven times. The information below would appear with a
model command that used this configuration of the data to construct a robust covariance matrix.
The commands are:

SAMPLE ; 1-5000 $
REJECT ; _groupti < 7 $
NAMELIST ; x = age,educ,hhninc,hhkids,married $
PROBIT ; Lhs = doctor ; Rhs = one,x
; Cluster = 7
; Stratum = 35
; Describe $

These results appear before any results of the probit command. They are produced by the ; Describe
specification in the command.
========================================================================
Summary of Sample Configuration for Two Level Stratified Data
========================================================================
Stratum # Stratum Number Groups Group Sizes
Size (obs) Sample FPC. 1 2 3 ... Mean
========== ========== ============= =================================
1 35 5 1.0000 7 7 7 ... 7.0
2 35 5 1.0000 7 7 7 ... 7.0
(Rows 3 – 28 omitted)
29 35 5 1.0000 7 7 7 ... 7.0

+---------------------------------------------------------------------+
| Covariance matrix for the model is adjusted for data clustering. |
| Sample of 1015 observations contained 145 clusters defined by |
| 7 observations (fixed number) in each cluster. |
| Sample of 1015 observations contained 29 strata defined by |
| 35 observations (fixed number) in each stratum. |
+---------------------------------------------------------------------+

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -621.15030
Restricted log likelihood -634.14416
Chi squared [ 5 d.f.] 25.98772
Significance level .00009
McFadden Pseudo R-squared .0204904
Estimation based on N = 1015, K = 6
Inf.Cr.AIC = 1254.301 AIC/N = 1.236
Hosmer-Lemeshow chi-squared = 18.58245
P-value= .01726 with deg.fr. = 8
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-117

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| .71039 2.41718 .29 .7688 -4.02720 5.44797
AGE| .00659 .03221 .20 .8378 -.05655 .06973
EDUC| -.05898 .14043 -.42 .6745 -.33421 .21625
HHNINC| -.13753 1.25599 -.11 .9128 -2.59921 2.32416
HHKIDS| -.11452 .56015 -.20 .8380 -1.21240 .98336
MARRIED| .29025 .82535 .35 .7251 -1.32741 1.90791
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N9.4 One and Two Way Fixed Effects Models

The fixed effects models are estimated by unconditional maximum likelihood. The
command for requesting the model is

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = independent variables - not including one
; Panel
; Fixed Effects or ; FEM $

NOTE: Your Rhs list should not include a constant term, as the fixed effects model fits a complete
set of constants for the set of groups. If you do include one in your Rhs list, it is automatically
removed prior to beginning estimation.

Further documentation and technical details on fixed effects models for binary choice appear in
Chapter E30.
The fixed effects model assumes a group specific effect:

Prob[yit = 1] = F(b′xit + αi)

where αi is the parameter to be estimated. You may also fit a two way fixed effects model

Prob[yit = 1] = F(b′xit + αi + γt)

where γt is an additional, time (period) specific effect. The time specific effect is requested by
adding
; Time

to the command if the panel is balanced, and

; Time = variable name

if the panel is unbalanced.

N9: Fixed and Random Effects Models for Binary Choice N-118

For the unbalanced panel, we assume that overall, the sample observation period is

t = 1,2,..., T

and that the ‘Time’ variable gives for the specific group, the particular values of t that apply to the
observations. Thus, suppose your overall sample is five periods. The first group is three
observations, periods 1, 2, 4, while the second group is four observations, 2, 3, 4, 5. Then, your
panel specification would be

; Pds = Ti, for example, where Ti = (3, 3, 3), (4, 4, 4, 4)

and ; Time = Pd, for example, where Pd = (1, 2, 4), (2, 3, 4, 5).

Results that are kept for this model are

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of b.
alphafe = estimated fixed effects if the command contains ; Parameters

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Last Model: b_variables

Last Function: None

The upper limit on the number of groups is 100,000. Partial effects are computed locally with
; Partial Effects in the command. The post estimation PARTIAL EFFECTS command does not
have the set of constant terms, some of which are infinite, so the probabilities cannot be computed.

Application

The gender and kids present dummy variables are time invariant and are omitted from the
model. Nonlinear models are like linear models in that time invariant variables will prevent
estimation. This is not due to the ‘within’ transformation producing columns of zeros. The within
transformation of the data is not used for nonlinear models. A similar effect does arise in the
derivatives of the log likelihood, however, which will halt estimation because of a singular Hessian.
The results of fitting models with no fixed effects, with the person specific effects and with
both person and time effects are listed below. The results are partially reordered to enable
comparison of the results, and some of the results from the pooled estimator are omitted.
N9: Fixed and Random Effects Models for Binary Choice N-119

The commands are:

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
NAMELIST ; x = age,educ,hhninc,newhsat $
PROBIT ; Lhs = doctor ; Rhs = x,one
; Partial Effects $
PROBIT ; Lhs = doctor ; Rhs = x
; FEM
; Panel
; Parameters
; Partial Effects $
PROBIT ; Lhs = doctor ; Rhs = x
; FEM
; Panel
; Time Effects
; Parameters
; Partial Effects $

These are the results for the pooled data without fixed effects.

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -16639.23971
Restricted log likelihood -18019.55173
Chi squared [ 4 d.f.] 2760.62404
Significance level .00000
McFadden Pseudo R-squared .0766008
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33288.479 AIC/N = 1.218
Hosmer-Lemeshow chi-squared = 20.51061
P-value= .00857 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .00856*** .00074 11.57 .0000 .00711 .01001
EDUC| -.01540*** .00358 -4.30 .0000 -.02241 -.00838
HHNINC| -.00668 .04657 -.14 .8859 -.09795 .08458
NEWHSAT| -.17499*** .00396 -44.21 .0000 -.18275 -.16723
Constant| 1.35879*** .06243 21.77 .0000 1.23644 1.48114
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-120

These are the estimates for the one way fixed effects model.

-----------------------------------------------------------------------------
FIXED EFFECTS Probit Model
Dependent variable DOCTOR
Log likelihood function -9187.45120
Estimation based on N = 27326, K =4251
Inf.Cr.AIC =26876.902 AIC/N = .984
Unbalanced panel has 7293 individuals
Skipped 3046 groups with inestimable ai
PROBIT (normal) probability model
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .04701*** .00438 10.74 .0000 .03844 .05559
EDUC| -.07187* .04111 -1.75 .0804 -.15244 .00870
HHNINC| .04883 .10782 .45 .6506 -.16249 .26015
NEWHSAT| -.18143*** .00805 -22.53 .0000 -.19721 -.16564
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Figure N9.2 Estimated Fixed Effects

N9: Fixed and Random Effects Models for Binary Choice N-121

Note that the results report that 3046 groups had inestimable fixed effects. These are
individuals for which the Lhs variable, doctor, was the same in every period, including 1525 groups
with Ti = 1. If there is no within group variation in the dependent variable for a group, then the fixed
effect for that group cannot be estimated, and the group must be dropped from the sample. The
; Parameters specification requests that the estimates of αi be kept in a matrix, alphafe. Groups for
which αi is not estimated are filled with the value -1.E20 if yit is always zero and +1.E20 if yit is
always one, as shown above.
The log likelihood function has increased from -16,639.24 to -9187.45 in computing the
fixed effects model. The chi squared statistic is twice the difference, or 14,903.57. This would far
exceed the critical value for 95% significance, so at least at first take, it would seem that the
hypothesis of no fixed effects should be rejected. There are two reasons why this test would be
invalid. First, because of the incidental parameters issue, the fixed effects estimator is inconsistent.
As such, the statistic just computed does not have precisely a chi squared distribution, even in large
samples. Second, the fixed effects estimator is based on a reduced sample. If the test were valid
otherwise, it would have to be based on the same data set. This can be accomplished by using the
commands

CREATE ; meandr = Group Mean(doctor, Str = id) $

REJECT ; meandr < .1 | meandr > .9 $
PROBIT ; Lhs = doctor ; Rhs = one,x $

(The mean value must be greater than zero and less than one. For groups of seven, it can be as high as
6/7 = .86.) Using the reduced sample, the log likelihood for the pooled sample would be -10,852.71.
The chi squared is 11,573.31 which is still extremely large. But, again, the statistic does not have the
large sample chi squared distribution that allows a formal test. It is a rough guide to the results, but not
precise as a formal rule for building the model.
In order to compute marginal effects, it is necessary to compute the index function, which
does require an αi. The mean of the estimated values is used for the computation. The results for the
pooled data are shown for comparison below the fixed effects results.

These are the partial effects for the fixed effects model.

-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Estimated E[y|means,mean alphai]= .625
Estimated scale factor for dE/dx= .379
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .01783*** 1.22903 6.39 .0000 .01237 .02330
EDUC| -.02726 -.49559 -1.40 .1628 -.06554 .01102
HHNINC| .01852 .01048 .45 .6542 -.06253 .09957
NEWHSAT| -.06882*** -.77347 -5.96 .0000 -.09144 -.04619
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-122

These are the partial effects for the pooled model.

-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00297*** .20554 11.66 .0000 .00247 .00347
EDUC| -.00534*** -.09618 -4.30 .0000 -.00778 -.00291
HHNINC| -.00232 -.00130 -.14 .8859 -.03401 .02937
NEWHSAT| -.06075*** -.65528 -49.40 .0000 -.06316 -.05834
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

These are the two way fixed effects estimates. The time effects, which are usually few in number,
are shown in the model results, unlike the group effects.

-----------------------------------------------------------------------------
FIXED EFFECTS Probit Model
Dependent variable DOCTOR
Log likelihood function -9175.69958
Estimation based on N = 27326, K =4257
Inf.Cr.AIC =26865.399 AIC/N = .983
Unbalanced panel has 7293 individuals
Skipped 3046 groups with inestimable ai
No. of period specific effects= 6
PROBIT (normal) probability model
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .03869*** .01310 2.95 .0031 .01301 .06437
EDUC| -.07985* .04130 -1.93 .0532 -.16080 .00109
HHNINC| .05329 .10807 .49 .6219 -.15852 .26510
NEWHSAT| -.18090*** .00806 -22.44 .0000 -.19670 -.16510
Period1| -.08649 .15610 -.55 .5795 -.39244 .21946
Period2| -.00782 .13926 -.06 .9552 -.28076 .26513
Period3| .08766 .12423 .71 .4804 -.15583 .33116
Period4| .03048 .10907 .28 .7799 -.18330 .24425
Period5| -.02437 .09372 -.26 .7948 -.20807 .15932
Period6| .05075 .07761 .65 .5131 -.10136 .20287
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-123

-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Estimated E[y|means,mean alphai]= .625
Estimated scale factor for dE/dx= .379
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .01467*** 1.01123 4.35 .0000 .00806 .02129
EDUC| -.03029 -.55056 -1.49 .1370 -.07021 .00964
HHNINC| .02021 .01144 .48 .6289 -.06176 .10218
NEWHSAT| -.06861*** -.77109 -4.34 .0000 -.09962 -.03761
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N9.5 Conditional MLE of the Fixed Effects Logit Model

Two nonlinear models, the binomial logit and Poisson regression can be estimated by
conditional maximum likelihood. This is a specialized approach that was devised to deal with the
problem of large numbers of incidental parameters discussed in the preceding section. (This model
was studied, among others, by Chamberlain (1980).) The log likelihood for the binomial logit model
with fixed effects is

∑ i 1=
∑ t 1 logΛ ( 2 yit − 1) (β′xit + αi ) 
N Ti
logL = =

The first term, 2yit - 1, makes the sign negative for yit = 0 and positive for yit = 1, and Λ(.) is the
logistic probability, Λ(z) = 1/[1 + exp(-z)]. Direct maximization of this log likelihood involves
estimation of N+K parameters, where N is the number of groups. As N may be extremely large, this
is a potentially difficult estimation problem. As we saw in the preceding section, direct estimation
with up to 100,000 coefficients is feasible. But, the method discussed here is not restricted – the
number of groups is unlimited because the fixed effects coefficients are not estimated. Rather, the
fixed effects are conditioned out of the log likelihood. The main appeal of this approach, however, is
that whereas the brute force estimator of the preceding section is subject to the incidental parameters
bias, the conditional estimator is not; it is consistent even for small T (even for T = 2).
The contribution to the likelihood function of the Ti observations for group i can be
conditioned on the sum of the observed outcomes to produce the conditional log likelihood,
Ti

∏t =1
exp[ yit β′xit ]
Lc = Ti

∑ all arrangements of Ti outcomes with the same sum ∏

s =1
exp[ yis β′xis ]

exp  ∑ t =i 1 yit β′xit 

T

=   .
∑ all arrangements of Ti outcomes with the same sum exp ∑ si=1 disβ′xis 
T
N9: Fixed and Random Effects Models for Binary Choice N-124

This function can be maximized with respect to the slope parameters, b, with no need to estimate the
fixed effects parameters. The number of terms in the denominator of the probability may be
T 
exceedingly large, as it is the sum of T* terms where T* is equal to the binomial coefficient  i  and
 Si 
Si is the sum of the binary outcomes for the ith group. This can be extremely large. The computation
of the denominator is accomplished by means of a recursion presented in Krailo and Pike (1984). Let
the denominator be denoted A(Ti,Si). The authors show that for any T and S the function obeys the
recursion
A(T,S) = A(T-1,S) + exp(xiT′b)A(T-1,S-1)

with initial conditions

A(T,s) = 0 if T < s and A(T,0) = 1.

This enables rapid computation of the denominator for Ti up to 200 which is the internal limit. (If
your model is this large, expect this computation to be quite time consuming. Although 200 periods
(or more) is technically feasible, the number of terms rises geometrically in Ti, and more than 20 or
30 or so is likely to test the limits of the program (as well as your patience). Note, as well that when
the sum the observations is zero or Ti, the conditional probability is one, since there is only a single
way that each of these can occur. Thus, groups with sums of zero or Ti fall out of the computation.
Estimation of this model is done with Newton’s method. When the data set is rich enough
both in terms of variation in xit and in Si, convergence will be quick and simple.

N9.5.1 Command
The command for estimation of the model by this method is

LOGIT ; Lhs = dependent variable

; Rhs = dependent variables (do not include one)
; Pds = fixed number of periods or variable for group sizes $

NOTE: You must omit the ; FEM from the logit command. This is the default panel data estimator
for the binary logit model. Use ; Fixed Effects or ; FEM to request the unconditional estimator
discussed in the previous section.

You may use weights with this estimator. Presumably, these would reflect replications of
the observations. Be sure that the weighting variable takes the same value for all observations within
a group. The specification would be

; Wts = variable, Noscale

The Noscaling option should be used here if the weights are replication factors. If not, then do be
aware that the scaling will make the weights sum to the sample size, not the number of groups.
N9: Fixed and Random Effects Models for Binary Choice N-125

Results that are retained with this estimator are the usual ones from estimation:

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of b
Scalars: kreg = number of variables in Rhs
nreg = number of observations
logl = log likelihood function
Last Model: b_variables

Last Function: None

N9.5.2 Application
The following will fit the binary logit model using the two methods noted. Bear in mind that
with Ti < 7, the unconditional estimator is inconsistent and in fact likely to be substantially biased.
The conditional estimator is consistent. Based on the simulation results cited earlier, the second
results should exceed the first by roughly 40%. Partial effects are shown as well.

NAMELIST ; x = age,educ,hhninc,newhsat $
LOGIT ; Lhs = doctor ; Rhs = x,one $
LOGIT ; Lhs = doctor ; Rhs = x
; Panel $ (Chamberlain conditional estimator)
LOGIT ; Lhs = doctor ; Rhs = x
; Panel ; FEM $ (unconditional estimator)
These are the pooled estimates.
-----------------------------------------------------------------------------
Binary Logit Model for Binary Choice
Dependent variable DOCTOR
Log likelihood function -16639.86860
Restricted log likelihood -18019.55173
Chi squared [ 4 d.f.] 2759.36627
Significance level .00000
McFadden Pseudo R-squared .0765659
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33289.737 AIC/N = 1.218
Hosmer-Lemeshow chi-squared = 23.04975
P-value= .00330 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
AGE| .01366*** .00121 11.26 .0000 .01128 .01604
EDUC| -.02604*** .00585 -4.45 .0000 -.03750 -.01458
HHNINC| -.01231 .07670 -.16 .8725 -.16264 .13801
NEWHSAT| -.29181*** .00681 -42.86 .0000 -.30515 -.27846
Constant| 2.28922*** .10379 22.06 .0000 2.08580 2.49265
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-126

These are the conditional maximum likelihood estimates followed by the unconditional fixed effects
estimates. For these data, the unconditional estimates are closer to the conditional ones than might
have been expected, but still noticeably higher as the received results would predict. The suggested
proportionality result also seems to be operating, but with an unbalanced panel, this would not
necessarily occur, and should not be used as any kind of firm rule (save, perhaps for the case of Ti = 2).
+--------------------------------------------------+
| Panel Data Binomial Logit Model |
| Number of individuals = 7293 |
| Number of periods =TI |
| Conditioning event is the sum of DOCTOR |
+--------------------------------------------------+
-----------------------------------------------------------------------------
Logit Model for Panel Data
Dependent variable DOCTOR
Log likelihood function -6092.58175
Estimation based on N = 27326, K = 4
Inf.Cr.AIC =12193.164 AIC/N = .446
Hosmer-Lemeshow chi-squared = *********
P-value= .00000 with deg.fr. = 8
Fixed Effect Logit Model for Panel Data
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .06391*** .00659 9.70 .0000 .05100 .07683
EDUC| -.09127 .05752 -1.59 .1126 -.20401 .02147
HHNINC| .06121 .16058 .38 .7031 -.25352 .37594
NEWHSAT| -.23717*** .01208 -19.63 .0000 -.26086 -.21349
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
FIXED EFFECTS Logit Model
Dependent variable DOCTOR
Log likelihood function -9279.06752
Estimation based on N = 27326, K =4251
Inf.Cr.AIC =27060.135 AIC/N = .990
Unbalanced panel has 7293 individuals
Skipped 3046 groups with inestimable ai
LOGIT (Logistic) probability model
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .07925*** .00738 10.74 .0000 .06479 .09372
EDUC| -.11803* .06779 -1.74 .0817 -.25090 .01484
HHNINC| .07814 .18102 .43 .6660 -.27665 .43294
NEWHSAT| -.30367*** .01376 -22.07 .0000 -.33064 -.27670
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.

When the panel is balanced, the estimator also produces a frequency count for the
conditioning sums. For example, if we restrict our sample to the individuals who are in the sample
for all seven periods, the following table will also appear with the results.
N9: Fixed and Random Effects Models for Binary Choice N-127

+--------------------------------------------------+
| Panel Data Binomial Logit Model |
| Number of individuals = 887 |
| Number of periods = 7 |
| Conditioning event is the sum of DOCTOR |
| Distribution of sums over the 7 periods: |
| Sum 0 1 2 3 4 5 6 |
| Number 48 73 82 100 115 116 151 |
| Pct. 5.41 8.23 9.24 11.27 12.97 13.08 17.02 |
| Sum 7 8 9 10 11 12 13 |
| Number 202 0 0 0 0 0 0 |
| Pct. 22.77 .00 .00 .00 .00 .00 .00 |
+--------------------------------------------------+

This count would be meaningless in an unbalanced panel, so it is omitted.

How should you choose which estimator to use? We should note that the two approaches
will generally give different numerical answers. The conditional and unconditional log likelihoods
are different. In general, you should use the conditional estimator if T is not relatively large. The
conditional estimator is less efficient by construction, but consistency trumps efficiency at this level.
In addition, if you have more than 100,000 groups, you must use the conditional estimator. If, on the
other hand, T is larger than, say, 10, and N is less than 100,000, then the unconditional estimator
might be preferred. The additional consideration discussed in the next section might also weigh in
favor of the unconditional estimator.

N9.5.3 Estimating the Individual Constant Terms

The conditional fixed effects estimator for the logit model specifically eliminates the fixed
effects, so they are not directly estimated. Without them, however, the parameter estimates are of
relatively little use. Fitted probabilities and marginal effects will both require some estimate of a
constant term. You can request post estimation computation of the fixed effects by using the
specification
; Parameters

This saves a matrix named alphafe in your matrix work area. This will be a vector with number of
elements equal to the number of groups, containing an ad hoc estimate of αi for the groups for which
there is within group variation in yit. We note how this is done. The logit model is

Prob[yit = 1|xit] = Λ(b′xit + αi) where Λ(z) = exp(z)/[1+exp(z)]

After estimation of b, we treat the b′xit part of this as known, and let zit = b′xit. These are now just
data. As such, the log likelihood for group i would be

log Li = Σt log Λ[(2yit – 1)(zit + αi)]

The likelihood equation for αi would be

Σt (yit – Pit) = 0 where Pit = Λ(zit + αi)

The implicit solution for αi is given by

Σt yit = Σt wit / (ai + wit) where wit = exp(zit) and ai = exp(-αi).

N9: Fixed and Random Effects Models for Binary Choice N-128

If yit is always zero or always one in every period, t, then there is no solution to maximizing this
function. The corresponding element of alphafe will be set equal to -1.d20 or +1.d20 But, if the yits
differ, then the αi that equates the left and right hand sides can be found by a straightforward search.
The remaining rows of alphafe will contain the individual specific solutions to these equations.
(This is the method that Heckman and MaCurdy (1980) suggested for estimation of the fixed effects
probit model.)
We emphasize, this is not the maximum likelihood estimator of αi because the conditional
estimator of b is not the unconditional MLE. Nor, in fact, is it consistent in N. It is consistent in Ti,
but that is not helpful here since Ti is fixed, and presumably small. This estimator is a means to an
end. The estimated marginal effects can be based on this estimator – it will give a reasonable
estimator of an overall average of the constant terms, which is all that is needed for the marginal
effects. Individual predicted probabilities remain ambiguous.

N9.5.4 A Hausman Test for Fixed Effects in the Logit Model

The fixed effects estimator is illustrated with the data used in the preceding examples: Note
that the first estimator is the pooled estimator. Under the alternative hypothesis of fixed effects, it is
inconsistent. Under the null, it is consistent and efficient. The second estimator is the conditional MLE
and the third one is the unconditional fixed effects estimator. The unconditional fixed estimator cannot
be used for formal testing because of the incidental parameters problem – it is inconsistent. The pooled
estimator and the conditional fixed effects estimator use different samples, so the likelihoods are not
comparable. Therefore, testing for the joint significance of the effects is problematic for the
conditional estimator. What one can do is use a Hausman test. The test is constructed as follows:

H0: There are no fixed effects; unconditional ML estimators are b0 and V0

H1: There are fixed effects: conditional ML estimators are b1 and V1

Under H0, b0 is consistent and efficient, while b1 is consistent but inefficient. Under H1, b0 is
inconsistent while b1 is consistent and efficient. The Hausman statistic would therefore be

H = (b1 - b0)′ [V1 - V0]-1(b1 - b0)

The statistic can be constructed as follows:

NAMELIST ; x = the independent variables, not including one $

LOGIT ; Lhs = ... ; Rhs = x, one $
CALC ; k = Col(x) $
MATRIX ; b0 = b(1:k) ; v0 = varb(1:k,1:k) $
LOGIT ; Lhs = ... ; Rhs = x ; Pds = ... ; FEM $
MATRIX ; b1 = b ; v1 = varb $
MATRIX ; d = b1 - b0 ; List ; h = d’ * Nvsm(v1, -v0) * d $

We apply this to our innovation data by defining x = imprtshr,fdishare,logsales,relsize,prod and the

dependent variable is innov. The remaining commands are generic.
N9: Fixed and Random Effects Models for Binary Choice N-129

The three sets of parameter estimates were given earlier. The Hausman statistic using the
procedure suggested above is

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
NAMELIST ; x = age,educ,hhninc,newhsat $
LOGIT ; Lhs = doctor ; Rhs = x, one $
CALC ; k = Col(x) $
MATRIX ; b0 = b(1:k) ; v0 = Varb(1:k,1:k) $
LOGIT ; Lhs = doctor ; Rhs = x ; Panel $
MATRIX ; b1 = b ; v1 = varb $
MATRIX ; d = b1 - b0 ; List ; h = d' * Nvsm(v1, -v0) * d $

The final result of the MATRIX command is

H| 1
--------+--------------
1| 98.1550

This statistic has four degrees of freedom. The critical value from the chi squared table is 9.49, so
based on this test, we would reject the null hypothesis of no fixed effects.

N9.6 Random Effects Models for Binary Choice

The five models we have developed here can also be fit with random effects instead of fixed
effects. The structure of the random effects model is

zit | ui = b′xit + eit + ui

where ui is the unobserved heterogeneity for the ith individual,

ui ~ N[0,σu2],

and eit is the stochastic term in the model that provides the conditional distribution.

Prob[yit = 1| xit, ui] = F(b′xit + ui), i = 1,...,N, t = 1,...,Ti.

where F(.) is the distribution discussed earlier (normal, logistic, extreme value, Gompertz). Note
that the unobserved heterogeneity, ui is the same in every period. The parameters of the model are
fit by maximum likelihood. As usual in binary choice models, the underlying variance,

σ2 = σu2 + σe2

is not identified. The reduced form parameter,

σ u2
ρ = ,
σ ε2 + σ u2
N9: Fixed and Random Effects Models for Binary Choice N-130

is estimated directly. With the normalization that we used earlier, σe2 = 1, we can determine

ρ
σu = .
1− ρ

Further discussion of the estimation of these structural parameters appears at the end of this section.
The model command for this form of the model is

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = independent variables - not including one
; Panel
; Random Effects $

NOTE: For this model, your Rhs list should include a constant term, one.

Partial effects are computed by setting the heterogeneity term, ui to its expected value of zero.
Restrictions may be tested and imposed exactly as in the model with no heterogeneity. Since
restrictions can be imposed on all parameters, including ρ, you can fix the value of ρ at any desired
value. Do note that forcing the ancillary parameter, in this case, ρ, to equal a slope parameter will
almost surely produce unsatisfactory results, and may impede or even prevent convergence of the
iterations.
Starting values for the iterations are obtained by fitting the basic model without random
effects. Thus, the initial results in the output for these models will be the binary choice models
discussed in the preceding sections. You may provide your own starting values for the parameters
with
; Start = ... the list of values for b, value for ρ

There is no natural moment based estimator for ρ, so a relatively low guess is used as the starting
value instead. The starting value for ρ is approximately .2 (q = [2ρ/(1-ρ)]1/2 ≈.29 – see the technical
details below. Maximum likelihood estimates are then computed and reported, along with the usual
diagnostic statistics. (An example appears below.) This model is fit by approximating the necessary
integrals in the log likelihood function by Hermite quadrature. An alternative approach to estimating
the same model is by Monte Carlo simulation. You can do exactly this by fitting the model as a
random parameters model with only a random constant term.
Your data might not be consistent with the random effects model. That is, there might be no
discernible evidence of random effects in your data. In this case, the estimate of ρ will turn out to be
negligible. If so, the estimation program issues a diagnostic and reverts back to the original,
uncorrelated formulation and reports (again) the results for the basic model.
N9: Fixed and Random Effects Models for Binary Choice N-131

Results that are kept for this model are

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of b
Scalars: kreg = number of variables in Rhs
nreg = number of observations
logl = log likelihood function
rho = estimated value of ρ
varrho = estimated asymptotic variance of estimator of ρ
Last Model: b_variables, ru
Last Function: Prob(y = 1|x,u=0) (Note: None if you use ; RPM to fit the RE model.)

The additional specification ; Par in the command requests that ρ be included in b and the
additional row and column corresponding to ρ be included in varb. If you have included ; Par, rho
and varrho will also appear at the appropriate places in b and varb.

NOTE: The hypothesis of no group effects can be tested with a Wald test (simple t test) or with a
likelihood ratio test. The Lagrange multiplier (LM) statistic developed by Greene and McKenzie
(2015) is reported with the other results, as shown in the example below.

Application
The following study fits the probit model under four sets of assumptions. The first uses the
pooled estimator, then corrects the standard errors for the clustering in the data. The second is the
unconditional fixed effects estimator. The third and fourth compute the random effects estimator,
first by quadrature, using the Butler and Moffitt method and the second using maximum simulated
likelihood with Halton draws. The output is trimmed in each model to compare only the estimates
and the marginal effects.

NAMELIST ; x = age,educ,hhninc,newhsat $
SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
PROBIT ; Lhs = doctor ; Rhs = x,one ; Partial Effects
; Cluster = id $
PROBIT ; Lhs = doctor ; Rhs = x ; Partial Effects
; Panel ; FEM $
PROBIT ; Lhs = doctor ; Rhs = x,one ; Partial Effects
; Panel ; Random Effects $

The random parameters model described in Chapter E31 provides an alternative estimator for the
random effects model based on maximum simulated likelihood rather than with Hermite quadrature.
The general syntax is used below for a probit model to illustrate the method.

PROBIT ; Lhs = doctor ; Rhs = x,one ; Partial Effects

; Panel ; RPM ; Fcn = one(n) ; Pts = 25 ; Halton $
CALC ; List ; b(6)^2/(1+b(6)^2) $
N9: Fixed and Random Effects Models for Binary Choice N-132

These are the pooled estimates with corrected standard errors.

+---------------------------------------------------------------------+
| Covariance matrix for the model is adjusted for data clustering. |
| Sample of 27326 observations contained 7293 clusters defined by |
| variable ID which identifies by a value a cluster ID. |
+---------------------------------------------------------------------+
-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -16639.23971
Restricted log likelihood -18019.55173
Chi squared [ 4 d.f.] 2760.62404
Significance level .00000
McFadden Pseudo R-squared .0766008
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33288.479 AIC/N = 1.218
Hosmer-Lemeshow chi-squared = 20.51061
P-value= .00857 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .00856*** .00098 8.76 .0000 .00664 .01047
EDUC| -.01540*** .00499 -3.09 .0020 -.02517 -.00562
HHNINC| -.00668 .05646 -.12 .9058 -.11735 .10398
NEWHSAT| -.17499*** .00490 -35.72 .0000 -.18460 -.16539
Constant| 1.35879*** .08475 16.03 .0000 1.19268 1.52491
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The unconditional fixed effects estimates appear next. They differ greatly from the pooled estimates.
It is worth noting that under the random effects assumption, neither the pooled nor these fixed effects
estimates are consistent.
-----------------------------------------------------------------------------
FIXED EFFECTS Probit Model
Dependent variable DOCTOR
Log likelihood function -9187.45120
Estimation based on N = 27326, K =4251
Inf.Cr.AIC =26876.902 AIC/N = .984
Unbalanced panel has 7293 individuals
Skipped 3046 groups with inestimable ai
PROBIT (normal) probability model
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .04701*** .00438 10.74 .0000 .03844 .05559
EDUC| -.07187* .04111 -1.75 .0804 -.15244 .00870
HHNINC| .04883 .10782 .45 .6506 -.16249 .26015
NEWHSAT| -.18143*** .00805 -22.53 .0000 -.19721 -.16564
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-133

These are the random effects estimates. The variance of u and correlation parameter ρ are given
explicitly in the results. In the MSL random effects estimates that appear next, only the standard
deviation of u is given. Squaring the 1.37554428 gives 1.892122, which is nearly the same as the
1.888060 given in the first results. In order to compare the first estimates to the MSL estimates, it is
necessary to divide the first by the estimate of 1+ρ. Thus, the scaled coefficient on age in the first
set of estimates would be 0.019322; that on educ would be -.027611, and so on. Thus, the two sets of
estimates are quite similar.

-----------------------------------------------------------------------------
Random Effects Binary Probit Model
Dependent variable DOCTOR
Log likelihood function -15614.49128
Restricted log likelihood -16639.23907
Chi squared [ 1](P= .000) 2049.49558
Significance level .00000
(Cannot compute pseudo R2. Use RHS=one
to obtain the required restricted logL)
Estimation based on N = 27326, K = 6
Inf.Cr.AIC = 31241.0 AIC/N = 1.143
Unbalanced panel has 7293 individuals
- ChiSqd[1] tests for random effects -
LM ChiSqd 807.866 P value .00000
LR ChiSqd 2049.496 P value .00000
Wald ChiSqd 1432.269 P value .00000
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .01305*** .00119 10.97 .0000 .01072 .01538
EDUC| -.01840*** .00594 -3.10 .0020 -.03005 -.00675
HHNINC| .06299 .06387 .99 .3240 -.06218 .18817
NEWHSAT| -.19418*** .00520 -37.32 .0000 -.20437 -.18398
Constant| 1.42666*** .09644 14.79 .0000 1.23765 1.61567
Rho| .39553*** .01045 37.84 .0000 .37504 .41601
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Random Coefficients Probit Model
Dependent variable DOCTOR
Log likelihood function -15619.14356
Restricted log likelihood -16639.23971
Chi squared [ 1 d.f.] 2040.19230
Significance level .00000
McFadden Pseudo R-squared .0613067
Estimation based on N = 27326, K = 6
Inf.Cr.AIC =31250.287 AIC/N = 1.144
Unbalanced panel has 7293 individuals
PROBIT (normal) probability model
Simulation based on 25 Halton draws
-----------------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-134

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .01288*** .00083 15.58 .0000 .01126 .01450
EDUC| -.01823*** .00395 -4.61 .0000 -.02598 -.01048
HHNINC| .06741 .05108 1.32 .1870 -.03271 .16752
NEWHSAT| -.19383*** .00435 -44.58 .0000 -.20235 -.18531
|Means for random parameters
Constant| 1.42554*** .06828 20.88 .0000 1.29172 1.55936
|Scale parameters for dists. of random parameters
Constant| .80930*** .01088 74.38 .0000 .78797 .83062
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The random parameters approach provides an alternative way to estimate a random effects
model. A comparison of the two sets of results illustrates the general result that both are consistent
estimators of the same parameters. We note, however, the Hermite quadrature approach produces an
estimator of ρ = σu2/(1 + σu2) while the RP approach produces an estimator of σu. To check the
consistency of the two approaches, we compute an estimate of ρ based on the RP results. The result
below demonstrates the near equivalence of the two approaches.

CALC ; List ; b(6)^2/(1+b(6)^2)$

[CALC] *Result*= .3957574

These are the four sets of estimated partial effects.

Pooled
-----------------------------------------------------------------------------
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00297*** .20554 8.83 .0000 .00231 .00363
EDUC| -.00534*** -.09618 -3.09 .0020 -.00874 -.00195
HHNINC| -.00232 -.00130 -.12 .9058 -.04074 .03610
NEWHSAT| -.06075*** -.65528 -39.87 .0000 -.06374 -.05777
--------+--------------------------------------------------------------------
Unconditional Fixed Effects
-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*]
Estimated E[y|means,mean alphai]= .625
Estimated scale factor for dE/dx= .379
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .01783*** 1.22903 6.39 .0000 .01237 .02330
EDUC| -.02726 -.49559 -1.40 .1628 -.06554 .01102
HHNINC| .01852 .01048 .45 .6542 -.06253 .09957
NEWHSAT| -.06882*** -.77347 -5.96 .0000 -.09144 -.04619
--------+--------------------------------------------------------------------
N9: Fixed and Random Effects Models for Binary Choice N-135

Random Effects
-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*]
Observations used for means are All Obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00376*** .25254 11.06 .0000 .00310 .00443
EDUC| -.00531*** -.09261 -3.10 .0020 -.00866 -.00195
HHNINC| .01817 .00986 .99 .3239 -.01793 .05426
NEWHSAT| -.05600*** -.58577 -37.33 .0000 -.05894 -.05306
--------+--------------------------------------------------------------------
Random Constant Term
-----------------------------------------------------------------------------
Partial derivatives of expected val. with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Scale Factor for Marginal Effects .3541
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00456*** .28882 11.14 .0000 .00376 .00536
EDUC| -.00646*** -.10635 -5.06 .0000 -.00896 -.00396
HHNINC| .02387 .01223 1.32 .1882 -.01168 .05942
NEWHSAT| -.06864*** -.67771 -33.24 .0000 -.07269 -.06459
--------+--------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-136

N10: Random Parameter Models for Binary

Choice
N10.1 Introduction
The probit and logit models are extended to panel data formats as internal procedures. Four
classes of models are supported:

• Fixed effects: Prob[yit = 1] = F(b′xit + αi),

αi correlated with xit.

• Random effects: Prob[yit = 1] = Prob[b′xit + eit + ui > 0],

ui uncorrelated with xit.

• Random parameters: Prob[yit = 1] = F(b i′xit),

bi | i ~ h(b|i) with mean vector b and covariance matrix Σ.

• Latent class: Prob[yit = 1|class j] = F(b j′xit),

Prob[class = j] = Fj(q).

The first two were developed in Chapter E30. This chapter documents the use of random parameters
(mixed) and latent class models for binary choice. Technical details on estimation of random
parameters are given in Chapter R24. Technical details for estimation of latent class models are
given in Chapter R25.

NOTE: None of these panel data models require balanced panels. The group sizes may always vary.

The random parameters and latent class models do not require panel data. You may fit them with a
cross section. If you omit ; Pds and ; Panel in these cases, the cross section case, Ti = 1, is assumed.
(You can also specify ; Pds = 1.) Note that this group of models (and all of the panel data models
described in the rest of this manual) does not use the ; Str = variable specification for indicating the
panel – that is only for REGRESS.
The probabilities and density functions supported here are as follows:

Probit

β 'x i exπ(−t 2 / 2)
F= ∫−∞ 2π
dt = Φ(b′xi), f = φ(b′xi)

Logit
exp(β′xi )
F= = Λ(b′xi), f = Λ(b′xi)[1 - Λ(b′xi)]
1 + exp(β′xi )
N10: Random Parameter Models for Binary Choice N-137

N10.2 Probit and Logit Models with Random Parameters

We have extended the random parameters model to the binary choice models as well as
many other models including the tobit and exponential regression models. Some of the relevant
background literature includes Revelt and Train (1998), Train (1998), Brownstone and Train (1999),
and Greene (2001). (In that literature, the models are described under the heading ‘mixed logit’
models. We will require a broader rubric for our purposes.) The structure of the random parameters
model is based on the conditional probability

Prob[yit = 1| xit, b i] = F(bi′xit), i = 1,...,N, t = 1,...,Ti.

where F(.) is the distribution discussed earlier (normal, logistic, extreme value, Gompertz). The model
assumes that parameters are randomly distributed with possibly heterogeneous (across individuals)

E[b i| zi] = b + Dzi,

(the second term is optional – the mean may be constant),

Var[b i| zi] = Σ.

The model is operationalized by writing

bi = b + Dzi + Γvi where vi ~ N[0,I].

As noted earlier, the heterogeneity term is optional. In addition, it may be assumed that some of the
parameters are nonrandom. It is convenient to analyze the model in this fully general form here.
One can easily accommodate nonrandom parameters just by placing rows of zeros in the appropriate
places in D and Γ. The command structure for these models makes this simple to do.

NOTE: If there is no heterogeneity in the mean, and only the constant term is considered random –
the model may specify that some parameters are nonrandom – then this model is equivalent to the
random effects model of the preceding section.

N10.2.1 Command for the Random Parameters Models

The basic model command for this form of the model is

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = independent variables
; Panel or Pds = fixed periods or count variable
; RPM
; Fcn = random parameters specification $

NOTE: For this model, your Rhs list should include a constant term.

NOTE: The ; Pds specification is optional. You may fit these models with cross section data.
N10: Random Parameter Models for Binary Choice N-138

Specifying Random Parameters

The ; Fcn = specification is used to define the random parameters. It is constructed from
the list of Rhs names as follows: Suppose your model is specified by

; Rhs = one, x1, x2, x3, x4

This involves five coefficients. Any or all of them may be random; any not specified as random are
assumed to be constant. For those that you wish to specify as random, use

; Fcn = variable name (distribution),

variable name (distribution), ...

Three distributions may be specified. All random variables have mean 0.

n = standard normal distribution, variance = 1,

t = triangular (tent shaped) distribution in [-1,+1], variance = 1/6,
u = standard uniform distribution [-1,1], variance = 1/3,
l = lognormal distribution, variance = exp(.5),
o = tent shaped distribution with one anchor at zero,
g = log gamma,
or c = variance = 0. (The parameter is not random.)

Each of these is scaled as it enters the distribution, so the variance is only that of the random draw
before multiplication. The normal distribution is used most often, but there are several other
possibilities. Numerous other formats for random parameters are described in Section R24.3. Those
results all apply to the binary choice models. To specify that the constant term and the coefficient on
x1 are each normally distributed with given mean and variance, use

; Fcn = one(n), x1(n).

This specifies that the first and second coefficients are random while the remainder are not. The
parameters estimated will be the mean and standard deviations of the distributions of these two
parameters and the fixed values of the other three.
The results include estimates of the means and standard deviations of the distributions of the
random parameters and the estimates of the nonrandom parameters. The log likelihood shown in the
results is conditioned on the random draws, so one might be cautious about using it to test
hypotheses, for example, that the parameters are random at all by comparing it to the log likelihood
from the basic model with all nonrandom coefficients. The test becomes valid as R increases, but the
50 used in our application is probably too few. With several hundred draws, one could reliably use
the simulated log likelihood for testing purposes.
N10: Random Parameter Models for Binary Choice N-139

Correlated Random Parameters

The preceding defines an estimator for a model in which the covariance matrix of the
random parameters is diagonal. To extend it to a model in which the parameters are freely
correlated, add
; Correlation (or just ; Cor)

to the command. An example appears below.

Heterogeneity in the Means

The preceding examples have specified that the mean of the random variable is fixed over
individuals. If there is measured heterogeneity in the means, in the form of

E[bki] = bk + Σm δkm zmi

where zm is a variable that is measured for each individual, then the command may be modified to

; RPM = list of variables in z

In the data set, these variables must be repeated for each observation in the group. In the application
below, we have specified that the random parameters have different means for individuals depending
on gender and marital status.

Autocorrelation

You may change the character of the heterogeneity from a time invariant effect to an AR(1)
process,
vkit = ρkvki,t-1 + wkit.

N10.2.2 Results from the Estimator and Applications

The results produced by this estimator begin with the familiar diagnostic statistics, likelihood
function, information criteria, etc. The coefficient estimates are possibly rearranged so that the
nonrandom parameters appear first. In the base case of a diagonal covariance matrix, the means of
the random parameters appear next, followed in the same order by the estimated scale parameters.
The example below illustrates. For normally distributed parameters, these are the standard
deviations. For other distributions, these scale factors are multiplied by the relevant standard
deviation to obtain the standard deviation of the parameter. For example, if we had specified

; Fcn = educ(u)

in the model command, then the parameter on educ would be defined to have mean 1.697 and
standard deviation .08084 times 1/sqr(6). (The uniform draw is transformed to be U[-1,+1].)
N10: Random Parameter Models for Binary Choice N-140

The commands are:

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
NAMELIST ; x = age,educ,hhninc,hsat $
LOGIT ; Lhs = doctor ; Rhs = x,one
; Partial Effects ; Panel ; RPM
; Fcn = one(n),hhninc(n),hsat(n)
; Pts = 25 ; Halton $
-----------------------------------------------------------------------------
Logit Regression Start Values for DOCTOR
Dependent variable DOCTOR
Log likelihood function -16639.59764
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33289.195 AIC/N = 1.218
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .01366*** .00121 11.25 .0000 .01128 .01603
EDUC| -.02603*** .00585 -4.45 .0000 -.03749 -.01457
Constant| 2.28946*** .10379 22.06 .0000 2.08604 2.49288
HHNINC| -.01221 .07670 -.16 .8735 -.16254 .13812
HSAT| -.29185*** .00681 -42.87 .0000 -.30519 -.27850
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Random Coefficients Logit Model
Dependent variable DOCTOR
Log likelihood function -15617.53717
Restricted log likelihood -16639.59764
Chi squared [ 3 d.f.] 2044.12094
Significance level .00000
McFadden Pseudo R-squared .0614234
Estimation based on N = 27326, K = 8
Inf.Cr.AIC =31251.074 AIC/N = 1.144
Unbalanced panel has 7293 individuals
LOGIT (Logistic) probability model
Simulation based on 25 Halton draws
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .01541*** .00100 15.39 .0000 .01344 .01737
EDUC| -.02538*** .00475 -5.34 .0000 -.03469 -.01607
|Means for random parameters
Constant| 1.77433*** .08285 21.42 .0000 1.61195 1.93671
HHNINC| .08517 .06181 1.38 .1682 -.03598 .20632
HSAT| -.23532*** .00541 -43.50 .0000 -.24592 -.22471
|Scale parameters for dists. of random parameters
Constant| 1.37499*** .01982 69.36 .0000 1.33614 1.41384
HHNINC| .18336*** .03792 4.84 .0000 .10904 .25768
HSAT| .00080 .00204 .39 .6960 -.00319 .00479
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-141

-----------------------------------------------------------------------------
Partial derivatives of expected val. with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Conditional Mean at Sample Point .6436
Scale Factor for Marginal Effects .2294
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00353*** .23902 15.53 .0000 .00309 .00398
EDUC| -.00582*** -.10241 -5.36 .0000 -.00795 -.00369
HHNINC| .01954 .01069 1.38 .1686 -.00827 .04735
HSAT| -.05398*** -.56914 -29.82 .0000 -.05753 -.05043
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

When the random parameters are specified to be correlated, the output is changed. The
parameter vector in this case is written

bi = b 0 + Γ vi

where Γ is a lower triangular Cholesky matrix. In this case, the nonrandom parameters and the
means of the random parameters are reported as before. The table then reports Γ in two parts. The
diagonal elements are reported first. These would correspond to the case above. The nonzero
elements of Γ below the diagonal are reported next, rowwise. In the example below, there are three
random parameters, so there are 1 + 2 elements below the main diagonal of Γ in the reported results.
The covariance matrix for the random parameters in this specification is

Var [ b i] = Ω = ΓAΓ′

where A is the known diagonal covariance matrix of vi. For normally distributed parameters, A = I.
This matrix is reported separately after the tabled coefficient estimates. Finally, the square roots of
the diagonal elements of the estimate of Ω are reported, followed by the correlation matrix derived
from Ω. The example below illustrates.

LOGIT ; Lhs = doctor ; Rhs = x,one

; Partial Effects
; Pds = _groupti
; RPM
; Fcn = one(n),hhninc(n),newhsat(n)
; Correlated
; Pts = 25
; Halton $
N10: Random Parameter Models for Binary Choice N-142

-----------------------------------------------------------------------------
Random Coefficients Logit Model
Dependent variable DOCTOR
Log likelihood function -15606.79747
Restricted log likelihood -16639.59764
Chi squared [ 6 d.f.] 2065.60035
Significance level .00000
McFadden Pseudo R-squared .0620688
Estimation based on N = 27326, K = 11
Inf.Cr.AIC =31235.595 AIC/N = 1.143
Unbalanced panel has 7293 individuals
LOGIT (Logistic) probability model
Simulation based on 25 Halton draws
-----------------------------------------------------------------------------

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .01471*** .00101 14.61 .0000 .01274 .01668
EDUC| -.02740*** .00475 -5.77 .0000 -.03670 -.01810
|Means for random parameters
Constant| 1.98083*** .08660 22.87 .0000 1.81111 2.15056
HHNINC| .09438 .06586 1.43 .1518 -.03470 .22346
HSAT| -.25657*** .00615 -41.74 .0000 -.26861 -.24452
|Diagonal elements of Cholesky matrix
Constant| 1.90753*** .07911 24.11 .0000 1.75248 2.06257
HHNINC| .91257*** .08028 11.37 .0000 .75522 1.06991
HSAT| .01770*** .00203 8.74 .0000 .01373 .02167
|Below diagonal elements of Cholesky matrix
lHHN_ONE| -.00234 .10500 -.02 .9822 -.20813 .20344
lHSA_ONE| -.08124*** .00932 -8.71 .0000 -.09951 -.06297
lHSA_HHN| .09466*** .00433 21.88 .0000 .08617 .10314
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Implied covariance matrix of random parameters

Var_Beta| 1 2 3
--------+------------------------------------------
1| 3.63867 -.00447279 -.154960
2| -.00447279 .832783 .0865698
3| -.154960 .0865698 .0158724

Implied standard deviations of random parameters

S.D_Beta| 1
--------+--------------
1| 1.90753
2| .912570
3| .125986

Implied correlation matrix of random parameters

Cor_Beta| 1 2 3
--------+------------------------------------------
1| 1.00000 -.00256946 -.644803
2| -.00256946 1.00000 .752973
3| -.644803 .752973 1.00000
N10: Random Parameter Models for Binary Choice N-143

-----------------------------------------------------------------------------
Partial derivatives of expected val. with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Conditional Mean at Sample Point .6464
Scale Factor for Marginal Effects .2286
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00336*** .22640 14.71 .0000 .00291 .00381
EDUC| -.00626*** -.10967 -5.78 .0000 -.00838 -.00414
HHNINC| .02157 .01175 1.43 .1522 -.00796 .05110
HSAT| -.05864*** -.61557 -27.65 .0000 -.06280 -.05448
--------+--------------------------------------------------------------------

Finally, if you specify that there is observable heterogeneity in the means of the parameters
with
; RPM = list of variables

then the model changes to

bi = b 0 + Dzi + Γ vi.

The elements of D, rowwise, are reported after the decomposition of Γ. The example below, which
contains gender and marital status, illustrates. Note that a compound name is created for the
elements of D.

LOGIT ; Lhs = doctor ; Rhs = x,one

; Partial Effects
; Panel
; RPM = female,married
; Fcn = one(n),hhninc(n),hsat(n)
; Correlated
; Pts = 25
; Halton $

-----------------------------------------------------------------------------
Random Coefficients Logit Model
Dependent variable DOCTOR
Log likelihood function -15470.04441
Restricted log likelihood -16639.59764
Chi squared [ 12 d.f.] 2339.10646
Significance level .00000
McFadden Pseudo R-squared .0702874
Estimation based on N = 27326, K = 17
Inf.Cr.AIC =30974.089 AIC/N = 1.134
Unbalanced panel has 7293 individuals
LOGIT (Logistic) probability model
Simulation based on 25 Halton draws
-----------------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-144

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .01375*** .00104 13.24 .0000 .01171 .01578
EDUC| -.00913* .00488 -1.87 .0613 -.01870 .00043
|Means for random parameters
Constant| 1.58591*** .12092 13.11 .0000 1.34890 1.82291
HHNINC| .10102 .12817 .79 .4306 -.15018 .35223
HSAT| -.25929*** .01173 -22.11 .0000 -.28228 -.23630
|Diagonal elements of Cholesky matrix
Constant| 1.85093*** .07867 23.53 .0000 1.69674 2.00512
HHNINC| 1.17355*** .08054 14.57 .0000 1.01570 1.33140
HSAT| .00147 .00202 .73 .4682 -.00250 .00543
|Below diagonal elements of Cholesky matrix
lHHN_ONE| .15728 .10367 1.52 .1293 -.04592 .36047
lHSA_ONE| -.06741*** .00926 -7.28 .0000 -.08555 -.04926
lHSA_HHN| .07996*** .00426 18.78 .0000 .07161 .08831
|Heterogeneity in the means of random parameters
cONE_FEM| .26949*** .09017 2.99 .0028 .09276 .44622
cONE_MAR| .11320 .10064 1.12 .2607 -.08404 .31044
cHHN_FEM| .10364 .12514 .83 .4075 -.14162 .34891
cHHN_MAR| -.08432 .13820 -.61 .5418 -.35520 .18655
cHSA_FEM| .03242*** .01081 3.00 .0027 .01124 .05360
cHSA_MAR| -.01361 .01218 -1.12 .2638 -.03748 .01026
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Implied covariance matrix of random parameters

Var_Beta| 1 2 3
--------+------------------------------------------
1| 3.42595 .291109 -.124767
2| .291109 1.40195 .0832340
3| -.124767 .0832340 .0109393

Implied standard deviations of random parameters

S.D_Beta| 1
--------+--------------
1| 1.85093
2| 1.18404
3| .104591

Implied correlation matrix of random parameters

Cor_Beta| 1 2 3
--------+------------------------------------------
1| 1.00000 .132831 -.644484
2| .132831 1.00000 .672107
3| -.644484 .672107 1.00000
N10: Random Parameter Models for Binary Choice N-145

-----------------------------------------------------------------------------
Partial derivatives of expected val. with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Conditional Mean at Sample Point .6687
Scale Factor for Marginal Effects .2215
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00305* .19821 1.89 .0591 -.00012 .00621
EDUC| -.00202 -.03425 -1.28 .1994 -.00511 .00107
HHNINC| .02238 .01178 .38 .7014 -.09203 .13679
HSAT| -.05744 -.58287 -.70 .4825 -.21776 .10288
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Results saved by this estimator are:

Matrices: b = estimate of q
varb = asymptotic covariance matrix for estimate of q.
gammaprm = the estimate of Γ
beta_i = individual specific parameters, if ; Par is requested
sdbeta_i = individual specific parameter standard deviations if ; Par
is requested

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Last Model: b_variables

Last Function: None

Simulation based estimation is time consuming. The sample size here is fairly large (27,326
observations). We limited the simulation to 25 Halton draws. The amount of computation rises
linearly with the number of draws. A typical application of the sort pursued here would use perhaps
300 draws, or 12 times what we used. Estimation of the last model required two minutes and 30
seconds, so in full production, estimation of this model might take 30 minutes. In general, you can
get an idea about estimation times by starting with a small model and a small number of draws. The
amount of computation rises linearly with the number of draws – that is the main consumer. It also
rises linearly with the number of random parameters. The time spent fitting the model will rise only
slightly with the number of nonrandom numbers. Finally, it will rise linearly with the number of
observations. Thus, a model with a doubled sample and twice as many draws will take four times as
long to estimate as one with the original sample and number of draws.
When you include ; Par in the model command, two additional matrices are created, beta_i
and sdbeta_i. Extensive detail on the computation of these matrices is provided in Section R24.5.
For the final specification described above, the results would be as shown in Figure N10.1.
N10: Random Parameter Models for Binary Choice N-146

Figure N10.1 Estimated Conditional Parameter Means

N10.2.3 Controlling the Simulation

R is the number of points in the simulation. Authors differ in the appropriate value. Train
recommends several hundred. Bhat suggests 1,000 is an appropriate value. The program default is
100. You can choose the value with

; Pts = number of draws, R

The value of 50 that we set in our experiments above was chosen purely to produce an example that
you could replicate without spending an inordinate amount of waiting for the results.
The standard approach to simulation estimation is to use random draws from the specified
distribution. As suggested immediately above, good performance in this connection requires very
large numbers of draws. The drawback to this approach is that with large samples and large models,
this entails a huge amount of computation and can be very time consuming. Some authors have
documented dramatic speed gains with no degradation in simulation performance through the use of
a small number of Halton draws instead of a large number of random draws. Authors (e.g., Bhat
(2001)) have found that a Halton sequence of draws with only one tenth the number of draws as a
random sequence is equally effective. To use this approach, add

; Halton

to your model command.

In order to replicate an estimation, you must use the same random draws. One implication
of this is that if you give the identical model command twice in sequence, you will not get the
identical set of results because the random draws in the sequences will be different. To obtain the
same results, you must reset the seed of the random number generator with a command such as

CALC ; Ran(seed value) $

N10: Random Parameter Models for Binary Choice N-147

(Note that we have used Ran(12345) before some of our earlier examples, precisely for this reason.
The specific value you use for the seed is not of consequence; any odd number will do.
The random sequence used for the model estimation must be the same in order to obtain
replicability. In addition, during estimation of a particular model, the same set of random draws
must be used for each person every time. That is, the sequence vi1, vi2, ..., viR used for each
individual must be same every time it is used to calculate a probability, derivative, or likelihood
function. (If this is not the case, the likelihood function will be discontinuous in the parameters, and
successful estimation becomes unlikely.) One way to achieve this which has been suggested in the
literature is to store the random numbers in advance, and simply draw from this reservoir of values
as needed. Because NLOGIT is able to use very large samples, this is not a practical solution,
especially if the number of draws is large as well. We achieve the same result by assigning to each
individual, i, in the sample, their own random generator seed which is a unique function of the global
random number seed, S, and their group number, i;

Seed(S,i) = S + 123.0 × i, then minus 1.0 if the result is even.

Since the global seed, S, is a positive odd number, this seed value is unique, at least within the
several million observation range of NLOGIT.

N10.2.4 The Parameter Vector and Starting Values

Starting values for the iterations are obtained by fitting the basic model without random
parameters. Other parameters are set to zero. Thus, the initial results in the output for these models
will be the binary choice models discussed in the preceding sections. You may provide your own
starting values for the parameters with

; Start = ... the list of values for q.

The parameter vector is laid out as follows, in this order:

α1, ..., αK are the K nonrandom parameters,

b1,...,bM are the M means of the distributions of the random parameters,
σ1,σ2,...,σM are the M scale parameters for the distributions of the random parameters.

These are the essential parameters. If you have specified that parameters are to be correlated, then
the σs are followed by the below diagonal elements of Γ. (The σs are the diagonal elements.) If you
have specified heterogeneity variables, z, then the preceding are followed by the rows of D.
Consider an example: The model specifies:

; RPM = z1,z2
; Rhs = one,x1,x2,x3,x4 ? base parameters b1, b2, b3, b4, b5
; Fcn = one(n),x2(n),x4(n)
; Cor
N10: Random Parameter Models for Binary Choice N-148

Then, after rearranging, the model becomes

Variable Parameter
x1 α1
x3 α2
one b1 + σ1vi1 + δ11zi1 + δ12zi2
x2 b2 + σ2vi2 + γ21vi1 + δ11zi1 + δ12zi2
x4 b3 + σ3vi3 + γ31vi1 + γ32vi2 + δ11zi1 + δ12zi2

and the parameter vector would be

q = α1, α2, b1, b2, b3, σ1, σ2, σ3, γ21, γ31, γ32, δ11, δ12, δ21, δ22, δ31, δ32.

You may use ; Rst and ; CML to impose restrictions on the parameters. Use the preceding as a
guide to the arrangement of the parameter vector. We do note, using ; Rst to impose fixed value,
such as zero restrictions, will generally work well. Other kinds of restrictions, particularly across the
parts of the parameter vector, will generally produce unfavorable results.
The variances of the underlying random variables are given earlier, 1 for the normal
distribution, 1/3 for the uniform, and 1/6 for the tent distribution. The σ parameters are only the
standard deviations for the normal distribution. For the other two distributions, σk is a scale
parameter. The standard deviation is obtained as σk/ 3 for the uniform distribution and σk/ 6 for
the triangular distribution. When the parameters are correlated, the implied covariance matrix is
adjusted accordingly. The correlation matrix is unchanged by this.

N10.2.5 A Dynamic Probit Model

We consider estimation of the dynamic (habit persistence) probit model

yit* = α + b′xit + γyi,t-1 + eit + σui, t = 0,...,Ti, i = 1,...,N

yit = 1(yit* > 0).

Simple estimation of the model by maximum likelihood is clearly inappropriate owing to the random
effect. ML random effects is likewise inconsistent because yi,t-1 will be correlated with the random
effect. Following Heckman (1981), a suggested formulation and procedure for estimation are as
follows: Treat the initial condition as an equilibrium, in which

yi0* = φ + δ′xi0 + ei0 + τui

yi0 = 1(yi0* > 0)

and retain the preceding model for periods 1,...,Ti. Note that the same random effect, ui appears
throughout, but the scaling parameter and the slope vector are different in the initial period. The
lagged value of yit does not appear in period 0. This model can be estimated in this form with the
random parameters estimator in NLOGIT. Use the following procedure.
N10: Random Parameter Models for Binary Choice N-149

Set up the variables:

dit = 1 in period 1, 0 in all other periods,

fit = 1 - dit = 1 in all periods except period 1,
xit = the set of regressors in the model, 0 in the first period,
xi0 = the set of regressors in the model in period 0, 0 in all other periods,
yi,-1 = yi,t-1 in periods 1,...,Ti, 0 in the first period.

Then, the encompassing model is

yit* = b′xit + δ′xi0 + φdit + αfit + γyi,-1 + eit + σfitui + τditui,

yit = 1(yit* > 0), t = 0,1,...,Ti.

The commands you might use to set up the data would follow these steps. First, use CREATE to set
up your group size count variable, _groupti.

CREATE ; yit = the dependent variable

; yit1 = yit[-1] ? Make sure that yit1 = 0 in the first period.
; t = Trn(-ti,1) or whatever means to set up 1,2,...Ti + 1
; dit = (t=1) ; fit = (t > 1) $
CREATE ; set up the xit and xi0 sets of variables $

The estimation command is a random parameters probit model. We make use of a special feature of
the RPM that allows the random component of the random parameters to be shared by more than one
parameter. This is precisely what is needed to have both τui and σui appear in the equation without
forcing τ = σ.

PROBIT ; Lhs = yit

; Rhs = xit,xi0,yit1,dit,fit
; Panel
; RPM
; Fcn = dit(n), fit(n)
; Common
; ... any other desired specifications for the estimation $

A refinement of this model assumes that ui = l′zi + wi for a set of time invariant variables. (See
Hyslop (1999) and Greene (2011.) One possibility is the vector of group means of the variables xit.
(Only the time varying variables would be included in these means.) These can be created and
included as additional Rhs variables.
N10: Random Parameter Models for Binary Choice N-150

N10.3 Latent Class Models for Binary Choice

The binary choice model for a panel of data, i = 1,...,N, t = 1,...,Ti is

Prob[Yit = yit | xit] = F(yit,b′xit) = P(i,t), yit = 0 or 1.

Henceforth, we use the term ‘group’ to indicate the Ti observations on respondent i in periods t =
1,...,Ti. Unobserved heterogeneity in the distribution of yit is assumed to impact the density in the
form of a random effect. The continuous distribution of the heterogeneity is approximated by using
a finite number of ‘points of support.’ The distribution is approximated by estimating the location of
the support points and the mass (probability) in each interval. In implementation, it is convenient
and useful to interpret this discrete approximation as producing a sorting of individuals (by
heterogeneity) into J classes, j = 1,...,J. (Since this is an approximation, J is chosen by the analyst.)
Thus, we modify the model for a latent sorting of yit into J ‘classes’ with a model which
allows for heterogeneity as follows: The probability of observing yit given that regime j applies is

P(i,t|j) = Prob[Yit = yit| xit, j]

where the density is now specific to the group. The analyst does not observe directly which class,
j = 1,...,J generated observation yit|j, and class membership must be estimated. Heckman and Singer
(1984) suggest a simple form of the class variation in which only the constant term varies across the
classes. This would produce the model

P(i,t|j) = F[yit, b′xit + δj], Prob[class = j] = Fj

We formulate this approximation more generally as,

P(i,t|j) = F[yit, b′xit + δj′xit], Fj = exp(qj) / Σj exp(qj), with qJ = 0.

In this formulation, each group has its own parameter vector, b j′ = b + δj, though the variables that
enter the mean are assumed to be the same. (This can be changed by imposing restrictions on the
full parameter vector, as described below.) This allows the Heckman and Singer formulation as a
special case by imposing restrictions on the parameters. You may also specify that the latent class
probabilities depend on person specific characteristics, so that

qij = qj′zi, qJ = 0.

The estimation command for this model is

PROBIT ; Lhs = dependent variable

or LOGIT ; Rhs = independent variables
; Panel or Pds = fixed periods or count variable
; LCM $
N10: Random Parameter Models for Binary Choice N-151

The default number of support points is five. You may set J from two to 30 classes with

; Pts = the value

Use ; LCM = list of variables in zi

to specify the multinomial logit form of the latent class probabilities.

Estimates retained by this model include

Matrices: b = full parameter vector, [b 1′, b2′,... F1,...,FJ]

varb = full covariance matrix
Note that b and varb involve J×(K+1) estimates.

Two additional matrices are created:

b_class = a J×K matrix with each row equal to the corresponding bj

class_pr = a J×1 vector containing the estimated class probabilities

If the command specifies ; Parameters, then the additional matrix created is:

beta_i = individual specific parameters

Scalars: kreg = number of variables in Rhs list

nreg = total number of observations used for estimation
logl = maximized value of the log likelihood function
exitcode = exit status of the estimation procedure

N10.3.1 Application
To illustrate the model, we will fit probit models with three latent classes as alternatives to
the continuously varying random parameters models in the preceding section. This model requires a
fairly rich data set – it will routinely fail to find a maximum if the number of observations in a group
is small. In addition, it will break down if you attempt to fit too many classes. (This point is
addressed in Heckman and Singer.)
The model estimates include the estimates of the prior probabilities of group membership. It
is also possible to compute the posterior probabilities for the groups, conditioned on the data. The
; List specification will request a listing of these. The final illustration below shows this feature for
a small subset of the data used above. The models use the following commands: The first is the
pooled probit estimator. The second is a basic, three class LCM. The third models the latent class
probabilities as functions of the gender and marital status dummy variables. The final model
command fits a comparable random parameters model. We will compare the two estimated models.
N10: Random Parameter Models for Binary Choice N-152

Fit the pooled probit model first, basic latent class, then latent class with the gender and
marital status dummy variables in the class probabilities.

PROBIT ; Lhs = doctor ; Rhs = x,one

; Partial Effects
; Cluster = id $
MATRIX ; betapool = b’ $
PROBIT ; Lhs = doctor ; Rhs = x,one
; Partial Effects
; Pds = _groupti
; LCM
; Pts = 3 $
PROBIT ; Lhs = doctor ; Rhs = x,one
; Partial Effects
; Pds = _groupti
; LCM = female,married
; Pts = 3
; Parameters $
Fit the random parameters probit model with heterogeneity in means.

PROBIT ; Lhs = doctor ; Rhs = x,one

; Partial Effects
; Pds = _groupti
; RPM = female,married
; Fcn = one(n),hhninc(n),newhsat(n)
; Correlated
; Pts = 25
; Halton
; Parameters $

These are the estimated parameters of the pooled probit model. The cluster correction is shown with
the pooled results.
+---------------------------------------------------------------------+
| Covariance matrix for the model is adjusted for data clustering. |
| Sample of 27326 observations contained 7293 clusters defined by |
| variable ID which identifies by a value a cluster ID. |
+---------------------------------------------------------------------+
-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable DOCTOR
Log likelihood function -16638.96591
Restricted log likelihood -18019.55173
Chi squared [ 4 d.f.] 2761.17165
Significance level .00000
McFadden Pseudo R-squared .0766160
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =33287.932 AIC/N = 1.218
Hosmer-Lemeshow chi-squared = 20.59314
P-value= .00831 with deg.fr. = 8
-----------------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-153

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
AGE| .00855*** .00098 8.75 .0000 .00664 .01047
EDUC| -.01539*** .00499 -3.08 .0020 -.02517 -.00561
HHNINC| -.00663 .05646 -.12 .9066 -.11729 .10404
HSAT| -.17502*** .00490 -35.72 .0000 -.18462 -.16542
Constant| 1.35894*** .08475 16.03 .0000 1.19282 1.52505
--------+--------------------------------------------------------------------

These are the estimates of the basic three class latent class model.

-----------------------------------------------------------------------------
Latent Class / Panel Probit Model
Dependent variable DOCTOR
Log likelihood function -15609.05992
Restricted log likelihood -16638.96591
Chi squared [ 13 d.f.] 2059.81198
Significance level .00000
McFadden Pseudo R-squared .0618972
Estimation based on N = 27326, K = 17
Inf.Cr.AIC =31252.120 AIC/N = 1.144
Unbalanced panel has 7293 individuals
PROBIT (normal) probability model
Model fit with 3 latent classes.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Model parameters for latent class 1
AGE| .01388*** .00228 6.10 .0000 .00942 .01835
EDUC| -.00381 .01146 -.33 .7399 -.02627 .01866
HHNINC| -.07299 .15239 -.48 .6320 -.37166 .22569
HSAT| -.20115*** .01709 -11.77 .0000 -.23466 -.16765
Constant| 2.08411*** .23986 8.69 .0000 1.61399 2.55424
|Model parameters for latent class 2
AGE| .01336*** .00183 7.29 .0000 .00977 .01696
EDUC| -.01886** .00815 -2.31 .0206 -.03483 -.00289
HHNINC| .06824 .10660 .64 .5221 -.14069 .27717
HSAT| -.20129*** .00994 -20.26 .0000 -.22076 -.18181
Constant| 1.15407*** .17393 6.64 .0000 .81317 1.49498
|Model parameters for latent class 3
AGE| .00547 .00464 1.18 .2390 -.00363 .01456
EDUC| -.04318** .01911 -2.26 .0239 -.08063 -.00572
HHNINC| .30044 .21747 1.38 .1671 -.12579 .72668
HSAT| -.14638*** .01965 -7.45 .0000 -.18489 -.10786
Constant| .24354 .31547 .77 .4401 -.37478 .86186
|Estimated prior probabilities for class membership
Class1Pr| .40689*** .04775 8.52 .0000 .31331 .50048
Class2Pr| .45729*** .03335 13.71 .0000 .39192 .52266
Class3Pr| .13581*** .02815 4.82 .0000 .08063 .19100
--------+--------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-154

The three class latent class model is extended to allow the prior class probabilities to differ by sex
and marital status.
-----------------------------------------------------------------------------
Latent Class / Panel Probit Model
Dependent variable DOCTOR
Log likelihood function -15471.73843
Restricted log likelihood -16638.96591
Chi squared [ 19 d.f.] 2334.45496
Significance level .00000
McFadden Pseudo R-squared .0701502
Estimation based on N = 27326, K = 21
Inf.Cr.AIC =30985.477 AIC/N = 1.134
Unbalanced panel has 7293 individuals
PROBIT (normal) probability model
Model fit with 3 latent classes.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Model parameters for latent class 1
AGE| .01225*** .00240 5.11 .0000 .00755 .01695
EDUC| .01438 .01311 1.10 .2725 -.01130 .04007
HHNINC| -.02303 .16581 -.14 .8895 -.34801 .30194
HSAT| -.17738*** .01802 -9.84 .0000 -.21271 -.14205
Constant| 1.76773*** .25126 7.04 .0000 1.27528 2.26018
|Model parameters for latent class 2
AGE| .00185 .00409 .45 .6508 -.00616 .00986
EDUC| -.03067** .01439 -2.13 .0331 -.05888 -.00245
HHNINC| .23788 .18111 1.31 .1890 -.11709 .59285
HSAT| -.15169*** .01623 -9.35 .0000 -.18349 -.11989
Constant| .44044* .26021 1.69 .0905 -.06957 .95045
|Model parameters for latent class 3
AGE| .01401*** .00199 7.02 .0000 .01010 .01791
EDUC| -.00399 .00847 -.47 .6372 -.02060 .01261
HHNINC| .03018 .11424 .26 .7916 -.19372 .25408
HSAT| -.21215*** .01178 -18.01 .0000 -.23524 -.18906
Constant| 1.13165*** .18329 6.17 .0000 .77241 1.49088
|Estimated prior probabilities for class membership
ONE_1| -.53375** .21925 -2.43 .0149 -.96347 -.10403
FEMALE_1| 1.18549*** .13400 8.85 .0000 .92284 1.44813
MARRIE_1| -.33518** .16234 -2.06 .0390 -.65336 -.01700
ONE_2| -.51961* .26512 -1.96 .0500 -1.03924 .00002
FEMALE_2| -.31028* .18197 -1.71 .0882 -.66694 .04638
MARRIE_2| -.42489** .18253 -2.33 .0199 -.78265 -.06713
ONE_3| 0.0 .....(Fixed Parameter).....
FEMALE_3| 0.0 .....(Fixed Parameter).....
MARRIE_3| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------
+------------------------------------------------------------+
| Prior class probabilities at data means for LCM variables |
| Class 1 Class 2 Class 3 Class 4 Class 5 |
| .36905 .17087 .46008 .00000 .00000 |
+------------------------------------------------------------+
N10: Random Parameter Models for Binary Choice N-155

Since the class probabilities now differ by observation, the program reports an average using
the data means. The earlier fixed prior class probabilities are shown below the averages for this
model. The extension brings only marginal changes in the averages, but this does not show the
variances across the different demographic segments (female/male, married/single) which may be
substantial.
These are the estimated ‘individual’ parameter vectors.

Figure N10.2 Latent Class Parameter Estimates

The random parameters model in which parameter means differ by sex and marital status and are
correlated with each other is comparable to the full latent class model shown above.

-----------------------------------------------------------------------------
Random Coefficients Probit Model
Dependent variable DOCTOR
Log likelihood function -15469.87914
Restricted log likelihood -16638.96591
Chi squared [ 12 d.f.] 2338.17354
Significance level .00000
McFadden Pseudo R-squared .0702620
Estimation based on N = 27326, K = 17
Inf.Cr.AIC =30973.758 AIC/N = 1.133
Unbalanced panel has 7293 individuals
PROBIT (normal) probability model
Simulation based on 25 Halton draws
-----------------------------------------------------------------------------
N10: Random Parameter Models for Binary Choice N-156

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .01161*** .00086 13.51 .0000 .00993 .01330
EDUC| -.00704* .00407 -1.73 .0833 -.01501 .00093
|Means for random parameters
Constant| 1.29395*** .09898 13.07 .0000 1.09995 1.48795
HHNINC| .08845 .10690 .83 .4080 -.12108 .29798
HSAT| -.21458*** .00954 -22.50 .0000 -.23327 -.19589
|Diagonal elements of Cholesky matrix
Constant| 1.04680*** .04364 23.98 .0000 .96126 1.13234
HHNINC| .69686*** .04676 14.90 .0000 .60521 .78851
HSAT| .00014 .00120 .12 .9049 -.00220 .00248
|Below diagonal elements of Cholesky matrix
lHHN_ONE| .10493* .05843 1.80 .0725 -.00960 .21946
lHSA_ONE| -.03295*** .00517 -6.37 .0000 -.04309 -.02282
lHSA_HHN| .04592*** .00248 18.54 .0000 .04107 .05078
|Heterogeneity in the means of random parameters
cONE_FEM| .20456*** .07264 2.82 .0049 .06218 .34694
cONE_MAR| .07909 .08153 .97 .3320 -.08070 .23888
cHHN_FEM| .08596 .10341 .83 .4059 -.11672 .28863
cHHN_MAR| -.07299 .11495 -.63 .5254 -.29828 .15230
cHSA_FEM| .02966*** .00873 3.40 .0007 .01256 .04677
cHSA_MAR| -.00931 .00991 -.94 .3474 -.02873 .01011
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Implied covariance matrix of random parameters

Var_Beta| 1 2 3
--------+------------------------------------------
1| 1.09579 .109842 -.0344941
2| .109842 .496629 .0285454
3| -.0344941 .0285454 .00319490

Implied standard deviations of random parameters

S.D_Beta| 1
--------+--------------
1| 1.04680
2| .704719
3| .0565235

Implied correlation matrix of random parameters

Cor_Beta| 1 2 3
--------+------------------------------------------
1| 1.00000 .148897 -.582977
2| .148897 1.00000 .716624
3| -.582977 .716624 1.00000

These are the estimated marginal effects from the three models estimated, the pooled probit model,
the three class latent class model and a comparable random parameters model, respectively.
N10: Random Parameter Models for Binary Choice N-157

Pooled
-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00297*** .20548 8.83 .0000 .00231 .00363
EDUC| -.00534*** -.09614 -3.09 .0020 -.00873 -.00195
HHNINC| -.00230 -.00129 -.12 .9066 -.04072 .03612
HSAT| -.06076*** -.65534 -39.87 .0000 -.06375 -.05777
--------+--------------------------------------------------------------------
3 Class Latent Class
--------+--------------------------------------------------------------------
AGE| .00446*** .28510 7.28 .0000 .00326 .00566
EDUC| -.00572*** -.09511 -2.64 .0082 -.00997 -.00148
HHNINC| .01510 .00780 .61 .5433 -.03360 .06381
HSAT| -.06917*** -.68884 -19.60 .0000 -.07609 -.06225
--------+--------------------------------------------------------------------
3 Class Heterogeneous Priors
-----------------------------------------------------------------------------
AGE| .00406*** .26197 7.00 .0000 .00292 .00520
EDUC| -.00064 -.01069 -.27 .7838 -.00519 .00391
HHNINC| .01657 .00865 .68 .4953 -.03106 .06420
HSAT| -.06804*** -.68420 -20.83 .0000 -.07444 -.06164
--------+--------------------------------------------------------------------
Random Parameters
-----------------------------------------------------------------------------
AGE| .00424*** .27768 3.18 .0015 .00162 .00685
EDUC| -.00257 -.04379 -1.48 .1385 -.00597 .00083
HHNINC| .03226 .01711 .55 .5814 -.08242 .14695
HSAT| -.07827 -.79992 -1.22 .2216 -.20379 .04724
--------+--------------------------------------------------------------------
N11: Semiparametric and Nonparametric Models for Binary Choice N-158

N11: Semiparametric and Nonparametric

Models for Binary Choice
N11.1 Introduction
This chapter will present three non- and semiparametric estimators for binary choice models.
Familiar parametric estimators of binary response models, such as the probit and logit are based on
the log likelihood criterion,

∑i =1
1 n
log L = log F ( yi | β' xi ) .
n
The Cramer-Rao theory justifies this procedure on the basis of efficiency of the parameter estimates.
But, it is to be noted that the criterion is not a function of the ability of the model to predict the
response. Moreover, in spite of the widely observed similarity of the predictions from the different
models, the issue of which parametric family (normal, logistic, etc.) is most appropriate has never
been settled, and there exist no formal tests to resolve the question in any given setting. Various
estimators have been suggested for the purpose of broadening the parametric family, so as to relax
the restrictive nature of the model specification. Two semiparametric estimators are presented in
NLOGIT, Manski’s (1975, 1985) and Manski and Thompson’s (1985, 1987) maximum score
(MSCORE) estimator and Klein and Spady’s (1993) kernel density estimator.
The MSCORE estimator is constructed specifically around the prediction criterion
Choose b to maximize S = Σi [yi* × zi*],
where yi* = sign (-1/1) of the dependent variable
zi* = the sign (-1/1) of b′xi.
Thus, the MSCORE estimator seeks to maximize the number of correct predictions by our familiar
prediction rule – predict yi = 1 when the estimated Prob[yi = 1] is greater than .5, assuming that the
true, underlying probability function is symmetric. In those settings, such as probit and logit, in
which the density is symmetric, the sign of the argument is sufficient to define whether the
probability is greater or less than .5. For the asymmetric distributions, this is not the case, which
suggests a limitation of the MSCORE approach. The estimator does allow another degree of
freedom in the choice of a quantile other than .5 for the prediction rule – see the definition below –
but this is only a partial solution unless one has prior knowledge about the underlying density.
Klein and Spady’s semiparametric density estimator is based on the specification

Prob[yi = 1] = P(b′xi)
where P is an unknown, continuous function of its argument with range [0,1]. The function P is not
specified a priori; it is estimated with the parameters. The probability function provides the location
for the index that would otherwise be provided by a constant term. The estimation criterion is
1 n
log L = ∑ [ yi log Pn (ββ
n i =1
′xi ) + (1 − yi ) log(1 − Pn ( ′xi ))]

where Pn is the estimator of P and is computed using a kernel density estimator.

N11: Semiparametric and Nonparametric Models for Binary Choice N-159

The third estimator is a nonparametric treatment of binary choice based on the index
function estimated from a parametric model such as a logit model.

N11.2 Maximum Score Estimation - MSCORE

Maximum score is a semiparametric approach to estimation which is based on a prediction
rule. The base case (quantile = ½) is

S = Σi [yi* × zi* ],

where yi* is the sign (-1/1) of the dependent variable and zi* is the counterpart for the fitted model;
zi* = the sign (-1/1) of b′xi. Thus, this base case is formulated precisely upon the ability of the sign
of the estimated index function to predict the sign of the dependent variable (which, in the binary
response models, is all that we observe). Formally, MSCORE maximizes the sample score function

MaxbeB Snα(b) = (1/n)Σi[yi* - (1-2α)]Sgn(b′xi),

where B = {b e RK : ║b║ = 1}.

The sample data consist of n observations [yi* ,xi] where yi* is the binary response. Input of yi is the
usual binary variable taking values zero and one; yi* is obtained internally by converting zeros to
minus ones. The quantile, α, is between zero and one and is provided by the user. The vector xi is
the usual set of K regressors, usually including a constant. An equivalent problem is to maximize the
normalized sample score function

SNα*(b) = (1/n)[Snα(b) / Wn + 1],

where Wn = (1/n)Σiwi
and wi = abs(yi* - (1-2α)).

This may then be rewritten as

Snα*(b) = Σi wi* × 1[yi* = Sgn(b′xi)],

where wi* = wi / Wn.

and 1[•] is the indicator function which equals 1 if the condition in the brackets is true and 0
otherwise. Thus, in the preceding, 1[•] equals 1 if the sign of the index function, b′xi, correctly
predicts yi*. The normalized sample score function is, thus, a weighted average of the prediction
indicators. If α = ½, then wi* equals 1/n, and the normalized score is the fraction of the observations
for which the response variable is correctly predicted. Maximum score estimation can therefore be
interpreted as the problem of finding the parameters that maximize a weighted average number of
correct predictions for the binary response.
The following shows how to use the MSCORE command and gives technical details about
the procedure. An application is given with the development of NPREG, which is a companion
program, in Section N11.4.
N11: Semiparametric and Nonparametric Models for Binary Choice N-160

N11.2.1 Command for MSCORE

The mandatory part of the command for invoking the maximum score estimator

MSCORE ; Lhs = y ; Rhs = x list of independent variables $

The first element of x should be one. The variable y is a binary dependent variable, coded 0/1. The
following are the optional specifications for this command. The default values given are used by
NLOGIT if the option is not specified on the command. MSCORE is designed for relatively small
problems. The internal limits are 15 parameters and 10,000 observations.

N11.2.2 Options Specific to the Maximum Score Estimator

Quantile

The quantile defines the way the score function is computed. The default of .5 dictates that
the score is to be calculated as (1/n) times the number of correctly predicted signs of the response
variable. You may choose any value between 0 and 1with

; Qnt = quantile (default = .5; this is α).

Number of Bootstrap Replications

Bootstrap estimates are computed as follows: After computing the point estimate,
MSCORE generates R bootstrap samples from the data by sampling n observations with
replacement. The entire point estimation procedure, including computation of starting values is
repeated for each one. Let b be the maximum score estimate, R be the number of bootstrap
replications, and di be the ith bootstrap estimate. The mean squared deviation matrix,

MSD = (1/R)Σi [(di - b)(di - b)′],

is computed from the bootstrap estimates. This is reported in the output as if it were the estimated
covariance matrix of the estimates. But, it must be noted that there is no theory to suggest that this is
correct. In purely practical terms, the deviations are from the point estimate, not the mean of the
bootstrap estimates. The results are merely suggestive. The use of ; Test: should also be done with
this in mind. Use
; Nbt = number of bootstraps (default = 20)

to set the number of bootstrap iterations.

N11: Semiparametric and Nonparametric Models for Binary Choice N-161

Analysis of Ties

The specification for analysis of ties is

; Ties to analyze ties (default = no)

If the ; Ties option is chosen, MSCORE reports information about regions of the parameter space
discovered during the endgame searches for which the sample score is tied with the score at the final
estimates. If a tie is found in a region, MSCORE records the endpoints of the interval, the current
search direction, and some information which records each observation’s contribution to the sample
score in the region. It is possible to determine whether ties found on separate great circle searches
represent disjoint regions or intersections of different great circles. Since the region containing the
final estimates is partially searched in each iteration, the tie checking procedure records extensive
information about this region. For each region, MSCORE reports the minimum and maximum
angular direction from the final estimates. These are labeled PSI-low and PSI-high. The parameter
values associated with these endpoints are also reported.
If tie regions are found that are far from the point estimate, it may be that the global
maximum remains to be found. If so, it may be useful to rerun the estimator using a starting value in
the tied region. The existence of many tie regions does not necessarily indicate an unreliable
estimate. Particularly in large samples, there may be a large number of disjoint regions in a small
neighborhood of the global maximum.

Number of Endgame Iterations

The number of endgame iterations is specified with

; End = number endgame iterations (default = 5).

A given set of great circle searches may miss a direction of increase in the score function. Moreover,
even if the trial maximum is a true local maximum, it may not be a global maximum. For these reasons,
upon finding a trial maximum, MSCORE conducts a user specified number of ‘endgame iterations.’
These are simply additional iterations of the maximization algorithm. The random search method is
such that with enough of these, the entire parameter space would ultimately be searched with
probability one. If the endgame iterations provide no improvement in the score, the trial maximum is
deemed the final estimate. If an improvement is made during an endgame search, the current estimate
is updated as usual and the search resumes. The logic of the algorithm depends on the endgame
searches to ensure that all regions of the parameter space are investigated with some probability. The
density of the coverage is an increasing function of the number of endgame searches.
There are no formal rules for the number of endgame searches. It should probably increase
with K and (perhaps a little less certainly) with n. But, because the step function more closely
approximates a continuous population score function, it may be that fewer endgame searches will be
needed as N increases.
N11: Semiparametric and Nonparametric Models for Binary Choice N-162

Starting Values

Starting values are specified with

; Start = starting values (default = none).

If starting values are not provided by the user, they are computed as follows: For each of the K
parameters, we form a vector equal to the kth column of an identity matrix. The sample score
function is evaluated at this vector, and the kth parameter is set equal to this value. At the
conclusion, the starting vector is normalized to unit length. If you do provide your own starting
values, they will be normalized to unit length before the iterations are begun.

Technical Output

Technical output is specified with

; Output = 4 or 5 for output of trace of bootstraps to output file

(default = neither).

This is used to control the amount of information about the bootstrap iterations that is produced.
This can generate hundreds or thousands of lines of output, depending on the number of bootstrap
estimates computed and the number of endgame searches requested. This information is displayed
on the screen, in order to trace the progress of execution. In general, the output is not especially
informative except in the aggregate. That is, individual lines of this trace are likely to be quite
similar. The default is not to retain information about individual bootstraps or endgame searches in
the file. Use ; Output = 4 to request only the bootstrap iterations (one line of output per). Use
; Output = 5 to include, in addition, the corresponding information about the endgame searches.

N11.2.3 General Options for MSCORE

The following general options used with the nonlinear estimators in NLOGIT are available
for MSCORE:
; Covariance Matrix to display MSE matrix (default = no),
same as ; Printvc
; List to display predicted values (default = no list)
; Keep = name to retain predictions in name (default = no)
; Res = name to retain fitted values in name (default = no)
; Test: spec to specify restriction (default = none)
; Maxit = n to set maximum iterations (default = 50)

Note the earlier caution about the MSD matrix when using the ; Test: option. The ; Rst = ... and
; CML: options for imposing restrictions are not available with this estimator.
N11: Semiparametric and Nonparametric Models for Binary Choice N-163

N11.2.4 Output from MSCORE

Output from MSCORE consists of the following, in the order in which it will appear on
your screen or your output file:

1. The iteration summary for the primary estimation procedure (this is labeled bootstrap sample
0’) and, if you have requested them, the bootstrap sample estimations. With each one, we
report the number of iterations, the number of completed ‘endgame iterations’ (see the
discussion above), the maximum normalized score, and the change in the normalized score.

2. Echo of input parameters in your command.

3. The score function and normalized score function evaluated at three different points:

a. naive, the first element of b is 1 or -1 and all other values are 0,

b. the starting values,
c. the final estimates.

4. The deviations of the bootstrap estimates from the point estimates are summarized in the
root mean square error and mean absolute angular deviation between them.

5. The point estimates of the parameters.

NOTE: The estimates are presented in NLOGIT’s standard format for parameter estimates.
If you have computed bootstrap estimates, the mean square deviation matrix (from the point
estimate) is reported as if it were an estimate of the covariance matrix of the estimates. This
includes ‘standard errors,’ ‘t ratios,’ and ‘prob. values.’ These may, in fact, not be
appropriate estimates of the asymptotic standard errors of these parameter estimates.
Discussion appears in the references below.

If you change the number of bootstrap estimates, you may observe large changes in these
standard errors. This is not to be interpreted as reflecting any changes in the precision of the
estimates. If anything, it reflects the unreliability of the bootstrap MSD matrix as an
estimate of the asymptotic covariance matrix of the estimates. It has been shown that the
asymptotic distribution of the maximum score estimator is not normal. (See Kim and
Pollard (1990).) Moreover, even under the best of circumstances, there is no guarantee that
the bootstrap estimates or functions of them (such as t ratios), converge to anything useful.

6. A cross tabulation of the predictions of the model vs. the actual values of the Lhs variable.

7. If the model has more than two parameters, and you have requested analysis of the ties, the
results of the endgame searches are reported last. Records of ties are recorded in your output
file if one is opened, but not displayed on your screen.
N11: Semiparametric and Nonparametric Models for Binary Choice N-164

The predicted values computed by MSCORE are the sign of b′xi, coded 0 or 1. Residuals
are yi - ŷ i, which will be 1, 0, or -1. The ; List specification also produces a listing of b′xi. The last
column of the listing, labeled Prob[y = 1] is the probabilities computed using the standard normal
distribution. Since the probit model has not been used to fit the model, these may be ignored.
Results which are saved by MSCORE are:

b = final estimates of parameters

varb = mean squared deviation matrix for bootstrap estimates
score = scalar, equal to the maximized value of the score function

The Last Model labels are b_variable. But, note once again, that the underlying theory needed to
justify use of the Wald statistic does not apply here.

N11.3 Klein and Spady’s Semiparametric Binary Choice Model

Klein and Spady’s semiparametric density estimator is based on the specification

Prob[yi = 1] = P(b′xi)

where P is an unknown, continuous function of its argument with range [0,1]. The function P is not
specified a priori; it is estimated with the parameters. The probability function provides the location
for the index that would otherwise be provided by a constant term. The estimation criterion is

1 n
log L = ∑ [ yi log Pn (ββ
n i =1
′xi ) + (1 − yi ) log(1 − Pn ( ′xi ))]

where Pn is the estimator of P and is computed using a kernel density estimator. The probability
function is estimated with a kernel estimator,

yj  β′(xi − x j ) 
∑
n
j =1
K 
h  h .
Pn(b′xi) =
1  β ′( x − x ) 
∑
n i j
j =1
K 
h  h 

Two kernel functions are provided, the logistic function, Λ(z) and the standard normal CDF, Φ(z).
As in the other semiparametric estimators, the bandwidth parameter is a crucial input. The
program default is n-(1/6), which ranges from .3 to about .6 for n ranging from 30 to 1000. You may
provide an alternative value.
N11: Semiparametric and Nonparametric Models for Binary Choice N-165

N11.3.1 Command
The command for this estimator is

SEMIPARAMETRIC
; Lhs = dependent, binary variable
; Rhs = independent variables $

Do not include one on the Rhs list. The function itself is playing the role of the constant. Optional
features include those specific to this model,

; Smooth = desired value for h

; Kernel = Normal – the logistic is standard
and the general ones available with other estimators,

; Partial Effects
; Prob = name to retain fitted probabilities
; Keep = name to retain predicted values
; Res = name to retain residuals
; Covariance Matrix to display the estimated asymptotic covariance matrix,
same as ; Printvc
The semiparametric log likelihood function is a continuous function of the parameters which is
maximized using NLOGIT’s standard tools for optimization. Thus, the options for controlling
optimization are available,
; Maxit = n to set maximum iterations
; Output = 1, 2, 3 to control intermediate output
; Alg = name to select algorithm

Restrictions may be imposed and tested with

; Test: spec to specify restriction (default = none)

; Rst = list to specify fixed value and equality restrictions
; CML: spec to specify other linear constraints

N11.3.2 Output
Output from this estimator includes the usual table of statistical results for a nonlinear
estimator. Note that the estimator constrains the constant term to zero and also normalizes one of the
slope coefficients to one for identification. This will be obvious in the results. Since probabilities
which are a continuous function of the parameters are computed, you may also request marginal
effects with
; Partial Effects
(In previous versions, the command was ; Marginal Effects. This form is still supported.) Partial
effects are computed using Pn(b′xi) and its derivatives (which are simple sums) computed at the
sample means.
N11: Semiparametric and Nonparametric Models for Binary Choice N-166

Results Kept by the Semiparametric Estimator

The model results kept by this estimator are
Matrices: b = final estimates of parameters
varb = mean squared deviation matrix for bootstrap estimates
Scalars: logl = log likelihood
kreg = number of Rhs variables
nreg = number of observations used to fit the function
exitcode = exit status for estimator
Last Model: The labels are b_variable

Last Function: None

N11.3.3 Application
The Klein and Spady estimator is computed with the binary logit model. We use only a
small subset of the data, the observations that are observed only once. The complete lack of
agreement of the two models is striking, though not unexpected.

REJECT ; _groupti > 1 $

SEMI ; Lhs = doctor
; Rhs = one,age,hhninc,hhkids,educ,married
; Partial Effects $
LOGIT ; Lhs = doctor
; Rhs = one,age,hhninc,hhkids,educ,married
; Partial Effects $
-----------------------------------------------------------------------------
Semiparametric Binary Choice Model
Dependent variable DOCTOR
Log likelihood function -1001.96124
Restricted log likelihood -1004.77427
Chi squared [ 4 d.f.] 5.62607
Significance level .22887
McFadden Pseudo R-squared .0027997
Estimation based on N = 1525, K = 4
Inf.Cr.AIC = 2011.922 AIC/N = 1.319
Hosmer-Lemeshow chi-squared = *********
P-value= .00000 with deg.fr. = 8
Logistic kernel fn. Bandwidth = .29475
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Odds Ratio Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Characteristics in numerator of Prob[Y = 1]
AGE| .98652 .02284 -.59 .5577 .94176 1.03128
HHNINC| .02962** .04607 -2.26 .0236 -.06067 .11991
HHKIDS| 3.16366 4.50864 .81 .4190 -5.67311 12.00042
EDUC| .96226 .11808 -.31 .7539 .73083 1.19368
MARRIED| 2.71828 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
N11: Semiparametric and Nonparametric Models for Binary Choice N-167

--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Odds ratio = exp(beta); z is computed for the original beta
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| -.00025 -.01488 -.59 .5523 -.00107 .00057
HHNINC| -.06479*** -.03782 -76.40 .0000 -.06645 -.06313
HHKIDS| .02120 .01063 .26 .7984 -.14148 .18388
EDUC| -.00071 -.01305 -.33 .7445 -.00497 .00355
MARRIED| .01841 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Binary Logit Model for Binary Choice
Dependent variable DOCTOR
Log likelihood function -996.30681
Restricted log likelihood -1004.77427
Chi squared [ 5 d.f.] 16.93492
Significance level .00462
McFadden Pseudo R-squared .0084272
Estimation based on N = 1525, K = 6
Inf.Cr.AIC = 2004.614 AIC/N = 1.315
Hosmer-Lemeshow chi-squared = 10.56919
P-value= .22732 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| .46605 .34260 1.36 .1737 -.20544 1.13754
AGE| .00509 .00448 1.14 .2556 -.00369 .01387
HHNINC| -.49045* .26581 -1.85 .0650 -1.01142 .03052
HHKIDS| -.36639*** .12639 -2.90 .0037 -.61410 -.11867
EDUC| .00783 .02419 .32 .7461 -.03957 .05523
MARRIED| .16046 .12452 1.29 .1975 -.08360 .40451
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N11: Semiparametric and Nonparametric Models for Binary Choice N-168

-----------------------------------------------------------------------------
Partial derivatives of E[y] = F[*] with
respect to the vector of characteristics
Average partial effects for sample obs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
DOCTOR| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00117 -.00127 1.14 .2554 -.00085 .00320
HHNINC| -.11304* .00087 -1.85 .0648 -.23301 .00694
HHKIDS| -.08606*** .00019 -2.87 .0041 -.14476 -.02736 #
EDUC| .00180 -.00053 .32 .7461 -.00912 .01273
MARRIED| .03702 -.00057 1.29 .1971 -.01924 .09327 #
--------+--------------------------------------------------------------------
# Partial effect for dummy variable is E[y|x,d=1] - E[y|x,d=0]
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N11.4 Nonparametric Binary Choice Model

The kernel density estimator is a device used to describe the distribution of a variable
nonparametrically, that is, without any assumption of the underlying distribution. This section
describes an extension to a simple regression function. The kernel density function estimates any
sufficiently smooth regression function, Fb(z) = E[δ|b′x=z], using the method of kernels, for any
parameter vector b. δ must be a response variable with bounded range [0,1]. In the special case in
which δ is a binary response taking values 0/1, NPREG estimates the probability of a positive
response conditional on the linear index b′x. With an appropriate choice of x and b, and by rescaling
the response, this estimator can estimate any sufficiently smooth univariate regression function with
known bounded range. One simple approach is to assume that x is a single variable and b equals 1.0,
in which case, the estimator describes E[yi|xi]. Alternatively, NPREG may be used with the
estimated index function, b′xi, from any binary choice estimator. The natural choice in this instance
would be MSCORE, since MSCORE does not compute the probabilities (that is, the conditional
mean). In principle, the estimated index function could come from any estimator, but from a probit
or other parametric model, this would be superfluous.
The regression function computed is

 z j − zi 
∑i =1 yi h K 
N 1

h 
F (z j ) = ., j = 1,...,M, i = 1,..., number of observations.
1  z j − zi 
∑i =1
N
K 
h  h 

The function is computed for a specified set of values zj, j = 1,...,M. Note that each value requires a
sum over the full sample of n values. The primary component of the computation is the kernel
function, K[.].
N11: Semiparametric and Nonparametric Models for Binary Choice N-169

Eight alternatives are provided:

1. Epanechnikov: K[z] = .75(1 - .2z2) / Sqr(5) if |z| <= 5, 0 else

2. Normal: K[z] = φ(z) (normal density)
3. Logit: K[z] = Λ(z)[1-Λ(z)] (default)
4. Uniform: K[z] = .5 if |z| < 1, 0 1 else
5. Beta: Z[z] = (1-z)(1+z)/24 if |z| < 1, 0 1 else
6. Cosine: K[z] = 1 + cos(2πz) if |z| < .5, 0 else
7. Triangle: K[z] = 1 - |z|, if |z| <= 1, 0 else
8. Parzen: K[z] = 4/3 - 8z2 + 8|z|3 if |z| <= .5, 8(1-|z|)3 else

The other essential part of the computation is the smoothing (bandwidth) parameter, h. Large values
of h stabilize the function, but tend to flatten it and reduce the resolution. Small values of h produce
greater detail, but also cause the estimator to become less stable.
The basic command is

NPREG ; Lhs = the dependent variable

; Rhs = the variable $

With no other options specified, the routine uses the logit kernel function, and uses a bandwidth
equal to
h = .9Q/n0.2 where Q = min(std.dev., range/1.5)

You may specify the kernel function to be used with

; Kernel = one of the names of the eight types of kernels listed above.

The bandwidth may be specified with

; Smooth = the bandwidth parameter.

There is no theory for choosing the right smoothing parameter, l. Large values will cause
the estimated function to flatten at the average value of yi. Values close to zero will cause the
function to pass through the points zi,yi and to become computationally unstable elsewhere. A choice
might be made on the basis of the CVMSPE. (See Wong (1983) for discussion.) A value that
minimizes CVMSPE(l) may work well in practice. Since CVMSPE is a saved result, you could
compute this for a number of values of l then retrieve the set of values to find the optimal one.
The default number of points specified is 100, with zj = a partition of the range of the
variable. You may specify the number of points, up to 200 with

; Pts = number of points to compute and plot.

The range of values plotted is the equally spaced grid from min(x)-h to max(x)+h, with the number
of points specified.
N11: Semiparametric and Nonparametric Models for Binary Choice N-170

N11.4.1 Output from NPREG

Output from KERNEL is a set of points for an estimated function, several descriptive
statistics, and a plot of the estimated regression function. The added specification

; List
displays the specific results, zi for the sample observations and the associated estimated regression
functions. These values are also placed in a two column matrix named kernel after estimation of the
function.
The cross validation mean squared prediction error (CVMSPE) is a goodness of fit measure.
Each observation, ‘i’ is excluded in turn from the sample. Using the reduced sample, the regression
function is reestimated at the point zi in order to provide a point prediction for yi. The average
squared prediction error defines the CVMSPE. The calculation is defined by

1  x − xi 
∑ j ≠i
yjK  j 
Fi* ( z ) =
h  h 
1  x j − xi 
∑ j ≠i
K
h  h 


Then, CVMSPE(h) = (1/n) Σi [yi - Fi*(xi)]2.

N11.4.2 Application
The following estimates the parameters of a regression function using MSCORE, then uses
NPREG to plot the regression function.
REJECT ; _groupti > 1 $
NAMELIST ; x = one,age,hhninc,hhkids,educ,married $
MSCORE ; Lhs = doctor ; Rhs =x $
CREATE ; xb = x'b $
NPREG ; Lhs = doctor ; Rhs = xb $
-----------------------------------------------------------------------------
Maximum Score Estimates of Linear Quantile
Regression Model from Binary Response Data
Quantile .500 Number of Parameters = 6
Observations input = 1525 Maximum Iterations = 500
End Game Iterations = 100 Bootstrap Estimates = 20
Check Ties? No
Save bootstraps? No
Start values from MSCORE (normalized)
Normal exit after 100 iterations.
Score functions: Naive At theta(0) Maximum
Raw .26033 .26033 .27738
Normalized .63016 .63016 .63869
Estimated MSEs from 20 bootstrap samples
(Nonconvergence in 0 cases)
Angular deviation (radians) of bootstraps from estimate
Mean square = 1.027841 Mean absolute = .979001
Standard errors below are based on bootstrap mean squared
deviations. These and the t-ratios are only approximations.
N11: Semiparametric and Nonparametric Models for Binary Choice N-171

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
DOCTOR| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Constant| .42253 .63272 .67 .5043 -.81758 1.66263
AGE| .01146 .03120 .37 .7134 -.04969 .07261
HHNINC| -.20766 .45880 -.45 .6508 -1.10689 .69157
HHKIDS| -.82224 .65955 -1.25 .2125 -2.11494 .47045
EDUC| .01446 .07191 .20 .8406 -.12648 .15541
MARRIED| .31926 .35336 .90 .3663 -.37331 1.01183
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
+---------------------------------------------------------+
|Predictions for Binary Choice Model. Predicted value is |
|1 when beta*x is greater than one, zero otherwise. |
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Value | |
|Value | 0 1 | Total Actual |
+------+----------------+----------------+----------------+
| 0 | 23 ( 1.5%)| 541 ( 35.5%)| 564 ( 37.0%)|
| 1 | 10 ( .7%)| 951 ( 62.4%)| 961 ( 63.0%)|
+------+----------------+----------------+----------------+
|Total | 33 ( 2.2%)| 1492 ( 97.8%)| 1525 (100.0%)|
+------+----------------+----------------+----------------+
+---------------------------------------------------------+
|Crosstab for Binary Choice Model. Predicted probability |
|vs. actual outcome. Entry = Sum[Y(i,j)*Prob(i,m)] 0,1. |
|Note, column or row total percentages may not sum to |
|100% because of rounding. Percentages are of full sample.|
+------+---------------------------------+----------------+
|Actual| Predicted Probability | |
|Value | Prob(y=0) Prob(y=1) | Total Actual |
+------+----------------+----------------+----------------+
| y=0 | 564 ( 37.0%)| 0 ( .0%)| 564 ( 37.0%)|
| y=1 | 961 ( 63.0%)| 0 ( .0%)| 961 ( 63.0%)|
+------+----------------+----------------+----------------+
|Total | 1525 (100.0%)| 0 ( .0%)| 1525 (100.0%)|
+------+----------------+----------------+----------------+
+---------------------------------------+
| Nonparametric Regression for DOCTOR |
| Observations = 1525 |
| Points plotted = 1525 |
| Bandwidth = .090121 |
| Statistics for abscissa values---- |
| Mean = .854823 |
| Standard Deviation = .433746 |
| Minimum = -.167791 |
| Maximum = 1.662874 |
| ---------------------------------- |
| Kernel Function = Logistic |
| Cross val. M.S.E. = .231635 |
| Results matrix = KERNEL |
+---------------------------------------+
N11: Semiparametric and Nonparametric Models for Binary Choice N-172

Figure N11.1 Nonparametric Regression

N12: Bivariate and Multivariate Probit and Partial Observability Models N-173

N12: Bivariate and Multivariate Probit and

Partial Observability Models
N12.1 Introduction
The basic formulation of the models in this chapter is the bivariate probit model:

zi1 = b 1′xi1 + ei1, yi1 = 1 if zi1 > 0, yi1 = 0 otherwise,

zi2 = b 2′xi2 + ei2, yi2 = 1 if zi2 > 0, yi2 = 0 otherwise,
[ei1,ei2] ~ bivariate normal (BVN) [0,0,1,1,ρ], -1 < ρ < 1,
individual observations on y1 and y2 are available for all i.

(This model is also available for grouped (proportions) data. See Section N12.2.2.) The model given
above would be estimated using a complete sample on [y1, y2, x1, x2] where y1 and y2 are binary
variables and xij are sets of regressors. This chapter will describe estimation of this model and
several variants:

• The disturbances in one or both equations may be heteroscedastic.

• The observation mechanism may be such that yi1 is not observed when yi2 equals zero.

• The observation mechanism may be such that only the product of yi1 and yi2 is observed.
That is, we only observe the compound outcomes ‘both variables equal one’ or ‘one or
both equal zero.’

• The basic model is extended to as many as 20 equations as a multivariate probit model.

NOTE: It is not necessary for there to be different variables in the two (or more) equations. The
Rh1 and Rh2 lists may be identical if your model specifies that. There is no issue of identifiability or
of estimability of the model – the variable lists are unrestricted. This is not a question of
identification by functional form. The analogous case is the SUR model which is also identified even
if the variables in the two equations are the same.

• Some extensions to a simultaneous equations model are easily programmed.

• The bivariate probit and partial observability models are extended to the random
parameters modeling framework for panel data.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-174

N12.2 Estimating the Bivariate Probit Model

The two equations can each be estimated consistently by individual single equation probit
methods (see Chapter E27). However, this is inefficient and incomplete in that it ignores the
correlation between the disturbances. Moreover, the correlation coefficient itself might be of interest.
The comparison is analogous to that between OLS and GLS in the multivariate regression model. The
model is estimated in NLOGIT using full information maximum likelihood. The essential command is

BIVARIATE PROBIT ; Lhs = y1,y2

(or just BIVARIATE) ; Rh1 = right hand side for equation 1
; Rh2 = right hand side for equation 2 $

N12.2.1 Options for the Bivariate Probit Model

Restrictions may be imposed both between and within equations by using
; Rst = list of specifications...
and ; CML: linear restrictions

You might, for example, force the coefficients in the two equations to be equal as follows:

NAMELIST ; x = ... $
CALC ; k = Col(x) $
BIVARIATE ; Lhs = y1,y2 ; Rh1 = x ; Rh2 = x ; Rst = k_b, k_b, r $

(The model is identified with the same variables in the two equations.)

NOTE: You should not use the name rho for ρ in your ; Rst specification; rho is the reserved name
for the scalar containing the most recently estimated value of ρ in whatever model estimated it. If it
has not been estimated recently, it is zero. Either way, when ; Rst contains the name rho, this is
equivalent to fixing ρ at the value then contained in the scalar rho. That is, rho is a value, not a
model parameter name such as b1. On the contrary, however, you might wish specifically to use rho
in your ; Rst specification. For example, to trace the maximized log likelihood over values of ρ, you
might base the study on a command set that includes

PROCEDURE $
BIVARIATE ; .... ; Rst = ..., rho $
...
ENDPROCEDURE $
EXECUTE ; rho = 0.0, .90, .10 $

This would estimate the bivariate probit model 10 times, with ρ fixed at 0, .1, .2, ..., .9. Presumably,
as part of the procedure, you would be capturing the values of logl and storing them for a later listing
or perhaps a plot of the values against the values of rho.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-175

If you use the constraints option, the parameter specification includes ρ. As such, you can
use this method to fix ρ to a particular value. This is a model for a voting choice and use of private
schools:
vote = f1(one,income,property_taxes)
private = f2(one,income,years,teacher).

Suppose it were desired to make the income coefficient the same in the two equations and, in a
second model, fix rho at 0.4. The commands could be

BIVARIATE ; Lhs = tax,priv

; Rh1 = one,inc,ptax ; Rh2 = one,inc,yrs,tch
; Rst = b10,bi,b12,b20,bi,b22,b23,r $

and BIVARIATE ; Lhs = tax,priv

; Rh1 = one,inc,ptax ; Rh2 = one,inc,yrs,tch
; Rst = b10,bi,b12,b20,bi,b22,b23,0.4 $

Choice Based Sampling

Any of the bivariate probit models may be estimated with choice based sampling. The feature
is requested with
; Wts = the appropriate weighting variable
; Choice Based

For this model, your weighting variable will take four values, for the four cells (0,0), (0,1), (1,0), and
(1,1);
wij = population proportion / sample proportion, i,j = 0,1.

The particular value corresponds to the outcome that actually occurs. You must provide the values.
You can obtain sample proportions you need if you do not already have them by computing a
crosstab for the two Lhs variables:

CROSSTAB ; Lhs = y1 ; Rhs = y2 $

The table proportions are exactly the proportions you will need. To use this estimator, it is assumed
that you know the population proportions.

Robust Covariance Matrix with Correction for Clustering

The standard errors for all bivariate probit models may be corrected for clustering in the
sample. Full details on the computation are given in Chapter R10, so we give only the final result
here. Assume that the data set is partitioned into G clusters of related observations (like a panel).
After estimation, let V be the estimated asymptotic covariance matrix which ignores the clustering.
Let gij denote the first derivatives of the log likelihood with respect to all model parameters for
observation (individual) i in cluster j.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-176

Then, the corrected asymptotic covariance matrix is

 G  G
Est.Asy.Var βˆ  = V   ∑ i 1 =
 G −1 
= ∑ (
ni
j 1 ∑
g ij =)(
ni
j 1 )
′
g ij  V


You specify the clusters with

; Cluster = either the fixed number of individuals in a group or the name of a variable
which identifies the group membership.

Any identifier which is common to all members in a cluster and different from other clusters may be
used. The controls for stratified and clustered data may be used as well. These are as follows:

; Cluster = the number of observations in a cluster (fixed) or the name of a

stratification variable which gives the cluster an identification. This
is the setup that is described above.
; Stratum = the number of observations in a stratum (fixed) or the name of a
stratification variable which gives the stratum an identification.
; Wts = the name of the usual weighting variable for model estimation if
weights are desired. This defines wics. This is the usual weighting
setup that has been used in all previous versions of LIMDEP and
NLOGIT.
; FPC = the name of a variable which gives the number of clusters in the
stratum. This number will be the same for all observations in a
stratum – repeated for all clusters in the stratum. If this number is
the same for all strata, then just give the number.
; Huber Use this switch to request hs. If omitted, hs = 1 is used.
; DFC Use this switch to request the use of d given above. If omitted,
d = 1 is used.

Note, these corrections will generally lead to larger standard errors compared to the uncorrected results.

N12.2.2 Proportions Data

Like other discrete choice models, this one may be fit with proportions data. Since this is a
bivariate model, you must provide the full set of four proportions variables, in the order

; Lhs = p00, p01, p10, p11.

(You may use your own names). Proportions must be strictly between zero and one, and the four
variables must add to 1.0.

NOTE: When you fit the model using proportions data, there is no cross tabulation of fitted and
actual values produced, and no fitted values or ‘residuals’ are computed.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-177

N12.2.3 Heteroscedasticity
All bivariate probit specifications, including the basic two equation model, the sample
selection model (Section N12.4), and the Meng and Schmidt partial observability model (Section
N12.7), may be fit with a multiplicative heteroscedasticity specification. The model is the same as
the univariate probit model;

ei ~ N{0, [exp( γi′zi)]2 }, i = 1 and/or 2.

Either or both equations may be specified in this fashion. Use

; Hf1 = list of variables if you wish to modify the first equation

; Hf2 = list of variables if you wish to modify the second equation

NOTE: Do not include one in either list. The model will become inestimable.

The model is unchanged otherwise, and the full set of options given earlier remains available. To
give starting values with this modification, supply the following values in the order given:

Θ = [b 1,b 2,γ1,γ2,ρ].

As before, all starting values are optional, and if you do provide the slopes, the starting value for ρ is
still optional. The internal starting values for the variance parameters are zero for both equations.
(This produces the original homoscedastic model.)

N12.2.4 Specification Tests

Wald, LM, and LR tests related to the slope parameters would follow the usual patterns
discussed in previous chapters. One might be interested in testing hypotheses about the correlation
coefficient. The Wald test for the hypothesis that ρ equals zero is part of the standard output for the
model – see the results below which include a ‘t’ statistic for this hypothesis. Likelihood ratio and
LM tests can be carried out as shown below:
The following routine will test the specification of the bivariate probit model against the null
hypothesis that two separate univariate probits apply. The test of the hypothesis that ρ equals zero is
sufficient for this. The first group of commands computes and saves the univariate probit
coefficients and log likelihoods.

NAMELIST ; x1 = ... Rhs for the first equation

; x2 = ... Rhs for the second equation $
PROBIT ; Lhs = y1 ; Rhs = x1 $
MATRIX ; b1 = b $
CALC ; l1 = logl $
PROBIT ; Lhs = y2 ; Rhs = x2 $
MATRIX ; b2 = b $
CALC ; l2 = logl $
N12: Bivariate and Multivariate Probit and Partial Observability Models N-178

To carry out the likelihood ratio test, we now fit the bivariate model, which is the unrestricted one.
The restricted model, with ρ = 0, is the two univariate models. The restricted log likelihood is the
sum of the two univariate values. The CALC command carries out the test. The BIVARIATE
command also produces a t statistic in the displayed output for the hypothesis that ρ = 0. To
automate the test, we can also use the automatically retained values rho and varrho. The second
CALC command carries out this test.

BIVARIATE ; Lhs = y1,y2 ; Rh1 = x1 ; Rh2 = x2 $

CALC ; lrtest = 2*(l1 + l2 - logl)
; pvalue = 1 - Chi(lrtest,1) $
CALC ; waldtest = rho^2 / varrho
; pvalue = 1 - Chi(waldtest,1) $

The Lagrange multiplier test is also simple to carry out using the built in procedure, as we have
already estimated the restricted model. The test is carried out with the model command that specifies
the starting values from the restricted model and restricts the maximum iterations to zero.

NAMELIST ; x1 = ... Rhs for the first equation

; x2 = ... Rhs for the second equation $
PROBIT ; Lhs = y1 ; Rhs = x1 $
MATRIX ; b1 = b $
PROBIT ; Lhs = y2 ; Rhs = x2 $
MATRIX ; b2 = b $
BIVARIATE ; Lhs = y1,y2 ; Rh1 = x1 ; Rh2 = x2
; Start = b1,b2,0 ; Maxit = 0 $

You can test the heteroscedasticity assumption by any of the three classical tests as well.
The LM test will be the simplest since it does not require estimation of the model with
heteroscedasticity. You can carry out the LM test as follows:

NAMELIST ; x1 = ... ; x2 = ... ; z1 = ... ; z2 = ... $

BIVARIATE ; Lhs = ... ; Rh1 = x1 ; Rh2 = x2 $
CALC ; h1 = Col(z1) ; h2 = Col(z2)
; k1 = Col(x1) ; k2 = Col(x2) ; k12 = k1+k2 $
MATRIX ; b1_b2 = b(1:k12) $
BIVARIATE ; Lhs = ...
; Rh1 = x1 ; Rh2 = x2 ? specify the two probit equations
; Hf1 = z1 ; Hf2 = z2 ? variables in the two variances
; Start = b1_b2, h1_0, h2_0, rho
; Maxit = 0 $

In this instance, the starting value for rho is the value that was estimated by the first model, which is
retained as a scalar value.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-179

N12.2.5 Model Results for the Bivariate Probit Model

The initial output for the bivariate probit models consists of the ordinary least squares results
if you request them with

; OLS

Final output includes the log likelihood value and the usual statistical results for the parameter
estimates.
The last output, requested with

; Summary

is a joint frequency table for four cells, with actual and predicted values shown. The predicted
outcome is the cell with the largest probability. Cell probabilities are computed using

Pi00 = 1 - Pi11 - Pi10 - Pi01 Pi01 = Φ [b 2′xi2] - Pi11

Pi10 = Φ [b 1′xi1] - Pi11 Pi11 = Φ2 [b 1′xi1, b 2′xi1, ρ]

A table which assesses the success of the model in predicting the two variables is presented as well.
An example appears below. The predictions and residuals are a bit different from the usual setup
(because this is a two equation model):

; Keep = name to retain the predicted y1

; Res = name to retain the predicted y2
; Prob = name to retain the probability for observed y1, y2 outcome
; Density = fitted bivariate normal density for observed outcome

Matrix results kept in the work areas automatically are b and varb. An extra matrix named
b_bprobt is also created. This is a two column matrix that collects the coefficients in the two
equations in a parameter matrix. The number of rows is the larger of the number of variables in x1
and x2. The coefficients are placed at the tops of the respective columns with the shorter column
padded with zeros.

NOTE: There is no correspondence between the coefficients in any particular row of b_bprobt. For
example, in the second row, the coefficient in the first column is that on the second variable in x1
and the coefficient in the second column is that on the second variable in x2. These may or may not
be the same.

The results saved by the binary choice models are:

Matrices: b = estimate of (b 1′,b 2′,ρ)′

varb = asymptotic covariance matrix

Scalars: kreg = number of parameters in model

nreg = number of observations
logl = log likelihood function
N12: Bivariate and Multivariate Probit and Partial Observability Models N-180

Variables: logl_obs = individual contribution to log likelihood

Last Model: b1_variables, b2_variables, c1_variables, c1_variables, r12

Last Function: Prob(y1 = 1,y2 = 1|x1,x2) = Φ2(b1′x1,b2′x2,r)

The saved scalars are nreg, kreg, logl, rho, varrho. The Last Model labels are b_variables and
b2_variables. If the heteroscedasticity specification is used, the additional coefficients are
c1_variables and c2_variables. To extract a vector that contains only the slopes, and not the
correlation, use

MATRIX ; {kb = kreg-1} ; b1b2 = b(1:kb) $

To extract the two parameter vectors separately, after defining the namelists, you can use

MATRIX ; {k1 = Col(x1) ; k12 = k1+1 ; kb = kreg-1}

; b1 = b(1:k1) ; b2 = b(k12:kb) $

You may use other names for the matrices. (Note that the MATRIX commands contain embedded
CALC commands contained in {}.) If the model specifies heteroscedasticity, similar constructions
can be used to extract the three or four parts of b.

N12.2.6 Partial Effects

Because it is a two equation model, it is unclear what should be an appropriate ‘marginal
effect’ in the bivariate probit model. (This is one of our frequently asked questions, as users are
often uncertain about what it is that they are looking for when they seek the ‘partial effects’ in the
model – effect of what? on what?) The literature is not necessarily helpful in this regard. The one
published result in the econometrics literature, Christofides, Stengos and Swidinsky (1997), plus an
error correction in a later issue, focuses on the joint probability of the two outcome variables
equaling one – which is not a conditional mean. The probability might be of interest. It can be
examined with the PARTIAL EFFECTS program. An example appears below. The marginal
means in the model are the univariate probabilities that the two variables equal one. These are also
not necessarily interesting, but, in any event, they can be computed using the univariate models.
NLOGIT analyzes the conditional mean function

E[y1 | y2 = 1, x1, x2] = Prob[y1 = 1,y2 = 1| x1,x2,ρ) / Prob[y2 = 1|x1].

This is the function analyzed in the bivariate probit marginal effects processor. The bivariate probit
estimator in NLOGIT allows either or both of the latent regressions to be heteroscedastic. The
reported effects for this model include the decomposition of the marginal effect into all four terms,
the regression part and the variance part, in each of the two latent models.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-181

The computations of the following marginal effects in the bivariate probit model are
included as an option with the estimator. There are two models, the base case of y1,y2 a pair of
correlated probit models, and y1|y2 = 1, the bivariate probit with sample selection. (See Section
N12.4 below.) The conditional mean computed for these two models would be identical,

E[y1|y2 = 1] = Φ2 [w1, w2 , ρ ] / Φ( w2 )

where Φ2 is the bivariate normal CDF and Φ is the univariate normal CDF. This model allows
multiplicative heteroscedasticity in either or both equations, so

w1 = b 1′x1 / exp(γ1′z1)

and likewise for w2. In the homoscedastic model, γ1 and/or γ2 is a zero vector. Four full sets of
marginal effects are reported, for x1, x2, z1, and z2. Note that the last two may be zero. The four
vectors may also have variables in common. For any variable which appears in more than one of the
parts, the marginal effect is the sum of the individual terms. A table is reported which displays these
total effects for every variable which appears in the model, along with estimated standard errors and
the usual statistical output. Formulas for the parts of these marginal effects are given below with the
technical details. For further details, see Greene (2012).
Note that you can get marginal effects for y2|y1 just by respecifying the model with y1 and y2
reversed (y2 now appears first) in the Lhs list of the command. You can also trick NLOGIT into
giving you marginal effects for y1|y2 = 0 (instead of y2 = 1) by computing z1 = 1-y1 and z2 = 1-y2, and
fitting the same bivariate probit model but with Lhs = z1,z2. You must now reverse the signs of the
marginal effects (and all slope coefficients) that are reported.
The example below was produced by a sampling experiment: Note that the model specifies
heteroscedasticity in the second equation though, in fact, there is none.

CALC ; Ran(12345) $
SAMPLE ; 1-500 $
CREATE ; u1 = Rnn(0,1) ; u2 = u1 + Rnn(0,1)
; z = Rnu(.2,.4) ; x1 = Rnn(0,1) ; x2 = Rnn(0,1)
; x3 = Rnn(0,1) ; y1 = (x1 + x2 + u1) > 0 ; y2 = (x1 + x3 + u2) > 0 $
BIVARIATE ; Lhs = y1,y2
; Rh1 = one,x1,x2 ; Rh2 = one,x1,x3
; Hf2 = z ; Partial Effects $

The first set of results is the model coefficients.

-----------------------------------------------------------------------------
FIML Estimates of Bivariate Probit Model
Dependent variable Y1Y2
Log likelihood function -416.31350
Estimation based on N = 500, K = 8
Inf.Cr.AIC = 848.627 AIC/N = 1.697
Disturbance model is multiplicative het.
Var. Parms follow 6 slope estimates.
For e(2), 1 estimates follow X3
-----------------------------------------------------------------------------
N12: Bivariate and Multivariate Probit and Partial Observability Models N-182

--------+--------------------------------------------------------------------
Y1| Standard Prob. 95% Confidence
Y2| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index equation for Y1
Constant| -.04292 .07362 -.58 .5599 -.18721 .10137
X1| 1.09235*** .08571 12.74 .0000 .92435 1.26035
X2| 1.06802*** .08946 11.94 .0000 .89268 1.24337
|Index equation for Y2
Constant| .01017 .06432 .16 .8744 -.11590 .13623
X1| .82908** .37815 2.19 .0283 .08792 1.57024
X3| .70123** .30512 2.30 .0215 .10321 1.29925
|Variance equation for Y2
Z| -.05575 1.45449 -.04 .9694 -2.90651 2.79500
|Disturbance correlation
RHO(1,2)| .66721*** .07731 8.63 .0000 .51568 .81874
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

This is the decomposition of the marginal effects for the four possible contributors to the effect.
+------------------------------------------------------+
| Partial Effects for Ey1|y2=1 |
+----------+---------------------+---------------------+
| | Regression Function | Heteroscedasticity |
| +---------------------+---------------------+
| | Direct | Indirect | Direct | Indirect |
| Variable | Efct x1 | Efct x2 | Efct h1 | Efct h2 |
+----------+----------+----------+----------+----------+
| X1 | .48383 | -.17370 | .00000 | .00000 |
| X2 | .47305 | .00000 | .00000 | .00000 |
| X3 | .00000 | -.14691 | .00000 | .00000 |
| Z | .00000 | .00000 | .00000 | .00092 |
+----------+----------+----------+----------+----------+

A table of the specific effects is produced for each contributor to the marginal effects. This first
table gives the total effects. The values here are the row total in the table above.
-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .661053
Observations used for means are All Obs.
Total effects reported = direct+indirect.
--------+--------------------------------------------------------------------
Y1| Partial Standard Prob. 95% Confidence
Y2| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X1| .31013*** .04356 7.12 .0000 .22476 .39550
X2| .47305*** .04338 10.91 .0000 .38804 .55807
X3| -.14691*** .02853 -5.15 .0000 -.20283 -.09099
Z| .00092 .02404 .04 .9694 -.04620 .04804
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N12: Bivariate and Multivariate Probit and Partial Observability Models N-183

The direct effects are the marginal effects of the variables (x1 and z1) that appear in the first equation.
-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .435447
Observations used for means are All Obs.
These are the direct marginal effects.
--------+--------------------------------------------------------------------
TAX| Partial Standard Prob. 95% Confidence
PRIV| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INC| .67814*** .24487 2.77 .0056 .19820 1.15807
PTAX| -.83030** .38146 -2.18 .0295 -1.57794 -.08266
YRS| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
-----------------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

The indirect effects are the effects of the variables that appear in the other (second) equation.

-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .661053
Observations used for means are All Obs.
These are the indirect marginal effects.
-----------------------------------------------------------------------------
--------+--------------------------------------------------------------------
Y1| Partial Standard Prob. 95% Confidence
E[y1|x,z| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X1| -.17370*** .03250 -5.34 .0000 -.23740 -.11000
X2| 0.0 .....(Fixed Parameter).....
X3| -.14691*** .02853 -5.15 .0000 -.20283 -.09099
Z| .00092 .02404 .04 .9694 -.04620 .04804
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

The marginal effects processor in the bivariate probit model detects when a regressor is a
dummy variable. In this case, the marginal effect is computed using differences, not derivatives.
The model results will contain a specific description. To illustrate this computation, we revisit the
German health care data. A description appears in Chapter E2. Here, we analyze the two health care
utilization variables, doctor = 1(docvis > 0) and hospital = 1(hospvis > 0) in a bivariate probit model.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-184

The model command is

SAMPLE ; All $
CREATE ; doctor = docvis > 0 ; hospital = hospvis > 0 $
BIVARIATE ; Lhs = doctor,hospital
; Rh1 = one,age,educ,hhninc,hhkids
; Rh2 = one,age,hhninc,hhkids
; Partial Effects $

The variable hhkids is a binary variable for whether there are children in the household. The
estimation results are as follows. This is similar to the preceding example. The final table contains
the result for the binary variable. In fact, the explicit treatment of the binary variable results in very
little change in the estimate.
-----------------------------------------------------------------------------
FIML Estimates of Bivariate Probit Model
Dependent variable DOCHOS
Log likelihood function -25552.65886
Estimation based on N = 27326, K = 10
Inf.Cr.AIC =51125.318 AIC/N = 1.871
--------+--------------------------------------------------------------------
DOCTOR| Standard Prob. 95% Confidence
HOSPITAL| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index equation for DOCTOR
Constant| .13653** .05618 2.43 .0151 .02642 .24663
AGE| .01353*** .00076 17.84 .0000 .01205 .01502
EDUC| -.02675*** .00345 -7.75 .0000 -.03352 -.01998
HHNINC| -.10245** .04541 -2.26 .0241 -.19144 -.01345
HHKIDS| -.12299*** .01670 -7.37 .0000 -.15571 -.09027
|Index equation for HOSPITAL
Constant| -1.54988*** .05325 -29.10 .0000 -1.65426 -1.44551
AGE| .00510*** .00100 5.08 .0000 .00313 .00707
HHNINC| -.05514 .05510 -1.00 .3169 -.16314 .05285
HHKIDS| -.02682 .02392 -1.12 .2622 -.07371 .02006
|Disturbance correlation
RHO(1,2)| .30251*** .01381 21.91 .0000 .27545 .32958
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

+--------------------------------+
| Partial Effects for Ey1|y2=1 |
+----------+----------+----------+
| | Direct | Indirect |
| Variable | Efct x1 | Efct x2 |
+----------+----------+----------+
| AGE | .00367 | -.00036 |
| EDUC | -.00726 | .00000 |
| HHNINC | -.02779 | .00385 |
| HHKIDS | -.03336 | .00187 |
+----------+----------+----------+
N12: Bivariate and Multivariate Probit and Partial Observability Models N-185

-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .822131
Observations used for means are All Obs.
Total effects reported = direct+indirect.
--------+--------------------------------------------------------------------
DOCTOR| Partial Standard Prob. 95% Confidence
HOSPITAL| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00332*** .00023 14.39 .0000 .00286 .00377
EDUC| -.00726*** .00096 -7.58 .0000 -.00913 -.00538
HHNINC| -.02394* .01225 -1.95 .0507 -.04796 .00008
HHKIDS| -.03149*** .00471 -6.69 .0000 -.04072 -.02226
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .822131
Observations used for means are All Obs.
These are the direct marginal effects.
--------+--------------------------------------------------------------------
DOCTOR| Partial Standard Prob. 95% Confidence
HOSPITAL| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| .00367*** .00022 16.44 .0000 .00323 .00411
EDUC| -.00726*** .00096 -7.58 .0000 -.00913 -.00538
HHNINC| -.02779** .01232 -2.25 .0241 -.05195 -.00364
HHKIDS| -.03336*** .00460 -7.26 .0000 -.04237 -.02436
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Partial derivatives of E[y1|y2=1] with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Effect shown is total of all parts above.
Estimate of E[y1|y2=1] = .822131
Observations used for means are All Obs.
These are the indirect marginal effects.
--------+--------------------------------------------------------------------
DOCTOR| Partial Standard Prob. 95% Confidence
E[y1|x,z| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AGE| -.00036*** .7075D-04 -5.03 .0000 -.00049 -.00022
EDUC| 0.0 .....(Fixed Parameter).....
HHNINC| .00385 .00385 1.00 .3167 -.00369 .01140
HHKIDS| .00187 .00167 1.12 .2620 -.00140 .00515
--------+--------------------------------------------------------------------
Note: nnnnn.D-xx or D+xx => multiply by 10 to -xx or +xx.
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------
N12: Bivariate and Multivariate Probit and Partial Observability Models N-186

+-----------------------------------------------------------+
| Analysis of dummy variables in the model. The effects are |
| computed using E[y1|y2=1,d=1] - E[y1|y2=1,d=0] where d is |
| the variable. Variances use the delta method. The effect |
| accounts for all appearances of the variable in the model.|
+-----------------------------------------------------------+
|Variable Effect Standard error t ratio |
+-----------------------------------------------------------+
HHKIDS -.031829 .004804 -6.625

N12.3 Tetrachoric Correlation

The tetrachoric correlation is a measure of the correlation between two binary variables. The
familiar Pearson, product moment correlation is inappropriate as it is used for continuous variables.
The tetrachoric correlation coefficient is equivalent to the correlation coefficient in the following
bivariate probit model:

y1* = µ + e1, y1 = 1(y1* > 0)

y2* = µ + e2, y2 = 1(y2* > 0)
(e1,e2) ~ N2[(0,0),(1,1,ρ)]

The applicable literature contains a number of approaches to estimation of this correlation

coefficient, some a bit ad hoc. We proceed directly to the implied maximum likelihood estimator.
You can fit this ‘model’ with

BIVARIATE ; Lhs = y1,y2 ; Rh1 = one ; Rh2 = one $

The reported estimate of ρ is the desired estimate. NLOGIT notices if your model does not contain
any covariates in the equation, and notes in the output that the estimator is a tetrachoric correlation.
The results below based on the German health care data show an example.

-----------------------------------------------------------------------------
FIML Estimation of Tetrachoric Correlation
Dependent variable DOCHOS
Log likelihood function -25898.27183
Estimation based on N = 27326, K = 3
Inf.Cr.AIC =51802.544 AIC/N = 1.896
--------+--------------------------------------------------------------------
DOCTOR| Standard Prob. 95% Confidence
HOSPITAL| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Estimated alpha for P[DOCTOR =1] = F(alpha)
Constant| .32949*** .00773 42.61 .0000 .31433 .34465
|Estimated alpha for P[HOSPITAL=1] = F(alpha)
Constant| -1.35540*** .01074 -126.15 .0000 -1.37646 -1.33434
|Tetrachoric Correlation between DOCTOR and HOSPITAL
RHO(1,2)| .31106*** .01357 22.92 .0000 .28446 .33766
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N12: Bivariate and Multivariate Probit and Partial Observability Models N-187

The preceding suggests an interpretation for the bivariate probit model; the correlation coefficient
reported is the conditional (on the independent variables) tetrachoric correlation.
The computation in the preceding can be generalized to a set of M binary variables, y1,...,yM.
The tetrachoric correlation matrix would be the M×M matrix, R, whose off diagonal elements are the
ρmn coefficients described immediately above. There are several ways to do this computation, again,
as suggested by a literature that contains recipes. Once again, the maximum likelihood estimator
turns out to be a useful device.
A direct approach would involve expanding the latent model to

y1* = µ + e1, y1 = 1(y1* > 0)

y2* = µ + e2, y2 = 1(y2* > 0)
...
yM* = µ + eM, yM = 1(yM* > 0)
(e1,e2,...,eM) ~ NM[0,R]

The appropriate estimator would be NLOGIT’s multivariate probit estimator, MPROBIT, which can
handle up to M = 20. The correlation matrix produced by this procedure is precisely the full
information MLE of the tetrachoric correlation matrix. However, for any M larger than two, this
requires use of the GHK simulator to maximize the simulated log likelihood, and is extremely slow.
The received estimators of this model estimate the correlations pairwise, as shown earlier. For this
purpose, the FIML estimator is unnecessary. The matrix can be obtained using bivariate probit
estimates. The following procedure would be useable:

NAMELIST ; y = y1,y2,...,ym $
CALC ; m = Col(y) $
MATRIX ; r = Iden(m) $
PROCEDURE $
DO FOR ; 20 ; i = 2,m $
CALC ; i1 = i - 1 $
DO FOR ; 10 ; j = 1,i1 $
BIVARIATE ; Lhs = y:i, y:j ; Rh1 = one ; Rh2 = one $
MATRIX ; r(i,j) = rho $
MATRIX ; r(j,i) = rho $
ENDDO ; 10 $
ENDDO ; 20 $
ENDPROCEDURE $
EXECUTE ; Quiet $

A final note, the preceding approach is not fully efficient. Each bivariate probit estimates (µm,µn)
which means that µm is estimated more than once when m > 1. A minimum distance estimator could
be used to reconcile these after all the bivariate probit estimates are computed. But, since the means
are nuisance parameters in this model, this seems unlikely to prove worth the effort.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-188

N12.4 Bivariate Probit Model with Sample Selection

In the bivariate probit setting, data on y1 might be observed only when y2 equals one. For
example, in modeling loan defaults with a sample of applicants, default will only occur among
applicants who are granted loans. Thus, in a bivariate probit model for the two outcomes, the
observed default data are nonrandomly selected from the set of applicants. The model is

zi1 = b′xi1 + ei1, yi1 = sgn(zi1),

zi2 = b′xi2 + ei2, yi2 = sgn(zi2),
ei1,ei2 ~ BVN(0,0,1,1,ρ),
(yi1,xi1) is observed only when yi2 = 1.

This is a type of sample selectivity model. The estimator was proposed by Wynand and van Praag
(1981). An extensive application which uses choice based sampling as well is Boyes, Hoffman, and
Low (1989). (See also Greene (1992 and 2011).) The sample selection model is obtained by adding

; Selection (or just ; Sel)

to the BIVARIATE PROBIT command. All other options and specifications are the same as
before. Except for the diagnostic table which indicates that this model has been chosen, the results
for the selection model are the same as for the basic model.

N12.5 Simultaneity in the Binary Variables

A simultaneous equations sort of model would appear as

zi1 = b 1′xi1 + γ1yi2 + ei1, yi1 = 1 if zi1 > 0, yi1 = 0 otherwise,

zi2 = b 2′xi2 + γ1yi1 + ei2, yi2 = 1 if zi2 > 0, yi2 = 0 otherwise,
[ei1,ei2] ~ bivariate normal (BVN) [0,0,1,1,ρ], -1 < ρ < 1,
individual observations on y1 and y2 are available for all i.

It would follow from the construction that

Prob[y1 = 1, y2 = 1] = Φ2 (b 1′x1+ γ1y2, b 2′x2+ γ2y1, ρ]

and likewise for the other cells, where y1 and y2 are two binary variables. Unfortunately, the model as
stated is not internally consistent, and is inestimable. Ultimately, it is not identifiable. As a practical
matter, you can verify this by attempting to devise a way to simulate a sample of observations that
conforms exactly to the assumptions of the model. In this case, there is none because there is no linear
reduced form for this model. (The approach suggested by Maddala (1983) is not consistent.) NLOGIT
will detect this condition and decline to attempt to do the estimation. For example:

BIVARIATE PROBIT ; Lhs = y1,y2 ; Rh1 = one,x1,x3,y2 ; Rh2 = one,x2,x3,y1 $

produces a diagnostic,
Error 809: Fully simultaneous BVP model is not identified
N12: Bivariate and Multivariate Probit and Partial Observability Models N-189

NOTE: Unlike the case in linear simultaneous equations models, nonidentifiability does not prevent
‘estimation’ in this model. (2SLS estimates cannot be computed when there are too few instrumental
variables, which is the signature of nonidentifiability in a linear context.) With the ‘fully
simultaneous bivariate probit model,’ it is possible to maximize what purports to be a log likelihood
function – numbers will be produced that might even look reasonable. However, as noted, the model
itself is nonsensical – it lacks internal coherency.

N12.6 Recursive Bivariate Probit Model

A slight modification of the model in the previous section is identified and used in many
recent applications. Consider the model for the probability of the event y1 = 0/1 and y2 = 0/1
assuming γ2 = 0.
Prob[y1 = 1, y2 = 1 | x1 , x2 ] = Φ2 (b 1′x1 + γ1, b 2′x2, ρ)
Prob[y1 = 1, y2 = 0 | x1 , x2 ] = Φ2 (b 1′x1, -b 2′x2, -ρ)
Prob[y1 = 0, y2 = 1 | x1 , x2 ] = Φ2 (-b 1′x1 + γ1, b 2′x2, -ρ)
Prob[y1 = 0, y2 = 0 | x1 , x2 ] = Φ2 (-b 1′x1, -b 2′x2, ρ)

This is a recursive simultaneous equations model. Surprisingly enough, it can be estimated by full
information maximum likelihood ignoring the simultaneity in the system;

BIVARIATE ; Lhs = y1, y2

; Rh1 = x1,y2 ; Rh2 = x2 $

(A proof of this result is suggested in Maddala (1983, p. 123) and pursued in Greene (1998).) An
application of the result to the gender economics study is given in Greene (1998). Some extensions
are presented in Greene (2003, 2011).
This model presents the same ambiguity in the conditional mean function and marginal
effects that were noted earlier in the bivariate probit model. The conditional mean for y1 is

E[y1 | y2 = 1, x1, x2] = Φ2 (b1′x1 + γ1, b 2′x2, ρ) / Φ(b 2′x2)

for which derivatives were given earlier. Given the form of this result, we can identify direct and
indirect effects in the conditional mean:

∂E[ y1 | y 2 =1, x1 , x 2 ] g1
= β1 = direct effects
∂x1 Φ (β ' x 2 )

= ∂E[ y1 | y2 1, x1 , x 2 ]  g 2 ′x1 , ′x 2 , ρ)φ( z2 ) 

Φ (ββ
=  − 2  β 2 = indirect effects
∂x 2  Φ (β′x 2 ) [Φ(β′x 2 )]
2

The unconditional mean function is

E[y1 | x1, x2] = Φ(b 2′x2) E[y1 | y2 = 1, x1, x2] + [1-Φ(b 2′x2)] E[y1 | y2 = 0, x1, x2]
= Φ2 (b 1′x1 + γ1, b 2′x2, ρ) + Φ2 (b 1′x1, -b 2′x2, -ρ)
N12: Bivariate and Multivariate Probit and Partial Observability Models N-190

Derivatives for marginal effects can be derived using the results given earlier. Analysis appears in
Greene (1998). The decomposition is done automatically when you specify a recursive bivariate
probit model – one in which the second Lhs variable appears in the Rhs of the first equation.
The following demonstrates this by extending the model. Note the appearance of priv on the
Rhs of the first equation, x1.

NAMELIST ; y = tax, priv

; x1 = one,inc,ptax,priv ; x2 = one,inc,yrs,ptax $
BIVARIATE ; Lhs = tax,priv ; Rh1 = x1 ; Rh2 = x2 ; Partial Effects $

-----------------------------------------------------------------------------
FIML - Recursive Bivariate Probit Model
Dependent variable PRITAX
Log likelihood function -74.21179
Estimation based on N = 80, K = 9
Inf.Cr.AIC = 166.424 AIC/N = 2.080
--------+--------------------------------------------------------------------
PRIV| Standard Prob. 95% Confidence
TAX| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index equation for PRIV
Constant| -2.81454 5.51612 -.51 .6099 -13.62594 7.99687
INC| .16264 .76312 .21 .8312 -1.33304 1.65832
YRS| -.03484 .04247 -.82 .4120 -.11808 .04840
PTAX| .04605 .98275 .05 .9626 -1.88011 1.97220
|Index equation for TAX
Constant| -.68059 4.05341 -.17 .8667 -8.62513 7.26394
INC| 1.22768 .81424 1.51 .1316 -.36820 2.82356
PTAX| -1.63160 .99598 -1.64 .1014 -3.58368 .32047
PRIV| .98178 .95912 1.02 .3060 -.89807 2.86162
|Disturbance correlation
RHO(1,2)| -.83119 .57072 -1.46 .1453 -1.94977 .28740
--------+--------------------------------------------------------------------

---------------------------------------------------------------
Decomposition of Partial Effects for Recursive Bivariate Probit
Model is PRIV = F(x1b1), TAX = F(x2b2+c*PRIV )
Conditional mean function is E[TAX |x1,x2] =
Phi2(x1b1,x2b2+gamma,rho) + Phi2(-x1b1,x2b2,-rho)
Partial effects for continuous variables are derivatives.
Partial effects for dummy variables (*) are first differences.
Direct effect is wrt x2, indirect is wrt x1, total is the sum.
---------------------------------------------------------------
Variable Direct Effect Indirect Effect Total Effect
---------+---------------+-----------------+-------------------
INC | .4787001 .0169062 .4956064
PTAX | -.6362002 .0047864 -.6314138
YRS | .0000000 -.0036217 -.0036217
---------+-----------------------------------------------------

The decomposition of the partial effects accounts for the direct and indirect influences. Note that
there is no partial effect given for priv because this variable is endogenous. It does not vary
‘partially.’
N12: Bivariate and Multivariate Probit and Partial Observability Models N-191

N12.7 Panel Data Bivariate Probit Models

The four bivariate probit models, bivariate probit, bivariate probit with selection, Poirier’s
partial observability and Abowd’s partial observability model have all been extended to the random
parameters form of the panel data models. (The fixed effects and latent class models are not
available.) Use of the random parameters formulation is described in detail in Chapter R24. We will
only sketch the extension here. The commands for the models are as follows, where [ ... ] indicates
an optional part of the specification:

BIVARIATE ; Lhs = y1, y2 ? Bivariate probit

; Rh1 = Rhs for equation 1
; Rh2 = Rhs for equation 2
[ ; Selection ] ? Partial observability

or PROBIT ; Lhs = y ? Probit model

; Rh1 = Rhs for equation 1
; Rh2 = Rhs for equation 2 ? Partial observability (Poirier)
[ ; Selection ] ? Abowd and Farber

Then, ; RPM [ = list for heterogeneity in the mean ]

; Pds = panel specification ? Optional if cross section
[ ; Pts = number of replications ]
[ ; Halton and other controls for the estimation ]
; Fcn = designation of random parameters $

For the random parameters specification, use

; name ( distribution ) distribution = n, u, t, l, c for the first equation

or ; name [ distribution ] for the second equation.

Note that random parameters in the second equation are designated by square brackets rather than
parentheses. This is necessary because the same variables can appear in both equations. Two other
specifications should be useful

; Cor allows the random parameters to be correlated.

; AR1 allows the random terms to evolve according to an AR(1) process
rather than be time invariant.

The two equation random parameters save the matrices b and varb and the scalar logl after
estimation. No other variables, partial effects, etc. are provided internally to the command. But, you
can use the estimation results directly in the SIMULATION, PARTIAL EFFECTS commands,
and so on. An example appears after the results of the simulation below.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-192

Application

To demonstrate this model, we will fit a true random effects model for a bivariate probit
outcome. Each equation has its own random effect, and the two are correlated. The model structure is

zit1 = b 1′xit1 + eit1 + ui1, yit1 = 1 if zit1 > 0, yit1 = 0 otherwise,

zit2 = b 2′xit2 + eit2 + ui2, yit2 = 1 if zit2 > 0, yit2 = 0 otherwise,
[eit1,eit2] ~ Bivariate normal (BVN) [0,0,1,1,ρ], -1 < ρ < 1,
[ui1,ui2] ~ Bivariate normal (BVN) [0,0,1,1,q], -1 < q < 1.

Individual observations on y1 and y2 are available for all i. Note, in the structure, the idiosyncratic eitj
creates the bivariate probit model, whereas the time invariant common effects, uij create the random
effects (random constants) model. Thus, there are two sources of correlation across the equations, the
correlation between the unique disturbances, ρ, and the correlation between the time invariant
disturbances, q. The data are generated artificially according to the assumptions of the model.

CALC ; Ran(12345) $
SAMPLE ; 1-200 $
CREATE ; x1 = Rnn(0,1) ; x2 = Rnn(0,1) ; x3 = Rnn(0,1) $
MATRIX ; u1i = Rndm(20) ; u2i = .5* Rndm(20) + .5* u1i $
CREATE ; i = Trn(10,0) ; u1 = u1i(i) ; u2 = u2i(i) $
CREATE ; e1 = Rnn(0,1) ; e2 = .7*Rnn(0,1) + .3*e1 $
CREATE ; y1 = (x1+e1 + u1) > 0
; y2 = (x2+x3+e2+u2) > 0 ; y12 = y1*y2 $
BIVARIATE ; Lhs = y1,y2 ; Rh1 = one,x1 ; Rh2 = one,x2,x3
; RPM ; Pds = 10 ; Pts = 25 ; Cor ; Halton
; Fcn = one(n), one[n] $
PROBIT ; Lhs = y12 ; Rh1 = one,x1 ; Rh2 = one,x2,x3
; RPM ; Pds = 10 ; Pts = 25 ; Cor ; Halton
; Fcn = one(n), one[n] ; Selection $
PROBIT ; Lhs = y12 ; Rh1 = one,x1 ; Rh2 = one,x2,x3
; RPM ; Pds = 10 ; Pts = 25 ; Cor ; Halton
; Fcn = one(n), one[n] $

Note that by construction, most of the cross equation correlation comes from the random effects, not
the disturbances. The second model is the Abowd/Farber version of the partial observability model.
The Poirier model is not estimable for this setup. It is easy to see why. The correlations in the Poirier
model are overspecified. Indeed, with ; Cor for the random effects, the Poirier model specifies two
separate sources of cross equation correlation. This is a weakly identified model. The implication can
be seen in the results below, where the estimator failed to converge for the probit model, and at the exit,
the estimate of ρ was nearly -1.0. This is the signature of a weakly identified (or unidentified) model.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-193

These are the estimates of the Meng and Schmidt model.

-----------------------------------------------------------------------------
Probit Regression Start Values for Y1
Dependent variable Y1
Log likelihood function -114.32973
--------+--------------------------------------------------------------------
Y1| Standard Prob. 95% Confidence
Y2| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X1| .65214*** .10287 6.34 .0000 .45052 .85375
Constant| -.12214 .09617 -1.27 .2041 -.31062 .06634
--------+--------------------------------------------------------------------
Probit Regression Start Values for Y2
Dependent variable Y2
Log likelihood function -83.99189
--------+--------------------------------------------------------------------
Y1| Standard Prob. 95% Confidence
Y2| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X2| .96584*** .14838 6.51 .0000 .67503 1.25665
X3| 1.00421*** .14562 6.90 .0000 .71880 1.28961
Constant| .17104 .11176 1.53 .1259 -.04801 .39009
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Random Coefficients BivProbt Model
Dependent variable Y1
Log likelihood function -163.43468
Estimation based on N = 200, K = 9
Inf.Cr.AIC = 344.869 AIC/N = 1.724
Sample is 10 pds and 20 individuals
Bivariate Probit model
Simulation based on 25 Halton draws
--------+--------------------------------------------------------------------
Y1| Standard Prob. 95% Confidence
Y2| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
X1_1| 1.08374*** .19408 5.58 .0000 .70335 1.46412
X2_2| 1.18264*** .22213 5.32 .0000 .74727 1.61800
X3_2| 1.18893*** .18946 6.28 .0000 .81758 1.56027
|Means for random parameters
ONE_1| -.05021 .12427 -.40 .6862 -.29377 .19335
ONE_2| .27827* .15481 1.80 .0723 -.02514 .58169
|Diagonal elements of Cholesky matrix
ONE_1| 1.08131*** .17778 6.08 .0000 .73288 1.42975
ONE_2| .42491*** .15811 2.69 .0072 .11503 .73480
|Below diagonal elements of Cholesky matrix
lONE_ONE| -.45867** .17845 -2.57 .0102 -.80842 -.10892
|Unconditional cross equation correlation
lONE_ONE| -.17471 .17798 -.98 .3263 -.52355 .17413
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N12: Bivariate and Multivariate Probit and Partial Observability Models N-194

Implied covariance matrix of random parameters

Var_Beta| 1 2
--------+----------------------------
1| 1.16924 -.495965
2| -.495965 .390927

Implied standard deviations of random parameters

S.D_Beta| 1
--------+--------------
1| 1.08131
2| .625242

Implied correlation matrix of random parameters

Cor_Beta| 1 2
--------+----------------------------
1| 1.00000 -.733586
2| -.733586 1.00000

These are the estimates of the Abowd and Farber model.

-----------------------------------------------------------------------------
Probit Regression Start Values for Y12
Dependent variable Y12
Log likelihood function -103.81770
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X1| .52842*** .10360 5.10 .0000 .32537 .73147
Constant| -.66498*** .10303 -6.45 .0000 -.86692 -.46304
--------+--------------------------------------------------------------------
Probit Regression Start Values for Y12
Dependent variable Y12
Log likelihood function -102.69669
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X2| .50336*** .11606 4.34 .0000 .27588 .73084
X3| .38430*** .11126 3.45 .0006 .16622 .60237
Constant| -.64606*** .10368 -6.23 .0000 -.84927 -.44286
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Random Coefficients PrshlObs Model
Dependent variable Y12
Log likelihood function -72.83435
Restricted log likelihood -102.69669
Chi squared [ 3 d.f.] 59.72467
Significance level .00000
McFadden Pseudo R-squared .2907819
Estimation based on N = 200, K = 8
Inf.Cr.AIC = 161.669 AIC/N = .808
Sample is 10 pds and 20 individuals
Partial observability probit model
Simulation based on 25 Halton draws
N12: Bivariate and Multivariate Probit and Partial Observability Models N-195

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
X1_1| 1.09511*** .23019 4.76 .0000 .64394 1.54629
X2_2| 2.26279*** .79573 2.84 .0045 .70319 3.82239
X3_2| 1.90015*** .70892 2.68 .0074 .51070 3.28960
|Means for random parameters
ONE_1| .09219 .22240 .41 .6785 -.34370 .52809
ONE_2| -.06872 .36077 -.19 .8489 -.77581 .63837
|Diagonal elements of Cholesky matrix
ONE_1| .59436** .23215 2.56 .0105 .13935 1.04937
ONE_2| 1.98257*** .73799 2.69 .0072 .53614 3.42900
|Below diagonal elements of Cholesky matrix
lONE_ONE| -.91612** .41168 -2.23 .0261 -1.72299 -.10925
|Unconditional cross equation correlation
lONE_ONE| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

Implied covariance matrix of random parameters

Var_Beta| 1 2
--------+----------------------------
1| .353265 -.544507
2| -.544507 4.76987

Implied standard deviations of random parameters

S.D_Beta| 1
--------+--------------
1| .594361
2| 2.18400

Implied correlation matrix of random parameters

Cor_Beta| 1 2
--------+----------------------------
1| 1.00000 -.419469
2| -.419469 1.00000

These are the estimates of the Poirier model.

-----------------------------------------------------------------------------
Probit Regression Start Values for Y12
Dependent variable Y12
Log likelihood function -103.81770
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X1| .52842*** .10360 5.10 .0000 .32537 .73147
Constant| -.66498*** .10303 -6.45 .0000 -.86692 -.46304
-----------------------------------------------------------------------------
Probit Regression Start Values for Y12
Dependent variable Y12
Log likelihood function -102.69669
N12: Bivariate and Multivariate Probit and Partial Observability Models N-196

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
X2| .50336*** .11606 4.34 .0000 .27588 .73084
X3| .38430*** .11126 3.45 .0006 .16622 .60237
Constant| -.64606*** .10368 -6.23 .0000 -.84927 -.44286
--------+--------------------------------------------------------------------

-----------------------------------------------------------------------------
Random Coefficients PrshlObs Model
Dependent variable Y12
Log likelihood function -70.16147
Sample is 10 pds and 20 individuals
Partial observability probit model
Simulation based on 25 Halton draws
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
Y12| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
X1_1| .95923*** .21311 4.50 .0000 .54154 1.37692
X2_2| 1.02185*** .28212 3.62 .0003 .46890 1.57480
X3_2| .77643*** .23096 3.36 .0008 .32376 1.22910
|Means for random parameters
ONE_1| .41477 .32108 1.29 .1964 -.21454 1.04407
ONE_2| .08625 .31520 .27 .7844 -.53153 .70402
|Diagonal elements of Cholesky matrix
ONE_1| .42395 .28240 1.50 .1333 -.12955 .97744
ONE_2| .98957*** .29127 3.40 .0007 .41869 1.56044
|Below diagonal elements of Cholesky matrix
lONE_ONE| -.62399** .31020 -2.01 .0443 -1.23197 -.01601
|Unconditional cross equation correlation
lONE_ONE| -.99693*** .01079 -92.41 .0000 -1.01808 -.97579
--------+--------------------------------------------------------------------

Implied covariance matrix of random parameters

Var_Beta| 1 2
--------+----------------------------
1| .179731 -.264539
2| -.264539 1.36861
Implied standard deviations of random parameters
S.D_Beta| 1
--------+--------------
1| .423947
2| 1.16988
Implied correlation matrix of random parameters
Cor_Beta| 1 2
--------+----------------------------
1| 1.00000 -.533382
2| -.533382 1.00000
N12: Bivariate and Multivariate Probit and Partial Observability Models N-197

N12.8 Simulation and Partial Effects

This is the model estimated at the beginning of the previous section.

y1* = a1 + b11 x1 + u1 + e1
y2* = a2 + b22 x2 + b23 x3 + u2 + e2.

The random effects, u1 and u2, are time invariant – the same value appears in each of the 10 periods
of the data. The model command is

BIVARIATE ; Lhs = y1,y2

; Rh1 = one,x1 ; Rh2 = one,x2,x3
; RPM ; Pds = 10 ; Pts = 25 ; Cor ; Halton
; Fcn = one(n), one[n] $

-----------------------------------------------------------------------------
Random Coefficients BivProbt Model
Bivariate Probit model
Simulation based on 25 Halton draws
--------+--------------------------------------------------------------------
Y1| Standard Prob. 95% Confidence
Y2| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
X1_1| 1.08374*** .19408 5.58 .0000 .70335 1.46412
X2_2| 1.18264*** .22213 5.32 .0000 .74727 1.61800
X3_2| 1.18893*** .18946 6.28 .0000 .81758 1.56027
|Means for random parameters
ONE_1| -.05021 .12427 -.40 .6862 -.29377 .19335
ONE_2| .27827* .15481 1.80 .0723 -.02514 .58169
|Diagonal elements of Cholesky matrix
ONE_1| 1.08131*** .17778 6.08 .0000 .73288 1.42975
ONE_2| .42491*** .15811 2.69 .0072 .11503 .73480
|Below diagonal elements of Cholesky matrix
lONE_ONE| -.45867** .17845 -2.57 .0102 -.80842 -.10892
|Unconditional cross equation correlation
lONE_ONE| -.17471 .17798 -.98 .3263 -.52355 .17413
--------+--------------------------------------------------------------------

Figure N12.1 Matrix Results

N12: Bivariate and Multivariate Probit and Partial Observability Models N-198

The estimator does not support predictions or partial effects. But, we can use the template
SIMULATE and PARTIAL EFFECTS programs to create our own by supplying our function and
estimates.. We will use the model exactly as shown in the results, with labels for the estimates in
order of their appearance: b11,b22,b23,a1,a2,c11,c22,c21,ro. For purposes of the exercise, we will
examine the bivariate normal probability P(y1=1,y2=1). With all the parts in place, other functions,
such as the conditional means, can be examined by making minor changes in the function definition.
For example, in the program below, partial effects are obtained simply by changing the command to
PARTIALS and changing ; Scenario: to ; Effects: x1.

? Create time invariant random effects. Used to create correlated u1 and u2

MATRIX ; mv1 = Rndm(20,1) ; mv2 = Rndm(20,1) $
CREATE ; index = Trn(10,0) $
CREATE ; v1 = mv1(index) ; v2 = mv2(index) $
? Simulate the joint probability and examine its behavior as x1 varies
SIMULATE ; Labels = b11,b22,b23,a1,a2,c11,c22,c21,ro
; Parameters = b ; Covariance = varb
; Function = xb1 = a1+b11*x1+c11*v1 |
xb2 = a2+b22*x2+b23*x3+c21*v1+c22*v2 |
Bvn(xb1,xb2,ro)
; Scenario: & x1 = -3(.2)3 ; Plot $
---------------------------------------------------------------------
Model Simulation Analysis for User Specified Function
---------------------------------------------------------------------
Simulations are computed by average over sample observations
---------------------------------------------------------------------
User Function Function Standard
(Delta method) Value Error |t| 95% Confidence Interval
---------------------------------------------------------------------
Avrg. Function .23829 .02576 9.25 .18780 .28878
X1 = -3.00 .00645 .00464 1.39 -.00266 .01555
X1 = -2.80 .00870 .00567 1.54 -.00240 .01981
(rows omitted)
X1 = 2.80 .51118 .03121 16.38 .45001 .57235
X1 = 3.00 .51513 .03049 16.90 .45538 .57488

Figure N12.2 Simulation of Estimated Model

N12: Bivariate and Multivariate Probit and Partial Observability Models N-199

N12.9 Multivariate Probit Model

The multivariate probit model is the extension to M equations of the bivariate probit model

yim* = b m′xim+ eim, m = 1,…,M

yim = 1 if yim* > 0, and 0 otherwise.
eim, m = 1,...,M ~ MVN [0,R]

where R is the correlation matrix. Each individual equation is a standard probit model. This
generalizes the bivariate probit model for up to M = 20 equations. Specify the model with the same
command structure as the SURE model, using the command MPROBIT,

MPROBIT ; Lhs = y1,y2,...,ym (list of up to 20 variables)

; Eq1 = list of Rhs variables in the first equation
; Eq2 = list of Rhs variables in the second equation
...
; EqM = list of Rhs variables for Mth equation $

The data for this model must be individual, not proportions and not frequencies. You may use

; Wts = name

as usual. Other options specific for this model in addition to the standard output options are

; Prob = name

which requests the estimator to save the predicted probability for the observed joint outcome, and

; Utility = name

where ‘name’ is an existing namelist to save the estimated utilities, Xmb m. Restrictions can be
imposed with
; Rst = list
and ; CML: specification for constraints

Note that either of these can be used to specify the correlation matrix. The list for ; Rst includes the
M(M-1)/2 below diagonal elements of R. You can use this to force correlations to equal each other,
or zero, or other values.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-200

N12.9.1 Retrievable Results

This model keeps the following retrievable results:

Matrices: b = estimate of (b 1′,b 2′,…,bM′ )′ = vector of slopes only

varb = asymptotic covariance matrix
omega = M×M correlation matrix of disturbances

Scalars: kreg = number of parameters in model

nreg = number of observations
logl = log likelihood function

Variables: logl_obs = individual contribution to log likelihood

Last Model: None

Last Function: None

N12.9.2 Partial Effects

You can obtain marginal effects for this model of the following form: The expected value of
y1 given that all other ys equal one is

E[y1|y2=1,...,yM=1] = Prob(y1=1,...,yM=1)/Prob(y2=1,...,yM=1) = P1...M / P2...M = E1.

The derivatives of this function are constructed as follows: Let x equal the union of all of the
regressors that appear in the model, and let γm be such that zm = x’γm = b m′xm. (γm will usually have
some zeros in it unless all regressors appear in all equations.) Then,

∂E1  1 ∂P1...M   1 ∂P2...M 

∑m=1 ∑m = 2
M M
=   γ m - E1  P  γ m
∂x  P2...M ∂z m   2...M ∂z m 

The relevant parts of this combination of the coefficient vectors are then extracted and reported for
the specific equations. Standard errors are obtained using the delta method, and all derivatives are
approximated numerically. All effects are computed at the means of the Rhs variables. Use

; Partial Effects

to request this computation. In the display of these results, derivatives with respect to the constant
term are set to zero.
N12: Bivariate and Multivariate Probit and Partial Observability Models N-201

Standard errors for these marginal effects cannot be computed directly. We report a
bootstrapped approximation computed as follows: Let the estimated set of marginal effects be
denoted d. This is computed using the parameter estimates from the model as given earlier. Let V
denote the estimated asymptotic covariance matrix for the coefficient estimates. An estimate of the
variance of the estimator of the marginal effects is obtained as the mean squared deviation of 50
random draws from the distribution of the underlying slope parameters. You can set the number of
bootstrap replications to use with

; Nbt = number of replications.

The draws are based on the asymptotic normal distribution with mean b and variance V. (The
estimated correlation parameters are taken as fixed.) Thus, the marginal effects at the data means are
computed 50 additional times with these new parameters, using

∑r =1 (d jr − d j )2
1 50
Est.Var[d j ] =
50

Note that the sums are centered at the original estimated marginal effect, not at the means of the
random draws.

N12.9.3 Sample Selection Model

There are two modifications of the multivariate probit model built into the estimator. The
first is a multivariate version of the selection model in Section N12.4. The model structure is

yi1* = b 1′xi1 + ei1

yi2* = b 2′xi2 + ei2
…
yi,M-1* = b M-1′xi,M-1 + eI,M-1
yiM* = b M′xiM + eiM
yim = 1 if yim* > 0, and 0 otherwise
eim, m = 1,...,M ~ MVN [0,R]
yi,1,yi,2,…,yi,M-1 only observed when yiM = 1

In the same fashion as earlier, the log likelihood is built up from the laws of probability. The
different terms in the likelihood function are

Prob(yiM = 1|xim)

for the nonselected case, then

Prob(Yi1 = yi1,…,Yi,M-1 = yi,M-1 , yiM = 1|xi1,…,xiM).

N12: Bivariate and Multivariate Probit and Partial Observability Models N-202

The last equation is the selection mechanism. This produces a difference in the likelihood that is
maximized (and, to some degree, in the interpretation of the model), but no essential difference in the
estimation results.
This form of the model is requested by adding

; Selection

to the MVPROBIT command. There are no other changes in the model specification, or the data.
Missing data may be coded as zeros or as missing.
N13: Ordered Choice Models N-203

N13: Ordered Choice Models

N13.1 Introduction
The basic ordered choice model is based on the latent regression,

yi* = b′xi + ei, ei ~ F(ei |q), E[ei|xi] = 0, Var[ei|xi] = 1,

The observation mechanism results from a complete censoring of the latent dependent variable as
follows:
yi = 0 if yi ≤ µ0,
= 1 if µ0 < yi ≤ µ1,
= 2 if µ1 < yi ≤ µ2,
...
= J if yi > µJ-1.

The latent ‘preference’ variable, yi* is not observed. The observed counterpart to yi* is yi. Five
stochastic specifications are provided for the basic model shown above. The ordered probit model
based on the normal distribution was developed by Zavoina and McElvey (1975). It applies in
applications such as surveys, in which the respondent expresses a preference with the above sort of
ordinal ranking. The variance of ei is assumed to be one, since as long as yi*, b, and ei are
unobserved, no scaling of the underlying model can be deduced from the observed data. (The
assumption of homoscedasticity is arguably a strong one. We will relax that assumption in Section
N14.2.) Since the µs are free parameters, there is no significance to the unit distance between the set
of observed values of y. They merely provide the coding. Estimates are obtained by maximum
likelihood. The probabilities which enter the log likelihood function are

Prob[yi = j] = Prob[yi* is in the jth range].

The model may be estimated either with individual data, with yi = 0, 1, 2, ... or with grouped data, in
which case each observation consists of a full set of J+1 proportions, p0i,...,pJi.

NOTE: If your data are not coded correctly, this estimator will abort with one of several possible
diagnostics – see below for discussion. Your dependent variable must be coded 0,1,...,J. We note
that this differs from some other econometric packages which use a different coding convention.

There are numerous variants and extensions of this model which can be estimated. The
underlying mathematical forms are shown below, where the CDF is denoted F(z) and the density is
f(z). (Familiar synonyms are given as well.) (See, as well, Chapters E34-E36.) The functional
forms of the two models considered here are
N13: Ordered Choice Models N-204

Probit

z exp(−t 2 / 2)
F(z) = ∫−∞
2p
dt = Φ(z), f(z) = φ(z),

Logit

exp( z )
F(z) = = Λ(z), f(z) = Λ(z)[1 - Λ(z)].
1 + exp( z )

The ordered probit model is an extension of the probit model for a binary outcome with normally
distributed disturbances. The ordered logit model results from the assumption that e has a standard
logistic distribution instead of a standard normal. A variety of additional specifications and extensions
are provided. Basic models are treated in this chapter. Extensions such as censoring and sample
selection are given in Chapter N14. Panel data models for ordered choice are discussed in Chapter
N15.

N13.2 Command for Ordered Probability Models

The essential command for estimating ordered probability models is

ORDERED ; Lhs = y or p0,p1,...pJ ; Rhs = regressors $

Note that the estimator accepts proportions data for a set of J proportions. The proportions would
sum to one at each observation. The probit model is the default specification. To estimate an
ordered logit model, add

; Model = Logit

to the command or change the verb to OLOGIT. The standardized logistic distribution (mean zero,
standard deviation approximately 1.81) is used as the basis of the model instead of the standard
normal.
This model must include a constant term, one, as the first Rhs variable. Since the equation
does include a constant term, one of the µs is not identified. We normalize µ0 to zero. (Consider the
special case of the binary probit model with something other than zero as its threshold value. If it
contains a constant, this cannot be estimated.) Data may be grouped or individual. (Survey data
might logically come in grouped form.) If you provide individual data, the dependent variable is
coded 0, 1, 2, ..., J. There must be at least three values. Otherwise, the binary probit model applies.
If the data are grouped, a full set of proportions, p0, p1, ..., pJ, which sum to one at every observation
must be provided. In the individual data case, the data are examined to determine the value of J,
which will be the largest observed value of y which appears in the sample. In the grouped data case,
J is one less than the number of Lhs variables you provide. Once again, we note that other programs
sometimes use different normalizations of the model. For example, if the constant term is forced to
equal zero, then one will instead, add a nonzero threshold parameter, µ0, which equals zero in the
presence of a nonzero constant term.
N13: Ordered Choice Models N-205

N13.3 Data Problems

If you are using individual data, the Lhs variable must be coded 0,1,...,J. All the values must
be present in the data. NLOGIT will look for empty cells. If there are any, estimation is halted. (If
value ‘j’ is not represented in the data, then the threshold parameter, µj is not estimable.) In this
circumstance, you will receive a diagnostic such as

ORDE,Panel,BIVA PROBIT:A cell has (almost) no observations.

Empty cell: Y never takes value 2

This diagnostic means exactly what it says. The ordered probability model cannot be estimated
unless all cells are represented in the data. Users frequently overlook the coding requirement,
y = 0,1,... If you have a dependent variable that is coded 1,2,..., you will see the following
diagnostic:

Models - Insufficient variation in dependent variable.

The reason this particular diagnostic shows up is that NLOGIT creates a new variable from your
dependent variable, say y, which equals zero when y equals zero, and one when y is greater than
zero. It then tries to obtain starting values for the model by fitting a regression model to this new
variable. If you have miscoded the Lhs variable, the transformed variable always equals one, which
explains the diagnostic. In fact, there is no variation in the transformed dependent variable. If this is
the case, you can simply use CREATE to subtract 1.0 from your dependent variable to use this
estimator.

N13.4 Output from the Ordered Probability Estimators

All of the ordered probit/logit models begin with an initial set of least squares results of
some sort. These are suppressed unless your command contains ; OLS. The iterations are then
followed by the maximum likelihood estimates in the usual tabular format. The final output includes
a listing of the cell frequencies for the outcomes. When the data are stratified, this output will also
include a table of the frequencies in the strata. The log likelihood function, and a log likelihood
computed assuming all slopes are zero are computed. For the latter, the threshold parameters are still
allowed to vary freely, so the model is simply one which assigns each cell a predicted probability
equal to the sample proportion. This appropriately measures the contribution of the nonconstant
regressors to the log likelihood function. As such, the chi squared statistic given is a valid test
statistic for the hypothesis that all slopes on the nonconstant regressors are zero.
The sample below shows the standard output for a model with six outcomes. These are the
German health care data used in several earlier examples. The dependent variable is the self reported
health satisfaction rating. For the purpose of a convenient sample application, we have truncated the
health satisfaction variable at five by discarding observations – in the original data set, it is coded
0,1,...,10.
N13: Ordered Choice Models N-206

HINT: The ordered logit model typically produces the same sort of scaling of the coefficient vector
that arises in the binary choice models discussed in Chapter E27. As before, the difference becomes
much less pronounced when the marginal effects are considered instead. We are unaware of a
convenient specification test for distinguishing between the probit and logit models. A test of
normality against the broader Pearson family of distributions is described in Glewwe (1997), but it is
not especially convenient. A test for skewness based on the Vuong test seems like a possibility.

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HSAT
Log likelihood function -11284.68638
Restricted log likelihood -11308.02002
Chi squared [ 4 d.f.] 46.66728
Significance level .00000
McFadden Pseudo R-squared .0020635
Estimation based on N = 8140, K = 9
Inf.Cr.AIC =22587.373 AIC/N = 2.775
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 1.32892*** .07276 18.27 .0000 1.18632 1.47152
FEMALE| .04526* .02546 1.78 .0755 -.00465 .09517
HHNINC| .35590*** .07832 4.54 .0000 .20240 .50940
HHKIDS| .10604*** .02665 3.98 .0001 .05381 .15827
EDUC| .00928 .00630 1.47 .1407 -.00307 .02162
|Threshold parameters for index
Mu(1)| .23635*** .01237 19.11 .0000 .21211 .26059
Mu(2)| .62954*** .01440 43.72 .0000 .60132 .65777
Mu(3)| 1.10764*** .01406 78.78 .0000 1.08008 1.13519
Mu(4)| 1.55676*** .01527 101.94 .0000 1.52683 1.58669
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

+--------------------------------------------------------------------+
| CELL FREQUENCIES FOR ORDERED CHOICES |
+--------------------------------------------------------------------+
| Frequency Cumulative < = Cumulative > = |
|Outcome Count Percent Count Percent Count Percent |
|----------- ------- --------- ------- --------- ------- --------- |
|HSAT=00 447 5.4914 447 5.4914 8140 100.0000 |
|HSAT=01 255 3.1327 702 8.6241 7693 94.5086 |
|HSAT=02 642 7.8870 1344 16.5111 7438 91.3759 |
|HSAT=03 1173 14.4103 2517 30.9214 6796 83.4889 |
|HSAT=04 1390 17.0762 3907 47.9975 5623 69.0786 |
|HSAT=05 4233 52.0025 8140 100.0000 4233 52.0025 |
+--------------------------------------------------------------------+
N13: Ordered Choice Models N-207

Cross tabulation of predictions and actual outcomes

+------+-----+-----+-----+-----+-----+-----+-----+
|y(i,j)| 0 | 1 | 2 | 3 | 4 | 5 |Total|
+------+-----+-----+-----+-----+-----+-----+-----+
| 0 | 0| 0| 0| 0| 0| 447| 447|
| 1 | 0| 0| 0| 0| 0| 255| 255|
| 2 | 0| 0| 0| 0| 0| 642| 642|
| 3 | 0| 0| 0| 0| 0| 1173| 1173|
| 4 | 0| 0| 0| 0| 0| 1390| 1390|
| 5 | 0| 0| 0| 0| 0| 4233| 4233|
+------+-----+-----+-----+-----+-----+-----+-----+
| Total| 0| 0| 0| 0| 0| 8140| 8140|
+------+-----+-----+-----+-----+-----+-----+-----+
Row = actual, Column = Prediction, Model = Probit
Prediction is number of the most probable cell.

Cross tabulation of outcomes and predicted probabilities.

+------+-----+-----+-----+-----+-----+-----+-----+
|y(i,j)| 0 | 1 | 2 | 3 | 4 | 5 |Total|
+------+-----+-----+-----+-----+-----+-----+-----+
| 0 | 26| 15| 36| 66| 77| 228| 447|
| 1 | 14| 8| 21| 37| 44| 131| 255|
| 2 | 36| 20| 51| 93| 110| 331| 642|
| 3 | 64| 37| 93| 170| 200| 609| 1173|
| 4 | 75| 43| 109| 200| 237| 725| 1390|
| 5 | 230| 132| 333| 610| 722| 2206| 4233|
+------+-----+-----+-----+-----+-----+-----+-----+
| Total| 445| 255| 644| 1176| 1389| 4230| 8140|
+------+-----+-----+-----+-----+-----+-----+-----+
Row = actual, Column = Prediction, Model = Probit
Value(j,m)=Sum(i=1,N)y(i,j)*p(i,m).
Column totals may not match cell sums because of rounding error.

The model output is followed by a (J+1)×(J+1) frequency table of predicted versus actual
values. (This table is not given when data are grouped or when there are more than 10 outcomes.)
The predicted outcome for this tabulation is the one with the largest predicted probability. Even
though the model appears to be highly significant, the table of predictions has seems to suggest a
lack of predictive power. Tables such as the one above are common with this model. The driver of
the result is the sample configuration of the data. Note in the frequency table that the sample is quite
unbalanced, and the highest outcome is quite likely to have the highest probability for every
observation. The estimation criterion for the ordered probability model is unrelated to its ability to
predict those cells, and you will rarely see a predictions table that closely matches the actual
outcomes. It often happens that even in a set of results with highly significant coefficients, only one
or a few of the outcomes are predicted by the model. The second table relates more closely to the
aggregate predictions of the model. The table entries are the sample proportions that would be
predicted for each outcome. For example, the first row of the table shows that 447 individuals in the
sample chose outcome 0. For every individual, the model produces a full set of J+1 probabilities.
For the 447 individuals, 8140 times the sum of the probabilities of outcome 0 equals 26, 8140 times
the sum of the probabilities of outcome 1 equals 15, and so on.
N13: Ordered Choice Models N-208

N13.4.1 Robust Covariance Matrix Estimation

The Sandwich Estimator

The standard robust covariance matrix is

 n  ∂ 2 log Fi    n  ∂ log Fi  ∂ log Fi ′   n  ∂ 2 log Fi

−1 −1

Est.Asy.Var 
β 
ˆ  = ∑ i 1 =
  ∑ i 1      ∑ i 1  ˆ ˆ ′ 
=
  ∂gggggg
ˆ ∂ ˆ ′      ∂ ˆ =
 ∂    ∂ ∂ 


where γ̂ indicates the full set of parameters in the model. To obtain this matrix with any of the
forms of the ordered choice models, use

; Robust

in the ORDERED command.

Clustering and Stratification

A related calculation is used when observations occur in groups which may be correlated.
This is rather like a panel; one might use this approach in a random effects kind of setting in which
observations have a common latent heterogeneity. The parameter estimator is unchanged in this
case, but an adjustment is made to the estimated asymptotic covariance matrix. Full details on this
estimator appear in Chapter R10. To specify this estimator, use

; Cluster = specification

where the specification is either a fixed number of observations or the name of a variable that
provides an identifier for the cluster, such as an id number. Note that if there is exactly one
observation per cluster, then this is G/(G-1) times the sandwich estimator discussed above. Also, if
you have fewer clusters than parameters, then this matrix is singular – it has rank equal to the
minimum of G and K, the number of parameters.
The extension of this estimator to stratified data is described in detail in Section R10.3. To
use this with the ; Cluster specification, add

; Stratum = specification.
N13: Ordered Choice Models N-209

N13.4.2 Saved Results

For each observation, the predicted probabilities for all J+1 outcomes are computed. Then if
you request ; List, the listing will contain

Predicted Y: Y with the largest probability.

Residual: the largest of the J+1 probabilities (i.e., Prob[y = fitted Y]).

∑
J
Var1: the estimate of E[yi] = i=0
i × Prob[Yi = i].

Var2: the probability estimated for the observed Y.

Estimation results kept by the estimator are as follows:

Matrices: b = estimate of b,
varb = estimated asymptotic covariance,
mu = J-1 estimated µs.

Scalars: kreg, nreg, and logl.

Last Model: The labels are b_variables, mu1, ...

Last Function: Prob(y = highest outcome | x).

The specification ; Par adds µ (the set of estimated threshold values) to b and varb. The additional
matrix, mu is kept regardless, but the estimated asymptotic covariance matrix is lost unless the
command contains ; Par. The Last Function is used in the SIMULATE and PARTIAL EFFECTS
routines. The default function is the probability of the highest outcome. You can specify a different
outcome in the command with

; Outcome = j

where j is the desired outcome. For example, in our earlier application in which outcomes are
0,1,2,3,4,5, the command might specify

PARTIAL EFFECTS ; Effects: hhninc ; Outcome = 3 $

and likewise for SIMULATE. A full examination of all outcomes is obtained by using

; Outcome = *
N13: Ordered Choice Models N-210

N13.5 Partial Effects and Simulations

There is potentially a large amount of output for the ordered choice model, in addition to the
basic model results. There is no single conditional mean because the outcomes are labels, not
measures. There are J+1 probabilities to analyze,

Prob[cell j] = F(µj - b′xi) - F(µj-1 - b′xi).

Typically, the highest or lowest cell is of interest. However, the PARTIAL EFFECTS (or just
PARTIALS) and SIMULATE commands can be used to examine any or all of them.
Marginal effects in the ordered probability models are also quite involved. Since there is no
meaningful conditional mean function to manipulate, we compute, instead, the effects of changes in
the covariates on the cell probabilities. These are:

∂Prob[cell j]/∂xi = [f(µj-1 - b′xi) - f(µj - b′xi)] × b,

where f(.) is the appropriate density for the standard normal, φ(•), logistic density, Λ(•)(1-Λ(•)),
Weibull, Gompertz or arctangent. Each vector is a multiple of the coefficient vector. But it is worth
noting that the magnitudes are likely to be very different. In at least one case, Prob[cell 0], and
probably more if there are more than three outcomes, the partial effects have exactly the opposite
signs from the estimated coefficients.

NOTE: This estimator segregates dummy variables for separate computation in the marginal
effects. The marginal effect for a dummy variable is the difference of the two probabilities, with and
without the variable.

Partial effects for the ordered probability models are obtained internally in the command by
adding
; Partial Effects

in the command. This produces a table oriented to the outcomes, such as the one below. A second
summary that is oriented to the variables rather than the outcomes is requested with

; Partial Effects ; Full

The internal results are computed at the means of the data. Partial effects can also be obtained with
the PARTIALS command. The third set of results below is obtained with

PARTIALS ; Effects: hhninc ; Outcome = * $

This command produces average partial effects by default, but you can request that they be
computed at the data means by adding ; Means to the command. Probabilities for particular
outcomes are obtained with the SIMULATE command. An example appears below.
N13: Ordered Choice Models N-211

-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HSAT| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
*FEMALE| -.00498* -.09207 -1.77 .0763 -.01049 .00053
HHNINC| -.03907*** -.23836 -4.53 .0000 -.05599 -.02216
*HHKIDS| -.01132*** -.20926 -4.08 .0000 -.01676 -.00588
EDUC| -.00102 -.20477 -1.47 .1409 -.00237 .00034
|--------------[Partial effects on Prob[Y=01] at means]--------------
*FEMALE| -.00210* -.06711 -1.78 .0758 -.00441 .00022
HHNINC| -.01647*** -.17397 -4.54 .0000 -.02358 -.00936
*HHKIDS| -.00483*** -.15473 -4.04 .0001 -.00718 -.00249
EDUC| -.00043 -.14945 -1.47 .1408 -.00100 .00014
|--------------[Partial effects on Prob[Y=02] at means]--------------
*FEMALE| -.00414* -.05244 -1.77 .0760 -.00872 .00043
HHNINC| -.03257*** -.13605 -4.50 .0000 -.04675 -.01838
*HHKIDS| -.00964*** -.12205 -3.98 .0001 -.01439 -.00489
EDUC| -.00085 -.11688 -1.47 .1412 -.00198 .00028
|--------------[Partial effects on Prob[Y=03] at means]--------------
*FEMALE| -.00473* -.03273 -1.77 .0764 -.00997 .00050
HHNINC| -.03727*** -.08501 -4.43 .0000 -.05375 -.02078
*HHKIDS| -.01121*** -.07751 -3.87 .0001 -.01689 -.00554
EDUC| -.00097 -.07303 -1.47 .1417 -.00227 .00032
|--------------[Partial effects on Prob[Y=04] at means]--------------
*FEMALE| -.00208* -.01214 -1.77 .0762 -.00438 .00022
HHNINC| -.01643*** -.03166 -4.34 .0000 -.02385 -.00901
*HHKIDS| -.00518*** -.03026 -3.66 .0002 -.00795 -.00241
EDUC| -.00043 -.02720 -1.47 .1427 -.00100 .00014
|--------------[Partial effects on Prob[Y=05] at means]--------------
*FEMALE| .01803* .03469 1.78 .0755 -.00185 .03792
HHNINC| .14181*** .09003 4.54 .0000 .08065 .20297
*HHKIDS| .04219*** .08116 3.99 .0001 .02145 .06292
EDUC| .00370 .07734 1.47 .1407 -.00122 .00861
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N13: Ordered Choice Models N-212

+----------------------------------------------------------------------+
| Summary of Marginal Effects for Ordered Probability Model (probit) |
| Effects computed at means. Effects for binary variables (*) are |
| computed as differences of probabilities, other variables at means. |
| Binary variables change only by 1 unit so s.d. changes are not shown.|
| Elasticities for binary variables = partial effect/probability = %chgP |
+----------------------------------------------------------------------+
+----------------------------------------------------------------------+
| Binary(0/1) Variable FEMALE Changes in *FEMALE % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.00498 -.00498 .00000 - -.00498 -.09207
Y = 01 -.00210 -.00708 .00498 - -.00210 -.06711
Y = 02 -.00414 -.01122 .00708 - -.00414 -.05244
Y = 03 -.00473 -.01595 .01122 - -.00473 -.03273
Y = 04 -.00208 -.01803 .01595 - -.00208 -.01214
Y = 05 .01803 .00000 .01803 - .01803 .03469
+----------------------------------------------------------------------+
| Continuous Variable HHNINC Changes in HHNINC % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.03907 -.03907 .00000 -.00655 -.11703 -.23836
Y = 01 -.01647 -.05555 .03907 -.00276 -.04933 -.17397
Y = 02 -.03257 -.08811 .05555 -.00546 -.09753 -.13605
Y = 03 -.03727 -.12538 .08811 -.00625 -.11161 -.08501
Y = 04 -.01643 -.14181 .12538 -.00275 -.04921 -.03166
Y = 05 .14181 .00000 .14181 .02377 .42472 .09003
+----------------------------------------------------------------------+
| Binary(0/1) Variable HHKIDS Changes in *HHKIDS % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.01132 -.01132 .00000 - -.01132 -.20926
Y = 01 -.00483 -.01615 .01132 - -.00483 -.15473
Y = 02 -.00964 -.02579 .01615 - -.00964 -.12205
Y = 03 -.01121 -.03701 .02579 - -.01121 -.07751
Y = 04 -.00518 -.04219 .03701 - -.00518 -.03026
Y = 05 .04219 .00000 .04219 - .04219 .08116
+----------------------------------------------------------------------+
| Continuous Variable EDUC Changes in EDUC % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.00102 -.00102 .00000 -.00212 -.01120 -.20477
Y = 01 -.00043 -.00145 .00102 -.00089 -.00472 -.14945
Y = 02 -.00085 -.00230 .00145 -.00177 -.00934 -.11688
Y = 03 -.00097 -.00327 .00230 -.00202 -.01069 -.07303
Y = 04 -.00043 -.00370 .00327 -.00089 -.00471 -.02720
Y = 05 .00370 .00000 .00370 .00770 .04066 .07734
------------------------------------------------------------------------
N13: Ordered Choice Models N-213

PARTIALS ; Effects: hhninc ; Outcome = * $

---------------------------------------------------------------------
Partial Effects Analysis for Ordered Probit Probability Y = 5
---------------------------------------------------------------------
Effects on function with respect to HHNINC
Results are computed by average over sample observations
Partial effects for continuous HHNINC computed by differentiation
Effect is computed as derivative = df(.)/dx
---------------------------------------------------------------------
df/dHHNINC Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
APE Prob(y= 0) -.03930 .00872 4.51 -.05640 -.02220
APE Prob(y= 1) -.01643 .00373 4.41 -.02374 -.00912
APE Prob(y= 2) -.03238 .00734 4.41 -.04677 -.01800
APE Prob(y= 3) -.03694 .00827 4.47 -.05315 -.02072
APE Prob(y= 4) -.01624 .00382 4.26 -.02372 -.00876
APE Prob(y= 5) .14129 .03099 4.56 .08055 .20204

SIMULATE ; Scenario: & hhninc = 0(.05)1 ; Plot(ci) ; Outcome = 4 $

---------------------------------------------------------------------
Model Simulation Analysis for Ordered Probit Probability Y = 4
---------------------------------------------------------------------
Simulations are computed by average over sample observations
---------------------------------------------------------------------
User Function Function Standard
(Delta method) Value Error |t| 95% Confidence Interval
---------------------------------------------------------------------
Avrg. Function .17068 .00988 17.27 .15131 .19005
HHNINC = .00 .17528 .01026 17.09 .15517 .19538
HHNINC = .05 .17477 .01021 17.11 .15476 .19479
HHNINC = .10 .17421 .01016 17.14 .15429 .19413
HHNINC = .15 .17360 .01011 17.17 .15379 .19342
HHNINC = .20 .17294 .01005 17.20 .15324 .19265
HHNINC = .25 .17223 .00999 17.23 .15264 .19182
HHNINC = .30 .17147 .00993 17.26 .15199 .19094
HHNINC = .35 .17065 .00987 17.28 .15130 .19001
HHNINC = .40 .16979 .00982 17.30 .15055 .18903
HHNINC = .45 .16888 .00976 17.30 .14975 .18801
HHNINC = .50 .16793 .00971 17.30 .14890 .18695
HHNINC = .55 .16692 .00966 17.28 .14799 .18586
HHNINC = .60 .16587 .00962 17.24 .14701 .18473
HHNINC = .65 .16478 .00959 17.18 .14598 .18358
HHNINC = .70 .16364 .00957 17.09 .14488 .18241
HHNINC = .75 .16246 .00957 16.98 .14371 .18122
HHNINC = .80 .16124 .00958 16.84 .14247 .18001
HHNINC = .85 .15998 .00960 16.66 .14116 .17880
HHNINC = .90 .15868 .00965 16.45 .13978 .17758
HHNINC = .95 .15734 .00971 16.21 .13832 .17637
HHNINC = 1.00 .15596 .00979 15.93 .13678 .17515
N13: Ordered Choice Models N-214

Figure N13.1 Ordered Probit Simulation

N14: Extended Ordered Choice Models N-215

N14: Extended Ordered Choice Models

N14.1 Introduction
The basic ordered choice model is based on the latent regression,

yi* = b′xi + ei, ei ~ F(ei |q), E[ei|xi] = 0, Var[ei|xi] = 1.

The observation mechanism results from a complete censoring of the latent dependent variable as
follows:
yi = 0 if yi ≤ µ0,
= 1 if µ0 < yi ≤ µ1,
= 2 if µ1 < yi ≤ µ2,
...
= J if yi > µJ-1.

The latent ‘preference’ variable, yi* is not observed. The observed counterpart to yi* is yi. The
probabilities which enter the log likelihood function are

Prob[yi = j] = Prob[yi* is in the jth range].

Estimation and analysis of the basic model are presented in Chapter N13 (and Chapter E34). A
variety of additional specifications and extensions are supported.

N14.2 Weighting and Heteroscedasticity

An ordered probit model with simple heteroscedasticity,

Var[ei] = wi2,

may be estimated with

ORDERED ; Rhs = ... ; Lhs = ...

; Wts = your weighting variable, wi
; Heteroscedastic $

Your command gives the name of the variable which carries the observed individual specific
standard deviations. This formulation does not add new parameters to the model, and only instructs
the estimator how the weighting variable is to be handled.
N14: Extended Ordered Choice Models N-216

This approach is different from estimating the model with weights. Without ; Het, this
model is treated as any other weighted log likelihood, and the estimator maximizes

∑
n
log L = i =1
wi log Pr ob(observed outcomei )

where Prob[cell j] = F(µj - b′xi) - F(µj-1 - b′xi).

With ; Het, the probabilities are built up from the heteroscedastic random variable, but the terms in
the log likelihood are unweighted. With this form of the command, using ; Het, the model is

Prob[cell j] = F[(µj - b′xi)/wi] - F[(µj-1 - b′xi)/wi]

∑
n
and log L = i =1
log Pr ob(observed outcomei )

N14.3 Multiplicative Heteroscedasticity

The model with multiplicative heteroscedasticity,

Var[ei] = [exp(γ′zi)]2,

is requested with

ORDERED ; Rhs = ... ; Lhs = ...

; Het
; Rh2 = list of variables in z $

NOTE: Do not include a constant (one) in z. A variable in z which has no variation, such as one,
will lead to a singular Hessian, and the estimator will fail to converge.

This formulation adds a vector of new parameters to the model. For purposes of starting values,
restrictions, and hypothesis tests, the full parameter vector becomes

Θ = [b1,...,bK,γ1,...,γL,µ1,...,µJ-1].

You can use ; Rst and ; CML: for imposing restrictions as usual. As always, restrictions that force
ancillary variance parameters (γh) to equal parameters in the conditional mean function (bk) will
rarely produce satisfactory results. In the saved results, the estimator of γ will always be included in
b and varb. Thus, if you want to extract parts of the parameter vector after estimation, you might use

NAMELIST ; x = ...
; z = ... $
ORDERED ; Lhs = y ; Rhs = x
; Rh2 = z ; Het $
CALC ; k = Col(x) ; k1 = k+1 ; kt = k + Col(z) $
MATRIX ; beta = b(1:k)
; gamma = b(k1:kt) $
N14: Extended Ordered Choice Models N-217

The µ threshold parameters are still the ancillary parameters. Marginal effects, fitted values, and so
on are requested exactly as before with this extension of the ordered probit model. In the Last Model
labels list, the variance parameters will be denoted c_variable, so with this model, the complete list
of labels is
Last Model = [B_...,C_...,MU1,...].

The Last Function for the model is the probability including the exponential heteroscedasticity model

 µ j − b ′x   µ j −1 − b′x 
=
Prob( y 1|=
x, z ) F  −F 
 exp( γ ′z )   exp( γ ′z ) 

N14.3.1 Testing for Heteroscedasticity

The model with homoscedastic disturbances is nested in this model (γ = 0) so the standard
tests, i.e., LM, likelihood ratio, and Wald, are available for testing the specification. The first two of
these will be very convenient. To carry out an LM test, you could use the following: First define the
two variable lists.

NAMELIST ; x = ...
; z = ... $

Fit the model without heteroscedasticity. This command saves b and mu needed later.

ORDERED ; Lhs = y ; Rhs = x $

Define the zero vector for the variance parameters.

MATRIX ; {h = Col(z)} ; gamma = Init (h,1,0) $

Now, fit the heteroscedastic model, but do not iterate. This displays the LM statistic.
ORDERED ; Lhs = y ; Rhs = x ; Rh2 = z ; Het
; Start = b,gamma,mu ; Maxit = 0 $
To use a likelihood ratio test, instead, the preceding is modified as follows:
1. Add CALC ; lr = logl $ after the first ORDERED command.

2. Omit ; Maxit = 0 from the second ORDERED command.

3. Add the command

CALC ; List ; chi = 2*(logl - lr) $

after the second ORDERED command; chi is the chi squared statistic. This can be referred
to the table with

CALC ; cstar = Ctb(.95,L) $

which provides the necessary critical value.
N14: Extended Ordered Choice Models N-218

The following experiment illustrates these computations. We test for heteroscedasticity in the
health satisfaction model, using the three standard tests in an ordered logit model as the platform. To
simplify it a bit, we use a restricted sample of only those individuals observed in all seven periods.

SAMPLE ; All $
REJECT ; _groupti < 7 $
ORDERED ; Lhs = newhsat
; Rhs = one,female,hhninc,hhkids,educ
; Logit $
CALC ; lr = logl $

This command carries out the LM test. The starting values are from the previous model for b and µ
and zeros for the elements of γ. The test is requested with ; Maxit = 0.

ORDERED ; Lhs = newhsat

; Rhs = one,female,hhninc,hhkids,educ
; Logit ; Het ; Rh2 = married,univ,working,female,hhninc
; Start = b,0,0,0,mu ; Maxit = 0 $

This command estimates the full heteroscedastic model. Based on these results, we then carry out
the likelihood ratio and Wald tests.

ORDERED ; Lhs = newhsat

; Rhs = one,female,hhninc,hhkids,educ
; Logit ; Het ; Rh2 = married,univ,working,female,hhninc $
CALC ; lu = logl $
CALC ; List ; lrtest = 2*(lu - lr) $
MATRIX ; gamma = b(6:10) ; vgamma = varb(6:10,6:10) $
MATRIX ; List
; waldstat = gamma'<vgamma>gamma $

As might be expected in a sample this large, the three tests give the same answer. The LM, LR and
Wald statistics obtained are 84.16200, 84.26808 and 83.90174, respectively.

The first set of results are for the restricted, homoscedastic model.

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable NEWHSAT
Log likelihood function -12971.89392
Restricted log likelihood -13138.97978
Chi squared [ 4 d.f.] 334.17171
Significance level .00000
McFadden Pseudo R-squared .0127168
Estimation based on N = 6209, K = 14
Inf.Cr.AIC =25971.788 AIC/N = 4.183
Underlying probabilities based on Logistic
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-219

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 3.02189*** .13081 23.10 .0000 2.76551 3.27827
FEMALE| -.31859*** .04729 -6.74 .0000 -.41129 -.22590
HHNINC| .23133* .13880 1.67 .0956 -.04072 .50338
HHKIDS| .47849*** .04529 10.56 .0000 .38972 .56726
EDUC| .10241*** .01122 9.12 .0000 .08041 .12441
|Threshold parameters for index
Mu(1)| .49176*** .05264 9.34 .0000 .38859 .59493
Mu(2)| 1.26288*** .05011 25.20 .0000 1.16468 1.36109
Mu(3)| 1.94907*** .04093 47.62 .0000 1.86886 2.02929
Mu(4)| 2.48180*** .03468 71.57 .0000 2.41383 2.54976
Mu(5)| 3.48744*** .02747 126.94 .0000 3.43360 3.54129
Mu(6)| 3.94860*** .02594 152.22 .0000 3.89776 3.99944
Mu(7)| 4.61859*** .02627 175.79 .0000 4.56710 4.67009
Mu(8)| 5.70197*** .03154 180.78 .0000 5.64015 5.76378
Mu(9)| 6.48830*** .04110 157.86 .0000 6.40774 6.56886
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The next set of results is the computation of the Lagrange multiplier statistic. This next command
does not reestimate the model. Note that the coefficient estimates are identical, save for the
parameters in the variance function. The estimated standard errors do change, however, because in
the restricted model above, the Hessian is computed and inverted just for the parameters estimated.
In the results below, the Hessian is computed as if the inserted zeros for γ were actually the
parameter estimates. These standard errors are not useful.
Maximum iterations reached. Exit iterations with status=1.
Maxit = 0. Computing LM statistic at starting values.
No iterations computed and no parameter update done.
-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable NEWHSAT
LM Stat. at start values 92.77220
LM statistic kept as scalar LMSTAT
Log likelihood function -12971.89392
Restricted log likelihood -13138.97978
Chi squared [ 9 d.f.] 334.17171
Significance level .00000
McFadden Pseudo R-squared .0127168
Estimation based on N = 6209, K = 19
Inf.Cr.AIC =25981.788 AIC/N = 4.185
Underlying probabilities based on Logistic
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 3.02189*** .18716 16.15 .0000 2.65507 3.38871
FEMALE| -.31859*** .04747 -6.71 .0000 -.41164 -.22555
HHNINC| .23133 .15162 1.53 .1271 -.06584 .52849
HHKIDS| .47849*** .05058 9.46 .0000 .37936 .57762
EDUC| .10241*** .01246 8.22 .0000 .07798 .12683
N14: Extended Ordered Choice Models N-220

|Variance function
MARRIED| 0.0 .02958 .00 1.0000 -.57975D-01 .57975D-01
UNIV| 0.0 .06508 .00 1.0000 -.12755D+00 .12755D+00
WORKING| 0.0 .02825 .00 1.0000 -.55371D-01 .55371D-01
FEMALE| 0.0 .02483 .00 1.0000 -.48663D-01 .48663D-01
HHNINC| 0.0 .07843 .00 1.0000 -.15372D+00 .15372D+00
|Threshold parameters for index
Mu(1)| .49176*** .06836 7.19 .0000 .35778 .62574
Mu(2)| 1.26288*** .09719 12.99 .0000 1.07240 1.45336
Mu(3)| 1.94907*** .11474 16.99 .0000 1.72420 2.17395
Mu(4)| 2.48180*** .12755 19.46 .0000 2.23181 2.73178
Mu(5)| 3.48744*** .15442 22.58 .0000 3.18479 3.79010
Mu(6)| 3.94860*** .16835 23.45 .0000 3.61864 4.27856
Mu(7)| 4.61859*** .18971 24.35 .0000 4.24677 4.99041
Mu(8)| 5.70197*** .22651 25.17 .0000 5.25801 6.14592
Mu(9)| 6.48830*** .25426 25.52 .0000 5.98996 6.98664
-----------------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

These are the estimates for the full heteroscedastic model. The test statistics appear after the
estimated parameters.
-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable NEWHSAT
Log likelihood function -12924.94799
Restricted log likelihood -13138.97978
Chi squared [ 9 d.f.] 428.06357
Significance level .00000
McFadden Pseudo R-squared .0162898
Estimation based on N = 6209, K = 19
Inf.Cr.AIC =25887.896 AIC/N = 4.169
Underlying probabilities based on Logistic
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.38708*** .14152 16.87 .0000 2.10971 2.66445
FEMALE| -.22820*** .03379 -6.75 .0000 -.29442 -.16199
HHNINC| .13810 .09576 1.44 .1492 -.04958 .32579
HHKIDS| .33481*** .03573 9.37 .0000 .26478 .40485
EDUC| .06415*** .00763 8.40 .0000 .04919 .07911
|Variance function
MARRIED| -.13333*** .03198 -4.17 .0000 -.19601 -.07066
UNIV| -.19916*** .05658 -3.52 .0004 -.31007 -.08826
WORKING| -.18323*** .02928 -6.26 .0000 -.24062 -.12584
FEMALE| -.03756 .02478 -1.52 .1296 -.08613 .01101
HHNINC| -.19768*** .07590 -2.60 .0092 -.34643 -.04893
|Threshold parameters for index
Mu(1)| .38333*** .05379 7.13 .0000 .27790 .48875
Mu(2)| .97539*** .07759 12.57 .0000 .82333 1.12746
Mu(3)| 1.48986*** .09299 16.02 .0000 1.30761 1.67211
Mu(4)| 1.88162*** .10423 18.05 .0000 1.67733 2.08590
Mu(5)| 2.60926*** .12681 20.58 .0000 2.36072 2.85779
Mu(6)| 2.93848*** .13795 21.30 .0000 2.66810 3.20885
Mu(7)| 3.41196*** .15468 22.06 .0000 3.10880 3.71512
Mu(8)| 4.16905*** .18272 22.82 .0000 3.81092 4.52718
Mu(9)| 4.72049*** .20380 23.16 .0000 4.32105 5.11992
N14: Extended Ordered Choice Models N-221

---------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The final results are the test statistics for the hypothesis of homoscedasticity. The three results are,
as expected, essentially the same.
LM Stat. at start values 92.77220 (from the earlier results)

[CALC] LRTEST = 93.8918620

WALDSTAT| 1
--------+--------------
1| 94.6903

N14.3.2 Partial Effects in the Heteroscedasticity Model

Partial effects in the ordered choice models with heteroscedasticity appear from two sources,
in the latent utility and in the variance function. When variables appear in both places, the total effect
is the sum of the two terms.

∂Prob( yi = µ j , s − b ′x i
 f ( a j −1, s ) − f ( a j , s )  b, a j , s =
j | xi , z i ) 1
=
 
∂xi wi exp( γ ′z i )

∂Prob( yi = j | xi , z i )  f ( a j −1, s ) a j −1, s − f ( a j , s ) a j , s 

=  zi .
∂z i  F ( a j , s ) − F ( a j −1.s ) 

Request the partial effects within the command with

; Partial Effects

The following results show the computation for the full model fit earlier. (Effects for
outcomes 0 to 7 are omitted below.)
+-------------------------------------------+
| Marginal Effects for OrdLogit |
| * Total effect = sum of terms |
+----------+----------+----------+----------+
| Variable | NEWHSA=8 | NEWHS=9 | NEWHS=10 |
+----------+----------+----------+----------+
| FEMALE | -.02676 | -.02181 | -.02998 |
| HHNINC | .01619 | .01320 | .01814 |
| HHKIDS | .03925 | .03200 | .04399 |
| EDUC | .00752 | .00613 | .00843 |
| MARRIED | .01949 | -.00278 | -.02676 |
| UNIV | .02911 | -.00415 | -.03997 |
| WORKING | .02678 | -.00382 | -.03677 |
| HHNINC | .02889 | -.00412 | -.03967 |
| FEMALE | .00549 | -.00078 | -.00754 |
| FEMALE *| -.02127 | -.02260 | -.03752 |
| HHNINC *| .04508 | .00908 | -.02153 |
+----------+----------+----------+----------+
N14: Extended Ordered Choice Models N-222

The PARTIAL EFFECTS (or just PARTIALS) and SIMULATE commands receive the
estimates form the heteroscedastic ordered choice model, so you can use them to analyze the
probabilities or partial effects. For example, to replace the preceding results, use

PARTIALS ; Effects: female / hhninc ; Outcome = * $

Three differences are first, this estimator uses average partial effects by default (or means if you
request them), second, it uses partial differences for dummy variables while the built in computation
uses scaled coefficients and, third, as seen below, the PARTIAL EFFECTS command produces
standard errors and confidence intervals for the partial effects.

---------------------------------------------------------------------
Partial Effects Analysis for Ordered Logit (Het) Prob[Y = 10]
---------------------------------------------------------------------
Effects on function with respect to FEMALE
Results are computed by average over sample observations
Partial effects for binary var FEMALE computed by first difference
---------------------------------------------------------------------
df/dFEMALE Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
APE Prob(y= 0) .00195 .00148 1.32 -.00096 .00485
APE Prob(y= 1) .00166 .00075 2.23 .00020 .00312
APE Prob(y= 2) .00534 .00170 3.14 .00201 .00867
APE Prob(y= 3) .00959 .00218 4.40 .00532 .01387
APE Prob(y= 4) .01189 .00210 5.66 .00778 .01601
APE Prob(y= 5) .03070 .00447 6.87 .02194 .03946
APE Prob(y= 6) .01222 .00255 4.79 .00721 .01722
APE Prob(y= 7) .00646 .00381 1.70 -.00100 .01393
APE Prob(y= 8) -.02026 .00510 3.97 -.03025 -.01027
APE Prob(y= 9) -.02224 .00323 6.89 -.02857 -.01591
APE Prob(y=10) -.03732 .00645 5.79 -.04996 -.02468

---------------------------------------------------------------------
Partial Effects Analysis for Ordered Logit (Het) Prob[Y = 10]
---------------------------------------------------------------------
Effects on function with respect to HHNINC
Results are computed by average over sample observations
Partial effects for continuous HHNINC computed by differentiation
Effect is computed as derivative = df(.)/dx
---------------------------------------------------------------------
df/dHHNINC Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
APE Prob(y= 0) -.01302 .00449 2.90 -.02183 -.00421
APE Prob(y= 1) -.00620 .00215 2.89 -.01041 -.00199
APE Prob(y= 2) -.01426 .00473 3.01 -.02354 -.00498
APE Prob(y= 3) -.01675 .00575 2.91 -.02803 -.00547
APE Prob(y= 4) -.01297 .00544 2.39 -.02362 -.00231
APE Prob(y= 5) -.00775 .01253 .62 -.03231 .01681
APE Prob(y= 6) .01008 .00739 1.36 -.00440 .02456
APE Prob(y= 7) .02766 .01108 2.50 .00593 .04938
APE Prob(y= 8) .04272 .01395 3.06 .01538 .07006
APE Prob(y= 9) .01063 .00909 1.17 -.00718 .02845
APE Prob(y=10) -.02014 .02072 .97 -.06076 .02047
N14: Extended Ordered Choice Models N-223

N14.4 Sample Selection and Treatment Effects

The following describes an ordered probit counterpart to the standard sample selection
model. This is only available for the ordered probit specification. The structural equations are, first,
the main equation, the ordered choice model,

yi* = b′xi + ei, ei ~ F(ei |q), E[ei] = 0, Var[ei] = 1,

yi = 0 if yi ≤ µ0,
= 1 if µ0 < yi ≤ µ1,
= 2 if µ1 < yi ≤ µ2,
...
= J if yi > µJ-1.

Second is the selection equation, a univariate probit model,

di* = α′zi + ui,

di = 1 if di* > 0 and 0 otherwise,

The observation mechanism is

[yi,xi] is observed if and only if di = 1.

ei,ui ~ N2[0,0,1,1,ρ]; there is ‘selectivity’ if ρ is not equal to zero.

This model is a straightforward generalization of the bivariate probit model with sample selection in
Section N12.4.
The treatment effects model includes di as an endogenous binary variable in the ordered
probit equation;
yi* = b′xi + γdi + ei, ei ~ F(ei |q), E[ei] = 0, Var[ei] = 1,
yi = j if µj-1 < yi* < µj, j = 0,1,…,J
di* = α′zi + ui,
di = 1 if di* > 0 and 0 otherwise,
ei,ui ~ N2[0,0,1,1,ρ]; di is endogenous if ρ is not equal to zero.

This model is a generalization of the recursive bivariate probit model in Section N12.6.
N14: Extended Ordered Choice Models N-224

N14.4.1 Command
These models require two passes to estimate. In the first, you fit a probit model for the
selection (or treatment) variable, d. You then pass these values to the ordered probit model using a
standard command for this operation, the ; Hold parameter in the probit command. The two
commands would be as follows: (This model is requested in the same fashion as NLOGIT’s other
sample selectivity models.) Estimate first stage probit model and hold results for next step in the
estimation.

PROBIT ; Lhs = d ; Rhs = Z list ; Hold $

Second, estimate the ordered probit model with selectivity.

ORDERED ; Lhs = y ; Rhs = X ; ... as usual ; Selection $

You need not make any other changes in the ordered probit command. For the treatment effects
case, the probit model is unchanged while the ORDERED command becomes

ORDERED ; Lhs = y ; Rhs = X,d ; ... as usual ; Selection ; All $

Note that the treatment variable now appears on the right hand side of the ordered choice model.
The ; Rst = ... and ; CML: options for imposing restrictions can be used freely with this
model to constrain b and α. The parameter vector is

Θ = [b1,...,bK,α1,...,αL,µ1,...,µJ-1,ρ].

The usual warning about cross equation restrictions apply. You may also give your own starting
values with ; Start = list ..., though the internal values will usually be preferable.

N14.4.2 Saved Results

All results kept for the basic model are also kept; b and varb still include only b, but ; Par
adds all of [µ,α,ρ] to the parameter vector. This model adds two additional scalars:

rho = estimate of ρ,
varrho = estimate of asymptotic variance of estimated ρ.

NOTE: The estimates of α update the estimates you stored with ; Hold when you fit the probit
model. Thus, for example, if you were to follow your ORDERED command immediately with the
identical command, the starting values used for α would be the MLEs from the prior ordered probit
command, not the ones from the original probit model that you fit earlier. Also, if you were to
follow this model command with a SELECTION model command, this estimate of α would be used
there, as well.
N14: Extended Ordered Choice Models N-225

With the corrected estimates of [b,µ] in hand, predictions for this model are computed in the same
manner as for the basic model without selection. The only difference is that no prediction for y is
computed in the selection model if d = 0.
The PARTIAL EFFECTS and SIMULATE commands are not available for these two
specifications (because they only operate on single equation models). An internal program for
partial effects is provided. An application below illustrates.

N14.4.3 Applications
To illustrate the computations of this model, we have fit an equation for insurance purchase,
then followed with an equation for health satisfaction in which insurance is taken to be a selection
mechanism. The treatment effects formulation is shown later.

PROBIT ; Lhs = public ; Rhs = one,age,hhninc,hhkids ; Hold $

ORDERED ; Lhs = newhsat ; Rhs = one,age,educ,hhninc,female
; Selection
; Partial Effects $

This is the initial probit equation.

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable PUBLIC
Log likelihood function -1868.84461
Restricted log likelihood -1976.59009
Chi squared [ 3 d.f.] 215.49097
Significance level .00000
McFadden Pseudo R-squared .0545108
Estimation based on N = 6209, K = 4
Inf.Cr.AIC = 3745.689 AIC/N = .603
Results retained for SELECTION model.
Hosmer-Lemeshow chi-squared = 46.95244
P-value= .00000 with deg.fr. = 8
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
PUBLIC| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 1.24898*** .13551 9.22 .0000 .98339 1.51458
AGE| .01695*** .00285 5.96 .0000 .01137 .02253
HHNINC| -1.73406*** .12491 -13.88 .0000 -1.97889 -1.48923
HHKIDS| -.07027 .04906 -1.43 .1521 -.16643 .02589
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-226

This ordered probit model is fit using the selected observations to obtain starting values for the full
model.
-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable NEWHSAT
Log likelihood function -13609.65952
Estimation based on N = 6209, K = 14
Inf.Cr.AIC =27247.319 AIC/N = 4.388
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.80968*** .11725 23.96 .0000 2.57986 3.03949
AGE| -.02310*** .00153 -15.13 .0000 -.02609 -.02011
EDUC| .04028*** .00808 4.99 .0000 .02445 .05611
HHNINC| .24424*** .08883 2.75 .0060 .07015 .41833
FEMALE| -.16710*** .02850 -5.86 .0000 -.22295 -.11124
|Threshold parameters for index
Mu(1)| .20275*** .02260 8.97 .0000 .15846 .24703
Mu(2)| .55416*** .02389 23.20 .0000 .50735 .60098
Mu(3)| .88530*** .02158 41.03 .0000 .84301 .92759
Mu(4)| 1.16592*** .01973 59.10 .0000 1.12726 1.20459
Mu(5)| 1.75777*** .01743 100.82 .0000 1.72360 1.79194
Mu(6)| 2.04344*** .01695 120.56 .0000 2.01022 2.07667
Mu(7)| 2.45759*** .01729 142.18 .0000 2.42371 2.49147
Mu(8)| 3.11320*** .01946 160.01 .0000 3.07507 3.15133
Mu(9)| 3.53306*** .02325 151.96 .0000 3.48749 3.57863
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

This is the full information maximum likelihood estimate of the full model
-----------------------------------------------------------------------------
Ordered Probit Model with Selection.
Dependent variable NEWHSAT
Log likelihood function -13607.57507
Restricted log likelihood -13609.65952
Chi squared [ 1 d.f.] 4.16889
Significance level .04117
McFadden Pseudo R-squared .0001532
Estimation based on N = 6209, K = 19
Inf.Cr.AIC =27253.150 AIC/N = 4.389
--------+--------------------------------------------------------------------
PUBLIC| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.57206*** .16019 16.06 .0000 2.25809 2.88604
AGE| -.01972*** .00194 -10.15 .0000 -.02353 -.01591
EDUC| .04014*** .00784 5.12 .0000 .02478 .05550
HHNINC| -.06053 .12872 -.47 .6382 -.31282 .19176
FEMALE| -.16256*** .02716 -5.99 .0000 -.21579 -.10933
N14: Extended Ordered Choice Models N-227

|Threshold parameters for index

Mu(1)| .19073*** .02687 7.10 .0000 .13807 .24340
Mu(2)| .52241*** .04182 12.49 .0000 .44044 .60437
Mu(3)| .83633*** .05229 15.99 .0000 .73385 .93881
Mu(4)| 1.10353*** .06012 18.35 .0000 .98569 1.22137
Mu(5)| 1.67048*** .07410 22.54 .0000 1.52524 1.81572
Mu(6)| 1.94557*** .07952 24.47 .0000 1.78972 2.10142
Mu(7)| 2.34576*** .08663 27.08 .0000 2.17597 2.51554
Mu(8)| 2.98257*** .09539 31.27 .0000 2.79561 3.16953
Mu(9)| 3.39287*** .09921 34.20 .0000 3.19843 3.58731
|Selection equation
Constant| 1.33407*** .13228 10.09 .0000 1.07481 1.59333
AGE| .01525*** .00287 5.32 .0000 .00963 .02087
HHNINC| -1.72207*** .09850 -17.48 .0000 -1.91514 -1.52901
HHKIDS| -.10648** .04594 -2.32 .0205 -.19653 -.01643
|Cor[u(probit),e(ordered probit)]
Rho(u,e)| .50973*** .14253 3.58 .0003 .23038 .78908
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The FIML results provide two test statistics for ‘selectivity.’ The z statistic on the estimate of ρ is
3.58, which is well over the critical value of 1.96. The likelihood ratio test can be carried out using
the initial results for the full model. The restricted value in

Log likelihood function -13607.57507

Restricted log likelihood -13609.65952

is based on the separate probit and ordered probit equations, which corresponds to the model with
ρ = 0. The LR statistic would be 2(-13607.57507 - (-13609.65952) = 4.169. The critical chi squared
with one degree of freedom would be 3.84, so the null hypothesis is rejected again.
A table of partial effects for the conditional model is produced for each outcome. Only the
last one is shown here.

-----------------------------------------------------------------------------
Partial effects of variables on P[NEWHSAT = 10|PUBLIC = 1]
--------+--------------------------------------------------------------------
PUBLIC| Partial Standard Prob. 95% Confidence
NEWHSAT| Effect Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Direct partial effect in ordered choice equation
AGE| -.00245*** .00033 -7.45 .0000 -.00310 -.00181
EDUC| .00499*** .00104 4.82 .0000 .00296 .00702
HHNINC| -.00753 .01591 -.47 .6360 -.03872 .02365
FEMALE| -.02022*** .00367 -5.52 .0000 -.02741 -.01304
|Indirect partial effect in sample selection equation
AGE| .00052*** .00016 3.19 .0014 .00020 .00084
HHNINC| -.05896*** .01285 -4.59 .0000 -.08414 -.03378
HHKIDS| -.00365** .00169 -2.16 .0307 -.00695 -.00034
|Full partial effect = direct effect + indirect effect
AGE| -.00193*** .00046 -4.17 .0000 -.00284 -.00102
HHNINC| -.06649** .02627 -2.53 .0114 -.11799 -.01499
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-228

The treatment effects model is obtained by adding public to the ; Rhs specification in the
ORDERED command and ; All to the command.
-----------------------------------------------------------------------------
Treatment Effects Model: Treatment=PUBLIC
Dependent variable NEWHSAT
Log likelihood function -14765.42035
Restricted log likelihood -14770.39033
Chi squared [ 1 d.f.] 9.93996
Significance level .00162
McFadden Pseudo R-squared .0003365
Estimation based on N = 6209, K = 20
Inf.Cr.AIC =29570.841 AIC/N = 4.763
--------+--------------------------------------------------------------------
PUBLIC| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.27014*** .22312 10.17 .0000 1.83283 2.70746
AGE| -.02027*** .00154 -13.13 .0000 -.02330 -.01724
EDUC| .03917*** .00692 5.66 .0000 .02561 .05273
HHNINC| .06610 .09022 .73 .4638 -.11072 .24292
FEMALE| -.14568*** .02612 -5.58 .0000 -.19687 -.09450
PUBLIC| .34172** .13586 2.52 .0119 .07544 .60801
|Threshold parameters for index
Mu(1)| .19408*** .02587 7.50 .0000 .14337 .24479
Mu(2)| .52700*** .03637 14.49 .0000 .45572 .59828
Mu(3)| .85528*** .04110 20.81 .0000 .77471 .93584
Mu(4)| 1.13190*** .04397 25.74 .0000 1.04573 1.21808
Mu(5)| 1.70234*** .04863 35.01 .0000 1.60703 1.79766
Mu(6)| 1.97911*** .05078 38.98 .0000 1.87959 2.07864
Mu(7)| 2.38797*** .05406 44.17 .0000 2.28201 2.49393
Mu(8)| 3.02974*** .05925 51.13 .0000 2.91361 3.14587
Mu(9)| 3.45667*** .06272 55.12 .0000 3.33375 3.57959

|Index function for probit equation

Constant| 1.26527*** .13081 9.67 .0000 1.00889 1.52164
AGE| .01641*** .00282 5.83 .0000 .01090 .02193
HHNINC| -1.68223*** .10083 -16.68 .0000 -1.87986 -1.48459
HHKIDS| -.09807** .04589 -2.14 .0326 -.18802 -.00812
|Cor[u(probit),e(ordered probit)]
Rho(1,2)| .41059*** .08110 5.06 .0000 .25164 .56955
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-229

N14.5 Hierarchical Ordered Probit Models

The hierarchical ordered probit model (or generalized ordered probit model) is a univariate
ordered probit model in which the threshold parameters depend on variables. (We opt for the
acronym HOPIT model as slightly more melodious than GOPIT. In the original proposal of this
model (Pudney and Shields (2000)), the thresholds were modeled as linear functions of the data,
producing the model
y* = b′x + e
y = 0 if y* < 0,
= 1 if 0 < y* < µ1,
= 2 if µ1 < y* < µ2,
...
µj = δj′z.

(There is no disturbance on the equation for the threshold variables.) The model has an inherent
identification problem, because in

Prob[y = j] = Φ(µj - b′x) - Φ(µj-1 - b′x),

if x and z have variables in common, then (with a sign change) the same model is produced whether
the common variable appears in µj or b′x. (Pudney and Shields note and discuss this.) The NLOGIT
implementation avoids this indeterminacy by using a different functional form. (That does imply
that we achieve identification through functional form.)
Two forms of the model are provided.

Form 1: µj = exp(qj + δ′z)

Form 2: µj = exp(qj + δj′z)

Note that in form 1, each µj has a different constant term, but the same coefficient vector, while in
form 2, each threshold parameter has its own parameter vector. (We note, for purposes of
estimation, it is always necessary for µj to be greater than µj-1. We are able to impose that on form 1
fairly easily by parameterizing qj in a way that does so. However, for form 2, this is much more
difficult to obtain, and users should expect to see diagnostics about unordered thresholds when they
use form 2.) The threshold coefficients will be difficult to compare between the original ordered
probit model and form 2 of the HOPIT model. For form 1, the model reverts to the unmodified
ordered probit model if the single vector δ equals 0.
The command for this model augments the usual ordered probit command with the
specification for the thresholds,

ORDERED ; Lhs = ... ; Rhs = ...

; HO1 = list of variables or ; HO2 = list of variables $
N14: Extended Ordered Choice Models N-230

The list of variables in the HO1 or HO2 part must not contain a constant term (one). All other
options for the ordered probit model are exactly as described previously, including fitted values,
restrictions, marginal effects, and so on, unchanged. This form of the ordered probit model can also
be combined with the sample selection corrected ordered probit model described in Section N14.3.
In the example below, the model is first fit to the health satisfaction variable with no
modification to the thresholds. In the HOPIT model fit next, the thresholds vary with whether or not
the family has kids in the household and with the number of types of insurance they have. For
purpose of a limited example, we use a subset of the sample.

SAMPLE ; All $
CREATE ; insuranc = public + addon $
ORDERED ; Lhs = hsat ; Rhs = one,age,educ,female,hhninc
; Partial Effects $
ORDERED ; Lhs = hsat ; Rhs = one,age,educ,female,hhninc
; HO1 = hhkids,insuranc
; Partial Effects $

These are the estimates for the base case. (We have omitted the partial effects.)

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HSAT
Log likelihood function -56876.85183
Restricted log likelihood -57836.42214
Chi squared [ 4 d.f.] 1919.14061
Significance level .00000
McFadden Pseudo R-squared .0165911
Estimation based on N = 27326, K = 14
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.68410*** .04392 61.12 .0000 2.59802 2.77018
AGE| -.02096*** .00056 -37.71 .0000 -.02205 -.01987
EDUC| .03341*** .00284 11.76 .0000 .02784 .03898
FEMALE| -.05800*** .01259 -4.61 .0000 -.08268 -.03332
HHNINC| .26478*** .03631 7.29 .0000 .19362 .33594
|Threshold parameters for index
Mu(1)| .19340*** .01002 19.30 .0000 .17376 .21305
Mu(2)| .49929*** .01087 45.93 .0000 .47799 .52060
Mu(3)| .83548*** .00990 84.39 .0000 .81608 .85489
Mu(4)| 1.10462*** .00908 121.63 .0000 1.08682 1.12242
Mu(5)| 1.66162*** .00801 207.44 .0000 1.64592 1.67732
Mu(6)| 1.93021*** .00774 249.46 .0000 1.91504 1.94537
Mu(7)| 2.33753*** .00777 300.92 .0000 2.32230 2.35275
Mu(8)| 2.99283*** .00851 351.70 .0000 2.97615 3.00951
Mu(9)| 3.45210*** .01017 339.31 .0000 3.43216 3.47204
--------+--------------------------------------------------------------------
N14: Extended Ordered Choice Models N-231

These are the estimates for the HO1 hierarchical model.

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HSAT
Log likelihood function -56868.23498
Restricted log likelihood -57836.42214
Chi squared [ 4 d.f.] 1936.37431
Underlying probabilities based on Normal
HOPIT (covariates in thresholds) model
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.66036*** .04828 55.10 .0000 2.56573 2.75499
AGE| -.02035*** .00058 -35.09 .0000 -.02149 -.01921
EDUC| .03313*** .00293 11.30 .0000 .02738 .03887
FEMALE| -.06072*** .01259 -4.83 .0000 -.08539 -.03606
HHNINC| .26373*** .03648 7.23 .0000 .19222 .33523
|Estimates of t(j) in mu(j)=exp[t(j)+d*z]
Theta(1)| -1.62461*** .06134 -26.49 .0000 -1.74484 -1.50439
Theta(2)| -.67653*** .03254 -20.79 .0000 -.74029 -.61276
Theta(3)| -.16186*** .02193 -7.38 .0000 -.20485 -.11888
Theta(4)| .11739*** .01750 6.71 .0000 .08309 .15170
Theta(5)| .52583*** .01258 41.79 .0000 .50117 .55049
Theta(6)| .67578*** .01122 60.25 .0000 .65379 .69776
Theta(7)| .86747*** .00979 88.62 .0000 .84828 .88665
Theta(8)| 1.11497*** .00843 132.20 .0000 1.09844 1.13150
Theta(9)| 1.25794*** .00787 159.74 .0000 1.24250 1.27337
|Threshold covariates mu(j)=exp[t(j)+d*z]
HHKIDS| -.01830*** .00526 -3.48 .0005 -.02862 -.00799
INSURANC| .15082D-04** .5872D-05 2.57 .0102 .35726D-05 .26592D-04
--------+--------------------------------------------------------------------
Note: nnnnn.D-xx or D+xx => multiply by 10 to -xx or +xx.
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
(Partial Effects for outcomes 0 – 9 are omitted)
-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HSAT| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=10] at means]--------------
AGE| -.00377*** -1.52276 -11.54 .0000 -.00441 -.00313
EDUC| .00614*** .64474 9.12 .0000 .00482 .00746
*FEMALE| -.01123 -.10424 -.50 .6182 -.05541 .03294
HHNINC| .04887*** .15964 3.51 .0004 .02161 .07613
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-232

N14.6 Zero Inflated Ordered Probit (ZIOP, ZIHOP) Models

Harris and Zhao (2007) have developed a zero inflated ordered probit (ZIOP) counterpart to
the zero inflated Poisson model. The ZIOP formulation would appear

d* = α′w + u, d = 1 (d* > 0)

y* = b′x + e, y = 0 if y* < 0 or d = 0
1 if 0 < y* < µ1 and d = 1,
2 if µ1 < y* < µ2 and d = 1,
and so on.

The first equation is assumed to be a probit model (based on the normal distribution) – this estimator
does not support a logit formulation. The correlation between u and e is ρ, which by default equals
zero, but may be estimated instead. The latent class nature of the formulation has the effect of
inflating the number of observed zeros, even if u and e are uncorrelated. The model with correlation
between u and e is an optional specification that analysts might want to test. The zero inflation
model may also be combined with the hierarchical (generalized) model discussed in the previous
section. Thus, it might also be specified as part of the model that

Form 1: µj = exp(qj + δ′z)

Form 2: µj = exp(qj + δj′z)

The command structure for ZIOP and ZIHOP models are

PROBIT ; Lhs = d ; Rhs = variables in w ; Hold $

ORDERED ; Lhs = y ; Rhs = variables in x
; ZIOP $

This form of the model imposes ρ = 0. To allow the correlation to be a free parameter, add

; Correlation

to the command.

NOTE: The ; HO1 and ; HO2 specifications discussed in the preceding section may also be used
with this model.

In the example below, we continue the analysis of the health care data. The (artificial)
model has the zero inflation probability based on the presence of ‘public’ insurance while the
ordered outcome continues to be the self reported health satisfaction. Here, we have used the entire
sample of 27,236 observations.
N14: Extended Ordered Choice Models N-233

The commands are:

SAMPLE ; All $
PROBIT ; Lhs = public
; Rhs = one,age,hhninc,hhkids,married ; Hold $
ORDERED ; Lhs = hsat
; Rhs = one,age,educ,female
; ZIO ; Correlated $

-----------------------------------------------------------------------------
Binomial Probit Model
Dependent variable PUBLIC
Log likelihood function -9229.32605
Restricted log likelihood -9711.25153
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
PUBLIC| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 1.51862*** .05021 30.25 .0000 1.42022 1.61702
AGE| .00553*** .00105 5.26 .0000 .00347 .00759
HHNINC| -1.55524*** .05120 -30.37 .0000 -1.65560 -1.45489
HHKIDS| -.08320*** .02370 -3.51 .0004 -.12966 -.03675
MARRIED| .10035*** .02694 3.72 .0002 .04754 .15316
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HSAT
Log likelihood function -56903.42663
Restricted log likelihood -57836.42214
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.70343*** .04379 61.73 .0000 2.61760 2.78926
AGE| -.02078*** .00056 -37.41 .0000 -.02186 -.01969
EDUC| .03881*** .00274 14.16 .0000 .03344 .04419
FEMALE| -.05742*** .01259 -4.56 .0000 -.08210 -.03274
|Threshold parameters for index
Mu(1)| .19279*** .00999 19.29 .0000 .17320 .21238
Mu(2)| .49771*** .01085 45.88 .0000 .47645 .51896
Mu(3)| .83298*** .00989 84.26 .0000 .81361 .85236
Mu(4)| 1.10156*** .00907 121.43 .0000 1.08378 1.11934
Mu(5)| 1.65744*** .00800 207.07 .0000 1.64175 1.67313
Mu(6)| 1.92551*** .00773 249.00 .0000 1.91036 1.94067
Mu(7)| 2.33231*** .00776 300.37 .0000 2.31709 2.34753
Mu(8)| 2.98735*** .00851 351.12 .0000 2.97067 3.00402
Mu(9)| 3.44694*** .01018 338.75 .0000 3.42700 3.46688
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N14: Extended Ordered Choice Models N-234

-----------------------------------------------------------------------------
Zero Inflated Ordered Probit Model.
Dependent variable HSAT
Log likelihood function -56895.22719
Restricted log likelihood -56903.42663
--------+--------------------------------------------------------------------
PUBLIC| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.77007*** .04944 56.03 .0000 2.67317 2.86697
AGE| -.02150*** .00057 -37.68 .0000 -.02262 -.02038
EDUC| .03769*** .00284 13.27 .0000 .03212 .04325
FEMALE| -.05844*** .01255 -4.66 .0000 -.08304 -.03384
|Threshold parameters for index
Mu(1)| .19868*** .01235 16.08 .0000 .17447 .22289
Mu(2)| .50918*** .01694 30.05 .0000 .47597 .54239
Mu(3)| .84768*** .01897 44.70 .0000 .81051 .88486
Mu(4)| 1.11767*** .01978 56.50 .0000 1.07890 1.15644
Mu(5)| 1.67504*** .02062 81.25 .0000 1.63463 1.71545
Mu(6)| 1.94359*** .02087 93.15 .0000 1.90269 1.98449
Mu(7)| 2.35098*** .02119 110.97 .0000 2.30946 2.39251
Mu(8)| 3.00678*** .02174 138.30 .0000 2.96417 3.04939
Mu(9)| 3.46677*** .02222 156.00 .0000 3.42322 3.51033
|Zero inflation probit probability
Constant| -.30749 1.71064 -.18 .8573 -3.66028 3.04530
AGE| .10718 .06555 1.63 .1021 -.02131 .23566
HHNINC| -.19155 .62143 -.31 .7579 -1.40954 1.02644
HHKIDS| -.59894** .24410 -2.45 .0141 -1.07737 -.12051
MARRIED| 1.06982 .94393 1.13 .2571 -.78024 2.91988
|Cor[u(probit),e(ordered probit)]
Rho(u,e)| -.90968 1.40561 -.65 .5175 -3.66462 1.84525
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N14.7 Bivariate Ordered Probit and Polychoric Correlation

The bivariate ordered probit model is analogous to the SUR model for the ordered probit
case:
yji* = bj′xji + eji
yji = 0 if yji* < 0,
1 if 0 < yji* < µ1,
2, ... and so on, j = 1,2,

for a pair of ordered probit models that are linked by Cor(e1i,e2i) = ρ. The model can be estimated
one equation at a time using the results described earlier. Full efficiency in estimation and an
estimate of ρ are achieved by full information maximum likelihood estimation. NLOGIT’s
implementation of the model uses FIML, rather than GMM. Either variable (but not both) may be
binary. If both are binary, the bivariate probit model should be used. (The development here draws
on Butler and Chatterjee (1997) who analyzed maximum likelihood and GMM estimators for the
bivariate extension of the ordered probit model.)
N14: Extended Ordered Choice Models N-235

The command structure requires prior estimation of the two univariate models to provide
starting values for the iterations. The third command then fits the bivariate model. We assume that
the first variable is multinomial.

ORDERED ; Lhs = y1 ; Rhs = ... $

MATRIX ; b1 = b ; mu1 = mu $

Use one of the following. If the second variable has more than two outcomes, use

ORDERED ; Lhs = y2 ; Rhs = ... $

MATRIX ; b2 = b ; mu2 = mu $

If the second variable is binary, use

PROBIT ; Lhs = y2 ; Rhs = ... $

MATRIX ; b2 = b $

Then, estimate the bivariate model with

ORDERED ; Lhs = y1,y2 ; Rh1 = ... ; Rh2 = ...

; Start = b1,mu1,b2,mu2, 0 $

The variable mu2 is omitted if y2 is binary. The final zero in the list of starting values is for ρ. You
may use some other value if you have one.
The standard options for estimation are available (iteration controls, technical output, cluster
corrections, etc.). You may also retain fitted values with ; Keep = yf1,yf2 (note that both names are
provided). Probabilities for the joint observed outcome are retained with ; Prob = name. Listings
of probabilities for outcomes are obtained with ; List as usual.
To illustrate the estimator, we use the health care utilization data analyzed earlier. The two
outcomes are y1 = health care satisfaction, taking values 0 to 5 (we reduced the sample) and y2 = the
number of types of health care insurance. Results for a bivariate ordered probit model appear below.
The initial univariate models are omitted.

SAMPLE ; All $
REJECT ; newhsat > 5 | _groupti < 7 $
ORDERED ; Lhs = newhsat ; Rhs = one,age,educ,female,hhninc $
MATRIX ; b1 = b ; mu1 = mu $
CREATE ; insuranc = public + addon $
CROSSTAB ; Lhs = newhsat ; Rhs = insuranc $
ORDERED ; Lhs = insuranc ; Rhs = one,age,educ,hhninc,hhkids $
MATRIX ; b2 = b ; mu2 = mu $
ORDERED ; Lhs = newhsat,insuranc
; Rh1 = one,age,educ,female,hhninc
; Rh2 = one,age,educ,hhninc,hhkids
; Start = b1,mu1,b2,mu2,0 $
N14: Extended Ordered Choice Models N-236

-----------------------------------------------------------------------------
Bivariate Ordered Probit Model
Dependent variable BivOrdPr
Log likelihood function -3099.59435
Restricted log likelihood -3100.36600
--------+--------------------------------------------------------------------
NEWHSAT| Standard Prob. 95% Confidence
INSURANC| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for Probability Model for NEWHSAT
Constant| 1.98379*** .23742 8.36 .0000 1.51846 2.44913
AGE| -.01233*** .00288 -4.28 .0000 -.01797 -.00668
EDUC| .01815 .01667 1.09 .2762 -.01452 .05082
FEMALE| .09626* .05301 1.82 .0694 -.00764 .20016
HHNINC| .13547 .17765 .76 .4457 -.21271 .48365
|Index function for Probability Model for INSURANC
Constant| 2.57737*** .38142 6.76 .0000 1.82980 3.32493
AGE| .01847*** .00609 3.03 .0024 .00654 .03040
EDUC| -.13925*** .02090 -6.66 .0000 -.18022 -.09828
HHNINC| -.63131* .33803 -1.87 .0618 -1.29383 .03121
HHKIDS| -.01720 .10527 -.16 .8702 -.22353 .18912
|Threshold Parameters for Probability Model for NEWHSAT
MU(01)| .24263*** .03171 7.65 .0000 .18048 .30479
MU(02)| .67851*** .04404 15.41 .0000 .59220 .76483
MU(03)| 1.15093*** .04917 23.41 .0000 1.05456 1.24730
MU(04)| 1.61433*** .05193 31.09 .0000 1.51255 1.71611
|Threshold Parameters for Probability Model for INSURANC
LMDA(01)| 4.07012*** .09615 42.33 .0000 3.88168 4.25856
|Disturbance Correlation = RHO(1,2)
RHO(1,2)| -.06225 .06013 -1.04 .3005 -.18010 .05560
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

+-----------------------------------------------------------------+
|Cross Tabulation |
|Row variable is NEWHSAT (Out of range 0-49: 0) |
|Number of Rows = 6 (NEWHSAT = 0 to 5) |
|Col variable is INSURANC (Out of range 0-49: 0) |
|Number of Cols = 3 (INSURANC = 0 to 2) |
|Chi-squared independence tests: |
|Chi-squared[ 10] = 17.61732 Prob value = .06177 |
|G-squared [ 10] = 27.62274 Prob value = .00207 |
+-----------------------------------------------------------------+
| INSURANC |
+--------+---------------------+------+ |
| NEWHSAT| 0 1 2| Total| |
+--------+---------------------+------+ |
| 0| 2 87 0| 89| |
| 1| 1 54 0| 55| |
| 2| 0 156 2| 158| |
| 3| 14 250 3| 267| |
| 4| 22 307 7| 336| |
| 5| 59 963 12| 1034| |
+--------+---------------------+------+ |
| Total| 98 1817 24| 1939| |
+-----------------------------------------------------------------+
N14: Extended Ordered Choice Models N-237

Polychoric Correlation

The polychoric correlation coefficient is used to quantify the correlation between discrete
variables that are qualitative measures. The standard interpretation is that the discrete variables are
discretized counterparts to underlying quantitative measures. We typically use ordered probit
models to analyze such data. The polychoric correlation measures the correlation between
y1 = 0,1,...,J1 and y2 = 0,1,...,J2. (Note, J1 need not equal J2.) One of the two variables may be binary
as well.
By this description, the polychoric correlation is simply the correlation coefficient in the
bivariate ordered probit model when the two equations contain only constant terms. Thus, to
compute the polychoric correlation for a pair of qualitative variables, you can use NLOGIT’s
bivariate ordered probit model. The commands are as follows: The first two model commands
compute the starting values, and the final one computes the correlation.

ORDERED ; Lhs = y1 ; Rhs = one $

MATRIX ; b1 = b ; mu1 = mu $
ORDERED ; Lhs = y2 ; Rhs = one $
MATRIX ; b2 = b ; mu2 = mu $

or PROBIT ; Lhs = y2 ; Rhs = one $

MATRIX ; b2 = b $

Then, ORDERED ; Lhs = y1,y2 ; Rh1 = one ; Rh2 = one

; Start = b1,mu1,b2,mu2,0 $

For a simple example, we compute the polychoric correlation between self reported health
status and sex in the health care usage data examined earlier. Results appear below. Note that the
‘model’ for sex is simply a computational device.

ORDERED ; Lhs = newhsat ; Rhs = one $

MATRIX ; b1 = b ; mu1 = mu $
PROBIT ; Lhs = female ; Rhs = one $
MATRIX ; b2 = b $
ORDERED ; Lhs = newhsat,female
; Rh1 = one ; Rh2 = one ; Start = b1,mu1,b2,0 $
N14: Extended Ordered Choice Models N-238

-----------------------------------------------------------------------------
Bivariate Ordered Probit Model
Dependent variable BivOrdPr
Log likelihood function -3976.40233
Restricted log likelihood -3977.17511
--------+--------------------------------------------------------------------
NEWHSAT| Standard Prob. 95% Confidence
FEMALE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Mean inverse probability for NEWHSAT
Constant| 1.68575*** .04935 34.16 .0000 1.58903 1.78248
|Mean inverse probability for FEMALE
Constant| .05109* .02849 1.79 .0729 -.00475 .10693
|Threshold Parameters for Probability Model for NEWHSAT
MU(01)| .24123*** .03150 7.66 .0000 .17950 .30296
MU(02)| .67373*** .04341 15.52 .0000 .58864 .75882
MU(03)| 1.14226*** .04824 23.68 .0000 1.04770 1.23681
MU(04)| 1.60213*** .05087 31.49 .0000 1.50242 1.70184
|Polychoric Correlation for NEWHSAT and FEMALE
RHO(1,2)| .03998 .03216 1.24 .2138 -.02305 .10302
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-239

N15: Panel Data Models for Ordered Choice

N15.1 Introduction
The basic ordered choice model is based on the following specification: There is a latent
regression,
yi* = b′xi + ei, ei~ F(ei |q), E[ei|xi] = 0, Var[ei|xi] = 1.

The observation mechanism results from a complete censoring of the latent dependent variable as
follows:
yi = 0 if yi ≤ µ0,
= 1 if µ0 < yi ≤ µ1,
= 2 if µ1 < yi ≤ µ2,
...
= J if yi > µJ-1.

The latent ‘preference’ variable, yi* is not observed. The observed counterpart to yi* is yi. Four
stochastic specifications are provided for the basic model shown above. The ordered probit model
based on the normal distribution was developed by Zavoina and McElvey (1975). It applies in
applications such as surveys, in which the respondent expresses a preference with the above sort of
ordinal ranking. The variance of ei is assumed to be one, since as long as yi*, b, and ei are
unobserved, no scaling of the underlying model can be deduced from the observed data. Estimates
are obtained by maximum likelihood. The probabilities which enter the log likelihood function are

Prob[yi = j] = Prob[yi* is in the jth range].

The model may be estimated either with individual data, with yi = 0, 1, 2, ... or with grouped data, in
which case each observation consists of a full set of J+1 proportions, p0i,...,pJi. This chapter gives the
panel data extensions of the ordered choice model.

NOTE: The panel data versions of the ordered choice models require individual data.

There are four classes of panel data models in NLOGIT, fixed effects, random effects,
random parameters, and latent class.
N15: Panel Data Models for Ordered Choice N-240

N15.2 Fixed Effects Ordered Choice Models

The fixed effects models are estimated by maximum likelihood. The command for
requesting the model is in two parts. You must fit the model without fixed effects first, to provide
the starting values, then the command for the fixed effects estimator follows. The first command and
the second must be identical, save for the panel specification in the second command and the
constant term in the first, as noted below.

ORDERED ; Lhs = dependent variable

; Rhs = independent variables
[ ; Model = Logit] $
ORDERED ; Lhs = dependent variable
; Rhs = independent variables
; Pds = fixed number of periods or count variable
; Fixed Effects
[ ; Model = Logit] $

NOTE: The Rhs in your first command must contain a constant term, one as the first variable. Your
Rhs list for a fixed effects model generally should not include a constant term as the fixed effects
model fits a complete set of constants for the set of groups. But, for the ordered probit model, you
must provide the identical Rhs list as in the first command, so for this model, do include one. It will
be removed prior to beginning estimation. When you set up your commands, leaving one in the Rhs
list will help insure that your model specification is correct. It will look correct. Note, it is crucial
that you fit the pooled model first so that NLOGIT can find the right starting values for the second
estimation step.

The fixed effects model assumes a group specific effect:

Prob[yit = j] = F( j,µ, b′xit + αi)

where αi is the parameter to be estimated. You may also fit a two way fixed effects model

Prob[yit = j] = F( j,µ, b′xit + αi + γt)

where γt is an additional, time (period) specific effect. The time specific effect is requested by
adding
; Time

to the command if the panel is balanced, and

; Time = variable name

if the panel is unbalanced. For the unbalanced panel, we assume that overall, the sample observation
period is t = 1,2,..., T and that the ‘Time’ variable gives for the specific group, the particular values
of t that apply to the observations. Thus, suppose your overall sample is five periods. The first
group is three observations, periods 1, 2, 4, while the second group is four observations, 2, 3, 4, 5.
N15: Panel Data Models for Ordered Choice N-241

Then, your panel specification would be

; Pds = Ti for example, where Ti = 3, 3, 3, 4, 4, 4, 4

and ; Time = Pd for example, where Pd = 1, 2, 4, 2, 3, 4, 5.

NOTE: See the discussion below on how this model is estimated. It places an important restriction
on the two way fixed effects model.

You must provide the starting values for the iterations by fitting the basic model without
fixed effects. You will have a constant term in these results even though it is dropped from the fixed
effects model. This is used to get the starting value for the fixed effects. Iterations begin with the
restricted model that forces all the fixed effects to equal the constant term in the restricted model.
Results that are kept for this model are

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of b.
alphafe = estimated fixed effects

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Last Model: b_variables

Last Function: None

The upper limit on the number of groups is 100,000.

NOTE: In the ordered probit model with fixed effects αi, the individual effect coefficient cannot be
estimated if the dependent variable within the group takes the same value in every period. The
results will indicate how many such groups had to be removed from the sample.

Application

We have fit a fixed effects ordered probit model with the German health care data used in
the previous examples. This is an unbalanced panel with 7,293 individuals. The health status
variable is coded 0 to 10. The model is fit using the commands below. We first fit the pooled
model, then the fixed effects model.

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
ORDERED ; Lhs = newhsat
; Rhs = one,hhninc,hhkids,educ ; Partial Effects $
ORDERED ; Lhs = newhsat
; Rhs = one,hhninc,hhkids,educ ; Partial Effects
; Fixed Effects ; Pds = _groupti $
N15: Panel Data Models for Ordered Choice N-242

-----------------------------------------------------------------------------
FIXED EFFECTS OrdPrb Model
Dependent variable NEWHSAT
Log likelihood function -42217.91813
Estimation based on N = 27326, K =5679
Inf.Cr.AIC =95793.836 AIC/N = 3.506
Probability model based on Normal
Unbalanced panel has 7293 individuals
Skipped 1626 groups with inestimable ai
Ordered probability model
Ordered probit (normal) model
LHS variable = values 0,1,...,10
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
HHNINC| -.38858*** .06374 -6.10 .0000 -.51351 -.26365
HHKIDS| .07337*** .02718 2.70 .0069 .02010 .12665
EDUC| -.04469* .02635 -1.70 .0898 -.09633 .00695
MU(1)| .32638*** .02045 15.96 .0000 .28630 .36646
MU(2)| .84692*** .02743 30.88 .0000 .79316 .90068
MU(3)| 1.39245*** .03005 46.34 .0000 1.33355 1.45135
MU(4)| 1.81634*** .03102 58.55 .0000 1.75554 1.87714
MU(5)| 2.68396*** .03226 83.19 .0000 2.62072 2.74719
MU(6)| 3.10845*** .03272 95.01 .0000 3.04432 3.17258
MU(7)| 3.76428*** .03340 112.69 .0000 3.69880 3.82975
MU(8)| 4.79590*** .03478 137.88 .0000 4.72773 4.86407
MU(9)| 5.50760*** .03610 152.55 .0000 5.43684 5.57836
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The results below compare the estimated partial effects for the outcome y = 10 for the fixed effects
model followed by the pooled model. The differences are large. Note that the educ coefficient is
significantly negative in the fixed effects model and significantly positive in the pooled model. The log
likelihood for the pooled model is -57420.08880, so the LR test statistic is about 30,000 with 7,293
degrees freedom. The critical chi squared for 7,292 degrees of freedom, given with the command

CALC ; List ; Ctb(.95,7292) $

is 7,491, which suggests that the fixed effects estimator, at least at this point is preferred. The
remains some question, however, because of the incidental parameters problem. Based on received
results, in the OP setting, the coefficient is biased away from zero, but not in sign, which still weighs
in favor of the FEM result.
N15: Panel Data Models for Ordered Choice N-243

-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
NEWHSAT| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=10] at means]--------------
HHNINC| .00025 .52441 .93 .3532 -.00028 .00078
*HHKIDS| .00469 .17144 1.46 .1431 -.00159 .01097
EDUC| -.00282*** -1.16548 -10.59 .0000 -.00334 -.00230
|--------------[Partial effects on Prob[Y=10] at means]--------------
HHNINC| .03739*** .11620 5.36 .0000 .02372 .05105
*HHKIDS| .04378*** .38649 16.73 .0000 .03865 .04891
EDUC| .00996*** .99545 18.30 .0000 .00889 .01103
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: nnnnn.D-xx or D+xx => multiply by 10 to -xx or +xx.
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N15.3 Random Effects Ordered Choice Models

The random effects model is

yit* = b‘xit + eit + ui

where i = 1,...,N indexes groups and t = 1,...,Ti indexes periods. (As always, the number of periods
may vary by individual.) The unique term, eit, is distributed as N[0,1], standard logistic, extreme
value, or Gompertz as specified in the general model discussed earlier. The group specific term, ui is
distributed as N[0,σ2] for all cases. Note that the unobserved heterogeneity, ui is the same in every
period. The parameters of the model are fit by maximum likelihood. As in the binary choice
models, the underlying variance, σ2 = σu2 + σe2 is not identified. The reduced form parameter,
ρ = σu2 / ( σε2 + σu2 ) , is estimated directly. With the normalization that we used earlier, σe2 = 1, we can
determine σu = ρ /(1 − ρ) . The ordered probability model with random effects is estimated in the
same fashion as the binary probability models with random effects. The heterogeneity is handled by
using Hermite quadrature to integrate the effect out of the joint density of the Ti observations for the
ith group. Technical details appear at the end of this section.
N15: Panel Data Models for Ordered Choice N-244

N15.3.1 Commands
The specification is for the ordered probability model. Use

ORDERED ; Lhs = ... ; Rhs = ...

; Panel spec.
[ ; Model = Logit, Comploglog, Arctangent or Gompertz] $

where the ; Pds specification follows the standard convention, fixed T or variable name for variable
T. The default is the ordered probit. Request the ordered logit just by adding ; Model = Logit etc. to
the command. The random effects model is the default panel data model for the ordered probability
models, so you need only include the ; Pds specification in the command.

NOTE: The random effect, ui is assumed to be normally distributed in all models. Thus, the logit,
arctangent, and other models contain a hybrid of distributions.

All other options are the same as were listed earlier for the pooled ordered probability
models.
Marginal effects are computed by setting the heterogeneity term, ui to its expected value of
zero. In order to do the computations of the marginal effects, it is also necessary to scale the
coefficients. The ordered probability model with the random effect in the equation is based on the
index function (µj - b′xi) / (1 + σu2).
This estimator can accommodate restrictions, so

; Rst = list
and ; CML: specification

are both available. Restrictions may be tested and imposed exactly as in the model with no
heterogeneity. Since restrictions can be imposed on all parameters, including ρ, you can fix the
value of ρ at any desired value. Do note that forcing the ancillary parameter, in this case, ρ, to equal
a slope parameter will almost surely produce unsatisfactory results, and may impede or even prevent
convergence of the iterations.
Starting values for the iterations are obtained by fitting the basic model without random
effects. Thus, the initial results in the output for these models will be the ordered choice models
discussed earlier. You may provide your own starting values for the parameters with

; Start = ... the list of values for b, values for µ, value for ρ

There is no natural moment based estimator for ρ, so a relatively low guess is used as the starting
value instead. The starting value for ρ is approximately .2 (q = [2ρ/(1-ρ)]1/2 ≈ .29 – see the technical
details below. Maximum likelihood estimates are then computed and reported, along with the usual
diagnostic statistics. (An example appears below.)
N15: Panel Data Models for Ordered Choice N-245

N15.3.2 Output and Results

Your data may not be consistent with the random effects model. That is, there may be no
discernible evidence of random effects in your data. In this case, the estimate of ρ will turn out to be
negligible. If so, the estimation program issues a diagnostic and reverts back to the original,
uncorrelated formulation and reports (again) the results for the basic model.
Results that are kept for this model are

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of b.

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function
rho = estimated value of ρ
varrho = estimated asymptotic variance of estimator of ρ.

Last Model: b_variables

Last Function: Prob(y = outcome | x)

The additional specification

; Par

in the command requests that µ and σu be included in b and the additional rows and columns be
included in varb. The Last Model is [b_variable,ru]. The PARTIAL EFFECTS and SIMULATE
commands use the same probability function as the pooled model. The default outcome is the
highest one, but you may use ; Outcome = j to specify a specific one, or ; Outcome = * for all.

NOTE: The hypothesis of no group effects can be tested with a Wald test (simple t test) or with a
likelihood ratio test. The LM approach, using ; Maxit = 0 with a zero starting value for ρ does not
work in this setting because with ρ = 0, the last row of the covariance matrix turns out to contain
zeros.

NOTE: This model is fit by approximating the necessary integrals in the log likelihood function by
Hermite quadrature. An alternative approach to estimating the same model is by Monte Carlo
simulation. You can do exactly this by fitting the model as a random parameters model with only a
random constant term.
N15: Panel Data Models for Ordered Choice N-246

N15.3.3 Application
In the following example, we fit random effects ordered probit models for the health status
data. The pooled estimator is fit with and without the clustered data correction. Then, the random
effects model is fit, first using the Butler and Moffitt method, then as a random parameters model
with a random constant term.

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
ORDERED ; Lhs = newhsat ; Rhs = one,hhninc,hhkids,educ $
ORDERED ; Lhs = newhsat ; Rhs = one,hhninc,hhkids,educ
; Cluster = id $
ORDERED ; Lhs = newhsat ; Rhs = one,hhninc,hhkids,educ
; Panel $
ORDERED ; Lhs = newhsat ; Rhs = one,hhninc,hhkids,educ $
ORDERED ; Lhs = newhsat ; Rhs = one,hhninc,hhkids,educ
; Panel ; RPM ; Fcn = one(n) ; Halton ; Pts = 25 $

The first pair of estimation results shown below compares the cluster estimator of the covariance
matrix to the pooled estimator which ignores the panel data structure. As can be seen in the results,
the robust standard errors are somewhat higher. The second set of results compares two estimators
of the random effects model. The first results are based on the quadrature estimator. The second uses
maximum simulated likelihood. These two estimators give almost the same results. They would be
closer still had we used a larger number of Halton draws. We set this to 25 to speed up the
computation. With, say, 250, the results of the two estimators would be extremely close.

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable NEWHSAT
Log likelihood function -57420.08880
Restricted log likelihood -57816.35761
Chi squared [ 3 d.f.] 792.53762
Significance level .00000
McFadden Pseudo R-squared .0068539
Estimation based on N = 27326, K = 13
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 1.42634*** .03136 45.48 .0000 1.36487 1.48781
HHNINC| .19469*** .03624 5.37 .0000 .12366 .26571
HHKIDS| .22199*** .01261 17.61 .0000 .19728 .24669
EDUC| .05187*** .00276 18.81 .0000 .04647 .05728
|Threshold parameters for index
Mu(1)| .19061*** .00988 19.29 .0000 .17123 .20998
Mu(2)| .49125*** .01073 45.80 .0000 .47023 .51228
Mu(3)| .82152*** .00979 83.95 .0000 .80233 .84070
Mu(4)| 1.08609*** .00898 120.91 .0000 1.06849 1.10370
Mu(5)| 1.63179*** .00793 205.69 .0000 1.61624 1.64734
Mu(6)| 1.88965*** .00767 246.35 .0000 1.87462 1.90469
N15: Panel Data Models for Ordered Choice N-247

Mu(7)| 2.28993*** .00770 297.40 .0000 2.27484 2.30503

Mu(8)| 2.92948*** .00843 347.32 .0000 2.91295 2.94601
Mu(9)| 3.38076*** .01008 335.50 .0000 3.36101 3.40051
|Index function for probability
Constant| 1.42634*** .05039 28.30 .0000 1.32757 1.52511
HHNINC| .19469*** .05008 3.89 .0001 .09653 .29284
HHKIDS| .22199*** .01886 11.77 .0000 .18503 .25894
EDUC| .05187*** .00432 12.00 .0000 .04340 .06035
|Threshold parameters for index
Mu(1)| .19061*** .02054 9.28 .0000 .15035 .23086
Mu(2)| .49125*** .03180 15.45 .0000 .42892 .55358
Mu(3)| .82152*** .03548 23.16 .0000 .75198 .89105
Mu(4)| 1.08609*** .03432 31.64 .0000 1.01882 1.15337
Mu(5)| 1.63179*** .03334 48.95 .0000 1.56644 1.69713
Mu(6)| 1.88965*** .03261 57.95 .0000 1.82574 1.95357
Mu(7)| 2.28993*** .02965 77.24 .0000 2.23183 2.34804
Mu(8)| 2.92948*** .02827 103.62 .0000 2.87407 2.98489
Mu(9)| 3.38076*** .02920 115.77 .0000 3.32353 3.43800
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Random Effects Ordered Probability Model
Dependent variable NEWHSAT
Log likelihood function -53631.92165
Underlying probabilities based on Normal
Unbalanced panel has 7293 individuals
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 2.19480*** .07252 30.27 .0000 2.05267 2.33692
HHNINC| -.03764 .04636 -.81 .4169 -.12850 .05323
HHKIDS| .18979*** .01866 10.17 .0000 .15322 .22635
EDUC| .07474*** .00609 12.27 .0000 .06280 .08668
|Threshold parameters for index model
Mu(01)| .27725*** .01553 17.85 .0000 .24680 .30769
Mu(02)| .71390*** .02041 34.98 .0000 .67391 .75390
Mu(03)| 1.18482*** .02235 53.01 .0000 1.14101 1.22863
Mu(04)| 1.55571*** .02305 67.49 .0000 1.51053 1.60089
Mu(05)| 2.32085*** .02394 96.95 .0000 2.27393 2.36777
Mu(06)| 2.68712*** .02427 110.74 .0000 2.63956 2.73469
Mu(07)| 3.25778*** .02467 132.08 .0000 3.20944 3.30612
Mu(08)| 4.16499*** .02560 162.70 .0000 4.11482 4.21517
Mu(09)| 4.79284*** .02605 183.99 .0000 4.74178 4.84390
|Std. Deviation of random effect
Sigma| 1.01361*** .01233 82.23 .0000 .98945 1.03778
--------+--------------------------------------------------------------------
Random Coefficients OrdProbs Model
Dependent variable NEWHSAT
Log likelihood function -53699.77298
Ordered probit (normal) model
Simulation based on 25 Halton draws
N15: Panel Data Models for Ordered Choice N-248

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
NEWHSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
HHNINC| -.02668 .03421 -.78 .4354 -.09373 .04037
HHKIDS| .18456*** .01227 15.05 .0000 .16052 .20860
EDUC| .07680*** .00278 27.58 .0000 .07134 .08226
|Means for random parameters
Constant| 2.13724*** .03627 58.93 .0000 2.06615 2.20832
|Scale parameters for dists. of random parameters
Constant| 1.04507*** .00729 143.43 .0000 1.03079 1.05935
|Threshold parameters for probabilities
MU(1)| .26755*** .01479 18.09 .0000 .23856 .29653
MU(2)| .69343*** .01916 36.20 .0000 .65588 .73097
MU(3)| 1.15786*** .02068 55.98 .0000 1.11732 1.19840
MU(4)| 1.52579*** .02116 72.09 .0000 1.48431 1.56728
MU(5)| 2.28879*** .02177 105.11 .0000 2.24612 2.33147
MU(6)| 2.65507*** .02203 120.53 .0000 2.61189 2.69824
MU(7)| 3.22614*** .02239 144.06 .0000 3.18225 3.27003
MU(8)| 4.13325*** .02334 177.07 .0000 4.08750 4.17900
MU(9)| 4.75862*** .02385 199.56 .0000 4.71188 4.80535
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N15.4 Random Parameters and Random Thresholds

Ordered Choice Models
The structure of the random parameters model is based on the conditional probability

Prob[yit = j| xit, b i] = F( j,µ, b i′xit + αi), i = 1,...,N, t = 1,...,Ti.

where F(.) is the distribution discussed earlier (normal, logistic, extreme value, Gompertz). The
model assumes that parameters are randomly distributed with possibly heterogeneous (across
individuals) parameters generated by

E[b i| zi] = b + Dzi,

(the second term is optional – the mean may be constant),

Var[b i| zi] = Σ.

The model is operationalized by writing

bi = b + Dzi + Γvi.

As noted earlier, the heterogeneity term is optional. In addition, it may be assumed that some of the
parameters are nonrandom. It is convenient to analyze the model in this fully general form here.
We accommodate nonrandom parameters just by placing rows of zeros in the appropriate places in D
and Γ.
N15: Panel Data Models for Ordered Choice N-249

NOTE: If there is no heterogeneity in the mean, and only the constant term is considered random –
the model may specify that some parameters are nonrandom – then this model is functionally
equivalent to the random effects model of the preceding section. The estimation technique is
different, however. An application appears in the previous section.

Two major extensions of the RP-OC model are provided. The threshold parameters, µij and
disturbance variance of ei may also be random, in the form

µij = µi,j-1 + exp(αj + δ′wi + quij), µ0 = 0, uij ~ N[0,1]

eit ~ N[0,σi2], σi = exp(γ′fi + τhi), hi ~ N[0,1]

N15.4.1 Model Commands

The basic model command for this form of the model is, as is the fixed effects estimator,
given in two parts. The model is fit conventionally first to provide the starting values, then fully
specified.

ORDERED ; Lhs = dependent variable

; Rhs = independent variables
[ ; Model = Logit ] $
ORDERED ; Lhs = dependent variable
; Rhs = independent variables
; Pds = fixed periods or count variable
; RPM
; Fcn = random parameters specification
[ ; Model = Logit ] $

NOTE: For this model, your Rhs list should include a constant term.

Starting values for the iterations are provided by the user by fitting the basic model without
random parameters first. Note in the applications below that the two random parameters ordered
probit estimators are each preceded by an otherwise identical fixed parameters version.

NOTE: The command cannot reuse an earlier set of results. You must refit the basic model without
random parameters each time. Thus,

ORDERED ; ... $
ORDERED ; RPM ; ... $
ORDERED ; RPM ; ... $

will not work properly. Each random parameters model must be preceded by a set of starting values.
N15: Panel Data Models for Ordered Choice N-250

Correlated Random Parameters

The preceding defines an estimator for a model in which the covariance matrix of the
random parameters is diagonal. To extend it to a model in which the parameters are freely
correlated, add
; Correlation (or just ; Cor)
to the command. Note that this formulation of the model has an ambiguous interpretation if your
parameters are not jointly normally distributed. A correlated mixture of several distributions is
difficult to interpret.

Heterogeneity in the Means

The preceding examples have specified that the mean of the random variable is fixed over
individuals. If there is measured heterogeneity in the means, in the form of

E[bki] = bk + Σmδkmzmi
where zm is a variable that is measured for each individual, then the command may be modified to
; RPM = list of variables in z.
In the data set, these variables must be repeated for each observation in the group.

Autocorrelation
You may change the character of the heterogeneity from a time invariant effect to an AR(1)
process, vkit = ρkvki,t-1 + wkit.

Controlling the Simulation

There are two parameters of the simulations that you can change. R is the number of points
in the simulation. Authors differ in the appropriate value. Train (2009) recommends several
hundred. Bhat suggests 1,000 is an appropriate value. The program default is 100. You can choose
the value with
; Pts = number of draws, R.
The value of 50 that we set in our experiments above was chosen purely to produce an example that
you could replicate without spending an inordinate amount of waiting for the results.
The standard approach to simulation estimation is to use random draws from the specified
distribution. As suggested immediately above, good performance in this connection requires very large
numbers of draws. The drawback to this approach is that with large samples and large models, this
entails a huge amount of computation and can be very time consuming. Some authors have
documented dramatic speed gains with no degradation in simulation performance through the use of a
small number of Halton draws instead of a large number of random draws. Some authors (e.g., Bhat
(2001)) have found that a Halton sequence of draws with only one tenth the number of draws as a
random sequence is equally effective. To use this approach, add

; Halton
to your model command.
N15: Panel Data Models for Ordered Choice N-251

In order to replicate an estimation, you must use the same random draws. One implication
of this is that if you give the identical model command twice in sequence, you will not get the
identical set of results because the random draws in the sequences will be different. To obtain the
same results, you must reset the seed of the random number generator with a command such as

CALC ; Ran (seed value) $

(Note that we have used ; Ran(12345) before each of our examples above, precisely for this reason.
The specific value you use for the seed is not of consequence; any odd number will do.
In this connection, we note a consideration which is crucial in this sort of estimation. The
random sequence used for the model estimation must be the same in order to obtain replicability. In
addition, during estimation of a particular model, the same set of random draws must be used for
each person every time. That is, the sequence vi1, vi2, ..., viR used for each individual must be the
same every time it is used to calculate a probability, derivative, or likelihood function. (If this is not
the case, the likelihood function will be discontinuous in the parameters, and successful estimation
becomes unlikely. This has been called simulation ‘noise’ or ‘buzz’ in the literature. ) One way to
achieve this which has been suggested in the literature is to store the random numbers in advance,
and simply draw from this reservoir of values as needed. Because NLOGIT is able to use very large
samples, this is not a practical solution, especially if the number of draws is large as well. We
achieve the same result by assigning to each individual, i, in the sample, their own random generator
seed which is a unique function of the global random number seed, S, and their group number, i;

Seed(S,i) = S + 123.0 ×i, then minus 1.0 if the result is even.

Since the global seed, S, is a positive odd number, this seed value is unique, at least within the
several million observation range of NLOGIT.

Specifying Random Parameters

The ; Fcn = specification is used to define the random parameters. It is constructed from
the list of Rhs names as follows: Suppose your model is specified by

; Rhs = one, x1, x2, x3, x4.

This involves five coefficients. Any or all of them may be random; any not specified as random are
assumed to be constant. For those that you wish to specify as random, use

; Fcn = variable name (distribution), variable name (distribution), ...

Numerous distributions may be specified. All random variables, vik, have mean zero. Distributions
can be specified with
c for constant (zero variance), vi = 0
n for normally distributed, vi = a standard normally distributed variable
u for uniform, vi= a standard uniform distributed variable in (-1,+1)
t for triangular (the ‘tent’ distribution)
l for lognormal
N15: Panel Data Models for Ordered Choice N-252

Each of these is scaled as it enters the distribution, so the variance is only that of the random draw
before multiplication. The latter two distributions are provided as one may wish to reduce the
amount of variation in the tails of the distribution of the parameters across individuals and to limit
the range of variation. (See Train, op. cit., for discussion.) To specify that the constant term and the
coefficient on x1 are normally distributed with fixed mean and variance, use

; Fcn = one(n), x1(n).

This specifies that the first and second coefficients are random while the remainder are not. The
parameters estimated will be the mean and standard deviations of the distributions of these two
parameters and the fixed values of the other three.

N15.4.2 Results
Results saved by this estimator are:

Matrices: b = estimate of q
varb = asymptotic covariance matrix for estimate of q.
beta_i = individual specific parameters, if ; Par is requested.

Scalars: kreg = number of variables in Rhs

nreg = number of observations
logl = log likelihood function

Last Model: b_variables

Last Function: Prob(yit = J|xit) = Probability of the highest cell.

May be changed with ; Outcome = j or ; Outcome = *.

N15.4.3 Application
The following example illustrates the random parameters ordered probit model. The data are
recoded to make a more compact example, and the sample is restricted to those groups that have
seven observations, to speed up the simulations. The first two ordered probit models are the fixed
parameters, pooled estimator followed by the random parameters case in which two of the five
coefficients are random. After the random parameters model is estimated, the individual specific
estimates of E[beduc|hs,x] are collected in a variable then a kernel estimator describes the distribution
of the conditional means across the sample. The results are rearranged to compare the coefficient
estimates then the partial effects across the specifications.
The results include estimates of the means and standard deviations of the distributions of the
random parameters and the estimates of the nonrandom parameters. The log likelihood shown is
conditioned on the random draws, so one might be cautious about using it to test hypotheses, for
example, that the parameters are random at all by comparing it to the log likelihood from the basic
model with all nonrandom coefficients.
N15: Panel Data Models for Ordered Choice N-253

The commands are:

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
NAMELIST ; x = one,age,educ,hhninc,handdum $
CREATE ; hs = newhsat $
RECODE ; hs ; 0/3 = 0 ; 4/6 = 1 ; 7/8 = 2 ; 9/10 = 3 $
HISTOGRAM ; Rhs = hs $
REJECT ; ti < 7 $
ORDERED ; Lhs = hs ; Rhs = x ; Partial Effects $
ORDERED ; Lhs = hs ; Rhs = x
; RPM ; Panel ; Fcn = age(n),educ(n) ; Halton ; Pts = 25
; Partial Effects ; Par $
SAMPLE ; 1-887 $
MATRIX ; mb_educ = beta_i(1:118,1:1) $
CREATE ; be_educ = mb_educ $
KERNEL ; Rhs = be_educ $
ORDERED ; Lhs = hs ; Rhs = x ; Partial Effects $
ORDERED ; Lhs = hs ; Rhs = x
; RPM ; Panel ; Fcn = age(n),educ(n) ; Halton ; Pts = 25
; Correlated ; Partial Effects ; Par $
+--------------------------------------------------------------------+
| CELL FREQUENCIES FOR ORDERED CHOICES |
+--------------------------------------------------------------------+
| Frequency Cumulative < = Cumulative > = |
|Outcome Count Percent Count Percent Count Percent |
|----------- ------- --------- ------- --------- ------- --------- |
|HS=00 569 9.1641 569 9.1641 6209 100.0000 |
|HS=01 2000 32.2113 2569 41.3754 5640 90.8359 |
|HS=02 2342 37.7194 4911 79.0949 3640 58.6246 |
|HS=03 1298 20.9051 6209 100.0000 1298 20.9051 |
+--------------------------------------------------------------------+

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HS
Log likelihood function -7679.52077
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HS| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Index function for probability
Constant| 1.72050*** .10585 16.25 .0000 1.51304 1.92796
AGE| -.02354*** .00155 -15.19 .0000 -.02658 -.02051
EDUC| .06417*** .00687 9.34 .0000 .05069 .07764
HHNINC| .26574*** .08773 3.03 .0025 .09381 .43768
HANDDUM| -.34752*** .03370 -10.31 .0000 -.41358 -.28146
|Threshold parameters for index
Mu(1)| 1.17217*** .01623 72.20 .0000 1.14035 1.20399
Mu(2)| 2.24966*** .01942 115.83 .0000 2.21160 2.28773
N15: Panel Data Models for Ordered Choice N-254

--------+--------------------------------------------------------------------
Random Coefficients OrdProbs Model
Dependent variable HS
Log likelihood function -6724.01324
Estimation based on N = 6209, K = 9
Unbalanced panel has 887 individuals
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HS| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
Constant| 2.56865*** .11016 23.32 .0000 2.35275 2.78455
HHNINC| .18922** .08693 2.18 .0295 .01884 .35960
HANDDUM| -.18622*** .03508 -5.31 .0000 -.25497 -.11747
|Means for random parameters
AGE| -.04128*** .00159 -26.01 .0000 -.04439 -.03817
EDUC| .10807*** .00748 14.45 .0000 .09341 .12273
|Scale parameters for dists. of random parameters
AGE| .01357*** .00034 39.55 .0000 .01289 .01424
EDUC| .08208*** .00155 53.01 .0000 .07905 .08512
|Threshold parameters for probabilities
MU(1)| 1.64297*** .02744 59.87 .0000 1.58918 1.69676
MU(2)| 3.17465*** .03234 98.16 .0000 3.11126 3.23804
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

-----------------------------------------------------------------------------
Random Coefficients OrdProbs Model
Dependent variable HS
Log likelihood function -994.76038
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HS| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
Constant| 2.97520*** .25659 11.60 .0000 2.47230 3.47811
HHNINC| .23351 .22085 1.06 .2903 -.19934 .66637
HANDDUM| -.25589*** .09735 -2.63 .0086 -.44670 -.06508
|Means for random parameters
AGE| -.04495*** .00386 -11.66 .0000 -.05250 -.03739
EDUC| .06925*** .01533 4.52 .0000 .03921 .09930
|Diagonal elements of Cholesky matrix
AGE| .00860*** .00262 3.29 .0010 .00347 .01373
EDUC| .04047*** .00337 12.02 .0000 .03388 .04707
|Below diagonal elements of Cholesky matrix
lEDU_AGE| .03878*** .01003 3.87 .0001 .01912 .05844
|Threshold parameters for probabilities
MU(1)| 1.65758*** .08339 19.88 .0000 1.49414 1.82102
MU(2)| 3.11571*** .09843 31.65 .0000 2.92279 3.30864
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-255

Implied covariance matrix of random parameters

Var_Beta| 1 2
--------+----------------------------
1| .739584E-04 .333495E-03
2| .333495E-03 .00314200
Implied standard deviations of random parameters
S.D_Beta| 1
--------+--------------
1| .00859991
2| .0560536
Implied correlation matrix of random parameters
Cor_Beta| 1 2
--------+----------------------------
1| 1.00000 .691818
2| .691818 1.00000

Figure N15.1 Estimators of E[b(educ)|y,x]

(Fixed parameters)
-----------------------------------------------------------------------------
Marginal effects for ordered probability model
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HS| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
AGE| .00353*** 1.93407 14.53 .0000 .00305 .00401
EDUC| -.00962*** -1.30082 -9.18 .0000 -.01168 -.00757
HHNINC| -.03986*** -.17200 -3.02 .0025 -.06570 -.01402
HANDDUM| .05213*** .13505 10.09 .0000 .04200 .06225
(outcomes 1 and 2 omitted)
|--------------[Partial effects on Prob[Y=03] at means]--------------
AGE| -.00654*** -1.46872 -14.52 .0000 -.00742 -.00566
EDUC| .01782*** .98783 9.17 .0000 .01401 .02163
HHNINC| .07381*** .13061 3.02 .0025 .02598 .12164
HANDDUM| -.09653*** -.10255 -10.15 .0000 -.11517 -.07788
--------+--------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-256

(Random parameters)
-----------------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
AGE| .00247*** 4.25914 16.65 .0000 .00218 .00276
EDUC| -.00647*** -2.75143 -12.52 .0000 -.00748 -.00546
HHNINC| -.01133** -.15380 -2.16 .0306 -.02159 -.00106
HANDDUM| .01115*** .09088 5.22 .0000 .00696 .01533
(Outcomes 1 and 2 omitted, effects reordered)
|--------------[Partial effects on Prob[Y=03] at means]--------------
AGE| -.00776*** -3.12921 -22.25 .0000 -.00844 -.00708
EDUC| .02031*** 2.02149 13.54 .0000 .01737 .02325
HHNINC| .03557** .11300 2.17 .0296 .00351 .06762
HANDDUM| -.03500*** -.06677 -5.27 .0000 -.04801 -.02199
--------+--------------------------------------------------------------------
(Correlated random parameters)
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
AGE| .00344*** 4.40201 6.82 .0000 .00245 .00443
EDUC| -.00530*** -1.78538 -4.17 .0000 -.00779 -.00281
HHNINC| -.01786 -.19039 -1.05 .2927 -.05114 .01541
HANDDUM| .01958*** .13543 2.67 .0077 .00519 .03397
|--------------[Partial effects on Prob[Y=03] at means]--------------
AGE| -.00772*** -3.51945 -9.49 .0000 -.00931 -.00612
EDUC| .01189*** 1.42743 4.34 .0000 .00653 .01726
HHNINC| .04010 .15222 1.06 .2906 -.03427 .11448
HANDDUM| -.04395** -.10827 -2.55 .0107 -.07768 -.01022
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N15.4.4 Random Parameters HOPIT Model

This model extends the hierarchical ordered probit model in several directions. The core
model is an ordered probit specification:
yit* = b′xit + eit,
yit = 0 if yit* < 0,
= 1 if 0 < yit* <µ1,
= 2 if µ1< yit* <µ2,
...
= J if yit* > µJ

as usual. The model is constructed to include random coefficients, b i, random variance

heterogeneity, σi, and random thresholds, µij. The random parameters form of the general model is
bi = b + Dhi + Γwi
where Γ is a diagonal matrix of standard deviations and wik ~ N[0,1], k = 1,…,K. The mean of the
random parameter vector is b + Dhi where hi may be a set of variables specified in the model. The
disturbance in the model may be heteroscedastic and distributed with random standard deviation as
well, with
εit ~ N[0,σi2], σi = exp[γ′zi + τvi] where vi ~ N[0,1].
N15: Panel Data Models for Ordered Choice N-257

Finally, the thresholds are formed as shown for the cross section variant of this model.

µij = µi,j-1 + exp(αj + δ′wi + qjuij], where uij ~ N[0,1]

µ0 = 0 and xit contains a constant term.

The various parts are optional. In addition, the model may be fit with cross section or panel data. As
usual, panel data are likely to be more effective. The command for this model is

ORDERED ; Lhs = … ; Rhs = …

; RPM for the random coefficients, b
or ; RPM = list of variables in hi
; RTM for the random thresholds model
; Limits = list of variables for the wi in the thresholds
; Random Effects to use a common ui in the thresholds
; RVM for the random term i, vi in σi
; Het ; Hfn = list of variables in zi for the heteroscedastic model $

When the model includes any of the three random components, the maximum simulated likelihood
estimator is used. The default model is an ordered probit specification. You may specify an ordered
logit model instead by adding

; Logit

to the command.
The simulation can be modified with

; Pts = the number of points or draws

and ; Halton

to indicate that Halton sequences rather than random draws be used for the simulations. Halton
sequences are recommended. The simulation is over the J elements in µij plus the element vi in σi
plus the K variables in the Rhs specification. If you specify a ‘random effects’ model, then the same
single random term appears in all of the threshold equations.
If you are using a panel data set, use either

SETPANEL ; Group = variable name

; Pds = variable name $

with ; Panel in the ORDERED command, or, if the Pds variable is already prepared, use

; Pds = the group count variable.

N15: Panel Data Models for Ordered Choice N-258

Partial effects for this model are computed internally and requested with

; Partial Effects.

This general form of the random parameters ordered probit model does not use the template
random parameters form described in Chapter R24. (Note that there is no ; Fcn = specification
component in the command.) As formulated, all parameters on the variables in the Rhs list are
assumed to be random. You can modify this by imposing a constraint that the corresponding
diagonal element of Γ, which is the standard deviation of the random part of that element of b i, be
equal to zero. To do this, include in the command

; Rh2 = list of variables with nonrandom parameters.

Thus, the full list of variables in the model is those in the Rhs list plus those in the Rh2 list. There is
no overlap – variables must appear in only one of these two lists.
Results saved by this estimator are:

Matrices: b = estimate of b
varb = asymptotic covariance matrix for estimate of q.
betartop = full set of parameter estimates , if ; Par is requested.
Scalars: kreg = number of variables in Rhs
nreg = number of observations
logl = log likelihood function

Last Model: b_variables

Last Function: None

The following application uses the subset of the GSOEP sample that have five observations
in each group. The application is further speeded up by using only 10 Halton draws in the
simulations. This is sufficient for a numerical example, but would be far too small for an actual
application. The estimated model allows for unobserved heterogeneity in all three places, the
parameters, thresholds and disturbance variance.

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
REJECT ; ti # 5 $
ORDERED ; Lhs = hsat ; Rhs = one,age,educ ; Rh2 = hhninc,married,hhkids
; RPM ; RTM ; RVM
; Limits = female ; Pts = 10
; Halton ; Panel ; Maxit = 25 $
N15: Panel Data Models for Ordered Choice N-259

-----------------------------------------------------------------------------
Random Thresholds Ordered Choice Model
Dependent variable HSAT
Log likelihood function -10134.79176
Restricted log likelihood -10899.81624
Chi squared [ 17 d.f.] 1530.04896
Significance level .00000
McFadden Pseudo R-squared .0701869
Estimation based on N = 5255, K = 29
Inf.Cr.AIC =20327.584 AIC/N = 3.868
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Latent Regression Equation
Constant| 4.17571*** .16744 24.94 .0000 3.84754 4.50388
AGE| -.04388*** .00218 -20.13 .0000 -.04815 -.03961
EDUC| .06261*** .00965 6.49 .0000 .04370 .08153
HHNINC| .35696*** .11753 3.04 .0024 .12662 .58731
MARRIED| .09078* .04999 1.82 .0694 -.00719 .18876
HHKIDS| -.09768** .04371 -2.23 .0254 -.18334 -.01201
|Intercept Terms in Random Thresholds
Alpha-01| -1.19538*** .13834 -8.64 .0000 -1.46653 -.92423
Alpha-02| -.69311*** .08966 -7.73 .0000 -.86884 -.51739
Alpha-03| -.70446*** .06420 -10.97 .0000 -.83029 -.57862
Alpha-04| -1.14567*** .08731 -13.12 .0000 -1.31679 -.97455
Alpha-05| -.19232*** .03307 -5.82 .0000 -.25713 -.12751
Alpha-06| -1.03759*** .05273 -19.68 .0000 -1.14094 -.93424
Alpha-07| -.58017*** .03466 -16.74 .0000 -.64810 -.51224
Alpha-08| -.04815* .02878 -1.67 .0943 -.10456 .00826
Alpha-09| -.39987*** .04048 -9.88 .0000 -.47920 -.32054
|Standard Deviations of Random Thresholds
Alpha-01| .24187*** .07688 3.15 .0017 .09118 .39256
Alpha-02| .34510*** .06721 5.14 .0000 .21338 .47682
Alpha-03| .19508** .08818 2.21 .0270 .02224 .36792
Alpha-04| .26252*** .08332 3.15 .0016 .09922 .42582
Alpha-05| .11536*** .03689 3.13 .0018 .04305 .18767
Alpha-06| .17729*** .06490 2.73 .0063 .05009 .30448
Alpha-07| .23047*** .03758 6.13 .0000 .15683 .30412
Alpha-08| .15433*** .02927 5.27 .0000 .09697 .21170
Alpha-09| .04443 .04045 1.10 .2721 -.03486 .12371
|Variables in Random Thresholds
FEMALE| -.03079** .01291 -2.38 .0171 -.05609 -.00549
|Standard Deviations of Random Regression Parameters
Constant| .06490 .05458 1.19 .2344 -.04208 .17187
AGE| .02166*** .00083 26.18 .0000 .02004 .02328
EDUC| .00519** .00234 2.22 .0264 .00061 .00977
HHNINC| 0.0 .....(Fixed Parameter).....
MARRIED| 0.0 .....(Fixed Parameter).....
HHKIDS| 0.0 .....(Fixed Parameter).....
|Latent Heterogeneity in Variance of Epsilon
Tau(v)| .29096*** .01860 15.65 .0000 .25451 .32741
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-260

+----------------------------------------------------------------------+
| Summary of Marginal Effects for Ordered Probability Model (probit) |
| Effects are computed by averaging over observs. during simulations. |
| Binary variables change only by 1 unit so s.d. changes are not shown.|
| Elasticities for binary variables = partial effect/probability = %chgP |
+----------------------------------------------------------------------+
+----------------------------------------------------------------------+
| Regression Variable AGE Changes in AGE % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 .00158 .00158 .00000 .01766 .06166 5.85945
Y = 01 .00057 .00215 -.00158 .00640 .02235 3.00925
Y = 02 .00128 .00343 -.00215 .01425 .04973 2.42584
Y = 03 .00168 .00511 -.00343 .01876 .06548 1.83159
Y = 04 .00130 .00641 -.00511 .01451 .05065 1.18846
Y = 05 .00336 .00977 -.00641 .03753 .13101 .94528
Y = 06 .00154 .01131 -.00977 .01720 .06003 .70612
Y = 07 .00046 .01176 -.01131 .00511 .01782 .12789
Y = 08 -.00304 .00872 -.01176 -.03401 -.11873 -.56476
Y = 09 -.00344 .00528 -.00872 -.03840 -.13403 -1.42223
Y = 10 -.00528 .00000 -.00528 -.05901 -.20598 -2.34240
+----------------------------------------------------------------------+
| Regression Variable EDUC Changes in EDUC % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.00226 -.00226 .00000 -.00540 -.02482 -2.13858
Y = 01 -.00082 -.00307 .00226 -.00196 -.00900 -1.09832
Y = 02 -.00182 -.00489 .00307 -.00435 -.02002 -.88538
Y = 03 -.00240 -.00729 .00489 -.00573 -.02636 -.66849
Y = 04 -.00185 -.00914 .00729 -.00443 -.02039 -.43376
Y = 05 -.00479 -.01394 .00914 -.01147 -.05273 -.34501
Y = 06 -.00220 -.01613 .01394 -.00525 -.02416 -.25772
Y = 07 -.00065 -.01679 .01613 -.00156 -.00717 -.04668
Y = 08 .00434 -.01244 .01679 .01039 .04779 .20613
Y = 09 .00490 -.00754 .01244 .01173 .05395 .51909
Y = 10 .00754 .00000 .00754 .01803 .08291 .85493
+----------------------------------------------------------------------+
| Regression Variable HHNINC Changes in HHNINC % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.01286 -.01286 .00000 -.00229 -.03857 -.37184
Y = 01 -.00466 -.01752 .01286 -.00083 -.01398 -.19097
Y = 02 -.01037 -.02790 .01752 -.00185 -.03111 -.15394
Y = 03 -.01366 -.04156 .02790 -.00244 -.04096 -.11623
Y = 04 -.01057 -.05213 .04156 -.00188 -.03168 -.07542
Y = 05 -.02733 -.07946 .05213 -.00487 -.08195 -.05999
Y = 06 -.01252 -.09198 .07946 -.00223 -.03755 -.04481
Y = 07 -.00372 -.09570 .09198 -.00066 -.01115 -.00812
Y = 08 .02477 -.07093 .09570 .00442 .07427 .03584
Y = 09 .02796 -.04297 .07093 .00499 .08384 .09025
Y = 10 .04297 .00000 .04297 .00766 .12884 .14865
+----------------------------------------------------------------------+
N15: Panel Data Models for Ordered Choice N-261

+----------------------------------------------------------------------+
| Regression Variable MARRIED Changes in MARRIED % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 -.00327 -.00327 .00000 -.00138 -.00327 -.20824
Y = 01 -.00119 -.00446 .00327 -.00050 -.00119 -.10695
Y = 02 -.00264 -.00710 .00446 -.00111 -.00264 -.08621
Y = 03 -.00347 -.01057 .00710 -.00147 -.00347 -.06509
Y = 04 -.00269 -.01326 .01057 -.00113 -.00269 -.04224
Y = 05 -.00695 -.02021 .01326 -.00293 -.00695 -.03359
Y = 06 -.00318 -.02339 .02021 -.00134 -.00318 -.02509
Y = 07 -.00095 -.02434 .02339 -.00040 -.00095 -.00455
Y = 08 .00630 -.01804 .02434 .00266 .00630 .02007
Y = 09 .00711 -.01093 .01804 .00300 .00711 .05054
Y = 10 .01093 .00000 .01093 .00461 .01093 .08325
+----------------------------------------------------------------------+
| Regression Variable HHKIDS Changes in HHKIDS % chg|
| ------------------------------ ------------------------------
Outcome Effect dPy<=nn/dX dPy>=nn/dX 1 StdDev Low to High Elast
------- ------------------------------ ------------------------------
Y = 00 .00352 .00352 .00000 .00173 .00352 .11752
Y = 01 .00128 .00480 -.00352 .00063 .00128 .06036
Y = 02 .00284 .00763 -.00480 .00139 .00284 .04865
Y = 03 .00374 .01137 -.00763 .00183 .00374 .03674
Y = 04 .00289 .01426 -.01137 .00142 .00289 .02384
Y = 05 .00748 .02174 -.01426 .00367 .00748 .01896
Y = 06 .00343 .02517 -.02174 .00168 .00343 .01416
Y = 07 .00102 .02619 -.02517 .00050 .00102 .00257
Y = 08 -.00678 .01941 -.02619 -.00332 -.00678 -.01133
Y = 09 -.00765 .01176 -.01941 -.00375 -.00765 -.02853
Y = 10 -.01176 .00000 -.01176 -.00577 -.01176 -.04698
------------------------------------------------------------------------
Indirect Partial Effects for Ordered Choice Model
Variables in thresholds
Outcome FEMALE
Y = 00 .000000
Y = 01 -.000468
Y = 02 -.001603
Y = 03 -.002728
Y = 04 -.002883
Y = 05 -.009219
Y = 06 -.005379
Y = 07 -.005158
Y = 08 .002091
Y = 09 .007557
Y = 10 .017791
N15: Panel Data Models for Ordered Choice N-262

N15.5 Latent Class Ordered Choice Models

The ordered choice model for a panel of data, i = 1,...,N, t = 1,...,Ti is

Prob[Yit = yit| xit] = F(yit, µ, b′xit) = P(i,t), yit = 0, 1,...,.

Henceforth, we use the term ‘group’ to indicate the Ti observations on respondent i in periods
t = 1,...,Ti. Unobserved heterogeneity in the distribution of Yit is assumed to impact the density in the
form of a random effect. The continuous distribution of the heterogeneity is approximated by using
a finite number of ‘points of support.’ The distribution is approximated by estimating the location of
the support points and the mass (probability) in each interval. In implementation, it is convenient
and useful to interpret this discrete approximation as producing a sorting of individuals (by
heterogeneity) into J classes, j = 1,...,J. (Since this is an approximation, J is chosen by the analyst.)
Thus, we modify the model for a latent sorting of yit into J ‘classes’ with a model which
allows for heterogeneity as follows: The probability of observing yit given that regime j applies is

P(i,t|j) = Prob[Yit = yit| xit, j]

where the density is now specific to the group. The analyst does not observe directly which class,
j = 1,...,J generated observation yit|j, and class membership must be estimated. Heckman and Singer
(1984) suggest a simple form of the class variation in which only the constant term varies across the
classes. This would produce the model

P(i,t|j) = F[yit, µ, b′xit + δj], Prob[class = j] = Fj

We formulate this approximation more generally as,

P(i,t|j) = F[yit, µ, b′xit + δj′xit], Fj = exp(qj) / Σj exp(qj), with qJ = 0.

In this formulation, each group has its own parameter vector, b j′ = b + δj, though the variables that
enter the mean are assumed to be the same. (This can be changed by imposing restrictions on the
full parameter vector, as described below.) This allows the Heckman and Singer formulation as a
special case by imposing restrictions on the parameters – each δj has only one nonzero element in the
location of the constant term. You may also specify that the latent class probabilities depend on
person specific characteristics, so that

qij = qj′zi, qJ = 0.

N15.5.1 Command
The estimation command for this model is
ORDERED ; Lhs = ...
; Rhs = independent variables
[; Model = Weibull, Logit or Gompertz]
; LCM (for latent class model)
[; LCM = list of variables in zi for multinomial logit class probabilities]
; Pds = panel data specification $
N15: Panel Data Models for Ordered Choice N-263

The default number of support points is five. You may set J to 2, 3, ..., 10 with
; Pts = the value you wish.
Some particular values computed for the latent class model are

; Group = the index of the most likely latent class

; Cprob = estimated posterior probability for the most likely
latent class
You can obtain a listing of these two results by using

; List.

You can use the ; Rst = list option to structure the latent class model so that different variables
appear in different classes. Alternatively, you can use this to force the Heckman and Singer form of
the model as follows, where we use a three class model as an example:

NAMELIST ; x = ... one, list of variables $

CALC ; k1 = Col(x) - 1 ; kmu = Max(y) - 1 $
ORDERED ; Lhs = ... ; Rhs = x ; LCM ; Pts = 3
; Rst = d1, k1_b, kmu_mu,
d2, k1_b, kmu_mu,
d3, k1_b, kmu_mu, t1,t2,t3 $

N15.5.2 Results
Results saved by this estimator are

Matrices: b = full parameter vector, [b 1′, b 2′,... F1,...,FJ]

varb = full covariance matrix

(Note that b and varb involve J×(K+#outcomes - 1) estimates.)

beta_ = individual specific parameters, if ; Par is requested

b_class = a J×K matrix with each row equal to the corresponding b j
class_pr = a J×1 vector containing the estimated class probabilities

Scalars: kreg = number of variables in Rhs list

nreg = total number of observations used for estimation
logl = maximized value of the log likelihood function
exitcode = exit status of the estimation procedure

Last Function: None

N15: Panel Data Models for Ordered Choice N-264

Application
To illustrate the model, we will fit an ordered probit model with three latent classes. We
have modified the health care data set to set up a compact example. (The latent class estimator is
actually unable to resolve more than one class with nine threshold parameters.) We have censored
the health satisfaction measure to three classes for purpose of this exercise. The ordered probit model
is the same one specified earlier. Some of the numerical results are omitted to simplify comparison
of the estimated models. The first set of commands creates the data set.

SAMPLE ; All $
SETPANEL ; Group = id ; Pds = ti $
CREATE ; health = newhsat $
RECODE ; health ; 0/4 = 0 ; 5/8 = 1 ; 9/10 = 2 $
NAMELIST ; x = one,hhninc,hhkids,educ $

We now fit the base case pooled model.

ORDERED ; Lhs = health ; Rhs = x ; Partial Effects $

This is a three class latent class model.

ORDERED ; Lhs = health ; Rhs = x ; Partial Effects

; LCM ; Pts = 3 ; Panel $

This fits two random effects models, the continuous, normally distributed effects model and
Heckman and Singer’s discrete approximation.

ORDERED ; Lhs = health ; Rhs = x ; Partial Effects ; Panel $

ORDERED ; Quiet ; Lhs = health ; Rhs = x $
ORDERED ; Lhs = health ; Rhs = x ; Partial Effects
; LCM ; Pts = 3 ; Panel
; Rst = alpha0,3_b,cmu,alpha1,3_b,cmu,
alpha2,3_b,cmu,theta0,theta1,theta2 $

This model specifies that the class probabilities depend on age and sex.
SAMPLE ; All $
ORDERED ; Quiet ; Lhs = health ; Rhs = x $
ORDERED ; Lhs = health ; Rhs = x ; Partial Effects
; LCM = one,age,female ; Pts = 3 ; Panel $
Finally, we use a small subsample to show the listing of the posterior class probabilities.

REJECT ; ti # 3 $
ORDERED ; Quiet ; Lhs = health ; Rhs = x $
ORDERED ; Lhs = health ; Rhs = x ; Partial Effects
; LCM = one,age,female ; Pts = 3 ; Panel ; List $

This is the base case, pooled ordered probit model, with no group effects followed by the estimates
of the parameters of the three class latent class model.
N15: Panel Data Models for Ordered Choice N-265

-----------------------------------------------------------------------------
Ordered Probability Model
Dependent variable HEALTH
Log likelihood function -24522.47670
Restricted log likelihood -24801.77601
Chi squared [ 3 d.f.] 558.59861
Significance level .00000
McFadden Pseudo R-squared .0112613
Estimation based on N = 27326, K = 5
Inf.Cr.AIC =49054.953 AIC/N = 1.795
Underlying probabilities based on Normal
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HEALTH| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Index function for probability
Constant| .38694*** .03538 10.94 .0000 .31761 .45628
HHNINC| .15134*** .04069 3.72 .0002 .07160 .23109
HHKIDS| .21408*** .01419 15.09 .0000 .18627 .24188
EDUC| .04904*** .00311 15.77 .0000 .04294 .05513
| Threshold parameters for index
Mu(1)| 1.83426*** .01130 162.26 .0000 1.81210 1.85641
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Latent Class / Panel OrdProbs Model
Dependent variable HEALTH
Log likelihood function -21956.55643
Estimation based on N = 27326, K = 17
Inf.Cr.AIC =43947.113 AIC/N = 1.608
Unbalanced panel has 7293 individuals
Ordered probability model
Ordered probit (normal) model
LHS variable = values 0,1,..., 2
Model fit with 3 latent classes.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HEALTH| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Model parameters for latent class 1
Constant| 1.16608*** .10831 10.77 .0000 .95379 1.37838
HHNINC| -.22927** .08945 -2.56 .0104 -.40458 -.05395
HHKIDS| .10979*** .03316 3.31 .0009 .04480 .17479
EDUC| .08077*** .00937 8.62 .0000 .06241 .09913
MU(1)| 1.73212*** .04607 37.60 .0000 1.64184 1.82241
| Model parameters for latent class 2
Constant| .62012*** .07038 8.81 .0000 .48218 .75805
HHNINC| -.06265 .07865 -.80 .4257 -.21681 .09151
HHKIDS| .24254*** .02664 9.11 .0000 .19034 .29475
EDUC| .06115*** .00621 9.85 .0000 .04899 .07332
MU(1)| 2.68221*** .02902 92.43 .0000 2.62533 2.73909
| Model parameters for latent class 3
Constant| -1.00572*** .11321 -8.88 .0000 -1.22762 -.78383
HHNINC| .52603*** .12473 4.22 .0000 .28157 .77050
HHKIDS| .24566*** .04766 5.15 .0000 .15225 .33908
EDUC| .05198*** .01000 5.20 .0000 .03239 .07157
MU(1)| 1.88097*** .06379 29.49 .0000 1.75595 2.00600
| Estimated prior probabilities for class membership
N15: Panel Data Models for Ordered Choice N-266

Class1Pr| .27635*** .00916 30.17 .0000 .25839 .29430

Class2Pr| .56896*** .01168 48.69 .0000 .54605 .59186
Class3Pr| .15470*** .00823 18.80 .0000 .13857 .17083
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

These are the estimated marginal effects for the two models presented above, with the pooled
estimates first followed by those derived from the latent class model.
-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HEALTH| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
HHNINC| -.03364*** -.08477 -3.72 .0002 -.05137 -.01591
*HHKIDS| -.04653*** -.33304 -15.36 .0000 -.05247 -.04060
EDUC| -.01090*** -.88316 -15.70 .0000 -.01226 -.00954
|--------------[Partial effects on Prob[Y=01] at means]--------------
HHNINC| -.01184*** -.00657 -3.63 .0003 -.01824 -.00545
*HHKIDS| -.01875*** -.02955 -11.05 .0000 -.02208 -.01542
EDUC| -.00384*** -.06848 -11.47 .0000 -.00449 -.00318
|--------------[Partial effects on Prob[Y=02] at means]--------------
HHNINC| .04548*** .07091 3.72 .0002 .02150 .06947
*HHKIDS| .06528*** .28908 14.74 .0000 .05660 .07396
EDUC| .01474*** .73880 15.58 .0000 .01288 .01659
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HEALTH| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
HHNINC| .00289 .01116 .34 .7345 -.01381 .01959
*HHKIDS| -.03296*** -.36179 -10.53 .0000 -.03910 -.02683
EDUC| -.01068*** -1.32670 -12.47 .0000 -.01236 -.00900
|--------------[Partial effects on Prob[Y=01] at means]--------------
HHNINC| .00154 .00073 .34 .7350 -.00738 .01046
*HHKIDS| -.01987*** -.02682 -7.68 .0000 -.02494 -.01479
EDUC| -.00569*** -.08698 -8.07 .0000 -.00707 -.00431
|--------------[Partial effects on Prob[Y=02] at means]--------------
HHNINC| -.00443 -.00928 -.34 .7347 -.03004 .02118
*HHKIDS| .05283*** .31427 10.18 .0000 .04265 .06300
EDUC| .01637*** 1.10240 12.05 .0000 .01371 .01903
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-267

This is the random effects model. It is comparable to the Heckman and Singer form that follows.
The first model with continuously distributed effects suggests a random constant term with mean
2.33642 and standard deviation 0.99095. From the Heckman and Singer model, using the three
estimated constants and the three estimated prior probabilities, we find a mean of 2.19016 and
standard deviation 0.90994. Since the remaining coefficients in the latent class model do not differ
across classes, they are directly comparable to the random effects model. The overall similarity is to
be expected, but there are some substantive differences. For example, the latent class model predicts
a much smaller influence of marital status than does the random effects model.
-----------------------------------------------------------------------------
Random Effects Ordered Probability Model
Dependent variable HEALTH
Log likelihood function -22042.38298
Restricted log likelihood -24522.47670
Chi squared [ 1 d.f.] 4960.18744
Significance level .00000
McFadden Pseudo R-squared .1011355
Estimation based on N = 27326, K = 6
Inf.Cr.AIC =44096.766 AIC/N = 1.614
Underlying probabilities based on Normal
Unbalanced panel has 7293 individuals
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HEALTH| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Index function for probability
Constant| .64927*** .07239 8.97 .0000 .50739 .79115
HHNINC| -.03500 .05665 -.62 .5367 -.14603 .07603
HHKIDS| .20576*** .02188 9.40 .0000 .16288 .24865
EDUC| .07118*** .00625 11.40 .0000 .05894 .08343
| Threshold parameters for index model
Mu(01)| 2.56175*** .01686 151.90 .0000 2.52870 2.59480
| Std. Deviation of random effect
Sigma| 1.00299*** .01483 67.63 .0000 .97392 1.03206
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Latent Class / Panel OrdProbs Model
Dependent variable HEALTH
Log likelihood function -22048.67454
Estimation based on N = 27326, K = 9
Inf.Cr.AIC =44115.349 AIC/N = 1.614
Unbalanced panel has 7293 individuals
Ordered probability model
Ordered probit (normal) model
LHS variable = values 0,1,..., 2
Model fit with 3 latent classes.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HEALTH| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Model parameters for latent class 1
Constant| 2.12385*** .06069 35.00 .0000 2.00490 2.24279
HHNINC| -.07289 .05188 -1.40 .1601 -.17458 .02880
HHKIDS| .20014*** .01936 10.34 .0000 .16220 .23808
EDUC| .05987*** .00507 11.81 .0000 .04994 .06981
MU(1)| 2.46535*** .01693 145.63 .0000 2.43217 2.49853
N15: Panel Data Models for Ordered Choice N-268

| Model parameters for latent class 2

Constant| -.95230*** .06385 -14.92 .0000 -1.07743 -.82717
HHNINC| -.07289 .05188 -1.40 .1601 -.17458 .02880
HHKIDS| .20014*** .01936 10.34 .0000 .16220 .23808
EDUC| .05987*** .00507 11.81 .0000 .04994 .06981
MU(1)| 2.46535*** .01693 145.63 .0000 2.43217 2.49853
| Model parameters for latent class 3
Constant| .56180*** .05806 9.68 .0000 .44801 .67560
HHNINC| -.07289 .05188 -1.40 .1601 -.17458 .02880
HHKIDS| .20014*** .01936 10.34 .0000 .16220 .23808
EDUC| .05987*** .00507 11.81 .0000 .04994 .06981
MU(1)| 2.46535*** .01693 145.63 .0000 2.43217 2.49853
| Estimated prior probabilities for class membership
Class1Pr| .23642*** .00833 28.38 .0000 .22009 .25275
Class2Pr| .13069*** .00723 18.07 .0000 .11652 .14487
Class3Pr| .63289*** .00995 63.60 .0000 .61338 .65239
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

The following takes a closer look at the distributions of heterogeneity implied by the continuous
random effects model and the discrete distribution implied by the Heckman and Singer model. The
program below plots the two distributions. The densities are evaluated at 500 points ranging from the
mean of the continuous distribution plus and minus three standard deviations. (The program could be
made generic based on the model results. We have used the actual values in a few commands.)

MATRIX ; ah = [2.12385/-.95230/.56180] $
MATRIX ; ph = [.23642/.13069/.63289] $
SAMPLE ; 1-500 $
CALC ; min = .64927 - 3*1.00299
; max = .64927 + 3*1.00929
; delta = .002 * (max-min) $
CREATE ; alpha = Trn(min,delta) $
CREATE ; Normal = 1/1.00929 * N01((alpha - .64927)/1.00929) $
CALC ; ahs1 = ah(2) ; ahs2 = ah(3) ; ahs3 = ah(1) $
CALC ; mid12 = .5*(ahs2+ahs1) ; mid23 = .5*(ahs2+ahs3) $
CALC ; dhs1 = ph(2)/(mid12-min) $
CALC ; dhs2 = ph(3)/(mid23-mid12) $
CALC ; dhs3 = ph(1)/(max-mid23) $
CREATE ; hecksing = dhs1*(alpha < mid12) +
dhs2*(alpha >= mid12) * (alpha < mid23) +
dhs3*(alpha >= mid23) $
PLOT ; Lhs = alpha ; Rhs = normal,hecksing
; Fill ; Limits = 0,.45 ; Endpoints = min,max
; Title = Discrete & Continuous Distributions of Heterogeneity
; Yaxis = RndmEfct $
N15: Panel Data Models for Ordered Choice N-269

Figure E15.2 Discrete and Continuous Distributions of Heterogeneity

These are the estimated marginal effects for the two models. Once again, they are quite similar, as
might be expected.
-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HEALTH| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|--------------[Partial effects on Prob[Y=00] at means]--------------
HHNINC| .00552 .01381 .62 .5368 -.01199 .02303
*HHKIDS| -.03196*** -.22713 -9.53 .0000 -.03853 -.02539
EDUC| -.01122*** -.90314 -11.26 .0000 -.01318 -.00927
|--------------[Partial effects on Prob[Y=01] at means]--------------
HHNINC| .00203 .00114 .62 .5350 -.00437 .00842
*HHKIDS| -.01283*** -.02046 -6.92 .0000 -.01646 -.00920
EDUC| -.00412*** -.07437 -8.19 .0000 -.00511 -.00313
|--------------[Partial effects on Prob[Y=02] at means]--------------
HHNINC| -.00754 -.01144 -.62 .5362 -.03145 .01636
*HHKIDS| .04479*** .19287 9.10 .0000 .03514 .05444
EDUC| .01534*** .74797 11.24 .0000 .01267 .01802
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

This is the Heckman and Singer form of the model.

-----------------------------------------------------------------------------
Marginal effects for ordered probability model
M.E.s for dummy variables are Pr[y|x=1]-Pr[y|x=0]
Names for dummy variables are marked by *.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HEALTH| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
N15: Panel Data Models for Ordered Choice N-270

|--------------[Partial effects on Prob[Y=00] at means]--------------

HHNINC| .00993 .04901 1.40 .1606 -.00394 .02380
*HHKIDS| -.02655*** -.37215 -10.42 .0000 -.03154 -.02155
EDUC| -.00816*** -1.29445 -11.47 .0000 -.00955 -.00676
|--------------[Partial effects on Prob[Y=01] at means]--------------
HHNINC| .00772 .00353 1.40 .1614 -.00308 .01852
*HHKIDS| -.02285*** -.02968 -7.96 .0000 -.02848 -.01723
EDUC| -.00634*** -.09323 -8.90 .0000 -.00774 -.00494
|--------------[Partial effects on Prob[Y=02] at means]--------------
HHNINC| -.01765 -.03913 -1.41 .1600 -.04227 .00697
*HHKIDS| .04940*** .31106 9.90 .0000 .03962 .05917
EDUC| .01450*** 1.03341 11.49 .0000 .01202 .01697
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

In the model below, the class probabilities depend on age and sex. These are averaged over the data
in the table at the end of the results. The constant probabilities from the model estimated earlier are
shown with them. An important feature to note here is that there is no natural ordering of classes in
the latent class model. The ordering of the second and third classes has changed from the earlier
model to this one.

-----------------------------------------------------------------------------
Latent Class / Panel OrdProbs Model
Dependent variable HEALTH
Log likelihood function -21779.75836
Estimation based on N = 27326, K = 21
Inf.Cr.AIC =43601.517 AIC/N = 1.596
Unbalanced panel has 7293 individuals
Ordered probability model
Ordered probit (normal) model
LHS variable = values 0,1,..., 2
Model fit with 3 latent classes.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HEALTH| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Model parameters for latent class 1
Constant| 1.41223*** .10283 13.73 .0000 1.21070 1.61377
HHNINC| -.24084*** .08785 -2.74 .0061 -.41301 -.06866
HHKIDS| .02548 .03257 .78 .4340 -.03836 .08932
EDUC| .06130*** .00862 7.11 .0000 .04441 .07819
MU(1)| 1.72679*** .04553 37.93 .0000 1.63756 1.81602
|Model parameters for latent class 2
Constant| -.80867*** .12257 -6.60 .0000 -1.04890 -.56845
HHNINC| .55004*** .12874 4.27 .0000 .29771 .80236
HHKIDS| .11778** .05227 2.25 .0242 .01533 .22023
EDUC| .03595*** .01105 3.25 .0011 .01430 .05760
MU(1)| 1.93880*** .06839 28.35 .0000 1.80477 2.07284
|Model parameters for latent class 3
Constant| .80114*** .07069 11.33 .0000 .66260 .93969
HHNINC| -.08541 .07783 -1.10 .2725 -.23796 .06713
HHKIDS| .16879*** .02640 6.39 .0000 .11706 .22052
EDUC| .04689*** .00614 7.64 .0000 .03487 .05892
MU(1)| 2.66629*** .02734 97.53 .0000 2.61270 2.71987
N15: Panel Data Models for Ordered Choice N-271

|Estimated prior probabilities for class membership

ONE_1| .81468*** .13922 5.85 .0000 .54181 1.08755
AGE_1| -.03807*** .00345 -11.05 .0000 -.04482 -.03131
FEMALE_1| -.13830* .07356 -1.88 .0601 -.28247 .00586
ONE_2| -3.09023*** .22351 -13.83 .0000 -3.52830 -2.65215
AGE_2| .04049*** .00447 9.07 .0000 .03174 .04924
FEMALE_2| -.01649 .09674 -.17 .8647 -.20609 .17312
ONE_3| 0.0 .....(Fixed Parameter).....
AGE_3| 0.0 .....(Fixed Parameter).....
FEMALE_3| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

+------------------------------------------------------------+
| Prior class probabilities at data means for LCM variables |
| Class 1 Class 2 Class 3 Class 4 Class 5 |
| .24199 .15782 .60019 .00000 .00000 |
+------------------------------------------------------------+

The model estimates include the estimates of the prior probabilities of group membership. It
is also possible to compute the posterior probabilities for the groups, conditioned on the data. The
; List specification will request a listing of these. The following illustration shows this feature for a
small subset of the data used above.

Predictions computed for the group with the largest posterior probability
Obs. Periods Fitted outcomes
=============================================================================
Ind.= 1 J* = 2 P(j)= .008 .881 .111
Ind.= 2 J* = 2 P(j)= .401 .491 .109
Ind.= 3 J* = 2 P(j)= .203 .737 .060
Ind.= 4 J* = 2 P(j)= .050 .909 .041
Ind.= 5 J* = 2 P(j)= .186 .702 .113
Ind.= 6 J* = 2 P(j)= .172 .735 .094
Ind.= 7 J* = 2 P(j)= .177 .735 .088
Ind.= 8 J* = 2 P(j)= .039 .869 .092
Ind.= 9 J* = 3 P(j)= .002 .334 .663
Ind.= 10 J* = 3 P(j)= .000 .003 .997
Ind.= 11 J* = 2 P(j)= .106 .836 .057
Ind.= 12 J* = 2 P(j)= .079 .758 .164
Ind.= 13 J* = 2 P(j)= .023 .928 .049
Ind.= 14 J* = 2 P(j)= .017 .959 .024
Ind.= 15 J* = 2 P(j)= .106 .829 .065
Ind.= 16 J* = 2 P(j)= .070 .895 .036
Ind.= 17 J* = 2 P(j)= .388 .497 .114
Ind.= 18 J* = 2 P(j)= .065 .842 .093
Ind.= 19 J* = 3 P(j)= .006 .111 .884
Ind.= 20 J* = 3 P(j)= .017 .391 .592
Ind.= 21 J* = 3 P(j)= .010 .353 .637
Ind.= 22 J* = 2 P(j)= .140 .735 .125
Ind.= 23 J* = 3 P(j)= .003 .422 .575
Ind.= 24 J* = 2 P(j)= .101 .826 .073
Ind.= 25 J* = 2 P(j)= .043 .920 .037
N16: The Multinomial Logit Model N-272

N16: The Multinomial Logit Model

N16.1 Introduction
This chapter and Chapter N17 describe two forms of the ‘multinomial logit’ model. These
models are also known variously as ‘conditional logit,’ ‘discrete choice,’ and ‘universal logit’
models, among other names. All of them can be viewed as special cases of a general model of utility
maximization: An individual is assumed to have preferences defined over a set of alternatives (travel
modes, occupations, food groups, etc.)

U(alternative 0) = b 0′xi0 + e i0
U(alternative 1) = b 1′xi1 + e i1
...
U(alternative J) = b J ′xiJ + eiJ
Observed Yi = choice j if Ui(alternative j) > Ui(alternative k) ∀ k ≠ j.

The ‘disturbances’ in this framework (individual heterogeneity terms) are assumed to be

independently and identically distributed with identical extreme value distribution; the CDF is

F(ej) = exp(-exp(-ej))

Based on this specification, the choice probabilities,

Prob[ choice j ] = Prob[Uj > Uk], ∀ k ≠ j

exp(β′j x ji )
= , j = 0,...,J,
∑ exp(β′m x mi )
J
m= 0

where ‘i’ indexes the observation, or individual, and ‘j’ and ‘m’ index the choices. We note at the
outset, the IID assumptions made about ej are quite stringent, and lead to the ‘Independence from
Irrelevant Alternatives’ or IIA implications that characterize the model. Much (perhaps all) of the
research on forms of this model consists of development of alternative functional forms and
stochastic specifications that avoid this feature.
The observed data consist of the Rhs vectors, xjt, and the outcome, or choice, yt. (We also
consider a number of variants.) There are many forms of the multinomial logit, or multinomial
choice model supported in NLOGIT and LIMDEP. LIMDEP contains two basic forms of the model.
The NLOGIT program provides the major extensions that are documented in this and the remaining
chapters of this manual.
N16: The Multinomial Logit Model N-273

This chapter will examine what we call the multinomial logit model. In this setting, it is
assumed that the Rhs variables consist of a set of individual specific characteristics, such as age,
education, marital status, etc. These are the same for all choices, so the choice subscript on x in the
formula above is dropped. The observation setting is the individual’s choice among a set of
alternatives, where it is assumed that the determinant of the choice is the characteristics of the
individual. An example might be a model of choice of occupation. (This is the model originally
devised by Nerlove and Press (1973).) For convenience at this point, we label this the multinomial
logit model. Essential features of the model and commands are documented here. This form of the
multinomial logit model is supported in LIMDEP as well as NLOGIT. Further details appear in
Chapter E37.
Chapter N17 will examine what we call (again, purely for convenience) the discrete choice
model and, also, to differentiate the command, the conditional logit model. In this framework, we
observe the attributes of the choices, as well as (or, possibly, instead of) the characteristics of the
individual. A well known example is travel mode choice. Samples of observations often consist of
the attributes of the different modes and the choice actually made. Sometimes, no characteristics of
the individuals are observed beyond their actual choice. Models may also contain mixtures of the
two types of choice determinants. (We emphasize, these naming distinctions are meaningless in the
modeling framework – we just use them here only to organize the applicable parts of LIMDEP and
NLOGIT. In practice, all of the models considered in this chapter and Chapter N17 are multinomial
logit models. The basic CLOGIT model is also supported by LIMDEP and discussed in Chapter E38.

N16.2 The Multinomial Logit Model – MLOGIT

The general form of the multinomial logit model is

exp(β′j xt )
Prob[ choice j ] = , j = 0,...,J,
∑ m=1 exp(β′m xt )
J

A possible J+1 unordered outcomes can occur. In order to identify the parameters of the model, we
impose the normalization b0 = 0. This model is typically employed for individual or grouped data in
which the ‘x’ variables are characteristics of the observed individual(s), not the choices. For present
purposes, that is the main distinction between this and the discrete choice model described in
Chapter N17. The characteristics are the same across all outcomes. The study of occupational
choice, by Schmidt and Strauss (1975) provides a well known application.
The data will appear as follows:

• Individual data: yt coded 0, 1, ..., J,

• Grouped data: y0t, y1t,...,yJt give proportions or shares.

In the grouped data case, a weighting variable, nt, may also be provided if the observations happen to
be frequencies. The proportions variables must range from zero to one and sum to one at each
observation. The full set must be provided, even though one is redundant. The data are inspected to
determine which specification is appropriate. The number of Lhs variables given and the coding of
the data provide the full set of information necessary to estimate the model, so no additional
information about the dependent variable is needed. There is a single line of data for each individual.
N16: The Multinomial Logit Model N-274

This model proliferates parameters. There are J×K nonzero parameters in all, since there is a
vector b j for each probability except the first. Consequently, even moderately sized models quickly
become very large ones if your outcome variable, y, takes many values. The maximum number of
parameters which can be estimated in a model is 150 as usual with the standard configuration.
However, if you are able to forego certain other optional features, the number of parameters can
increase to 300. The model size is detected internally. If your configuration contains more than 150
parameters, the following options and features become unavailable:

• marginal effects
• choice based sampling
• ; Rst = list for imposing restrictions
• ; CML: specification for imposing linear constraints
• ; Hold for using the multinomial logit model as a sample selection equation

In addition, if your model size exceeds 150 parameters, the matrices b and varb cannot be retained.
(But, see below for another way to retrieve large parameter matrices.)
The choice set should be restricted to no more than 25 choices. If you have more than 25
choices, the number of characteristics that may be used becomes very small. Nonetheless, it is
possible to fit models with up to 500 choices by using CLOGIT, as discussed in Chapter N17.

N16.3 Model Command for the Multinomial Logit Model

The command for fitting this form of multinomial logit model is

MLOGIT ; Lhs = y or y0,y1,...yJ

; Rhs = regressors $

(The command may also be LOGIT, which is what has always been used in previous versions of
LIMDEP.) All general options for controlling output and iterations are available except ; Keep =
name. (A program which can be used to obtain the fitted probabilities is listed below.) There are
internally computed predictions for the multinomial logit model.

N16.3.1 Imposing Constraints on Parameters

The ; Rst = list form of restrictions is supported for imposing constraints on model
parameters, either fixed value or equality. One possible application of the constrained model
involves making the entire vector of coefficients in one probability equal that in another. You can do
this as follows:

NAMELIST ; x = the entire set of Rhs variables $

CALC ; k = Col(x) $
MLOGIT ; Lhs = y
; Rhs = x
; Rst = k_b, k_b, ... , k_b $
N16: The Multinomial Logit Model N-275

This would force the corresponding coefficients in all probabilities to be equal. You could also
apply this to some, but not all of the outcomes, as in

; Rst = k_b, k_b, k_b2, k_b3

HINT: The coefficients in this model are not the marginal effects. But, forcing the coefficient on a
characteristic in probability j to equal its counterpart in probability m also forces the two marginal
effects to be equal.

N16.3.2 Starting Values

The parameter vector for this model is a J×K column vector,

Θ = [b 1′ ,b 2′ , ...,b J′ ]′ .

You may provide starting values with ; Start = list.

N16.4 Robust Covariance Matrix

You can compute a ‘robust covariance matrix’ for the MLE. (The misspecification to which
the matrix is robust is left unspecified in most cases.) The desired robust covariance matrix would
result in the preceding computation if wi equals one for all observations. This suggests a simple way
to obtain it, just by specifying

; Robust.

The estimator of the asymptotic covariance matrix produced with this request is the standard
‘sandwich’ estimator,

V = [-H]-1 (G′G) [-H]-1

where H is the estimated second derivatives matrix of the log likelihood and G is the matrix with
rows equal to the first derivatives, usually labeled the OPG or ‘outer product of gradients’ estimator.
N16: The Multinomial Logit Model N-276

N16.5 Cluster Correction

A related calculation is used when observations occur in groups which may be correlated.
This is rather like a panel; one might use this approach in a random effects kind of setting in which
observations have a common latent heterogeneity. The parameter estimator is unchanged in this
case, but an adjustment is made to the estimated asymptotic covariance matrix. The calculation is
done as follows: Suppose the n observations are assembled in C clusters of observations, in which
the number of observations in the cth cluster is nc. Thus,

∑
C
c =1
nc = n.

Denote by b the full set of model parameters, [b 1′, ..., b J′]′. Let the observation specific gradients
and Hessians for individual i in cluster c be

∂ log Lic
gic =
∂β

∂ 2 log Lic
Hic = .
∂ββ ∂ '

The uncorrected estimator of the asymptotic covariance matrix based on the Hessian is

( −∑ )
−1
∑
C nc
VH = -H-1 = =c 1 =i 1
H ic

The corrected asymptotic covariance matrix is

∧ 
Est.Asy.Var β  = VH
 
C  C
C −1 

= ∑ ∑
c 1=
nc
g
i 1 ic ( )( ∑
=i 1
nc
)
g ic ' VH


Note that if there is exactly one observation per cluster, then this is C/(C-1) times the sandwich
(robust) estimator discussed above. Also, if you have fewer clusters than parameters, then this
matrix is singular – it has rank equal to the minimum of C and JK, the number of parameters. This
estimator is requested with

; Cluster = specification

where the specification is either a fixed number of observations per cluster, or an identifier that
distinguishes clusters, such as an identification number. This estimator can also be extended to
stratified as well as clustered data, using

; Stratum = specification.

The full description of using these procedures appears in Chapter R10.

N16: The Multinomial Logit Model N-277

N16.6 Choice Based Sampling

The choice based sampling methodology for individual data can be applied here. You must
provide a weighting variable which gives the sampling ratios. The variable gives the ratio of the
true, population proportion to the sample proportions. This presumes that you know the population
proportions, φ0,...,φJ. If you know the sample proportions, f0,...,fJ, as well, then you can calculate the
necessary ratios, w0,...,wJ = φj/fj needed for the calculations to follow. With these in hand, you can
create the weights using RECODE as follows:

CREATE ; wts = y (your dependent variable) $

RECODE ; wts ; 0 = weight for 0 ; 1 = weight for 1 ; ... $

A convenient way to do the same computation is to create a vector with the weights,

MATRIX ; cbwt = [w0, w1,...,wJ] $

then you can use the following:

CREATE ; yplus1 = y + 1 ; wts = cbwt(yplus1) $ Zero is not a valid subscript.

Regardless, you must have the population proportions in hand. If you do not know the appropriate
sample proportions, there is a special MATRIX function, Prpn(variable), for this purpose, which
you can use as follows:

CREATE ; yplus1 = y + 1 $
MATRIX ; f = Prpn (yplus1) $

Since you have φj in hand, you can now proceed as follows:

MATRIX ; phi = [ φ0,...,φJ] $ You provide the values.

MATRIX ; cbwt = diag(f) ; cbwt = phi * <cbwt> $
CREATE ; wts = cbwt(yplus1) $

(Note, the Prpn(variable) function is used specifically for this purpose. It creates a vector with one
column and number of rows equal to the minimum of 100 and the maximum of yplus1. Values
larger than 100 or less than one are discarded, and not counted in the proportions.)
Be sure to provide a sampling ratio for every outcome. With the weights in place, your
MLOGIT command is

MLOGIT ; Lhs = y ; Rhs = regressors

; Wts = weights ; Choice Based Sampling $

This adjustment changes the estimator in two ways. First, the observations are weighted in
computing the parameter estimates. Second, after estimation, the standard errors are adjusted. The
estimator of the asymptotic covariance matrix for the choice based sampling case is

Asy.Var[bCBWT] = (-H)-1BHHH (-H)-1

N16: The Multinomial Logit Model N-278

where the weighted matrices are constructed from the Hessian and first derivatives using

∂2log L/∂b l∂b m′ = Σt wt{-[1(l=m)Pl - PlPm ]}X′X.

∂log L/∂b j = Σt wi(dtj - tij)xt where dtj = 1 if person t makes choice j;

BHHH(in blocks) = Σt wi(dtl - Ptl)(dtm - Ptm)xtxt′

and wt = population frequency for choice made by individual t

divided by sample proportion for choice made by
individual t.

N16.7 Output for the Logit Models

Initial ordinary least squares results are used for the starting values for this model. For
individual data, J binary variables are implied by the model. These are used in a least squares
regression. For the grouped data case, a minimum chi squared, generalized least squares estimate is
obtained by the weighted regression of

qij = log(Pij / Pi0)

on the regressors, with weights wij = (niPijPi0)1/2 (ni may be 1.0). The OLS estimates based on the
individual data are inconsistent, but the grouped data estimates are consistent (and, in the binomial
case, efficient). The least squares estimates are included in the displayed results by including

; OLS

in the model command. The iterations are followed by the maximum likelihood estimates with the
usual diagnostic statistics. An example is shown below.

NOTE: Minimum chi squared (MCS) is an estimator, not a model. Moreover, the MCS estimator
has the same properties as, but is different from the maximum likelihood estimator. Since the MCS
estimator in NLOGIT is not iterated, it should not be used as the final results of estimation. Without
iteration, the MCS estimator is not a fixed point – the weights are functions only of the sample
proportions, not the parameters. For current purposes, these are only useful as starting values.

Standard output for the logit model will begin with a table such as the following which
results from estimation of a model in which the dependent variable takes values 0,1,2,3,4,5:

SAMPLE ; All $
REJECT ; hsat > 5 $
MLOGIT ; Lhs = hsat ; Rhs = one,educ,hhninc,age,hhkids $
(This is based on the health satisfaction variable analyzed in the preceding chapter. We reduced the
sample to those with hsat reported zero to five. We would note, though these make for a fine
numerical example, the multinomial logit model would be inappropriate for these ordered data.) The
restricted log likelihood is computed for a model in which one is the only Rhs variable. In this case,

log L0 = Σj nj logPj
N16: The Multinomial Logit Model N-279

where nj is the number of individuals who choose outcome j and Pj = nj/n = the jth sample
proportion. The chi squared statistic is 2(log L - log L0). If your model does not contain a constant
term, this statistic need not be positive, in which case it is not reported. But, even if it is computable,
the statistic is meaningless if your model does not contain a constant.
The diagnostic statistics are followed by the coefficient estimates: These are b 1,...,bJ. Recall
b0 is normalized to zero, and not reported.
-----------------------------------------------------------------------------
Multinomial Logit Model
Dependent variable HSAT
Log likelihood function -11246.96937
Restricted log likelihood -11308.02002
Chi squared [ 20 d.f.] 122.10132
Significance level .00000
McFadden Pseudo R-squared .0053989
Estimation based on N = 8140, K = 25
Inf.Cr.AIC =22543.939 AIC/N = 2.770
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -1.77566** .69486 -2.56 .0106 -3.13756 -.41376
EDUC| .07326 .04476 1.64 .1017 -.01447 .16099
HHNINC| .28572 .58129 .49 .6231 -.85359 1.42503
AGE| .00566 .00838 .68 .4996 -.01077 .02209
HHKIDS| .27188 .19642 1.38 .1663 -.11311 .65686
|Characteristics in numerator of Prob[Y = 2]
Constant| -.54217 .54866 -.99 .3231 -1.61752 .53318
EDUC| .06152* .03617 1.70 .0890 -.00937 .13240
HHNINC| .85929* .44943 1.91 .0559 -.02158 1.74017
AGE| -.00090 .00651 -.14 .8903 -.01365 .01185
HHKIDS| .13921 .15530 .90 .3700 -.16517 .44359
|Characteristics in numerator of Prob[Y = 3]
Constant| -.25433 .49206 -.52 .6053 -1.21876 .71010
EDUC| .10996*** .03247 3.39 .0007 .04632 .17359
HHNINC| 1.54517*** .40167 3.85 .0001 .75791 2.33242
AGE| -.00955 .00584 -1.64 .1017 -.02099 .00189
HHKIDS| .08178 .14014 .58 .5595 -.19289 .35645
|Characteristics in numerator of Prob[Y = 4]
Constant| .09378 .48301 .19 .8461 -.85291 1.04047
EDUC| .10453*** .03202 3.26 .0011 .04178 .16729
HHNINC| 1.74362*** .39382 4.43 .0000 .97175 2.51550
AGE| -.01430** .00571 -2.50 .0123 -.02550 -.00310
HHKIDS| .19549 .13660 1.43 .1524 -.07224 .46321
|Characteristics in numerator of Prob[Y = 5]
Constant| 1.58459*** .45170 3.51 .0005 .69927 2.46991
EDUC| .07527** .03035 2.48 .0131 .01579 .13475
HHNINC| 1.64030*** .37209 4.41 .0000 .91101 2.36959
AGE| -.01481*** .00526 -2.82 .0049 -.02512 -.00450
HHKIDS| .19988 .12655 1.58 .1142 -.04815 .44791
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N16: The Multinomial Logit Model N-280

The statistical output for the coefficient estimates is followed by a table of predicted and
actual frequencies, such as the following: This table is requested by adding

; Summary

to the MLOGIT command.

Frequencies of actual & predicted outcomes
Predicted outcome has maximum probability.

Frequencies of actual & predicted outcomes

Predicted outcome has maximum probability.

Predicted
------ ------------------------------ + -----
Actual 0 1 2 3 4 5 | Total
------ ------------------------------ + -----
0 0 0 0 0 0 447 | 447
1 0 0 0 0 0 255 | 255
2 0 0 0 0 0 642 | 642
3 0 0 0 0 0 1173 | 1173
4 0 0 0 0 0 1390 | 1390
5 0 0 0 0 0 4233 | 4233
------ ------------------------------ + -----
Total 0 0 0 0 0 8140 | 8140

The prediction for any observation is the cell with the largest predicted probability for that
observation.

NOTE: If you have more than three outcomes, it is very common, as occurred above, for the model
to predict zero outcomes in one or more of the cells. Even in a model with very high t ratios and
great statistical significance, it takes a very well developed model to make predictions in all cells.

The ; List specification produces a listing such as the following:

Predicted Values (* => observation was not in estimating sample.)

Observation Observed Y Predicted Y Residual MaxPr(i) Prob[Y*=y]
20 2.0000000 5.0000000 .000000 .6845695 .0631146
24 .000000 4.0000000 .000000 .3196778 .0885942
38 5.0000000 5.0000000 .000000 .6041918 .6041918
39 2.0000000 5.0000000 .000000 .6439476 .1224276
57 5.0000000 5.0000000 .000000 .5050133 .5050133
59 5.0000000 5.0000000 .000000 .4284611 .4284611
60 5.0000000 5.0000000 .000000 .4173034 .4173034

In the listing, the MaxPr(i) is the probability attached to the outcome with the largest predicted
probability; the outcome is shown as the Predicted Y. The last column shows the predicted
probability for the observed outcome. Residuals are not computed – there is no significance to the
reported zero. (The results above illustrate the format of the table. They were complete with a small
handful of observations, not the 8,140 used to fit the model shown earlier.)
N16: The Multinomial Logit Model N-281

The results kept for further use are:

Matrices: b and varb.

b_logit = (J+1)×K.
This additional matrix contains the parameters arranged so that b j′ is the jth
row. The first row is zero. This matrix can be used to obtain fitted
probabilities, as discussed below.

Scalars: kreg, nreg, logl, and exitcode.

Labels for WALD are constructed from the outcome and variable numbers. For example, if
there are three outcomes and ; Rhs = one,x1,x2, the labels will be

Last Model: [b1_1,b1_2,b1_3,b2_1,b2_2,b2_3].

Last Function: Prob(y = J|x).

You may specify other outcomes in the PARTIALS and SIMULATE commands.

N16.8 Partial Effects

The partial effects in this model are

δj = ∂Pj/∂x, j = 0,1,...,J.

For the present, ignore the normalization b 0 = 0. The notation Pj is used for Prob[y = j]. After some
tedious algebra, we find

δj = Pj(b j - β )

∑ j =0
J
where β = Pj b j.

It follows that neither the sign nor the magnitude of δj need bear any relationship to those of b j.
(This is worth bearing in mind when reporting results.) The asymptotic covariance matrix for the
estimator of δj would be computed using

Asy.Var. δ j  = Gj Asy.Var β  Gj′

∧ ∧

   

where b is the full parameter vector. It can be shown that

∧ ∧
Asy.Var. δ j  = Σl Σm Vjl Asy.Cov.[ β l, β m′]Vjm′, j=0,...,J,
∧

 
where Vjl = [1(j = l) - Pl ]{PjI + δjx′} - Pjδlx′

and 1(j = l) = 1 if j = l, and 0 otherwise.

N16: The Multinomial Logit Model N-282

N16.8.1 Computation of Partial Effects with the Model

This full set of results is produced automatically when your LOGIT command includes

; Partial Effects.

NOTE: Marginal effects are computed at the sample averages of the Rhs variables in the model.

There is no conditional mean function in this model, so marginal effects are interpreted a bit
differently from the usual case. What is reported are the derivatives and elasticities of the
probabilities. (Note this is the same as the ordered probability models.) These derivatives are saved
in a matrix named partials which has J+1 rows and K columns. Each row is the vector of partial
effects of the corresponding probability. Since the probabilities will always sum to one, the column
sums in this matrix will always be zero. That is,

MATRIX ; List ; 1 ’ partials $

will display a row matrix of zeros. The elasticities of the probabilities, (∂Pj/∂xk)×(xk/Pj) are placed in
a (J+1)×K matrix named elast_ml. The format of the results is illustrated in the example below.

-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
Observations used for means are All Obs.
A full set is given for the entire set of
outcomes, HSAT = 0 to HSAT = 5
Probabilities at the mean values of X are
0= .052 1= .030 2= .078 3= .145 4= .171
5= .523
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HSAT| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Marginal effects on Prob[Y = 0]
EDUC| -.00415*** -.87310 -2.87 .0042 -.00699 -.00131
HHNINC| -.07533*** -.48081 -4.28 .0000 -.10982 -.04085
AGE| .00059** .53969 2.36 .0184 .00010 .00109
HHKIDS| -.00875 -.05610 -1.44 .1505 -.02067 .00317
|Marginal effects on Prob[Y = 1]
EDUC| -.00021 -.07636 -.21 .8331 -.00220 .00178
HHNINC| -.03570*** -.38652 -2.64 .0083 -.06222 -.00918
AGE| .00052*** .80559 2.62 .0087 .00013 .00091
HHKIDS| .00313 .03408 .68 .4994 -.00596 .01222
|Marginal effects on Prob[Y = 2]
EDUC| -.00147 -.20405 -.92 .3557 -.00458 .00165
HHNINC| -.04677** -.19725 -2.31 .0211 -.08652 -.00703
AGE| .00083*** .49750 2.67 .0075 .00022 .00144
HHKIDS| -.00234 -.00993 -.32 .7478 -.01662 .01194
N16: The Multinomial Logit Model N-283

|Marginal effects on Prob[Y = 3]

EDUC| .00430** .32277 2.29 .0218 .00063 .00797
HHNINC| .01276 .02908 .53 .5938 -.03413 .05965
AGE| .00028 .09081 .70 .4822 -.00050 .00106
HHKIDS| -.01265 -.02898 -1.35 .1760 -.03097 .00567
|Marginal effects on Prob[Y = 4]
EDUC| .00416** .26381 2.07 .0385 .00022 .00810
HHNINC| .04913** .09457 1.98 .0482 .00040 .09787
AGE| -.00048 -.13248 -1.14 .2552 -.00132 .00035
HHKIDS| .00452 .00874 .46 .6444 -.01466 .02370
|Marginal effects on Prob[Y = 5]
EDUC| -.00262 -.05450 -.94 .3475 -.00809 .00285
HHNINC| .09591*** .06048 2.78 .0054 .02827 .16355
AGE| -.00174*** -.15634 -3.07 .0021 -.00285 -.00063
HHKIDS| .01609 .01020 1.23 .2205 -.00965 .04183
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Marginal Effects Averaged Over Individuals

--------+---------+---------+---------+---------+---------+---------+
Variable| Y=00 | Y=01 | Y=02 | Y=03 | Y=04 | Y=05 |
--------+---------+---------+---------+---------+---------+---------+
ONE | -.0377 | -.0772 | -.0975 | -.1380 | -.1051 | .4556 |
EDUC | -.0044 | -.0002 | -.0014 | .0043 | .0042 | -.0025 |
HHNINC | -.0786 | -.0361 | -.0459 | .0136 | .0494 | .0977 |
AGE | .0006 | .0005 | .0008 | .0003 | -.0005 | -.0018 |
HHKIDS | -.0092 | .0033 | -.0023 | -.0125 | .0045 | .0162 |
--------+---------+---------+---------+---------+---------+---------+

Averages of Individual Elasticities of Probabilities

--------+---------+---------+---------+---------+---------+---------+
Variable| Y=00 | Y=01 | Y=02 | Y=03 | Y=04 | Y=05 |
--------+---------+---------+---------+---------+---------+---------+
ONE | -.7050 | -2.4807 | -1.2472 | -.9593 | -.6112 | .8796 |
EDUC | -.8732 | -.0764 | -.2041 | .3227 | .2638 | -.0545 |
HHNINC | -.4847 | -.3904 | -.2011 | .0252 | .0907 | .0566 |
AGE | .5315 | .7974 | .4894 | .0827 | -.1406 | -.1645 |
HHKIDS | -.0571 | .0330 | -.0110 | -.0300 | .0077 | .0092 |
--------+---------+---------+---------+---------+---------+---------+
N16: The Multinomial Logit Model N-284

Figure N16.1 Matrices Created by MLOGIT

Marginal effects are computed by averaging the effects over individuals rather than computing them
at the means. The difference between the two is likely to be quite small. Current practice favors the
averaged individual effects, rather than the effects computed at the means. MLOGIT also reports
elasticities with the marginal effects. An example appears above.
N16: The Multinomial Logit Model N-285

N16.8.2 Partial Effects Using the PARTIALS EFFECTS Command

The ; Partials specification in the MLOGIT command computes the partial effects at the
means of the variables. The post estimation command, PARTIAL EFFECTS (or just PARTIALS),
can be used to compute average partial effects, and to compute various simulations of the outcome.
For example, we compute the partial effects on Prob(hsat = 5|x) for the model estimated above with

SAMPLE ; All $
REJECT ; hsat > 5 $
LOGIT ; Lhs = hsat ; Rhs = one,educ,hhninc,age,hhkids ; Partials $
PARTIALS ; Effects: educ / hhninc / age / hhkids ; Summary $
The first results below are those reported earlier. The second set are the average partial effects. (The
similarity is striking.)
-----------------------------------------------------------------------------
Partial derivatives of probabilities with
respect to the vector of characteristics.
They are computed at the means of the Xs.
--------+--------------------------------------------------------------------
| Partial Prob. 95% Confidence
HSAT| Effect Elasticity z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Marginal effects on Prob[Y = 5]
EDUC| -.00262 -.05450 -.94 .3475 -.00809 .00285
HHNINC| .09591*** .06048 2.78 .0054 .02827 .16355
AGE| -.00174*** -.15634 -3.07 .0021 -.00285 -.00063
HHKIDS| .01609 .01020 1.23 .2205 -.00965 .04183
--------+--------------------------------------------------------------------
z, prob values and confidence intervals are given for the partial effect
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
---------------------------------------------------------------------
Partial Effects for Multinomial Logit Probability Y = 5
Partial Effects Averaged Over Observations
* ==> Partial Effect for a Binary Variable
---------------------------------------------------------------------
Partial Standard
(Delta method) Effect Error |t| 95% Confidence Interval
---------------------------------------------------------------------
EDUC -.00249 .00279 .89 -.00796 .00298
HHNINC .09767 .03445 2.84 .03015 .16519
AGE -.00175 .00056 3.11 -.00286 -.00065
* HHKIDS .01592 .01310 1.22 -.00976 .04160
---------------------------------------------------------------------

The various optional specifications in PARTIALS may be used here. For example,

PARTIALS ; Effects: hhkids & hhninc=.05(.5)3 ; Outcome = 4 ; Plot $

plots the effect of hhkids on Prob(hsat=4) for several values of hhninc. The PARTIALS command
will also report elasticities with respect to continuous variables such as hhnnc by enclosing the name
in brackets, such as

PARTIALS ; Effects: <hhninc> $

N16: The Multinomial Logit Model N-286

N16.9 Predicted Probabilities

Predicted probabilities can be computed automatically for the multinomial logit model.
Since there are multiple outcomes, this must be handled a bit differently from other models. The
procedure is as follows: Request the computation with

; Prob = name

as you would normally for a discrete choice model. However, for this model, NLOGIT does the
following:

1. A namelist is created with name consisting of up to the first four letters of ‘name’ and prob
is appended to it. Thus, if you use ; Prob = pfit, the namelist will be named pfitprob.

2. The set of variables, one for each outcome, are named with the same convention, with prjj
instead of prob.

For example, in a five outcome model, the specification

; Prob = job

produces a namelist

jpbprob = jobpr00, jobpr01, jobpr02, jobpr03, jobpr04.

For our running example,

; Prob = hsat

produces the namelist named hsatprob and variables hsatpr00, hsatpr01, …, hsatpr05. The variables
will then contain the respective probabilities. You may also use

; Fill

with this procedure to compute probabilities for observations that were not in the sample.
Observations which contain missing data are bypassed as usual.
N16: The Multinomial Logit Model N-287

N16.10 Generalized Maximum Entropy (GME) Estimation

This is an alternative estimator for the multinomial logit model. The GME criterion is based
on the entropy of the probability distribution,

E(p0,...,pJ) = -Σj pj lnpj.

The implementation of the GME estimator in NLOGIT’s multinomial logit model is done by
augmenting the likelihood function with a term that accounts for the entropy of the choice
probability set. Let

H = the number of support points for the entropy distribution.

and V = an H specific set of weights. These are

V = -1/ N , +1/ N for H = 2

= -1/ N , 0, +1/ N for H = 3

= -1/ N , -.5/ N , [0], +.5/ N , +1/ N for H = 4 or 5

= ... [0], +.33/ N , +.67/ N , +1/ N for H = 6 or 7

= ... [0], +.25/ N , +.50/ N , +.75/ N , +1/ N for H = 8 or 9

(You may optionally choose to scale the entire V by 1/ N ). Then,

Ψij = ∑ exp[Vhβ′j xi ]
H
h =1

Then, the additional term which augments the contribution to the log likelihood for individual i is

∑
J
FΨi = j =0
ln Ψ ij

This estimator is invoked simply by adding

; GME = the number of support points, H

to the LOGIT command. You may choose to scale the weighting vector with

; Scale

You may also choose the GME estimator in the command builder.
N16: The Multinomial Logit Model N-288

In the example below, we have treated the self reported health satisfaction measure as a
discrete choice (doubtlessly inappropriately – just for the purpose of a numerical example). The first
set of estimates given are the GME results. The model is refit by maximum likelihood in the second
set. As can be seen, the GME estimator triggers some additional results in the table of summary
statistics. It also brings some relatively modest changes in the estimated parameters.

-----------------------------------------------------------------------------
Generalized Maximum Entropy (Logit)
Dependent variable HSAT
Log likelihood function -106287.21094
Estimation based on N = 8140, K = 25
Number of support points = 7
Weights in support scaled to 1/sqr(N)
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -1.76249** .69184 -2.55 .0108 -3.11848 -.40650
EDUC| .07199 .04453 1.62 .1059 -.01529 .15926
HHNINC| .26975 .57843 .47 .6410 -.86396 1.40346
AGE| .00570 .00835 .68 .4951 -.01067 .02207
HHKIDS| .26950 .19568 1.38 .1684 -.11402 .65302
|Characteristics in numerator of Prob[Y = 2]
Constant| -.53230 .54599 -.97 .3296 -1.60243 .53782
EDUC| .06033* .03595 1.68 .0933 -.01012 .13078
HHNINC| .84177* .44699 1.88 .0597 -.03432 1.71786
AGE| -.00083 .00648 -.13 .8986 -.01353 .01188
HHKIDS| .13734 .15466 .89 .3745 -.16579 .44047
|Characteristics in numerator of Prob[Y = 3]
Constant| -.24497 .48927 -.50 .6166 -1.20392 .71398
EDUC| .10879*** .03223 3.38 .0007 .04562 .17197
HHNINC| 1.52790*** .39910 3.83 .0001 .74567 2.31013
AGE| -.00948 .00581 -1.63 .1030 -.02087 .00191
HHKIDS| .07994 .13948 .57 .5666 -.19344 .35332
|Characteristics in numerator of Prob[Y = 4]
Constant| .10311 .48018 .21 .8300 -.83803 1.04426
EDUC| .10338*** .03178 3.25 .0011 .04108 .16567
HHNINC| 1.72645*** .39122 4.41 .0000 .95966 2.49323
AGE| -.01423** .00569 -2.50 .0124 -.02538 -.00308
HHKIDS| .19367 .13593 1.42 .1542 -.07276 .46009
|Characteristics in numerator of Prob[Y = 5]
Constant| 1.59393*** .44877 3.55 .0004 .71437 2.47350
EDUC| .07412** .03010 2.46 .0138 .01512 .13312
HHNINC| 1.62344*** .36941 4.39 .0000 .89940 2.34748
AGE| -.01474*** .00523 -2.82 .0049 -.02500 -.00448
HHKIDS| .19810 .12585 1.57 .1155 -.04857 .44477
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N16: The Multinomial Logit Model N-289

+--------------------------------------------------------------------+
| Information Statistics for Discrete Choice Model. |
| M=Model MC=Constants Only M0=No Model |
| Criterion F (log L) -106287.21094 -106347.98256 -109623.17376 |
| LR Statistic vs. MC 121.54324 .00000 .00000 |
| Degrees of Freedom 20.00000 .00000 .00000 |
| Prob. Value for LR .00000 .00000 .00000 |
| Entropy for probs. 11250.94128 11311.43749 14584.92208 |
| Normalized Entropy .77141 .77556 1.00000 |
| Entropy Ratio Stat. 6667.96160 6546.96918 .00000 |
| Bayes Info Criterion 26.13692 26.15185 26.95656 |
| BIC(no model) - BIC .81965 .80472 .00000 |
| Pseudo R-squared .22859 .00000 .00000 |
| Pct. Correct Pred. 52.00246 52.00246 16.66667 |
| Means: y=0 y=1 y=2 y=3 y=4 y=5 y=6 y>=7 |
| Outcome .0549 .0313 .0789 .1441 .1708 .5200 .0000 .0000 |
| Pred.Pr .0552 .0314 .0788 .1440 .1707 .5199 .0000 .0000 |
| Notes: Entropy computed as Sum(i)Sum(j)Pfit(i,j)*logPfit(i,j). |
| Normalized entropy is computed against M0. |
| Entropy ratio statistic is computed against M0. |
| BIC = 2*criterion - log(N)*degrees of freedom. |
| If the model has only constants or if it has no constants, |
| the statistics reported here are not useable. |
+--------------------------------------------------------------------+

N16.11 Technical Details on Optimization

Newton’s method is used to obtain the estimates in all cases. The log likelihood function for
the multinomial logit model is

log L = ΣtΣjdtj logPtj,

where Ptj is the probability defined earlier and dtj = 1 if yt = j, 0 otherwise, j = 0,...,J or dtj equals the
proportion for choice j for individual t in the grouped data case. The first and second derivatives are

∂log L/∂b j = Σt (dtj - Ptj)xt.

∂ log L/∂b l∂b m′ = Σt -[1(l=m)Ptl - PtlPtm ]xtxt′.
2

The negative inverse of the Hessian provides the asymptotic covariance matrix.
The log likelihood function for the multinomial logit model is globally concave. With the
exception of OLS and possibly the Poisson regression model, this is the most benign optimization
problem in NLOGIT, and convergence should always be routine. As such, you should not need to
change the default algorithm or the convergence criteria. If you do observe convergence problems,
such as more than a handful of iterations, you should suspect the data. Occasionally, a data set will
contain some peculiarities that impede Newton’s method. In most cases, switching the algorithm to
BFGS with
; Alg = BFGS

will solve the problem.

N16: The Multinomial Logit Model N-290

N16.12 Sequential Logit Model

The sequential logit model treats the outcome of a multinomial choice as a sequence of
transitions. The outcomes are ordered by some construction (which might suggest an ordered choice
model as a natural formulation.) Consider the outcome, level of schooling: less than high school (y =
0), high school (y =1), college (y = 2) and advanced degree (y = 3). The sequential logit model treats
the level as a sequence of transitions starting with transition from level 0, which everyone enters so
P0 = 1. The probability of stopping at level 0 is the probability of entering and not passing level 1,
which is P0(1 – P1). The probability of stopping at level 1 (high school) is the probability of leaving
level 1 and not leaving level 2, P0P1(1 – P2). For level 2, the probability is, P0P1P2(1 – P3) and
finally for level 3, P0P1P2P3 This is treated as a sequence of conditional binary logit models;

Prob(observed level = j) = Probability(exit levels m = 1,…,j) ×

Probability(no exit at level j+1).

The outcome variable must be coded 0,1,… as for other forms of the multinomial logit model. By
this formulation, for the outcomes listed above,

Prob(observed level = j) = P0,1,…,j(1 – PJ+1).

The command for the sequential logit model is

SEQLOGIT ; Lhs = outcome variable

; Rhs = exogenous influences $

The standard options for nonlinear models, including ; Cluster = specification, are available for this
model. The outcomes are labeled ‘name = 0’, ‘name = 1,’ and so on in the estimation results. You
may provide a set of labels with

; Choices = a list of labels for the outcomes.

For purposes of partial effects and simulations of the outcome variable, the default function is an
expected value

E[y] = 0×Prob(y = 0) + 1×Prob(y = 1) + …

In a particular application, the outcomes might represent quantifiable levels, such as years of
education. For this case, you may supply a set of levels to be used instead of (0,1,…,J) with

; Levels = list of values.

N16: The Multinomial Logit Model N-291

The following example is based on the health care data used in the previous example.

CREATE ; edlevel = (educ > 10) + (educ > 12) + (educ > 16) $
SEQLOGIT ; Lhs = edlevel ; Rhs = one,income,married
; Choices = lths,hs,college,graduate
; Levels = 10,12,16,18 $
SIMULATE $

(We have used a rather arbitrary coding of the years of education variable for purposes of this
numerical example.)

-----------------------------------------------------------------------------
Sequential Multinomial Logit Model
Dependent variable EDLEVEL
Log likelihood function -27123.76693
Restricted log likelihood -28275.56899
Chi squared [ 6](P= .000) 2303.60411
Significance level .00000
McFadden Pseudo R-squared .0407349
Estimation based on N = 27326, K = 9
Inf.Cr.AIC = 54265.5 AIC/N = 1.986
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
EDLEVEL| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Transition from LTHS to HS etc............................
Constant| .48944*** .04215 11.61 .0000 .40684 .57204
INCOME| 3.11877*** .11521 27.07 .0000 2.89296 3.34458
MARRIED| -.29636*** .03638 -8.15 .0000 -.36766 -.22505
|Transition from HS to COLLEGE etc............................
Constant| -1.65743*** .04400 -37.67 .0000 -1.74367 -1.57119
INCOME| 2.62319*** .09537 27.51 .0000 2.43627 2.81011
MARRIED| -.94624*** .03830 -24.71 .0000 -1.02131 -.87118
|Transition from COLLEGE to GRADUATE etc............................
Constant| -1.31118*** .08099 -16.19 .0000 -1.46992 -1.15244
INCOME| 2.23477*** .17356 12.88 .0000 1.89460 2.57494
MARRIED| .03204 .06843 .47 .6397 -.10209 .16616
--------+--------------------------------------------------------------------
***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
---------------------------------------------------------------------
Model Simulation Analysis for E[Outcome] in Sequential Logit Model
---------------------------------------------------------------------
Simulations are computed by average over sample observations
---------------------------------------------------------------------
User Function Function Standard
(Delta method) Value Error |t| 95% Confidence Interval
---------------------------------------------------------------------
Avrg. Function 15.03525 .02874 523.12 14.97892 15.09158
N16: The Multinomial Logit Model N-292

N16.13 Panel Data Multinomial Logit Models

The random parameters model described in Chapter R24 is useful for constructing two types
of panel data structures for the multinomial logit model, random effects and a dynamic model.

N16.13.1 Random Effects and Common (True) Random Effects

The structural equations of the multinomial logit model are

Uijt = b j′xit + eijt, t = 1,...,Ti, j = 0,1,...,J, i = 1,...,N,

where Uijt gives the utility of choice j by person i in period t – we assume a panel data application
with t = 1,...,Ti. The model about to be described can be applied to cross sections, where Ti = 1.
Note also that as usual, we assume that panels may be unbalanced. We also assume that eijt has a
type 1 extreme value (Gumbel) distribution and that the J random terms are independent. Finally,
we assume that the individual makes the choice with maximum utility. Under these (IIA inducing)
assumptions, the probability that individual i makes choice j in period t is

exp(β′j xit )
Pijt = .
∑ exp(β′j xit )
J
j =0

Note that this is the MLOGIT form of the model – the Rhs data are in the form of individual
characteristics, not attributes of the choices. That would be handled by CLOGIT. We now suppose
that individual i has latent, unobserved, time invariant heterogeneity that enters the utility functions
in the form of a random effect, so that

Uijt = b j′xit + αij + eijt, t = 1,...,Ti, j = 0,1,...,J, i = 1,...,N.

The resulting choice probabilities, conditioned on the random effects, are

exp(β′j xit + α ij )
Pijt | αi1,...,αiJ = .
∑ exp(β′j xit + α ij )
J
j =0

To complete the model, we assume that heterogeneity is normally distributed with zero means and
(J+1)×(J+1) covariance matrix, Σ. For identification purposes, one of the coefficient vectors must
be normalized to zero and one of the αijs is set to zero. We normalize the first element – subscript 0 –
to zero. For convenience, this normalization is left implicit in what follows. It is automatically
imposed by the software. To allow the remaining random effects to be freely correlated, we write
the J×1 vector of nonzero αs as

αi = Γ vi

where Γ is a lower triangular matrix to be estimated and vi is a standard normally distributed (mean
zero, covariance matrix, I) vector.
N16: The Multinomial Logit Model N-293

The preceding extends the random effects model to the multinomial logit framework. It is
also of the form of NLOGIT’s other random parameter models, which is how we do the estimation,
by maximum simulated likelihood. There are two additional versions of the essential structure:

1. Independent effects: Γ = A diagonal matrix.

2. True random effects: Γ = A diagonal matrix,

and vji = vi = the same random variable in all utility functions.

Thus, in the second case, the preference heterogeneity is a choice invariant characteristic of the
person.
The command structure for this model has two parts. In the first, the logit model is fit
without the effects in order to obtain the starting values. In the second, we use a standard form of the
random parameters model;

MLOGIT ; Lhs = dependent variable

; Rhs = list of variables including one $
MLOGIT ; Lhs = dependent variable
; Rhs = list of variables including one
; RPM ; Fcn = remnl
[; Halton]
[; Pts = ...]
; Pds = panel specification $

(In earlier versions of NLOGIT, the ; Fcn = remnl would have been ; Fcn = one(n) instead. You
may still use this syntax.) The items in the square brackets are optional. This requests the type 1,
independent effects model. To estimate the second model, type 2, true random effects model, add

; Common Effect

to the commands. To fit the general model with freely correlated effects, use, instead,

; Correlated.

To illustrate this estimator, we constructed an example using the health care data. The Lhs
variable is health satisfaction. We restricted the sample by first, keeping only groups with Ti = 7.
We then eliminated all observations with Lhs variable greater than four. This leaves a dependent
variable that takes five outcomes, 0,1,2,3,4, and a total sample of 905 observations in 394 groups
ranging in size from one to seven. So, the resulting panel is unbalanced. The Rhs variables are one,
age, income and hhkids that is kids in the household. We fit all three models described above.
N16: The Multinomial Logit Model N-294

The commands are as follows:

REJECT ; _groupti < 7 $

REJECT ; hsat > 4 $
SETPANEL ; Group = it ; Pds = ti $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids
; RPM ; Fcn = remnl ; Common ; Halton ; Pts = 50 ; Panel $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids ; Quiet $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids
; RPM ; Fcn = remnl ; Halton ; Pts = 50 ; Panel $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids ; Quiet $
MLOGIT ; Lhs = hsat ; Rhs = one,age,hhninc,hhkids
; RPM ; Fcn = remnl ; Correlated ; Halton ; Pts = 50 ; Panel $

These are the initial values, without latent effects.

-----------------------------------------------------------------------------
Multinomial Logit Model
Dependent variable HSAT
Log likelihood function -1289.68419
Restricted log likelihood -1295.05441
Chi squared [ 12 d.f.] 10.74042
Significance level .55129
McFadden Pseudo R-squared .0041467
Estimation based on N = 905, K = 16
Inf.Cr.AIC = 2611.368 AIC/N = 2.885
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[Y = 1]
Constant| -.97586 1.20831 -.81 .4193 -3.34410 1.39238
AGE| .00500 .02273 .22 .8259 -.03954 .04954
HHNINC| .29496 1.23304 .24 .8109 -2.12176 2.71167
HHKIDS| .47793 .42941 1.11 .2657 -.36370 1.31957
|Characteristics in numerator of Prob[Y = 2]
Constant| -.58489 .93591 -.62 .5320 -2.41923 1.24946
AGE| .01279 .01758 .73 .4667 -.02166 .04724
HHNINC| 1.48473 .93548 1.59 .1125 -.34877 3.31823
HHKIDS| .22135 .33932 .65 .5142 -.44370 .88641
|Characteristics in numerator of Prob[Y = 3]
Constant| 1.05098 .84361 1.25 .2128 -.60247 2.70442
AGE| -.00744 .01590 -.47 .6400 -.03860 .02373
HHNINC| 1.28703 .87733 1.47 .1424 -.43251 3.00657
HHKIDS| -.03754 .31211 -.12 .9043 -.64926 .57419
|Characteristics in numerator of Prob[Y = 4]
Constant| .56268 .83149 .68 .4986 -1.06700 2.19237
AGE| .00343 .01564 .22 .8263 -.02723 .03409
HHNINC| 1.55568* .85486 1.82 .0688 -.11982 3.23118
HHKIDS| .30585 .30374 1.01 .3140 -.28946 .90116
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N16: The Multinomial Logit Model N-295

This model has a separate, independent effect in each utility function.

+---------------------------------------------+
| Random Coefficients MltLogit Model |
| Dependent variable HSAT |
| Log likelihood function -1232.79687 |
| Estimation based on N = 905, K = 20 |
| Inf.Cr.AIC = 2505.594 AIC/N = 2.769 |
| Unbalanced panel has 394 individuals |
+---------------------------------------------+

-----------------------------------------------------------------------------
Random Coefficients MltLogit Model
All parameters have the same random effect
Multinomial logit with random effects
Simulation based on 50 Halton draws
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| .00522 .01994 .26 .7936 -.03387 .04431
HHNINC| .18002 1.04166 .17 .8628 -1.86160 2.22165
HHKIDS| .48013 .38705 1.24 .2148 -.27848 1.23874
AGE| .02077 .01814 1.15 .2520 -.01477 .05632
HHNINC| 1.20948 .82664 1.46 .1434 -.41070 2.82967
HHKIDS| .23686 .35048 .68 .4992 -.45007 .92379
AGE| .00077 .01694 .05 .9636 -.03243 .03397
HHNINC| .96235 .86369 1.11 .2652 -.73045 2.65516
HHKIDS| -.01765 .35090 -.05 .9599 -.70539 .67010
AGE| .01048 .01741 .60 .5472 -.02364 .04460
HHNINC| 1.19343 .87672 1.36 .1734 -.52492 2.91177
HHKIDS| .31389 .34815 .90 .3673 -.36847 .99625
|Means for random parameters
Constant| -.97734 1.00299 -.97 .3298 -2.94317 .98849
Constant| .23872 .96599 .25 .8048 -1.65459 2.13202
Constant| 2.06626** .88897 2.32 .0201 .32392 3.80860
Constant| 1.56019* .90344 1.73 .0842 -.21052 3.33089
|Scale parameters for dists. of random parameters
Constant| .02031 .19069 .11 .9152 -.35343 .39406
Constant| 1.22214*** .17722 6.90 .0000 .87480 1.56948
Constant| 1.73095*** .17833 9.71 .0000 1.38142 2.08048
Constant| 2.55108*** .18704 13.64 .0000 2.18448 2.91768
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N16: The Multinomial Logit Model N-296

This model has the same latent effect in each utility function, though different scale factors.

-----------------------------------------------------------------------------
Random Coefficients MltLogit Model
Dependent variable HSAT
Log likelihood function -1258.50063
Estimation based on N = 905, K = 20
Inf.Cr.AIC = 2557.001 AIC/N = 2.825
Unbalanced panel has 394 individuals
Multinomial logit with random effects
Simulation based on 50 Halton draws
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Nonrandom parameters
AGE| -.00209 .02263 -.09 .9264 -.04644 .04226
HHNINC| .48018 1.17852 .41 .6837 -1.82968 2.79003
HHKIDS| .29347 .43402 .68 .4989 -.55720 1.14414
AGE| .01538 .01558 .99 .3234 -.01515 .04591
HHNINC| 1.34339* .70838 1.90 .0579 -.04501 2.73178
HHKIDS| .21473 .32248 .67 .5055 -.41733 .84679
AGE| -.00776 .01237 -.63 .5304 -.03201 .01649
HHNINC| 1.19572* .65055 1.84 .0661 -.07933 2.47077
HHKIDS| -.05011 .29433 -.17 .8648 -.62699 .52676
AGE| .00310 .01324 .23 .8149 -.02286 .02906
HHNINC| 1.44279** .70145 2.06 .0397 .06796 2.81761
HHKIDS| .31137 .29645 1.05 .2936 -.26967 .89241
|Means for random parameters
Constant| -1.47532 1.20016 -1.23 .2190 -3.82759 .87696
Constant| -.70734 .82080 -.86 .3888 -2.31608 .90140
Constant| 1.09794* .62345 1.76 .0782 -.12401 2.31988
Constant| .64952 .67371 .96 .3350 -.67094 1.96998
|Scale parameters for dists. of random parameters
Constant| 1.38963*** .18611 7.47 .0000 1.02486 1.75439
Constant| .40740*** .09464 4.30 .0000 .22192 .59289
Constant| .26460*** .07701 3.44 .0006 .11367 .41553
Constant| 1.27599*** .10406 12.26 .0000 1.07203 1.47995
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

This model has separate, correlated effects in all utility functions.

-----------------------------------------------------------------------------
Random Coefficients MltLogit Model
Dependent variable HSAT
Log likelihood function -1228.68780
Estimation based on N = 905, K = 26
Inf.Cr.AIC = 2509.376 AIC/N = 2.773
Unbalanced panel has 394 individuals
Multinomial logit with random effects
Simulation based on 50 Halton draws
-----------------------------------------------------------------------------
N16: The Multinomial Logit Model N-297

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
HSAT| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
| Nonrandom parameters
AGE| -.00277 .01900 -.15 .8840 -.04001 .03447
HHNINC| .18258 1.05908 .17 .8631 -1.89318 2.25833
HHKIDS| .44728 .39924 1.12 .2626 -.33522 1.22978
AGE| .01952 .01979 .99 .3239 -.01927 .05832
HHNINC| .99148 .88908 1.12 .2648 -.75109 2.73405
HHKIDS| .19586 .36220 .54 .5887 -.51404 .90577
AGE| -.00134 .01802 -.07 .9407 -.03667 .03398
HHNINC| .74182 .88342 .84 .4011 -.98965 2.47329
HHKIDS| -.06698 .35619 -.19 .8508 -.76510 .63114
AGE| .00795 .01824 .44 .6631 -.02780 .04369
HHNINC| .95944 .89476 1.07 .2836 -.79425 2.71313
HHKIDS| .26625 .34917 .76 .4457 -.41811 .95061
| Means for random parameters
Constant| -1.44262 .98772 -1.46 .1441 -3.37851 .49327
Constant| .03520 1.05196 .03 .9733 -2.02660 2.09700
Constant| 2.00734** .94721 2.12 .0341 .15083 3.86384
Constant| 1.54147 .94470 1.63 .1027 -.31011 3.39305
| Diagonal elements of Cholesky matrix
Constant| .77973*** .21166 3.68 .0002 .36489 1.19458
Constant| 1.02801*** .14489 7.10 .0000 .74403 1.31199
Constant| .22445** .09346 2.40 .0163 .04127 .40763
Constant| .18188** .08031 2.26 .0235 .02447 .33929
| Below diagonal elements of Cholesky matrix
lONE_ONE| .50481*** .18120 2.79 .0053 .14966 .85995
lONE_ONE| 1.08605*** .17694 6.14 .0000 .73926 1.43284
lONE_ONE| .94188*** .13768 6.84 .0000 .67204 1.21172
lONE_ONE| 1.88987*** .18720 10.10 .0000 1.52296 2.25677
lONE_ONE| 1.07104*** .14041 7.63 .0000 .79584 1.34624
lONE_ONE| .37947*** .09765 3.89 .0001 .18807 .57086
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Implied covariance matrix of random parameters
Var_Beta| 1 2 3 4
--------+--------------------------------------------------------------------
1| .607984 .393614 .846831 1.47359
2| .393614 1.31163 1.51651 2.05506
3| .846831 1.51651 2.11703 3.14646
4| 1.47359 2.05506 3.14646 4.89580
Implied standard deviations of random parameters
S.D_Beta| 1
--------+--------------
1| .779734
2| 1.14527
3| 1.45500
4| 2.21265
Implied correlation matrix of random parameters
Cor_Beta| 1 2 3 4
--------+--------------------------------------------------------------------
1| 1.00000 .440776 .746426 .854121
2| .440776 1.00000 .910072 .810972
3| .746426 .910072 1.00000 .977343
4| .854121 .810972 .977343 1.00000
N16: The Multinomial Logit Model N-298

N16.13.2 Dynamic Multinomial Logit Model

The preceding random effects model can be modified to produce the dynamic multinomial
logit model analyzed in Gong, van Soest and Villagomez (2000). Then

exp(β′j xit + γ ′j z it + α ij )
Pijt | αi1,...,αiJ = t = 1,...,Ti, j = 0,1,...,J, i = 1,...,N
∑ exp(β′j xit + γ ′j z it + α ij )
J
j =1

where zit contains lagged values of the dependent variables (these are binary choice indicators for the
choice made in period t) and possibly interactions with other variables. The zit variables are now
endogenous, and conventional maximum likelihood estimation is inconsistent. The authors argue
that Heckman’s treatment of initial conditions is sufficient to produce a consistent estimator. The
core of the treatment is to treat the first period as an equilibrium, with no lagged effects,

exp(δ′j xi 0 + θij )
Pij0 | qi1,...,qiJ = , t = 0, j = 0,1,...,J, i = 1,...,N
∑ exp(δ′j xi 0 + θij )
J
j =1

where the vector of effects, q, is built from the same primitives as α in the later choice probabilities.
Thus, αi = Γvi and q = Φ vi, for the same vi, but different lower triangular scaling matrices. This
treatment slightly less than doubles the size of the model – it amounts to a separate treatment for the
first period.) Full information maximum likelihood estimates of the model parameters,
(b 1,...,b J,γ1,...,γJ,δ1,...,δJ,Γ,Φ) are obtained by maximum simulated likelihood, by modifying the
random effects model. The likelihood function for individual i consists of the period 0 probability as
shown above times the product of the period 1,2,...,Ti probabilities defined earlier.
In order to use this procedure, you must create the lagged values of the variables, and the
products with other variables if any are to be present – that is, the elements of zit. Then, starting
values for both parameter vectors must be provided for the iterations. The program below shows the
several steps involved. In terms of the broad command structure, the essential new ingredient will be
the addition of
; Rh2 = the variables in z

to the model definition. However, again, several steps must precede this, as shown in the command
set below.
To construct this estimator in generic form, we assume the dependent variable is named y
and the independent variables are to be contained in a namelist x. Several commands remain
application specific. These are modified for the specific model. We need a time variable first. For
convenience, periods are numbered 1,...,T with t = 1 being the initial period.

NAMELIST ; x = the x variables in the model, including one $

SAMPLE ; All $
CREATE ; time = Trn(-T,0) $ Fixed number of periods
or CREATE ; time = Ndx(ID,1) $ Unbalanced panel, variable T(i)
N16: The Multinomial Logit Model N-299

Compute the binary variables for the outcomes - endogenous variables.

CREATE ; dit1 = (y=1) ; dit2 = (y=2) ; dit3 = (y=3) ... and so on ... $

Create lagged values of the dummy variables and interactions of lagged dummy variables with other
variables in the model if desired. You will name variables according to your application. This is just
a template. (And repeat likewise for a second, third, ... x variable.)

CREATE ; dit1lag = dit1[-1] ; dit2lag = dit2[-1]

; dit3lag = dit3[-1] ... and so on $
CREATE ; d1x1lag = dit1lag*x1 ; d2x1lag = dit2lag*x1 ... $
NAMELIST ; z = dit1L,dit2L,...,d1x1L,...,... for the z variables $

Fit the time invariant model for the first period and retain the coefficients.

REJECT ; time > 1 $

MLOGIT ; Lhs = y ; Rhs = x $
MATRIX ; delta = b $

Fit the dynamic part for 2,...,Ti and again, save the coefficients.

INCLUDE ; New ; T > 1 $

MLOGIT ; Lhs = y ; Rhs = x,z $
MATRIX ; betagama = b $

The full model for all periods is a random parameters model.

SAMPLE ; All $
MLOGIT ; Lhs = y ; Rhs = x
; Rh2 = z ? This indicates the dynamic MNL model.
; Start = delta,betagama
; RPM ; (options including ; Halton, ; Pts = replications)
; Panel specification
; Fcn = one(n) ; Common $ (; Correlated may be specified)
N17: Conditional Logit Model N-300

N17: Conditional Logit Model

N17.1 Introduction
An individual is assumed to have preferences defined over a set of alternatives (travel
modes, occupations, food groups, etc.)

U(alternative 1) = b 1′xi1 + γ1′zi + e i1

...
U(alternative J) = b J ′xiJ + γJ′zi + eiJ
Observed Yi = choice j if Ui(alternative j) > Ui(alternative k) ∀ k ≠ j.

In this expanded specification, we use xij to denote the attributes of choice j that face individual i –
attributes generally differ across choices and across individuals. We use zi to denote characteristics
of individual i, such as age, income, gender, etc. Characteristics differ across individuals, but not
across choices. The ‘disturbances’ in this framework (individual heterogeneity terms) are assumed
to be independently and identically distributed with identical extreme value distribution; the CDF is

F(ej) = exp(-exp(-ej)).

Based on this specification, the choice probabilities,

Prob[ choice j ] = Prob[Uj > Uk], ∀ k ≠ j

exp(β′ x ji + γ ′j z i )
= , j = 1,...,J,
∑ exp(β′ x mi + γ ′m z i )
J
m=1

where ‘i’ indexes the observation, or individual, and ‘j’ and ‘m’ index the choices. We note at the
outset, the IID assumptions made about ej are quite stringent, and lead to the ‘Independence from
Irrelevant Alternatives’ or IIA implications that characterize the model. Much (perhaps all) the
research on forms of this model consists of development of alternative functional forms and
stochastic specifications that avoid this feature.
The observed data consist of the vectors, xjt and zi and the outcome, or choice, yi. (We also
consider a number of variants.) A well known example is travel mode choice. Samples of
observations often consist of the attributes of the different modes and the choice actually made.
Usually, no characteristics of the individuals are observed beyond their actual choice, though survey
data may include familiar sociodemographics such as age and income. Models may also contain
mixtures of the two types of choice determinants. Chapters E38-E40 present the various aspects of
this model contained in LIMDEP. This chapter describes the basic model specification and
estimation. Other features of the model, including those extensions contained in LIMDEP and
NLOGIT are described in Chapters N18-N22.
N17: Conditional Logit Model N-301

N17.2 The Conditional Logit Model – CLOGIT

In the multinomial logit model described in Chapter N16, there is a single vector of
characteristics that describes the individual, and a set of J parameter vectors. In the ‘discrete choice’
setting of this chapter, these are essentially reversed. The J (not J+1 – we will be changing the
notation slightly here) alternatives are each characterized by a set of K ‘attributes,’ xij. Respondent ‘i’
chooses among the J alternatives. In the example we will use throughout this discussion, a sampled
individual making a trip between Sydney and Melbourne chooses one of four modes of travel, air,
train, bus or car. The attributes include cost, travel time and terminal time, which differ by mode,
and characterize the choice, not the person. The data also include a characteristic of the chooser,
household income. It will emerge shortly however, that MLOGIT and CLOGIT are not different
models at all. The estimator described here accommodates both cases, and mixtures of the two. For
example, for the commuting application just noted, we also have income for the person and traveling
party size, both of which are choice invariant.
For the present, we develop the model with a single parameter vector, b. The model
underlying the observed data is assumed to be the following random utility specification:

U(choice j for individual i) = Uij = b′xij + γ′zi + eij, j = 1,...,J.

The random, individual specific terms, (ei1,ei2,...,eiJ) are once again assumed to be independently
distributed across the utilities, each with the same type 1 extreme value distribution

F(eij) = exp(-exp(-eij)).

Under these assumptions, the probability that individual i chooses alternative j is

Prob[Uij > Uim] for all m ≠ j.

It has been shown that for independent extreme value (Gumbel) distributions, as above, this
probability is
exp ( β′xij + γ ′j z i )
Prob[yi = j] =
∑ exp ( β′xim + γ ′m z i )
J
m=1

where yi is the index of the choice made. As before, we note at the outset that the IID assumptions
made about ej are quite stringent, and induce the ‘Independence from Irrelevant Alternatives’ or IIA
features that characterize the model. We will return to this restriction later in Chapter E40.
Regardless of the number of choices, there is a single vector of K parameters to be estimated. This
model does not suffer from the proliferation of parameters that appears in the MLOGIT model
described in Section N16.2.
For convenience in what follows, we will refer to the estimator as CLOGIT, keeping in
mind, this refers to a command and class of models in LIMDEP and NLOGIT, not a separate
program.
N17: Conditional Logit Model N-302

The basic setup for this model consists of observations on n individuals, each of whom
makes a single choice among Ji choices, or alternatives. There is a subscript on Ji because
ultimately, we will not restrict the choice sets to have the same number of choices for every
individual. The data will typically consist of the choices and observations on K ‘attributes’ for each
choice. The attributes that describe each choice, i.e., the variables that enter the utility functions,
may be the same for all choices, or may be defined differently for each utility function. The
estimator described in this chapter allows a large number of variations of this basic model. In the
discrete choice framework, the observed ‘dependent variable’ usually consists of an indicator of
which among Ji alternatives was most preferred by the respondent. All that is known about the
others is that they were judged inferior to the one chosen. But, there are cases in which information
is more complete and consists of a subjective ranking of all Ji alternatives by the individual.
CLOGIT allows specification of the model for estimation with ‘ranks data.’ In addition, in some
settings, the sample data might consist of aggregates for the choices, such as proportions (market
shares) or frequency counts. CLOGIT will accommodate these cases as well.

N17.3 Clogit Data for the Applications

The documentation of the CLOGIT program below includes numerous applications based on
the data set clogit.dat, that is distributed with LIMDEP and NLOGIT These data provide a compact
illustration of how data should be arranged for the CLOGIT model. The data set is a survey of the
transport mode chosen by a sample of 210 travelers between Sydney and Melbourne (about 500
miles) and other points in nonmetropolitan New South Wales. As will be shown, clogit data will
generally consist of a record (row of data) for each alternative in the choice set, for each individual.
Thus, the data file contains 210 observations, or 840 records. The variables in the data set are as
follows:

Original Data

mode = 0/1 for four alternatives: air, train, bus, car

(this variable equals one for the choice made, labeled choice below),
ttme = terminal waiting time,
invc = invehicle cost for all stages,
invt = invehicle time for all stages,
gc = generalized cost measure = Invc + Invt × value of time,
chair = dummy variable for chosen mode is air,
hinc = household income in thousands,
psize = traveling party size.

Transformed Variables

aasc = choice specific dummy for air (generated internally),

tasc = choice specific dummy for train,
basc = choice specific dummy for bus,
casc = choice specific dummy for car,
hinca = hinc × aasc,
psizea = psize × aasc.
N17: Conditional Logit Model N-303

The table below lists the first 10 observations in the data set. In the terms used here, each
‘observation’ is a block of four rows. The mode chosen in each block is boldfaced.

mode choice ttme invc invt gc chair hinc psize aasc tasc basc casc hinca psizea obs.

Air 0 69 59 100 70 0 35 1 1 0 0 0 35 1 i=1

Train 0 34 31 372 71 0 35 1 0 1 0 0 0 0
Bus 0 35 25 417 70 0 35 1 0 0 1 0 0 0
Car 1 0 10 180 30 0 35 1 0 0 0 1 0 0
Air 0 64 58 68 68 0 30 2 1 0 0 0 30 2 i=2
Train 0 44 31 354 84 0 30 2 0 1 0 0 0 0
Bus 0 53 25 399 85 0 30 2 0 0 1 0 0 0
Car 1 0 11 255 50 0 30 2 0 0 0 1 0 0
Air 0 69 115 125 129 0 40 1 1 0 0 0 40 1 i=3
Train 0 34 98 892 195 0 40 1 0 1 0 0 0 0
Bus 0 35 53 882 149 0 40 1 0 0 1 0 0 0
Car 1 0 23 720 101 0 40 1 0 0 0 1 0 0
Air 0 64 49 68 59 0 70 3 1 0 0 0 70 3 i=4
Train 0 44 26 354 79 0 70 3 0 1 0 0 0 0
Bus 0 53 21 399 81 0 70 3 0 0 1 0 0 0
Car 1 0 5 180 32 0 70 3 0 0 0 1 0 0
Air 0 64 60 144 82 0 45 2 1 0 0 0 45 2 i=5
Train 0 44 32 404 93 0 45 2 0 1 0 0 0 0
Bus 0 53 26 449 94 0 45 2 0 0 1 0 0 0
Car 1 0 8 600 99 0 45 2 0 0 0 1 0 0

Air 0 69 59 100 70 0 20 1 1 0 0 0 20 1 i=6

Train 1 40 20 345 57 0 20 1 0 1 0 0 0 0
Bus 0 35 13 417 58 0 20 1 0 0 1 0 0 0
Car 0 0 12 284 43 0 20 1 0 0 0 1 0 0
Air 1 45 148 115 160 1 45 1 1 0 0 0 45 1 i=7
Train 0 34 111 945 213 1 45 1 0 1 0 0 0 0
Bus 0 35 66 935 167 1 45 1 0 0 1 0 0 0
Car 0 0 36 821 125 1 45 1 0 0 0 1 0 0
Air 0 69 121 152 137 0 12 1 1 0 0 0 12 1 i=8
Train 0 34 52 889 149 0 12 1 0 1 0 0 0 0
Bus 0 35 50 879 146 0 12 1 0 0 1 0 0 0
Car 1 0 50 780 135 0 12 1 0 0 0 1 0 0
Air 0 69 59 100 70 0 40 1 1 0 0 0 40 1 i=9
Train 0 34 31 372 71 0 40 1 0 1 0 0 0 0
Bus 0 35 25 417 70 0 40 1 0 0 1 0 0 0
Car 1 0 17 210 40 0 40 1 0 0 0 1 0 0
Air 0 69 58 68 65 0 70 2 1 0 0 0 70 2 i=10
Train 0 34 31 357 69 0 70 2 0 1 0 0 0 0
Bus 0 35 25 402 68 0 70 2 0 0 1 0 0 0
Car 1 0 7 210 30 0 70 2 0 0 0 1 0 0
N17: Conditional Logit Model N-304

N17.3.1 Setting Up the Data

The clogit data are arranged as follows, where we use a specific set of values for the problem
to illustrate. Suppose you observe 25 individuals. Each individual in the sample faces three choices
and there are two attributes, q and w. For each observation, we also observe which choice was made.
Suppose further that in the first three observations, the choices made were two, three, and one,
respectively. The data matrix would consist of 75 rows, with 25 blocks of three rows. Within each
block, there would be the set of attributes and a variable y, which, at each row, takes the value one if
the alternative is chosen and zero if not. Thus, within each block of J rows, y will be one once and
only once. For the hypothetical case, then, we have:

y q w
i=1 0 q1,1 w1,1
>1 q2,1 w2,1
0 q3,1 w3,1

i=2 0 q1,2 w1,2
0 q2,2 w2,2
>1 q3,2 w3,2

i=3 >1 q1,3 w1,3
0 q2,3 w2,3
0 q3,3 w3,3

and so on, continuing to i = 25, where ‘>’marks the row of the respondent’s actual choice. The
clogit.dat data set shown earlier illustrates the general construction of the data set. Note that for
purposes of CLOGIT, the data are set up in the same fashion as a panel data set in other settings.
When you IMPORT or READ the data for this model, the data set is not treated any
differently. Nobs would be the total number of rows in the data set, in the hypothetical case, 75, not
25, and 840 for clogit.dat. The separation of the data set into the above groupings would be done at
the time this particular model is estimated – that is, after the data are read into the program.

NOTE: Missing values are handled automatically by this estimator. Do not reset the sample or use
SKIP with CLOGIT. Observations which have missing values are bypassed as a group. We note
an implication of this: the multiple imputation programs in LIMDEP and NLOGIT cannot be used to
fill missing values in a multinomial choice setting.

Thus far, it is assumed that the observed outcome is an indicator of which choice was made
among a fixed set of up to 500 choices. There are numerous possible variations:

• Data on the observed outcome may be in the form of frequencies, market shares or ranks.
• The number of choices may differ across observations.

See Chapters N18 and N20 for further details on choice sets and data types also fixed and variable
number of choices and restricting the choice set during estimation.
N17: Conditional Logit Model N-305

N17.4 Command for the Discrete Choice Model

The essential command for the discrete choice models is

CLOGIT ; Lhs = variable which indicates the choice made

; Choices = a set of J names for the set of choices
; Rhs = choice varying attributes in the utility functions
; Rh2 = choice invariant variables, including one for ASCs $

(The commands DISCRETE CHOICE and NLOGIT in this form may also be used.)
The command builder for this model is found in Model:Discrete Choice/Discrete Choice.
The model and the choice set are set up on the Main page. The Rhs variables (attributes) and Rh2
variables (characteristics) are defined on the Options page. Note in the two windows on the
Options page, the Rhs of the model is defined in the left window and the Rh2 variables are specified
in the right window.
A set of exactly J choice labels must be provided in the command. These are used to label
the choices in the output. The number you provide is used to determine the number of choices there
are in the model. Therefore, the set of the right number of labels is essential. Use any descriptor of
eight or fewer characters desired – these do not have to be valid names, just a set of labels, separated
in the list by commas.

The internal limit on J, the number of choices, is 500.

There are K attributes (Rhs variables) measured for the choices. The next chapter will
describe variations of this for different formulations and options. The total number of parameters in
the utility functions will include K1 for the Rhs variables and (J-1)K2 for the Rh2 variables. The total
number of utility function parameters is thus K = K1 + (J-1)K2.

The internal limit on K, the number of utility function parameters, is 300.

The random utility model specified by this setup is precisely of the form

Ui,j = b1xi,1 + b2xi,2 + ... + bK1xi,K1 + γ1,jzi,1 + ... + γK2,jzi,K2 + ei,j

where the x variables are given by the Rhs list and the z variables are in the Rh2 list. By this
specification, the same attributes and the same characteristics appear in all equations, at the same
position. The parameters, bk appear in all equations, and so on. There are various ways to change
this specification of the utility functions – i.e., the Rhs of the equations that underlie the model, and
several different ways to specify the choice set. These will be discussed at various points below.
N17: Conditional Logit Model N-306

Figure N17.1 Command Builder for the Conditional Logit Model

N17: Conditional Logit Model N-307

N17.5 Results for the Conditional Logit Model

The output for the CLOGIT estimator may contain a description of the model before the
statistical results. The description consists of a table that shows the sample proportions (and a ‘tree’
structure that is not useful here) and one that lists the components of the utility functions. You can
request these two listings by adding

; Show Model

to your CLOGIT command. Starting values for the iterations are either zeros or the values you
provide with ; Start = list. As such, there is no initial listing of OLS results. Output begins with the
final results for the model. Here is a sample: The command is

CLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc
; Rh2 = one,hinc
; Show Model $

The full set of results is as follows:

Sample proportions are marginal, not conditional.

Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .27619| 1.000|
|TRAIN .30000| 1.000|
|BUS .14286| 1.000|
|CAR .28095| 1.000|
+----------------+------+---
+---------------------------------------------------------------+
| Model Specification: Table entry is the attribute that |
| multiplies the indicated parameter. |
+--------+------+-----------------------------------------------+
| Choice |******| Parameter |
| |Row 1| INVC INVT GC A_AIR AIR_HIN1 |
| |Row 2| A_TRAIN TRA_HIN2 A_BUS BUS_HIN3 |
+--------+------+-----------------------------------------------+
|AIR | 1| INVC INVT GC Constant HINC |
| | 2| none none none none |
|TRAIN | 1| INVC INVT GC none none |
| | 2| Constant HINC none none |
|BUS | 1| INVC INVT GC none none |
| | 2| none none Constant HINC |
|CAR | 1| INVC INVT GC none none |
| | 2| none none none none |
+---------------------------------------------------------------+
Normal exit: 5 iterations. Status=0, F= 246.1098
N17: Conditional Logit Model N-308

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -246.10979
Estimation based on N = 210, K = 9
Inf.Cr.AIC = 510.220 AIC/N = 2.430
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -283.7588 .1327 .1201
Chi-squared[ 6] = 75.29796
Prob [ chi squared > value ] = .00000
Response data are given as ind. choices
Number of obs.= 210, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.04613*** .01665 -2.77 .0056 -.07876 -.01349
INVT| -.00839*** .00214 -3.92 .0001 -.01258 -.00419
GC| .03633** .01478 2.46 .0139 .00737 .06530
A_AIR| -1.31602* .72323 -1.82 .0688 -2.73353 .10148
AIR_HIN1| .00649 .01079 .60 .5477 -.01467 .02765
A_TRAIN| 2.10710*** .43180 4.88 .0000 1.26079 2.95341
TRA_HIN2| -.05058*** .01207 -4.19 .0000 -.07424 -.02693
A_BUS| .86502* .50319 1.72 .0856 -.12120 1.85125
BUS_HIN3| -.03316** .01299 -2.55 .0107 -.05862 -.00770
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

NOTE: (This is one of our frequently asked questions.) The ‘R-squareds’ shown in the output are
R2s in name only. They do not measure the fit of the model to the data. It has become common for
researchers to report these with results as a measure of the improvement that the model gives over
one that contains only a constant. But, users are cautioned not to interpret these measures as
suggesting how well the model predicts the outcome variable. It is essentially unrelated to this.

To underscore the point, we will examine in detail the computations in the diagnostic
measures shown in the box that precedes the coefficient estimates. Consider the example below,
which was produced by fitting a model with five coefficients subject to two restrictions, or three free
coefficients – npfree = 3. (The effect is achieved by specifying ; Choices = air,(train),(bus),car.

+------------------------------------------------------+
|WARNING: Bad observations were found in the sample. |
|Found 93 bad observations among 210 individuals. |
|You can use ;CheckData to get a list of these points. |
+------------------------------------------------------+
Sample proportions are marginal, not conditional.
Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .49573| 1.000|
|TRAIN .00000| 1.000|*
|BUS .00000| 1.000|*
|CAR .50427| 1.000|
+----------------+------+---
N17: Conditional Logit Model N-309

+---------------------------------------------------------------+
| Model Specification: Table entry is the attribute that |
| multiplies the indicated parameter. |
+--------+------+-----------------------------------------------+
| Choice |******| Parameter |
| |Row 1| GC TTME A_AIR A_TRAIN A_BUS |
+--------+------+-----------------------------------------------+
|AIR | 1| GC TTME Constant none none |
|TRAIN | 1| GC TTME none Constant none |
|BUS | 1| GC TTME none none Constant |
|CAR | 1| GC TTME none none none |
+---------------------------------------------------------------+
Normal exit from iterations. Exit status=0.
-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -62.58418
Estimation based on N = 117, K = 3
Inf.Cr.AIC = 131.168 AIC/N = 1.121
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -81.0939 .2283 .2079
Chi-squared[ 2] = 37.01953
Prob [ chi squared > value ] = .00000
Response data are given as ind. choices
Number of obs.= 210, skipped 93 obs
Restricted choice set. Excluded choices are
TRAIN BUS
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
GC| .01320* .00695 1.90 .0574 -.00042 .02682
TTME| -.07141*** .01605 -4.45 .0000 -.10286 -.03996
A_AIR| 3.96117*** .98004 4.04 .0001 2.04032 5.88201
A_TRAIN| 0.0 .....(Fixed Parameter).....
A_BUS| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------

There are 210 individuals in the data set, but this model was fit to a restricted choice set which
reduced the data set to n = 210 - 93 = 117 useable observations. The original choice set had Ji = 4
choices, but two were excluded, leaving Ji = 2 in the sample. The log likelihood of -62.58418 is
computed as shown in Section N23.6. The ‘constants only’ log likelihood is obtained by setting each
choice probability to the sample share for each outcome in the choice set. For this application, those
are 0.49573 for air and 0.50427 for car. (This computation cannot be done if the choice set varies by
person or if weights or frequencies are used.) Thus, the log likelihood for the restricted model is

Log L0 = 117 ( 0.49573 × log 0.49573 + 0.50427 × log 0.50427 ) = -81.09395.

The ‘R2’ is 1 - (-62.54818/-81.0939) = 0.22869 (including some rounding error). The adjustment factor
is
K = (Σi Ji - n) / [(Σi Ji - n) - npfree] = (234 - 117)/(234 - 117 - 3) = 1.02632.

and the ‘Adjusted R2’ is 1 - K(log L /LogL0)

Adjusted R2 = 1 - 1.02632 (-62.54818/-81.0939) = 0.20794.

N17: Conditional Logit Model N-310

Results kept by this estimator are:

Matrices: b and varb = coefficient vector and asymptotic covariance matrix

Scalars: logl = log likelihood function

nreg = N, the number of observational units
kreg = the number of Rhs variables

Last Model: b_variable = the labels kept for the WALD command

NOTE: This estimator does not use PARTIALS or SIMULATE after estimation. Self contained
routines are contained in the estimator. These are described in Chapters N21 and N22.

In the Last Model, groups of coefficients for variables that are interacted with constants get
labels choice_variable, as in trai_gco. (Note that the names are truncated – up to four characters for
the choice and three for the attribute.) The alternative specific constants are a_choice, with names
truncated to no more than six characters. For example, the sum of the three estimated choice specific
constants could be analyzed as follows:

WALD ; Fn1 = a_air + a_train + a_bus $

-----------------------------------------------------------------------------
WALD procedure. Estimates and standard errors for nonlinear
functions and joint test of nonlinear restrictions.
Wald Statistic = 16.33643
Prob. from Chi-squared[ 1] = .00005
Functions are computed at means of variables
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
WaldFcns| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Fncn(1)| 3.96117*** .98004 4.04 .0001 2.04032 5.88201
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

N17.5.1 Robust Standard Errors

The ‘cluster’ estimator is available in CLOGIT. However, this routine does not support
hierarchical samples. There may be only one level of clustering. Also, the cluster specification is
defined with respect to the CLOGIT groups of data, not the data set. CLOGIT sorts out how many
clusters there are and how they are delineated. But, since the row count of the data set is used in
constructing the estimator, you must treat a group of NALT observations as one. For example, our
sample data used in this section contain 210 groups of four rows of data. Each group of four is an
observation. Suppose that these data were grouped in clusters of three choice situations. The
estimation command with the cluster estimator would appear

CLOGIT ; ... (the model) ; Cluster = 3 $

N17: Conditional Logit Model N-311

The relevant part of the output would appear as follows:

+---------------------------------------------------------------------+
| Covariance matrix for the model is adjusted for data clustering. |
| Sample of 210 observations contained 70 clusters defined by |
| 3 observations (fixed number) in each cluster. |
+---------------------------------------------------------------------+
-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Estimation based on N = 210, K = 9
Number of obs.= 210, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.04613** .01836 -2.51 .0120 -.08211 -.01014
(rows omitted)
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

Use ; Cluster as per the other models in LIMDEP and NLOGIT – the same construction is
used here.

N17.5.2 Descriptive Statistics

Request a set of descriptive statistics for your model by adding

; Describe

to the model command. For each alternative, a table is given which lists the nonzero terms in the
utility function and the means and standard deviations for the variables that appear in the utility
function. Values are given for all observations and for the individuals that chose that alternative.
For the example shown above, the following tables would be produced:

CLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc ; Rh2 = one,hinc
; Describe $

+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative AIR |
| Utility Function | | 58.0 observs. |
| Coefficient | All 210.0 obs.|that chose AIR |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 85.252 27.409| 97.569 31.733 |
| INVT -.0084 INVT | 133.710 48.521| 124.828 50.288 |
| GC .0363 GC | 102.648 30.575| 113.552 33.198 |
| A_AIR -1.3160 ONE | 1.000 .000| 1.000 .000 |
| AIR_HIN1 .0065 HINC | 34.548 19.711| 41.724 19.115 |
+-------------------------------------------------------------------------+
N17: Conditional Logit Model N-312

+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative TRAIN |
| Utility Function | | 63.0 observs. |
| Coefficient | All 210.0 obs.|that chose TRAIN |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 51.338 27.032| 37.460 20.676 |
| INVT -.0084 INVT | 608.286 251.797| 532.667 249.360 |
| GC .0363 GC | 130.200 58.235| 106.619 49.601 |
| A_TRAIN 2.1071 ONE | 1.000 .000| 1.000 .000 |
| TRA_HIN2 -.0506 HINC | 34.548 19.711| 23.063 17.287 |
+-------------------------------------------------------------------------+
+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative BUS |
| Utility Function | | 30.0 observs. |
| Coefficient | All 210.0 obs.|that chose BUS |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 33.457 12.591| 33.733 11.023 |
| INVT -.0084 INVT | 629.462 235.408| 618.833 273.610 |
| GC .0363 GC | 115.257 44.934| 108.133 43.244 |
| A_BUS .8650 ONE | 1.000 .000| 1.000 .000 |
| BUS_HIN3 -.0332 HINC | 34.548 19.711| 29.700 16.851 |
+-------------------------------------------------------------------------+
+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative CAR |
| Utility Function | | 59.0 observs. |
| Coefficient | All 210.0 obs.|that chose CAR |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 20.995 14.678| 15.644 9.629 |
| INVT -.0084 INVT | 573.205 274.855| 527.373 301.131 |
| GC .0363 GC | 95.414 46.827| 89.085 49.833 |
+-------------------------------------------------------------------------+

You may also request a cross tabulation of the model predictions against the actual choices.
(The predictions are obtained as the integer part of Σt P̂ jt yjt.) Add

; Crosstab

to your model command. For the same model, this would produce
+-------------------------------------------------------+
| Cross tabulation of actual choice vs. predicted P(j) |
| Row indicator is actual, column is predicted. |
| Predicted total is F(k,j,i)=Sum(i=1,...,N) P(k,j,i). |
| Column totals may be subject to rounding error. |
+-------------------------------------------------------+
--------+----------------------------------------------------------------------
NLOGIT Cross Tabulation for 4 outcome Multinomial Choice Model
CrossTab| AIR TRAIN BUS CAR Total
--------+----------------------------------------------------------------------
AIR| 19 13 8 18 58
TRAIN| 12 30 9 12 63
BUS| 10 8 6 6 30
CAR| 17 12 7 23 59
--------+----------------------------------------------------------------------
Total| 58 63 30 59 210
N17: Conditional Logit Model N-313

+-------------------------------------------------------+
| Cross tabulation of actual y(ij) vs. predicted y(ij) |
| Row indicator is actual, column is predicted. |
| Predicted total is N(k,j,i)=Sum(i=1,...,N) Y(k,j,i). |
| Predicted y(ij)=1 is the j with largest probability. |
+-------------------------------------------------------+
--------+----------------------------------------------------------------------
NLOGIT Cross Tabulation for 4 outcome Multinomial Choice Model
CrossTab| AIR TRAIN BUS CAR Total
--------+----------------------------------------------------------------------
AIR| 23 15 0 20 58
TRAIN| 8 49 0 6 63
BUS| 13 12 1 4 30
CAR| 15 13 0 31 59
--------+----------------------------------------------------------------------
Total| 59 89 1 61 210

N17.6 Estimating and Fixing Coefficients

Maximum likelihood estimates are obtained by Newton’s method. Since this is a
particularly well behaved estimation problem, zeros are used for the start values with little loss in
computational efficiency. The gradient and Hessian used in iterations and for the asymptotic
covariance matrix are computed as follows:

dji = 1 if individual i makes choice j and 0 otherwise

exp ( β′x ji )
Pji = Prob[yi = j] = Prob[dji = 1] =
∑ exp ( β′x mi )
Ji
m=1

∑ i 1=
∑ j 1 d ji log Pji
n Ji
Log L = =

= ∑ j i=1 Pji x ji ,
J
xi

∂ log L
∑ i 1=
∑ ji 1 d ji (x ji − xi ) ,
n J
==
∂β

∂ 2 log L
∑ i 1=
∑ j i 1 Pji (x ji − xi )(x ji − xit )′ ,
n J
==
∂ββ ∂ ′

Occasionally, a data set will be such that Newton’s method does not work – this tends to occur when
the log likelihood is flat in a broad range of the parameter space. There is no way that you can
discern this from looking at the data, however. If Newton’s method fails to converge in a small
number of iterations, unless the data make estimation impossible, you should be able to estimate the
model by using
; Alg = BFGS

as an alternative. The BFGS algorithm will take slightly longer, but for most data sets, the difference
will be a few seconds. If this method fails as well, you should conclude that your model is
inestimable.
N17: Conditional Logit Model N-314

You may provide your own starting values with

; Start = list of K values

If you have requested a set of alternative specific constants, you must provide starting values for
them as well. Regardless of where ‘one’ appears in the Rhs list, the ASCs will be the last J-1
coefficients corresponding to that list. If you have Rh2 variables, the coefficients will follow the Rhs
coefficients, including the list of ASCs.
Coefficients may be fixed at specific values during optimization. Use

; Fix = variable name [ value ]

for example, ; Fix = ttme [ .01 ]

The following results are obtained from

CLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = gc,ttme
; Rh2 = one
; Fix = ttme[.01] $

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -287.31412
Estimation based on N = 210, K = 4
Inf.Cr.AIC = 582.628 AIC/N = 2.774
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -283.7588 -.0125-.0190
Response data are given as ind. choices
Number of obs.= 210, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
GC| -.02118*** .00403 -5.26 .0000 -.02908 -.01329
TTME| .01000 .....(Fixed Parameter).....
A_AIR| -.53263*** .19044 -2.80 .0052 -.90589 -.15937
A_TRAIN| .40186* .22238 1.81 .0708 -.03400 .83773
A_BUS| -.66610*** .23961 -2.78 .0054 -1.13572 -.19648
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------
N17: Conditional Logit Model N-315

N17.7 Generalized Maximum Entropy Estimator

The CLOGIT multinomial logit model may be estimated using the generalized maximum
entropy estimator described in Section N16.10 for the MLOGIT model. The estimator is the same –
the difference between there and here is only the constraint on the parameter vectors – there is only a
single parameter vector in the CLOGIT model. The computations are identical; the only difference
is the format of the data. The estimator is requested by adding

; GME
or ; GME = number of support points

to the CLOGIT command. In the application below, we reestimate the model used in several
examples, using GME instead of MLE. The MLE is shown at the end of the results for ease of
comparison. The command would be

CLOGIT ; Lhs = mode

; Rhs = one,gc,ttme
; Choices = air,train,bus,car
; GME = 5 $

-----------------------------------------------------------------------------
Generalized Maximum Entropy LOGIT Estimator
Dependent variable Choice
Log likelihood function -1556.27248
Estimation based on N = 210, K = 5
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
GC| -.01014*** .00356 -2.85 .0044 -.01711 -.00316
TTME| -.09407*** .01002 -9.38 .0000 -.11371 -.07442
A_AIR| 5.62289*** .63242 8.89 .0000 4.38337 6.86241
A_TRAIN| 3.68504*** .41687 8.84 .0000 2.86800 4.50209
A_BUS| 3.10729*** .43557 7.13 .0000 2.25360 3.96098
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
N17: Conditional Logit Model N-316

+--------------------------------------------------------------------+
| Information Statistics for Conditional Logit Model fit by GME |
| Number of support points =5. Weights in support scaled to 1/sqr(N) |
| M=Model MC=Constants Only M0=No Model |
| Criterion Function -1556.27248 -1635.80211 -2516.41511 |
| LR Statistic vs. MC 159.05926 .00000 .00000 |
| Degrees of Freedom 2.00000 .00000 .00000 |
| Prob. Value for LR .00000 .00000 .00000 |
| Entropy for probs. 207.71575 283.75877 291.12182 |
| Normalized Entropy .71350 .97471 1.00000 |
| Entropy Ratio Stat. 166.81214 14.72609 .00000 |
| Bayes Info Criterion 3133.93338 3292.99265 5054.21865 |
| BIC - BIC(no model) 1920.28527 1761.22600 .00000 |
| Pseudo R-squared .04862 .00000 .00000 |
| Pct. Correct Prec. 70.47619 30.00000 25.00000 |
| Notes: Entropy computed as Sum(i)Sum(j)Pfit(i,j)*logPfit(i,j). |
| Normalized entropy is computed against M0. |
| Entropy ratio statistic is computed against M0. |
| BIC = 2*criterion - log(N)*degrees of freedom. |
| If the model has only constants or if it has no constants, |
| the statistics reported here are not useable. |
| If choice sets vary in size, MC and M0 are inexact. |
+--------------------------------------------------------------------+

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -199.97662
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
GC| -.01578*** .00438 -3.60 .0003 -.02437 -.00719
TTME| -.09709*** .01044 -9.30 .0000 -.11754 -.07664
A_AIR| 5.77636*** .65592 8.81 .0000 4.49078 7.06193
A_TRAIN| 3.92300*** .44199 8.88 .0000 3.05671 4.78929
A_BUS| 3.21073*** .44965 7.14 .0000 2.32943 4.09204
--------+--------------------------------------------------------------------
N17: Conditional Logit Model N-317

N17.8 MLOGIT and CLOGIT

When there are no choice varying attributes, CLOGIT is the same model as MLOGIT.
From Chapter N16, the functional form for MLOGIT is

exp(β′j xi )
Prob(yi = j|xi) = , j = 0,...,J,
∑ exp(β′m xi )
J
m=1

From the introduction in this chapter,

exp(β′ x ji + γ ′j z i )
Prob(choice = j | Xi,zi) = , j = 1,...,J.
∑ exp(β′ x mi + γ ′m z i )
J
m= 0

In the second equation, if b equals zero – there are no choice varying attributes – then the second
probability is the same as the first, after a simple renaming of the parts; γj in the second replacing b j
in the first, and zi replacing xi. (The alternatives are renumbered, indexing from 1 to J rather than
from 0 to J.) The following illustrates the result:

? CLOGIT using the original data

CLOGIT ; Lhs = mode ; Choices = air,train,bus,car
; Rhs = one ; Rh2 = hinc
; Effects: hinc(*) $
? Create the dependent variable for MLOGIT, using the first row of clogit data
CREATE ; pick = mode*(0*aasc+1*tasc+2*basc+3*casc) $
CREATE ; choice = 3 - (pick+pick[+1]+pick[+2]+pick[+3]) $
? Use only the first row for MLOGIT
MLOGIT ; If[aasc = 1 ] ; Lhs=choice ; Rhs=one,hinc
; Partial Effects
; Labels = car,bus,train,air $

We have normalized MLOGIT so that choice = 0 means pick car and choice = 3 means pick air.
The elasticities then correspond to those in the CLOGIT results, and the coefficients are the same.
N17: Conditional Logit Model N-318

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -261.74506
Estimation based on N = 210, K = 6
Number of obs.= 210, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
A_AIR| .04252 .45456 .09 .9255 -.84840 .93345
A_TRAIN| 2.00595*** .42180 4.76 .0000 1.17923 2.83266
A_BUS| .64169 .49249 1.30 .1926 -.32358 1.60696
AIR_HIN1| -.00142 .00989 -.14 .8858 -.02081 .01797
TRA_HIN2| -.06048*** .01169 -5.17 .0000 -.08339 -.03756
BUS_HIN3| -.03677*** .01282 -2.87 .0041 -.06190 -.01165
--------+--------------------------------------------------------------------

--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Elasticity of Choice Probabilities with Respect to HINC
--------+-----------------------------------
| AIR TRAIN BUS CAR
--------+-----------------------------------
HINC| .5418 -1.4986 -.6796 .5908

-----------------------------------------------------------------------------
Multinomial Logit Model
Dependent variable CHOICE
Log likelihood function -261.74506
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
CHOICE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
|Characteristics in numerator of Prob[BUS ]
Constant| .64169 .49249 1.30 .1926 -.32358 1.60696
HINC| -.03677*** .01282 -2.87 .0041 -.06190 -.01165
|Characteristics in numerator of Prob[TRAIN ]
Constant| 2.00595*** .42180 4.76 .0000 1.17923 2.83266
HINC| -.06048*** .01169 -5.17 .0000 -.08339 -.03756
|Characteristics in numerator of Prob[AIR ]
Constant| .04252 .45456 .09 .9255 -.84840 .93345
HINC| -.00142 .00989 -.14 .8858 -.02081 .01797
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------
Averages of Individual Elasticities of Probabilities
--------+---------+---------+---------+---------+
Variable| CAR | BUS | TRAIN | AIR |
--------+---------+---------+---------+---------+
HINC | .5908 | -.6796 | -1.4986 | .5418 |
--------+---------+---------+---------+---------+
N18: Data Setup for NLOGIT N-319

N18: Data Setup for NLOGIT

N18.1 Introduction
In general, the data for the models described in Chapters N23-N33 will be arranged in a
format that is set up to work well with the specific NLOGIT estimators. In almost all cases, the data
used for all models that you fit with NLOGIT will be set up as if they were a panel. That is, each
individual choice situation will have a set of observations, with one ‘line’ of data for each choice in
the choice set. Thus, in the analogy to a panel, the ‘group’ is a person and the group size would be
the number of choices. You will use this arrangement in nearly all cases. This chapter will explain
the various aspects of setting up the data for the NLOGIT models. We note one specific feature of
the data set that is unusual is the ‘ignored value code,’ -888, described in Section N18.9. This
special code is used to signal values that are deliberately omitted from the data set by the observed
individual – they are ‘missing values,’ with a specific understanding for why they are missing.

N18.2 Basic Data Setup for NLOGIT

In the base case, the data are arranged as follows, where we use a specific set of values for
the problem to illustrate. Suppose you observe 25 individuals. Each individual in the sample faces
three choices and there are two attributes, q and w. For each observation, we also observe which
choice was made. Suppose further that in the first three observations, the choices made were two,
three, and one, respectively. The data matrix would consist of 75 rows, with 25 blocks of three rows.
Within each block, there would be the set of attributes and a variable y, which, at each row, takes the
value one if the alternative is chosen and zero if not. Thus, within each block of J rows, y will be
one once and only once. For the hypothetical case, then, we have:

YQW
i=1 0 q1,1 w1,1
1 q2,1 w2,1
0 q3,1 w3,1

i=2 0 q1,2 w1,2
0 q2,2 w2,2
1 q3,2 w3,2

i=3 1 q1,3 w1,3
0 q2,3 w2,3
0 q3,3 w3,3

and so on, continuing to i = 25, where the arrow marks the row of the respondent’s actual choice.
When you read these data, the data set is not treated any differently from any other panel.
Nobs would be the total number of rows in the data set, in the hypothetical case, 75, not 25. The
separation of the data set into the above groupings would be done at the time your particular model is
estimated.
N18: Data Setup for NLOGIT N-320

NOTE: Missing values are handled automatically by estimation programs in NLOGIT. You should
not reset the sample or use SKIP with the NLOGIT models. Observations that have missing values
are bypassed as a group.

Thus far, it is assumed that the observed outcome is an indicator of which choice was made
among a fixed set of up to 100 choices. Numerous variations on this are possible:

• Data on the observed outcome may be in the form of frequencies, market shares, or ranks.
These possibilities are discussed further in Section N18.3.

• The number of choices may differ across observations. This is discussed further in Section
N20.2.

The preceding described the base case model for a fixed number of choices using individual
level data. There are several alternative formulations that might apply to the data set you are using.

N18.3 Types of Data on the Choice Variable

We allow several types of data on the choice variable, y. If you have grouped data, the
values of y will be proportions or frequencies, instead of individual choices. In the first case, within
each observation (J data points), the values of y will sum to one when summed down the J rows.
(This will be the only difference in the grouped data treatment.) In the second case, y will simply be
a set of nonnegative integers. An example of a setting in which such data might arise would be in
marketing, where the proportions might be market shares of several brands of a commodity. Or, the
data might be counts of responses to particular questions in a survey in which groups of people in
different locations or at different times were surveyed. Finally, y might be a set of ranks, in which
case, instead of zeros and ones, y would take values 1,2,...,J (not necessarily in that order) within,
and reading down, each block.
More specifically, data on the dependent (Lhs) variable may come in these four forms:

• Individual Data: The Lhs variable consists of zeros and a single one which indicates the
choice that the individual made. When data are individual, the observations on the Lhs
variable will sum exactly to 1.0 for every person in the sample. A sum of 0.0 or some other
value will only arise if a data error has occurred. Individual choice data may also be
simulated. See Section N18.3.1 below.

• Proportions Data: The Lhs variable consists of a set of sample proportions. Values range
from zero to one, and again, they sum to 1.0 over the set of choices in the choice set.
Observed proportions may equal 1.0 or 0.0 for some individuals.

• Frequency Data: The Lhs variable consists of a set of frequency counts for the outcomes.
Frequencies are nonnegative integers for the outcomes in the choice set and may be zero.
N18: Data Setup for NLOGIT N-321

• Ranks Data: The Lhs variable consists of a complete set of ranks of the alternatives in the
individual’s choice set. Thus, if there are J alternatives available, the observation will
consist of a full set of the integers 1,...,J not necessarily in that order, which indicate the
individual’s ranking of the alternatives. The number of choices may still differ by
observation. Thus, we might have [(unranked),0,1,0,0,0] in the usual case, and [(ranked)
4,1,3,2,5] with ranks data. Note that the positions of the ones are the same for both sets, by
definition. (See Beggs, Cardell, and Hausman (1981).) You may also have partial rankings.
For example, suppose respondents are given 10 choices and asked to rank their top three.
Then, the remaining six choices should be coded 4.0. A set of ranks might appear thusly:
[1,4,2,4,3,4,4,4,4,4]. The ties must only appear at the lowest level. Ties in the data are
detected automatically. No indication is needed. For later reference, we note the following
for the model based on ranks data:

° You may have observation weights, but no choice based sampling.

° The IIA test described in Section N21.4.1 is not available.
° The number of choices may be fixed or variable, as described above.
° You may keep probabilities or inclusive values as described in Chapter N21.
° Ranks data may only be used with the conditional logit model (CLOGIT) and the
mixed logit (random parameters) model (RPLOGIT).

• Best/Worst Data: This would be a variant of ranks data. When data are in the form of best
and worst, there will be three values of the outcome variable. The choice variable for the
best (most preferred) outcome is coded 1 as usual. The least favored outcome is coded with
any value larger than 1, such as 2, 9, or any other value. Outcomes between these that are
not chosen either best or worst will be coded 0 as usual.

The first three data types are detected automatically by NLOGIT. You do not have to give
any additional information about the data set, since the type of data being provided can usually be
deduced from the values. (See below for one exception.) The ranks data are an exception for which
you would use

NLOGIT ; ... as usual ...; Ranks $

If you are using frequency or proportions data, and your data contain zeros or ones, certain
kinds of observations cannot be distinguished from erroneous individual data, and they may be
flagged as such. For example, in a frequency data set, the observation [0,0,1,1,0,0] is a valid
observation, but for individual data, it looks like a badly coded observation. In order to avoid this
kind of ambiguity, if you have frequency data containing zeros, add

; Frequencies

to your NLOGIT command. (You may use this in any event to be sure that the data are always
recognized correctly.) If you have proportions data, instead, you may use

; Shares

to be sure that the data are correctly marked. (Again, this will only be relevant if your data contain
zeros and/or ones.)
N18: Data Setup for NLOGIT N-322

Best/worst data can come in three forms. The simple case is that in which the chooser
simultaneously identifies the most and least favored alternatives. For this case, use

; Best worst.

If the chooser identifies the best alternative and then chooses the worst among those that remain,
then this choice is sequential. Indicate this with

; Best worst ; Seq = bw.

Lastly, if the worst alternative is indicated first and the best is chosen from among the remainder, use

; Best worst ; Seq = wb.

The actual data will look the same for the three cases. The difference that is implied relates to the
way the likelihood is formulated for estimation. We note, in practical terms, the three are likely to
produce similar, if not indistinguishable results. This makes sense. Most of the information about
preferences is provided by the most favored alternative. This will be the same in all cases. The
change from simultaneous choice to sequential is likely only to lead to marginal changes in the
model results.
Data are checked for validity and consistency. An unrecognizable mixture of the three types
will cause an error. For example, a mixture of frequency and proportions data cannot be properly
analyzed. For the ranks data, an error will occur if the set of ranks is miscoded or incomplete or if
ties are detected at any ranks other than the lowest.

N18.3.1 Unlabeled Choice Sets

In some situations, particularly in choice experiments and survey data, the choices will not
be a well defined set of alternatives such as (air, train, bus, car), but, rather will simply be a set of
unordered choices distinguished only by the different attributes. For example, in a marketing
experiment, the choice set might consist of (first, second, third, none of these). When the choice set
does not have natural labels, you may use

; Choices = number_name

to define the list. For our example, we might use

; Choices = 3_brand,none

which produces the list (brand1,brand2,brand3,none).

N18: Data Setup for NLOGIT N-323

N18.3.2 Simulated Choice Data

For some kinds of experiments and simulations, you might want to draw a random sample of
choices given known utility functions. NLOGIT allows simulation of the Lhs variable in a choice
model using
Y = j* from Max(Uij),

where Uij = vij + a simulated random term. You must provide the utility values as the Lhs variable.
The choice outcome is then simulated by adding a type 1 extreme value error term to each utility
value, and choosing the j associated with the largest simulated utility. Request this computation by
adding
; MCS (for Monte Carlo Simulation)

to the NLOGIT or CLOGIT command. (The utilities are not lost. You can reuse them, for
example to do another simulation. On the other hand, the simulated data are lost at the end of the
estimation.) Keep in mind, if you want to reuse the data for a simulation, you have to reset the seed
for the random number generator. You might for example want to fit different models with the same
simulated data set. For example, suppose you wanted to compare the results of two different nesting
specifications using the simulated data. The utilities are in variable utility.
The command set might appear as follows:

CALC ; Ran(56791) $
NLOGIT ; Lhs = utility ; Choices = air,train,bus,car
; Tree = (air,train,bus),(car)
; ... $
CALC ; Ran(56791) $
NLOGIT ; Lhs = utility ; Choices = air,train,bus,car
; Tree = (train,bus),(air,car)
; ... $

N18.3.3 Checking Data Validity

NLOGIT does a full check of the data for bad observations (usually coding errors or missing
values) before estimation is done. The program output will contain a simple count of the number of
invalid observations that have been bypassed. For example, we sprinkled some missing values into
the clogit.dat data set, and fit a model. The initial output contains the count:

+------------------------------------------------------+
|WARNING: Bad observations were found in the sample. |
|Found 3 bad observations among 210 individuals. |
|You can use ;CheckData to get a list of these points. |
+------------------------------------------------------+
N18: Data Setup for NLOGIT N-324

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -181.67965
Estimation based on N = 207, K = 7
Inf.Cr.AIC = 377.359 AIC/N = 1.823
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -279.9949 .3511 .3437
Chi-squared[ 4] = 196.63055
Prob [ chi squared > value ] = .00000
Response data are given as ind. choices
Number of obs.= 210, skipped 3 obs
-----------------------------------------------------------------------------

You may request the program to show you exactly where the problem observations are by adding

; Check Data

to the command. A complete listing of the bad observations is produced – note in a large data set,
this could be quite long. For the preceding, we obtained

+----------------------------------------------------------+
| Inspecting the data set before estimation. |
| These errors mark observations which will be skipped. |
| Row Individual = 1st row then group number of data block |
+----------------------------------------------------------+
1 1 Individual data, LHS variable is not 0 or 1
9 3 Missing value found for characteristic or attribute in utility
17 5 Missing value found for LHS variable

N18.4 Weighting
You can, in principle, use any weighting variable you wish with this model to weight
observations. The model does not require that weights be the same for all outcomes for a given
observation. For example, in a grouped data case, you might have at hand the total number of
observations which gave rise to each of the proportions in the proportions data. If so, you could use
the information to replicate each observation the appropriate number of times. In this case, use the

; Wts = name

option on the CLOGIT command, as you would with any other model. Normally, this variable
would take the same value for each of the J data vectors associated with observation i. (Suppose
instead of 0,1,0 for the first observation, we observed .4, .5, .1 based on 200 observations. Then,
‘name’ would take the value 200 for the first three observations, etc.) (Of course, you could achieve
the same result by providing the frequencies as the Lhs variable.)
N18: Data Setup for NLOGIT N-325

N18.5 Choice Based Sampling

The weighting may be based on the outcomes. For example, suppose the model predicts
mode of travel, car, train, or horse. The true population proportions are known to be .6, .35, and .05.
But, we deliberately oversample the last category so that the sample proportions are, say, .5, .3, and
.2. In estimation, to account for the nonrandom sampling, we would use a weighting scheme which
gives observations in which outcome 1 (car) received a weight of .6/.5 = 1.2, outcome 2 (train),
.35/.3 = 1.16667, and outcome 3 (horse), .05/.2 = .25. Notice that regardless of the number of
observations, the weighting variable in this scenario takes only J values, where J is the number of
outcomes. The Lerman-Manski (1981) correction to the variance matrix of the estimates is used at
convergence to obtain the appropriate standard errors. The covariance matrix used is V = H-1DH-1,
where H is the weighted Hessian and D is the weighted sum of the outer products of the first
derivatives, as opposed to V = H-1 which would be used normally.
To request this procedure, it is only necessary for you to provide the J population weights.
Everything else is automated. The weights are provided after the labels for the outcomes following a
slash. The following example is consistent with the discussion above. The unweighted specification
would be

CLOGIT ; ... ; Choices = car,train,horse $

The choice based sampling weights would be provided in

CLOGIT ; ... ; Choices = car,train,horse / .6,.35,.05 $

Notice that you only provide the population weights. The program obtains the sample proportions
and computes the appropriate weights for the estimator. This is a bit different from the earlier
applications (probit and logit), and it is the only estimator in NLOGIT for which you provide only the
population weights, as opposed to the sampling ratios.
Everything else is the same as before. Note, you do not use a weighting (; Wts) variable
here. Your population weights must sum to 1.0; if not, an error occurs and estimation is halted. If
you provide population weights, you must give a full set. Thus, if your list has the slash
specification, the number of values after the slash must match exactly the number of labels before it.
The data used in our examples are choice based. The example below shows the use of this
option to make the appropriate corrections to the estimates:

CLOGIT ; Lhs = mode

; Rhs = invc,invt,gc,ttme
; Rh2 = one
; Choices = air,train,bus,car / .14,.13,.09,.64
; Show $

The ; Show parameter requests the display of the table below. Otherwise, only the note in the box of
diagnostic statistics indicates use of the choice based sampling estimator.)
N18: Data Setup for NLOGIT N-326

Sample proportions are marginal, not conditional.

Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .27619| .507|
|TRAIN .30000| .433|
|BUS .14286| .630|
|CAR .28095| 2.278|
+----------------+------+---
+---------------------------------------------------------------+
| Model Specification: Table entry is the attribute that |
| multiplies the indicated parameter. |
+--------+------+-----------------------------------------------+
| Choice |******| Parameter |
| |Row 1| INVC INVT GC TTME A_AIR |
| |Row 2| A_TRAIN A_BUS |
+--------+------+-----------------------------------------------+
|AIR | 1| INVC INVT GC TTME Constant |
| | 2| none none |
|TRAIN | 1| INVC INVT GC TTME none |
| | 2| Constant none |
|BUS | 1| INVC INVT GC TTME none |
| | 2| none Constant |
|CAR | 1| INVC INVT GC TTME none |
| | 2| none none |
+---------------------------------------------------------------+
Normal exit: 6 iterations. Status=0, F= 132.5388

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -132.53879
Estimation based on N = 210, K = 7
Vars. corrected for choice based sampling
Response data are given as ind. choices
Number of obs.= 210, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.11080*** .02336 -4.74 .0000 -.15659 -.06502
INVT| -.01736*** .00299 -5.81 .0000 -.02322 -.01151
GC| .09787*** .01967 4.98 .0000 .05931 .13643
TTME| -.13929*** .02589 -5.38 .0000 -.19003 -.08855
A_AIR| 5.68250*** 1.58789 3.58 .0003 2.57029 8.79472
A_TRAIN| 4.09890*** .90704 4.52 .0000 2.32113 5.87667
A_BUS| 3.91452*** .92554 4.23 .0000 2.10050 5.72854
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

These are the parameter estimates computed without the correction for choice based sampling. This
is not only a correction to the covariance matrix. The parameter estimates will change as well.
N18: Data Setup for NLOGIT N-327

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.08493*** .01938 -4.38 .0000 -.12292 -.04694
INVT| -.01333*** .00252 -5.30 .0000 -.01827 -.00840
GC| .06930*** .01743 3.97 .0001 .03513 .10346
TTME| -.10365*** .01094 -9.48 .0000 -.12509 -.08221
A_AIR| 5.20474*** .90521 5.75 .0000 3.43056 6.97893
A_TRAIN| 4.36060*** .51067 8.54 .0000 3.35972 5.36149
A_BUS| 3.76323*** .50626 7.43 .0000 2.77098 4.75548
--------+--------------------------------------------------------------------

N18.6 Entering Data on a Single Line

Data for NLOGIT are generally provided as if in a panel data set, in blocks of Ji observations
per individual, where Ji is the number of choices. The following describes an alternative format in
which data for these models are provided in one line per individual. This construction can only be used
for discrete choice models with a fixed number of alternatives available to each individual. This
feature is not available for cases in which the choice set varies across individuals. (We have seen this
arrangement of data called the ‘wide form,’ with the data arranged as earlier in the ‘long form.’)
In general, discrete choice models require that the data set be arranged with a line of data
(observation) for each alternative in the model, essentially as a panel. For purposes of the
discussion, it will be useful to consider an example. Suppose individuals choose among four
alternatives, (air,train,bus,car), and the attributes are cost and traveltime, which vary across choice,
and income which is fixed. The actual data for an observation would consist of four variables on
four records, arranged as follows: (The yj variable consists of three zeros and a one to indicate the
choice made.)
The arrangement is

Choice Cost Time Income

Air  yair costair timeair income 
Train y costtrain timetrain income 
 train
Bus  ybus costbus timebus income 
 
Car  ycar costcar timecar income 

The model observation would be constructed from the four variables, and would, with alternative
specific constants for the first three alternatives, ultimately appear as follows:

choice cost time constants income

 yair ca ta 1 0 0 income 0 0 
y ct tt 0 1 0 0 income 0 
Xi =  train
 ybus cb tb 0 0 1 0 0 income 
 
 ycar cc tc 0 0 0 0 0 0 
N18: Data Setup for NLOGIT N-328

This setup normally requires four lines of data. But, an alternative way to arrange the same
data would be in a single line of data, consisting of

Choice(coded 0,1,2,3) ca ct cb cc ta tt tb tc one income

from which it would be straightforward to construct the observation above.

The command for this arrangement will contain the following to set this up: First, the choice
set is specified as follows:

; Lhs = the name of the choice variable (here, choice)

; Choices = the list of J choice labels [coding of Lhs variable]

The coding is contained in square brackets. If the dependent variable is coded as consecutive
integers, such as 0,1,2,3, then just put the first value in the brackets. Thus, 0,1,2,3 is indicated with
[0], while 1,2,3,4 is [1]. For our example, this is going to appear

; Lhs = choice
; Choices = air,train,bus,car [0]

If the coding is some other set of integers, put the set of integers in the square brackets. Suppose, for
example, in our model, we eliminated train as a choice. Then, the coding might be [0,2,3].

NOTE: It is only the square brackets in the ; Choices specification which indicates that you will be
using this data arrangement instead of the standard one.

Second, for variables which provide attributes which vary by choice, such as cost and time above, a;
Rhs specification must contain blocks of J variable names. For the example, this might be

; Rhs = cair,ctrain,cbus,ccar,tair,ttrain,tbus,tcar

For variables which are to be interacted with alternative specific constants, as well as the constants
themselves, use ; Rh2 instead of ; Rhs. Thus, for the example above, we might use

; Rh2 = one,income

NOTE: To request a set of alternative specific constants, include one in the Rh2 list, not the Rhs
list.

Notice that when these interactions are created, the last one in the set is dropped. In the example
above, only three constants and three income terms appear in the four choice model.
Third, for the Rhs groups, a name is created for the group, attrib01, attrib02, and so on. If
you would like to provide your own names for the blocks, use

; Attr = list of k labels

N18: Data Setup for NLOGIT N-329

To combine all of these in our example, we might use

; Lhs = mode
; Choices = air,train,bus,car [ 0 ]
; Rhs = cair,ctrain,cbus,ccar,tair,ttrain,tbus,tcar
; Rh2 = one,income
; Attr = cost,time
The following options are unavailable when data are arranged on a single line:
• Data scaling: See Section N18.10.
• Ranks data: See Section N18.3.
• Keeping predictions, probabilities, inclusive values, etc. See the relevant parts of Chapter N21.
• Model: U(...) = spec...: You must use ; Rhs and/or ; Rh2. See Chapter N19.

N18.7 Converting One Line Data Sets for NLOGIT

Data for the several discrete choice models in NLOGIT are assumed to be arranged in a
‘stack’ for each observation. For example, suppose you are studying mode choice for transportation
(of course), and your observation consists of the following (as in the preceding example):
• The choice variable, choice = 1, 2, or 3 for car, train, bus,
• For each mode, time, cost – note that this differs by choice,
• For the individual, age, income – note that this does not differ by choice.
NLOGIT would usually expect each observation in the sample to consist of three rows, such as the
following
choice time cost age income
car 0 44 125 37 56.5 
train 1 29 40 37 56.5 
.

bus 0 56 25 37 56.6 

Suppose that your data were arranged not in this fashion, but in a single observation, as in

choicei ctime ttime btime ccost tcost bcost agei incomei 

 2 56.6 
.
 44 29 56 125 40 25 37

The estimator in NLOGIT can handle either arrangement, but for several purposes it will usually be
more convenient to use the first. You can convert this one line observation to the three record format
in order to use NLOGIT’s estimation programs. There are two ways to do so. NLOGIT provides a
command that does the full conversion of the data set internally for you – essentially it creates a new
data set for you. The second way to convert the data set is to write a new data file (using NLOGIT’s
commands) containing the necessary variables, and read in the newly created data set. You could
use this operation to create a data set for export as well. We note, there are relatively few
commercial packages available that do the kinds of modeling that you will do with NLOGIT – for
several of the models, NLOGIT is unique. As far as we are aware, other software generally use the
more cumbersome single line format. You will find the operation useful when you import data from
other programs into NLOGIT.
N18: Data Setup for NLOGIT N-330

N18.7.1 Converting the Data Set to Multiple Line Format

The single line format for multinomial choice modeling is clumsy, and will become extremely
unwieldy if the choice set has more than a few alternatives or the model has more than two or three
attributes. A utility program is provided for you to convert single line choice data to the more
convenient format.
We wish to transform the data set so that one observation in the second form shown above
becomes three observations in the first form above. The general command is

NLCONVERT ; Lhs = one or more choice variables

; Choices = the J names for the choices in the choice set
; Rhs = K sets of J variable names – the attributes
; Rh2 = M characteristics variables
; Names = names for new choice variables,
names for new attribute variables,
names for new characteristic variables $

For the example above, the command would be

NLCONVERT ; Lhs = choicei

; Choices = car,train,bus
; Rhs = ctime,ttime,btime,ccost,tcost,bcost
; Rh2 = agei,incomei
; Names = choice,time,cost,age,income $

This command is set up to resemble a model command to make it simple to construct. But, it does
nothing but rearrange the data set.
Some points to note about NLCONVERT are:

• It is only for choice settings with fixed numbers of choices for every observation

• You can recode more than one choice variable with the other data

• You can rearrange the entire data set, not just the variables for a particular model. The
appearance of the command as a model command is only for convenience.

• After the data are converted, the new data are placed at the top of the data array, regardless
of where they were before. You can, for example, convert rows 201 to 250 in your data set.
If this is a three choice setting, the new data will be observations 1 to 150.
N18: Data Setup for NLOGIT N-331

There are also several conventions that must be followed:

• The new names must not be in use for anything else already in your project, including other
variables. NLCONVERT cannot replace existing variables.

• You must provide the ; Names and ; Choices specifications. These are mandatory.

• You must provide at least one of ; Rhs or ; Rh2 variable. Either is optional, but at least one
of the two must be present.

• Note that the count of Rhs variables is an exact multiple of the number of choices in the ;
Choices list.

• The number of names in the ; Names list is the sum of

° the number of Lhs variables

° the number of sets of Rhs variables
° the number of Rh2 variables.

Note that the count of Rhs variables is an exact multiple of the number of choices in the ; Choices
list.
When NLCONVERT is executed, the sample is reset to the number of observations in the
new sample. There is an additional option with NLCONVERT. After the data are converted, you
can discard the original data set with

; Clear

This leaves the entire data set consisting of the variables that are in your ; Names list. (Use this with
caution. The operation cannot be reversed.)
To illustrate the operation of this command, suppose the data set consists of these three
observations:

choicei1 choicei 2 ctime ttime btime ccost tcost bcost agei incomei 
 2 3 44 29 56 125 40 25 37 56.6 
 .
 1 1 19 44 20 160 18 50 42 98.6 
 
 3 2 28 55 15 85 50 9 10 22.0 

We wish to convert this data set to NLOGIT’s multiple line format. There are three choices in the
choice set, so there will be three rows of data for each observation.
N18: Data Setup for NLOGIT N-332

The command and the results are as follows:

IMPORT $
choicei1,choicei2,ctime,ttime,btime,ccost,tcost,bcost,agei,incomei
2,3,44,29,56,125,40,25,37,56.6
1,1,19,44,20,160,18,50,42,98.6
3,2,28,55,15, 85,50, 9,10,22.0

ENDDATA $
NLCONVERT ; Lhs = choicei1,choicei2
; Choices = car,train,bus
; Rhs = ctime,ttime,btime,ccost,tcost,bcost
; Rh2 = agei,incomei
; Names = choice1,choice2,time,cost,age,income ; Clear $

Figure N18.1 Converted Data Set

=================================================================
Data Conversion from One Line Format for NLOGIT
Original data were cleared. This is now the whole data set.
The new sample contains 9 observations.
=================================================================
Choice set in new data set has 3 choices:
CAR TRAIN BUS
-----------------------------------------------------------------
There were 2 choice variables coded 1,..., 3 converted to binary
Old variable = CHOICEI1, New variable = CHOICE1
Old variable = CHOICEI2, New variable = CHOICE2
-----------------------------------------------------------------
There were 2 sets of variables on attributes converted. Each
set of 3 variables is converted to one new variable
New Attribute variable TIME is constructed from
CTIME TTIME BTIME
New Attribute variable COST is constructed from
CCOST TCOST BCOST
-----------------------------------------------------------------
There were 2 characteristics that are the same for all choices.
Old variable = AGEI , New variable = AGE
Old variable = INCOMEI , New variable = INCOME
=================================================================
N18: Data Setup for NLOGIT N-333

N18.7.2 Writing a Multiple Line Data File for NLOGIT

If you need to create a data file in the multiple line format, you can, of course, use
NLCONVERT, then just use WRITE to create the file. The following shows a way that you can
bypass NLCONVERT if you wish. The first command creates the three choice variables (one will
appear in each row of the new data set).

CREATE ; car = (choice=1) ; train = (choice=2) ; bus = (choice=3) $

The next command writes out the 15 variables, but only allows five items to appear on each line,
which is what you need to recreate the data file.

WRITE ; car, ctime, ccost, age, income,

train, ttime, tcost, age, income,
bus, btime, bcost, age, income
; File = whatever you choose
; Format = (exactly 5 format codes, not 15) $

For example, ; Format = ( 5F10.3). See Chapter R3 for discussion of using formats for reading and
writing data files.
The WRITE command takes advantage of a very useful feature of this type of formatting.
The WRITE command instructs NLOGIT to write 15 values, but it provides only five format codes.
What happens is that the program will write the first five values according to the format given, then
start over in the same format, on a new line. That is exactly what we want. This WRITE command
writes three lines per observation. When it is done, the data can be read back into NLOGIT with no
further processing necessary, in the format required for NLOGIT.

N18.8 Merging Invariant Variables into a Panel

Some panel data sets contain variables that do not vary across the observations in a group. A
common example is the data shown in the preceding two sections. Some variables in the data set
will be attributes of the choices, and, as such, will be different for each choice. Others may be
characteristics of the individual, and will, therefore, be repeated on each record in the panel.
NLOGIT allows you to keep separate data files for the variable and invariant data. This may result in
a large amount of space saving. The data may be merged when they are read into NLOGIT, rather
than in the data set. For example, consider a panel with three individuals, and a variable number of
observations per individual, two, then three, then two. The two data sets might look like

File=var.dat File=invar.dat
Variable data Invariant data
xyniz
ind=1 1.1 4 2 ind=1 100.7
1.2 2 2 ind=2 93.6
ind=2 3.7 8 3 ind=3 88.2
4.9 3 3
5.0 1 3
ind=3 0.1 2 2
1.2 5 2
N18: Data Setup for NLOGIT N-334

Note the usual count variable for handling panels. To merge these files, use this setup

READ ; File = var.dat ; Nobs = 7 ; Nvar = 3 ; Names = x,y,ni $

This reads the original panel data set. Now, to expand the invariant data, the syntax is

READ ; File = invar.dat ; Nobs = 3 ; Nvar = 1 ; Names = z ; Group = ni $

The new feature is the ; Group = ... specification. ; Group specifies either a count variable, as
above, or a fixed group size, as usual for NLOGIT’s handling of panel data sets. The resulting data
will be
x y ni z
ind=1 1.1 4 2 100.7
1.2 2 2 100.7
ind=2 3.7 8 3 93.6
4.9 3 3 93.6
5.0 1 3 93.6
ind=3 0.1 2 2 88.2
1.2 5 2 88.2

Note the following checks and errors:

• Nobs must be given on the second READ command.

• Nobs must match exactly the number of groups in the existing data set.

• The existing panel must be properly blocked out by the ; Groups variable or by a constant
group size.

• This form may not be used with spreadsheet files.

• This form may not be used to read data ; By Variables.

• This form may not be used with the APPEND command.

• The first data set could be read with a simple IMPORT ; File = var.dat $ command,
however, the second requires a fully specified READ command because of the merging
feature.
N18: Data Setup for NLOGIT N-335

N18.9 Modeling Choice Strategy

In some occasions in survey data, particularly in stated preference experiments, respondents
will indicate that they did not consider certain attributes among a set of attributes in making their
choices. When this aspect of the data is known, it has been conventional to insert zeros for the
attribute in the choice model, thereby to remove that attribute from the utility function. However, in
fact, that does not remove the attribute from the choice probability; it forces it to enter with a
peculiar, possibly extreme value. Consider, for example, a price variable. If a respondent indicates
that they ignored price in a choice, then setting the price to zero in the choice set would force an
extreme value on the choice process. Hensher, Rose, and Greene (2005) argued that if a respondent
truly ignores an attribute in a choice situation, then what should be zero in the choice model is not
the attribute, but its coefficient in the utility function. That restriction definitely removes the
attribute from the choice consideration by taking it out of the model altogether.
Accommodating this idea requires, in essence, that there be a possibly different model for
each respondent. That is, one with possibly different zero restrictions imposed for different
individuals. NLOGIT allows you to automate precisely this formulation in all discrete choice models
with a special data coding.
For respondents which ignore attributes (it must be known in the data) simply code the
attribute with value -888 for this respondent.
With this data convention, the program autodetects this feature and adjusts the model accordingly.
You do not have to add any other codes to any NLOGIT commands to signal this aspect of the data.
The model output will contain a diagnostic box noting when this option is being used when NLOGIT
finds these values in the data. Some aspects of this convention are:

• At least some respondents must actually consider the attribute. It cannot be omitted from the
model for everyone.
• In the multinomial, multiperiod probit model, if an attribute is ever ignored, it must be
ignored in all periods. This is not the case for LCM or RPL which use repeated choice
situation data. A respondent may ignore attributes in some choice situations (say the later
ones in an experiment) and not in others (say the early ones).
• In nested logit models, this feature can only be used at the lowest, twig level of the tree. It
will not be picked up if it used at branch or higher levels. For example, in nested logit
models, one often puts the demographic data in the model at the branch level. This feature
will not be picked up in branch level variables.
• In computing elasticities, if ; Means is used, it may distort the means slightly. How much so
depends on how many observations are in use and how often the attribute is ignored. No
generalizations are possible.
• In computing descriptive statistics with the ; Describe option, this may distort the means
because the -888 values are not skipped, they are changed to 0.0. Output will contain a
warning to this effect if it is noticed.
• In models that can produce person specific parameters (mixed logit, latent class), the saved
parameters for the individual will contain the requested zeros if the indicated attribute is
noted as not used.
N18: Data Setup for NLOGIT N-336

N18.10 Scaling the Data

In some applications involving stated preference data, it is useful to estimate the model with
different scales of the same data. That is, if all of the data on all attributes are collected in a matrix,
X, then we estimate the discrete choice model with the data set

X* = qX,

for different values (near 1.0) of the scalar q. There are two ways to do this. Suppose the attributes
in X are named x1, x2, ..., xk. To set up the procedure, we create a placeholder for X*:

CREATE ; x1s = x1 ; x2s = x2 ; ... $

Now, define the matrices:

NAMELIST ; x = x1, x2,..., xk

; xs = x1s, x2s, ... , xks $

Finally, define a procedure which sets up the NLOGIT estimation in terms of the variables in xs
instead of x, along with a MATRIX command that does the scaling:

PROCEDURE $
CREATE ; xs = x $
MATRIX ; xs = Xmlt(theta) $
NLOGIT ; ... $
ENDPROCEDURE $

Now, the model can be fit with any desired scaling of the data with the command

EXECUTE ; theta = the desired value $

NLOGIT also provides a more fully automated procedure for scaling when you wish to
change only some of the variables in a model. You can specify as part of the command

; Scale ( list of variables ) = qlow , qhigh , number of points.

This requests NLOGIT to examine ‘number of points’ equally spaced values ranging from qlow to
qhigh. The value associated with the highest value of the log likelihood is then used to reestimate the
model. (No output is produced during the search.) You may also specify a second round, finer
search with
; Scale (list of variables ) = qlow , qhigh , number of points , nfine.

If you specify the second round search (nfine), evenly spaced points ranging from the adjacent
values below and above the value found in the first search are examined to try to improve the value
of the log likelihood. For example, if you specify the grid .5,1.5,11,11, the first search will examine
the values .5, .6, ..., 1.5. If the best value were found at, say, 1.2, then the finer search would
examine 1.10, 1.12, .., 1.30.
N18: Data Setup for NLOGIT N-337

N18.11 Data for the Applications

The documentation of the NLOGIT program in the chapters to follow includes numerous
applications based on the data set clogit.dat, that is distributed with NLOGIT. These data are a
survey of the transport mode chosen by a sample of 210 travelers between Sydney and Melbourne
(about 500 miles) and other points in nonmetropolitan New South Wales. Data for NLOGIT will
generally consist of a record (row of data) for each alternative in the choice set, for each individual.
Thus, the data file contains 210 observations, or 840 records. The variables in the data set are as
follows:

Original Data

mode = 0/1 for four alternatives: air, train, bus, car

(this variable equals one for the choice made, labeled choice below),
ttme = terminal waiting time,
invc = invehicle cost for all stages,
invt = invehicle time for all stages,
gc = generalized cost measure = Invc + Invt × value of time,
chair = dummy variable for chosen mode is air,
hinc = household income in thousands,
psize = traveling party size.

Transformed variables

aasc = choice specific dummy for air (generated internally),

tasc = choice specific dummy for train,
basc = choice specific dummy for bus,
casc = choice specific dummy for car,
hinca = hinc×aasc,
psizea = psize×aasc.
N18: Data Setup for NLOGIT N-338

The table below lists the first 10 observations in the data set. In the terms used here, each
‘observation’ is a block of four rows. The mode chosen in each block is boldfaced.

mode choice ttme invc invt gc chair hinc psize aasc tasc basc casc hinca psizea obs.
Air 0 69 59 100 70 0 35 1 1 0 0 0 35 1 i=1
Train 0 34 31 372 71 0 35 1 0 1 0 0 0 0
Bus 0 35 25 417 70 0 35 1 0 0 1 0 0 0
Car 1 0 10 180 30 0 35 1 0 0 0 1 0 0
Air 0 64 58 68 68 0 30 2 1 0 0 0 30 2 i=2
Train 0 44 31 354 84 0 30 2 0 1 0 0 0 0
Bus 0 53 25 399 85 0 30 2 0 0 1 0 0 0
Car 1 0 11 255 50 0 30 2 0 0 0 1 0 0
Air 0 69 115 125 129 0 40 1 1 0 0 0 40 1 i=3
Train 0 34 98 892 195 0 40 1 0 1 0 0 0 0
Bus 0 35 53 882 149 0 40 1 0 0 1 0 0 0
Car 1 0 23 720 101 0 40 1 0 0 0 1 0 0
Air 0 64 49 68 59 0 70 3 1 0 0 0 70 3 i=4
Train 0 44 26 354 79 0 70 3 0 1 0 0 0 0
Bus 0 53 21 399 81 0 70 3 0 0 1 0 0 0
Car 1 0 5 180 32 0 0 3 0 0 0 1 0 0
Air 0 64 60 144 82 0 45 2 1 0 0 0 45 2 i=5
Train 0 44 32 404 93 0 45 2 0 1 0 0 0 0
Bus 0 53 26 449 94 0 45 2 0 0 1 0 0 0
Car 1 0 8 600 99 0 45 2 0 0 0 1 0 0
Air 0 69 59 100 70 0 20 1 1 0 0 0 20 1 i=6
Train 1 40 20 345 57 0 20 1 0 1 0 0 0 0
Bus 0 35 13 417 58 0 20 1 0 0 1 0 0 0
Car 0 0 12 284 43 0 20 1 0 0 0 1 0 0
Air 1 45 148 115 160 1 45 1 1 0 0 0 45 1 i=7
Train 0 34 111 945 213 1 45 1 0 1 0 0 0 0
Bus 0 35 66 935 167 1 45 1 0 0 1 0 0 0
Car 0 0 36 821 125 1 45 1 0 0 0 1 0 0
Air 0 69 121 152 137 0 12 1 1 0 0 0 12 1 i=8
Train 0 34 52 889 149 0 12 1 0 1 0 0 0 0
Bus 0 35 50 879 146 0 12 1 0 0 1 0 0 0
Car 1 0 50 780 135 0 12 1 0 0 0 1 0 0
Air 0 69 59 100 70 0 40 1 1 0 0 0 40 1 i=9
Train 0 34 31 372 71 0 40 1 0 1 0 0 0 0
Bus 0 35 25 417 70 0 40 1 0 0 1 0 0 0
Car 1 0 17 210 40 0 40 1 0 0 0 1 0 0
Air 0 69 58 68 65 0 70 2 1 0 0 0 70 2 i=10
Train 0 34 31 357 69 0 70 2 0 1 0 0 0 0
Bus 0 35 25 402 68 0 70 2 0 0 1 0 0 0
Car 1 0 7 210 30 0 70 2 0 0 0 1 0 0
N18: Data Setup for NLOGIT N-339

N18.12 Merging Revealed Preference (RP) and Stated

Preference (SP) Data Sets
For applications in which you wish to merge RP and SP data sets, we assume that a data set
is built up for each individual in the sample from an RP observation and one or more SP
observations, for the same person. To construct the data for the simulation, you will require two
variables:

1. a numeric identification (id) that is the same for the RP and SP observations,
2. a treatment or choice set type index, coded 0 for the RP observation and 1,...,T (may vary by
person) for the SP data.

It is assumed that there is exactly one RP observation and any number up to T SP observations. The
type code need not obey any particular convention; you may code it any way you wish. What is
essential is that this type code equal zero for the RP observation and some positive value for the SP
observation(s). The SP observations may have the same or different values for this coding. From
this information NLOGIT can deduce the form of the choice set.

NOTE: This feature of the simulator cannot be used if the data are already arranged as
RP,SP1,RP,SP2,RP,SP3,RP… That is, the RP observation must not be repeated.

The ; Choices = list specification in the model command must include the full universal
choice list for both RP and SP. In most applications of this sort, the RP observations will use one
subset and the SP observations will use the remainder and there will be no overlap. For example, the
universal choice set might include a set of, say, five RP choices and 15 SP choices in which each RP
choice setting involves some smaller number, say four, of the latter. However, this partitioning is
not necessary. For example, you might have survey data in which variants on an existing choice set
are presented to individuals, for example, as in ‘would you choose option A,B,C... if price were
changed by ...?’. The additional specification for NLOGIT will be

; MergeSPRP (id = name of unique identifier,

type = the name of the treatment indicator variable)

where id is the unique identifying variable that links the SP and RP observations (or any
observations associated with the same id from two data sets).
The effect of the preceding specification is to expand each observation into T combined sets
of data, in the form shown above. (NLOGIT wants to do the expansion itself.) This does not
actually modify your data set. The observations are created temporarily during the computations.
N19: NLOGIT Commands and Results N-340

N19: NLOGIT Commands and Results

N19.1 Introduction
This and the next three chapters will describe the common features of the NLOGIT models
and commands used to fit them. Section N19.2 presents the generic command structure for
NLOGIT. The specification of models for NLOGIT follows the general pattern for model commands
in LIMDEP. Section N19.3 describes optional command specifications. The different models, such
as nested logit, mixed logit and multivariate probit, are requested by modifying the basic command.
Section N19.4 describes output features, such as estimation results and elasticities, that are common
to all the models. The subsequent chapters, N20-N22 will provide greater details on the model
specifications, including choice sets and utility functions in Chapter N20 and partial effects,
hypothesis tests and model simulations in Chapter N21.
NLOGIT is built around estimation of the parameters of the random utility model for discrete
choice,
U(choice j for individual i) = Uij = b ij′xij + eij, j = 1,...,Ji,

in which individual i makes choice j if Uij is the largest among the Ji utilities in the choice set. The
parameters in the model are the weights in the utility functions and the deeper parameters of the
distribution of the random terms. In some cases, the ‘taste’ parameters in the utility functions might
vary across individuals and in most cases, they will vary across choices. The latter is simple to
accommodate just by merging all parameters into one grand b and redefining x with some zeros in
the appropriate places. But, for the former case, we will be interested in a lower level
parameterization that involves what are sometimes labeled the ‘hyperparameters.’ Thus, it might be
the extreme case (as in the random parameters logit model) that b ij = f(zi, D, Γ, b, vi) where D, Γ, b
are lower level parameters, zi is observed data, and vi is a set of latent unobserved variables. The
parameters of the random terms will generally be few in number, usually consisting of a small
number of scaling parameters as in the heteroscedastic logit model, but they might be quite
numerous, again in the random parameters model. In all cases, the main function of the routines is
estimation of the structural parameters, then use of the estimated model for analysis of individual and
aggregate behavior.

N19.2 NLOGIT Commands

The essential command for the set of discrete choice models in NLOGIT is the same for all,
with the exception of the model name:

Model ; Lhs = variable which indicates the choice made

; Choices = a set of J names for the set of choices
(utility functions)
; Rhs = choice varying attributes in the utility functions
; Rh2 = choice invariant variables, including one for ASCs $
(or)
; Model: utility specifications… $
N19: NLOGIT Commands and Results N-341

The various models are as follows, where either of the two forms given may be used:

Model Command Alternative Command Form

Conditional Logit CLOGIT NLOGIT
Random Regret Logit RRLOGIT NLOGIT ; RRM
Scaled Multinomial Logit SMNLOGIT NLOGIT ; SMNL
Error Components Logit ECLOGIT NLOGIT ; ECM = ...
Heteroscedastic Extreme Value HLOGIT NLOGIT ; HET
Nested Logit NLOGIT NLOGIT ; Tree = ...
Generalized Nested Logit GNLOGIT NLOGIT ; GNL
Random Parameters Logit RPLOGIT NLOGIT ; RPL
Generalized Mixed Logit GMCLOGIT NLOGIT ; GMX
Nonlinear Random Parameters NLRPLOGIT NLOGIT ; NLRP =
Latent Class Logit LCLOGIT NLOGIT ; LCM
Latent Class Random Parameters LCRPLOGIT NLOGIT ; RPL ; LCM
Multinomial Probit MNPROBIT NLOGIT ; MNP

The description to follow in the rest of this chapter applies equally to all models. For convenience,
we will use the generic NLOGIT command in most of the discussion, while you can use the specific
model names in your estimation commands.
The command builders for these models can be found in Model:Discrete Choice. There
are several model options as shown in Figure N19.1

Figure N19.1 Command Builders for NLOGIT Models

N19: NLOGIT Commands and Results N-342

The Main and Options pages of the command builder for the conditional logit model are shown in
Figures N19.2, N19.3 and N19.4. (Some features of the models, and the ECM model, are not
provided by the command builders. Most of the features of these models are much easier to specify
in the editor using the command mode of entry.) The model and the choice set are set up on the
Main page. The Rhs variables (attributes) and Rh2 variables (characteristics) are defined on the
Options page. Note in the two windows on the Options page, the Rhs variables of the model are
defined in the left window and the Rh2 variables are specified in the right window.

Figure N19.2 Discrete Choice Command Builder Main Page

A set of exactly J choice labels must be provided in the command. These are used to label
the choices in the output. The number you provide is used to determine the number of choices there
are in the model. Therefore, the set of the right number of labels is essential. Use any descriptor of
eight or fewer characters desired – these do not have to be valid names, just a set of labels, separated
in the list by commas.

The internal limit on J, the number of choices, is 100.

There are K attributes (Rhs variables) measured for the choices. The sections below will
describe variations of this for different formulations and options. The total number of parameters in
the utility functions will include K1 for the Rhs variables and (J-1)K2 for the Rh2 variables. The total
number of utility function parameters is thus K = K1 + (J-1)K2.

The internal limit on K, the number of utility function parameters, is 100.

N19: NLOGIT Commands and Results N-343

Figure N19.3 Specifying Choices on Command Builder Main Page

Figure N19.4 Options Page of Command Builder for Conditional Logit Model

The random utility model specified by this setup is precisely of the form
Ui,j = b1xi,1 + b2xi,2 + ... + bK1xi,K1 + γ1,jzi,1 + ... + γK2,jzi,K2 + ei,j,
where the x variables are given by the Rhs list and the z variables are in the Rh2 list. By this
specification, the same attributes and the same characteristics appear in all equations, at the same
position. The parameters, bk appear in all equations, and so on. There are various ways to change
this specification of the utility functions – i.e., the Rhs of the equations that underlie the model, and
several different ways to specify the choice set. These will be discussed at several points below.
N19: NLOGIT Commands and Results N-344

N19.3 Other Optional Specifications on NLOGIT Commands

The NLOGIT command operates like other LIMDEP model commands. The following lists
command features and options that are used with this command. There are numerous additional
command specifications that are used with the specific models fit with NLOGIT, such as ; RPM to
specify a random parameters model, and ; Umax which is a technical specification if it is necessary
to control the accumulation of rounding error in estimating certain models.

Controlling Output from Model Commands

; Par saves person specific parameter vectors, used with the random parameters
logit model and heteroscedastic extreme value model.
; Effects: spec displays partial effects and elasticities of probabilities.
; Table = name adds model results to stored tables.

Robust Asymptotic Covariance Matrices

; Covariance Matrix displays estimated asymptotic covariance matrix (normally not shown).
; Cluster = spec computes robust cluster corrected asymptotic covariance matrix.
; Robust computes robust sandwich estimator for asymptotic covariance matrix.

Optimization Controls for Nonlinear Optimization

; Start = list provides starting values for a nonlinear model.

; Tlg [ = value]
sets the convergence value for convergence on the gradient.
; Tlf [ = value]
sets the convergence value for function convergence.
; Tlb[ = value]sets the convergence value for convergence on change in parameters.
; Alg = name specifies optimization method. Newton’s method is best. BFGS is
occasionally needed.
; Maxit = n sets the maximum iterations.
; Output = n requests technical output during iterations; the level ‘n’ is 1, 2, 3 or 4.
; Set keeps current setting of optimization parameters as permanent.
; Umax = value specifies maximum allowed value for utility in MNL probability.
(Generally not needed save for extreme data. Default = 300.)

Predictions and Residuals

; List lists predicted probabilities and predicted outcomes with model results.
; Keep = name keeps fitted values as a new (or replacement) variable in data set.
(Several other similar specifications are used with NLOGIT.)
; Prob = name keeps probabilities as a new (or replacement) variable.

Hypothesis Tests and Restrictions

; CML: spec defines a constrained maximum likelihood estimator.

; Test: spec defines a Wald test of linear restrictions.
; Wald: spec defines a Wald test of linear restrictions, same as ; Test: spec.
; Rst = list imposes fixed value and equality constraints.
; Maxit = 0 ; Start = the restricted values specifies Lagrange multiplier test.
N19: NLOGIT Commands and Results N-345

N19.4 Estimation Results

This section will detail the common results produced by the different models in NLOGIT.

N19.4.1 Descriptive Headers for NLOGIT Models

The output for the NLOGIT estimators may contain a description of the model before the
statistical results. The description consists of a table that shows the sample proportions and the tree
structure if you fit a nested logit model, and a table that lists the components of the utility functions.
You can request these listings by adding

; Show Model

to your NLOGIT command. (We used this device in several earlier examples.) Starting values for
the iterations are either zeros or the values you provide with ; Start = list. As such, there is no initial
listing of OLS results. Output begins with the final results for the model. Here is a sample: The
command is

NLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc
; Rh2 = one,hinc
; Show Model $

Sample proportions are marginal, not conditional.

Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .27619| 1.000|
|TRAIN .30000| 1.000|
|BUS .14286| 1.000|
|CAR .28095| 1.000|
+----------------+------+---
+---------------------------------------------------------------+
| Model Specification: Table entry is the attribute that |
| multiplies the indicated parameter. |
+--------+------+-----------------------------------------------+
| Choice |******| Parameter |
| |Row 1| INVC INVT GC A_AIR AIR_HIN1 |
| |Row 2| A_TRAIN TRA_HIN2 A_BUS BUS_HIN3 |
+--------+------+-----------------------------------------------+
|AIR | 1| INVC INVT GC Constant HINC |
| | 2| none none none none |
|TRAIN | 1| INVC INVT GC none none |
| | 2| Constant HINC none none |
|BUS | 1| INVC INVT GC none none |
| | 2| none none Constant HINC |
|CAR | 1| INVC INVT GC none none |
| | 2| none none none none |
+---------------------------------------------------------------+
N19: NLOGIT Commands and Results N-346

The initial header includes a display of the tree structure when you fit a nested logit model. For
example, the command

NLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc
; Rh2 = one,hinc
; Tree = Public[(air),(train,bus)],Private[(car)]
; Show Model $

produces the header:

Tree Structure Specified for the Nested Logit Model

Sample proportions are marginal, not conditional.
Choices marked with * are excluded for the IIA test.
----------------+----------------+----------------+----------------+------+---
Trunk (prop.)|Limb (prop.)|Branch (prop.)|Choice (prop.)|Weight|IIA
----------------+----------------+----------------+----------------+------+---
Trunk{1} 1.00000|PUBLIC .71905|B(1|1,1) .27619|AIR .27619| 1.000|
| |B(2|1,1) .44286|TRAIN .30000| 1.000|
| | |BUS .14286| 1.000|
|PRIVATE .28095|B(1|2,1) .28095|CAR .28095| 1.000|
----------------+----------------+----------------+----------------+------+---

(Note, this particular model is not identified – we specified it only for purpose of illustrating the
display of its tree structure.)

N19.4.2 Standard Model Results

Estimation results for the model commands consist of the initial display of diagnostic
followed by notes about the model, then the estimated coefficients. The preceding command,
without the tree structure or the initial echo of the model specification,

NLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc
; Rh2 = one,hinc $

produces the following results:

Normal exit from iterations. Exit status=0.

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -246.10979
Estimation based on N = 210, K = 9
Inf.Cr.AIC = 510.2 AIC/N = 2.430
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -283.7588 .1327 .1201
Chi-squared[ 6] = 75.29796
Prob [ chi squared > value ] = .00000
Response data are given as ind. choices
Number of obs.= 210, skipped 0 obs
-----------------------------------------------------------------------------
N19: NLOGIT Commands and Results N-347

--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.04613*** .01665 -2.77 .0056 -.07876 -.01349
INVT| -.00839*** .00214 -3.92 .0001 -.01258 -.00419
GC| .03633** .01478 2.46 .0139 .00737 .06530
A_AIR| -1.31602* .72323 -1.82 .0688 -2.73353 .10148
AIR_HIN1| .00649 .01079 .60 .5477 -.01467 .02765
A_TRAIN| 2.10710*** .43180 4.88 .0000 1.26079 2.95341
TRA_HIN2| -.05058*** .01207 -4.19 .0000 -.07424 -.02693
A_BUS| .86502* .50319 1.72 .0856 -.12120 1.85125
BUS_HIN3| -.03316** .01299 -2.55 .0107 -.05862 -.00770
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
-----------------------------------------------------------------------------

NOTE: (This is one of our frequently asked questions.) The ‘R-squareds’ shown in the output are
R2s in name only. They do not measure the fit of the model to the data. It has become common for
researchers to report these with results as a measure of the improvement that the model gives over
one that contains only a constant. But, users are cautioned not to interpret these measures as
suggesting how well the model predicts the outcome variable. It is essentially unrelated to this.

To underscore the point, we will examine in detail the computations in the diagnostic
measures shown in the box that precedes the coefficient estimates. Consider the example below,
which was produced by fitting a model with five coefficients subject to two restrictions, or three free
coefficients – npfree = 3. The effect is achieved by specifying

NLOGIT ; Lhs = mode ; Show

; Choices = air,(train),(bus),car
; Rhs = gc,ttme ; Rh2 = one $

+------------------------------------------------------+
|WARNING: Bad observations were found in the sample. |
|Found 93 bad observations among 210 individuals. |
|You can use ;CheckData to get a list of these points. |
+------------------------------------------------------+
Sample proportions are marginal, not conditional.
Sample proportions are marginal, not conditional.
Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .49573| 1.000|
|TRAIN .00000| 1.000|*
|BUS .00000| 1.000|*
|CAR .50427| 1.000|
+----------------+------+---
N19: NLOGIT Commands and Results N-348

+---------------------------------------------------------------+
| Model Specification: Table entry is the attribute that |
| multiplies the indicated parameter. |
+--------+------+-----------------------------------------------+
| Choice |******| Parameter |
| |Row 1| GC TTME A_AIR A_TRAIN A_BUS |
+--------+------+-----------------------------------------------+
+--------+------+-----------------------------------------------+
|AIR | 1| GC TTME Constant none none |
|TRAIN | 1| GC TTME none Constant none |
|BUS | 1| GC TTME none none Constant |
|CAR | 1| GC TTME none none none |
+---------------------------------------------------------------+
Normal exit: 6 iterations. Status=0, F= 62.58418

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -62.58418
Estimation based on N = 117, K = 3
Inf.Cr.AIC = 131.2 AIC/N = 1.121
R2=1-LogL/LogL* Log-L fncn R-sqrd R2Adj
Constants only -81.0939 .2283 .2079
Chi-squared[ 2] = 37.01953
Prob [ chi squared > value ] = .00000
Response data are given as ind. choices
Number of obs.= 210, skipped 93 obs
Restricted choice set. Excluded choices are
TRAIN BUS
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
GC| .01320* .00695 1.90 .0574 -.00042 .02682
TTME| -.07141*** .01605 -4.45 .0000 -.10286 -.03996
A_AIR| 3.96117*** .98004 4.04 .0001 2.04032 5.88201
A_TRAIN| 0.0 .....(Fixed Parameter).....
A_BUS| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------

There are 210 individuals in the data set, but this model was fit to a restricted choice set which
reduced the data set to n = 210 - 93 = 117 useable observations. The original choice set had Ji = 4
choices, but two were excluded, leaving Ji = 2 in the sample. The log likelihood is -62.58418. The
‘constants only’ log likelihood is obtained by setting each choice probability to the sample share for
each outcome in the choice set. For this application, those are 0.49573 for air and 0.50427 for car.
(This computation cannot be done if the choice set varies by person or if weights or frequencies are
used.)
Thus, the log likelihood for the restricted model is

Log L0 = 117 ( 0.49573 × log 0.49573 + 0.50427 × log 0.50427 ) = -81.09395.

N19: NLOGIT Commands and Results N-349

The ‘R2’ is 1 - (-62.54818/-81.0939) = 0.22829 (including some rounding error). The adjustment
factor is

K = (Σi Ji - n) / [(Σi Ji - n) - npfree] = (234 - 117)/(234 - 117 - 3) = 1.02632.

and the ‘Adjusted R2’ is 1 - K(log L /LogL0);

Adjusted R2 = 1 - 1.02632 (-62.54818/-81.0939) = 0.20794.

N19.4.3 Retained Results

Results kept by this estimator are:

Matrices: b and varb = coefficient vector and asymptotic covariance matrix

Scalars: logl = log likelihood function

nreg = N, the number of observational units
kreg = the number of Rhs variables
lastiter = number of iterations completed for estimation
opt_exit = exit code. 0.0 indicates successful estimation
modelaic = AIC for estimated model
modelbic = BIC for estimate model

Last Model: b_variable = the labels kept for the WALD command

In the Last Model, groups of coefficients for variables that are interacted with constants get
labels choice_variable, as in trai_gco. (Note that the names are truncated – up to four characters for
the choice and three for the attribute.) The alternative specific constants are a_choice, with names
truncated to no more than six characters. For example, the sum of the three estimated choice specific
constants could be analyzed as follows:

NLOGIT ; Lhs = mode ; Show

; Choices = air,train,bus,car
; Rhs = gc,ttme ; Rh2 = one $
WALD ; Fn1 = a_air + a_train + a_bus $

-----------------------------------------------------------------------------
WALD procedure. Estimates and standard errors
for nonlinear functions and joint test of
nonlinear restrictions.
Wald Statistic = 78.54713
Prob. from Chi-squared[ 1] = .00000
Functions are computed at means of variables
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
WaldFcns| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
Fncn(1)| 12.9101*** 1.45668 8.86 .0000 10.0550 15.7651
--------+--------------------------------------------------------------------
N19: NLOGIT Commands and Results N-350

N19.4.4 Descriptive Statistics for Alternatives

You may request a set of descriptive statistics for your model by adding

; Describe

to the model command. For each alternative, a table is given which lists the nonzero terms in the
utility function and the means and standard deviations for the variables that appear in the utility
function. Values are given for all observations and for the individuals that chose that alternative.
For the example shown above, the following tables would be produced:

NLOGIT ; Lhs = mode ; Choices = air,train,bus,car

; Rhs = invc,invt,gc ; Rh2 = one,hinc
; Show Model
; Describe $

+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative AIR |
| Utility Function | | 58.0 observs. |
| Coefficient | All 210.0 obs.|that chose AIR |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 85.252 27.409| 97.569 31.733 |
| INVT -.0084 INVT | 133.710 48.521| 124.828 50.288 |
| GC .0363 GC | 102.648 30.575| 113.552 33.198 |
| A_AIR -1.3160 ONE | 1.000 .000| 1.000 .000 |
| AIR_HIN1 .0065 HINC | 34.548 19.711| 41.724 19.115 |
+-------------------------------------------------------------------------+
+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative TRAIN |
| Utility Function | | 63.0 observs. |
| Coefficient | All 210.0 obs.|that chose TRAIN |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 51.338 27.032| 37.460 20.676 |
| INVT -.0084 INVT | 608.286 251.797| 532.667 249.360 |
| GC .0363 GC | 130.200 58.235| 106.619 49.601 |
| A_TRAIN 2.1071 ONE | 1.000 .000| 1.000 .000 |
| TRA_HIN2 -.0506 HINC | 34.548 19.711| 23.063 17.287 |
+-------------------------------------------------------------------------+
+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative BUS |
| Utility Function | | 30.0 observs. |
| Coefficient | All 210.0 obs.|that chose BUS |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 33.457 12.591| 33.733 11.023 |
| INVT -.0084 INVT | 629.462 235.408| 618.833 273.610 |
| GC .0363 GC | 115.257 44.934| 108.133 43.244 |
| A_BUS .8650 ONE | 1.000 .000| 1.000 .000 |
| BUS_HIN3 -.0332 HINC | 34.548 19.711| 29.700 16.851 |
+-------------------------------------------------------------------------+
N19: NLOGIT Commands and Results N-351

+-------------------------------------------------------------------------+
| Descriptive Statistics for Alternative CAR |
| Utility Function | | 59.0 observs. |
| Coefficient | All 210.0 obs.|that chose CAR |
| Name Value Variable | Mean Std. Dev.|Mean Std. Dev. |
| ------------------- -------- | -------------------+------------------- |
| INVC -.0461 INVC | 20.995 14.678| 15.644 9.629 |
| INVT -.0084 INVT | 573.205 274.855| 527.373 301.131 |
| GC .0363 GC | 95.414 46.827| 89.085 49.833 |
+-------------------------------------------------------------------------+

You may also request a cross tabulation of the model predictions against the actual choices.
(The predictions are obtained as the integer part of Σt P̂ jt yjt.) Add

; Crosstab

to your model command. For the same model, this would produce the two sets of results below.
Note the first cross tabulation is based on the fitted probabilities while the second is based on the
observed choices.

+-------------------------------------------------------+
| Cross tabulation of actual choice vs. predicted P(j) |
| Row indicator is actual, column is predicted. |
| Predicted total is F(k,j,i)=Sum(i=1,...,N) P(k,j,i). |
| Column totals may be subject to rounding error. |
+-------------------------------------------------------+
--------+----------------------------------------------------------------------
NLOGIT Cross Tabulation for 4 outcome Multinomial Choice Model
CrossTab| AIR TRAIN BUS CAR Total
--------+----------------------------------------------------------------------
AIR| 19 13 8 18 58
TRAIN| 12 30 9 12 63
BUS| 10 8 6 6 30
CAR| 17 12 7 23 59
--------+----------------------------------------------------------------------
Total| 58 63 30 59 210

+-------------------------------------------------------+
| Cross tabulation of actual y(ij) vs. predicted y(ij) |
| Row indicator is actual, column is predicted. |
| Predicted total is N(k,j,i)=Sum(i=1,...,N) Y(k,j,i). |
| Predicted y(ij)=1 is the j with largest probability. |
+-------------------------------------------------------+
--------+----------------------------------------------------------------------
NLOGIT Cross Tabulation for 4 outcome Multinomial Choice Model
CrossTab| AIR TRAIN BUS CAR Total
--------+----------------------------------------------------------------------
AIR| 23 15 0 20 58
TRAIN| 8 49 0 6 63
BUS| 13 12 1 4 30
CAR| 15 13 0 31 59
--------+----------------------------------------------------------------------
Total| 59 89 1 61 210
N19: NLOGIT Commands and Results N-352

N19.5 Calibrating a Model

When the data consists of two subsets, for example an RP data set and a counterpart SP data
set, it is sometimes useful to fit the model with one of the data sets, then refit the second one while
retaining the original coefficients, and just adjusting the constants. Consider the application below:

SAMPLE ; 1-420 $
CLOGIT ; Lhs = mode ; Choices = air,train,bus,car
; Model: U(air) = aa + gc * gc + ttme * ttme + invt * invt /
U(train) = at + gc * gc + ttme * ttme /
U(bus) = ab + gc * gc + ttme * ttme /
U(car) = + gc * gc + ttme * ttme $
SAMPLE ; 421-840 $
CLOGIT ; Lhs = mode ; Choices = air,train,bus,car
; Model: U(air) = aa + gc[ ] * gc + ttme[ ] * ttme + invt[ ] * invt /
U(train) = at + gc * gc + ttme * ttme /
U(bus) = ab + gc * gc + ttme * ttme /
U(car) = + gc * gc + ttme * ttme $

The model is first fit with the first half of the data set (observations 1 - 105). Then, for the second
estimation, we want to refit the model, but only recompute the constant terms but keep the
previously estimates slope parameters. The device to use for the second model is the ‘[ ]’
specification, which indicates that you wish to use the previously estimated parameters. The
commands above will, in principle, produce the desired result, with one consideration. Newton’s
method is very sensitive to the starting values for this model, and with the constraints imposed in the
second model, will generally fail to converge. (See the example below.) The practical solution is to
change the algorithm to BFGS, which will then produce the desired result. You can do this just by
adding
; Alg = BFGS

to the second command. An additional detail is that the second model will now replace the first as
the ‘previous’ model. So, if you want to do a second calibration, you have to refit the first model.
To preempt this, you can use

; Calibrate

in the second command. This specification changes the algorithm and also instructs NLOGIT not to
replace the previous estimates with the current ones. Three notes about this procedure:

• You may use this device with any discrete choice model that you fit with NLOGIT.
• The second sample must have the same configuration as the first.
• The device can only be used to fix the utility function parameters.

The third point implies that if you do this with a random parameters model, the random parameters
will become fixed – have the variances fixed at zero.
N19: NLOGIT Commands and Results N-353

The commands above (with the addition of ; Calibrate to the second CLOGIT command)
produce the following results: (Some parts of the results are omitted.) The note before the second
set of results has been produced because the estimator converges very quickly – this will usually
happen when the model contains only the alternative specific constants.

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -93.51621
Estimation based on N = 105, K = 6
Inf.Cr.AIC = 199.0 AIC/N = 1.896
Number of obs.= 105, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AA| 7.94929*** 1.44243 5.51 .0000 5.12217 10.77641
GC| -.01705*** .00626 -2.72 .0064 -.02931 -.00478
TTME| -.08983*** .01452 -6.19 .0000 -.11829 -.06136
INVT| -.01974** .00775 -2.55 .0109 -.03494 -.00455
AT| 4.31669*** .64859 6.66 .0000 3.04549 5.58790
AB| 2.60715*** .72991 3.57 .0004 1.17656 4.03774
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AA|-.22520D+33*** 1.00000 ******** .0000 -.22520D+33 -.22520D+33
GC| -.01705 .....(Fixed Parameter).....
TTME| -.08983 .....(Fixed Parameter).....
INVT| -.01974 .....(Fixed Parameter).....
AT| .24951D+34 .....(Fixed Parameter).....
AB| .68897D+33 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -97.65109
Number of obs.= 105, skipped 0 obs
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
AA| 8.06593*** .29707 27.15 .0000 7.48368 8.64817
GC| -.01705 .....(Fixed Parameter).....
TTME| -.08983 .....(Fixed Parameter).....
INVT| -.01974 .....(Fixed Parameter).....
AT| 2.94882*** .34838 8.46 .0000 2.26600 3.63164
AB| 3.09656*** .31503 9.83 .0000 2.47910 3.71402
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.
-----------------------------------------------------------------------------
N20: Choice Sets and Utility Functions N-354

N20: Choice Sets and Utility Functions

N20.1 Introduction
Chapter N17 described how to fit the generic form of the multinomial logit model for
multinomial choice. This chapter presents some modifications of the basic command that
accommodate more general choice sets (possibly varying across individuals) and a convenient
alternative command format that allows more general specifications of the utility functions.

N20.2 Choice Sets

Every multinomial model fit by NLOGIT must include a specification for the choice variable
and a definition of the choice set. The basic formulation would appear as

; Lhs = the dependent, or choice variable

; Choices = the names of the choices in the model

Several variations on this formula appear in Sections N20.3 and N20.4. In general, your dependent
variable is the name of a variable which indicates by a one or zero whether a particular alternative is
selected, or it gives the proportion or frequency of individuals sampled that selected a particular
alternative. When they are enumerated, the ; Choices list gives names and possibly sampling
weights for the set of alternatives.
All command builders begin with these two specifications. The discrete choice and nested
logit models allow the full set of variants discussed in this section while the other command builders
expect the simple form with a fixed choice set. The Main page of the conditional logit command
builder shown in Figure N20.1 illustrates. (A similar Main page is used for the nested logit
command builder.) The command builder allows you to specify the choice variable and type of
choice set in the three sections of this dialog box.

NOTE: The command builder for the multinomial probit, HEV and RPL models requires you to
provide a fixed sized choice set. This is a limitation of the command builder window, not the
estimator. With the exception of the multinomial probit model, this is not a requirement of the
models themselves. Only the multinomial probit model requires the number of choices to be fixed.
For the HEV and RPL models, if you build your command in the text editor, rather than with the
command builder, you may specify a variable choice set, as described in Section N20.2.1.
N20: Choice Sets and Utility Functions N-355

Figure N20.1 Main Page of Command Builder for Conditional Logit Model

In the standard case, data on the Lhs variable will consist of a column of J-1 zeros and a one
for the choice made, when reading down the J rows of data for the individual. We allow other types
of data on the choice variable. If you have grouped data, the values will be proportions or
frequencies, instead. For proportions data, within each observation (J data points), the values of the
Lhs variable will sum to one when summed down the J rows. (This will be the only difference in the
grouped data treatment.) With frequencies, the values will simply be a set of nonnegative integers.
An example of a setting in which such data might arise would be in marketing, where the proportions
might be market shares of several brands of a commodity. Alternatively, the choice variable might
be a set of ranks, in which case, instead of zeros and ones, the Lhs variable would take values
1,2,...,J (not necessarily in that order) within, and reading down, each block.
The following modifications apply to all multinomial models that are fit with NLOGIT. We
use NLOGIT as the generic verb for this description. Any of the others described in the next
chapter will be treated the same. Note, as well, the NLOGIT commands, which do not contain any
additional model specifications, will be equivalent to and act like CLOGIT commands. That is, the
command, NLOGIT, with no additional model specifications is equivalent to CLOGIT. (It is also
the same as DISCRETE CHOICE, which although no longer used by NLOGIT, remains acceptable
as the basic model verb.)
N20: Choice Sets and Utility Functions N-356

N20.2.1 Fixed and Variable Numbers of Choices

When every individual in the sample chooses from the same choice set, and all alternatives
are available to all individuals, then the data set will appear as in the example developed in Chapter
N17, and will consist of n sets of J ‘observations.’ You indicate this case with a command such as:

NLOGIT ; Lhs = the choice variable

; Choices = ... a list of J names for the choices
; ... the rest of the command $
For example,

NLOGIT ; Lhs = mode

; Choices = air,train,bus,car
; etc. $

A fixed choice set can be specified in the command builder as shown in Figure N20.2.

Figure N20.2 Fixed Choice Set Specified in Command Builder

There are many cases in which the choice set will vary from one individual to another. We consider
the random choice model first in which the number of choices is not constant from one observation
to the next. Ranks data are considered later.
N20: Choice Sets and Utility Functions N-357

Two possible arrangements that might produce variable sized choice sets are as follows:

• There is a universal choice set, from which individuals make their choice. But, not all
choices are available to all individuals. Consider, for example, the choice of travel mode
among train, bus, car, ferry. If respondents are observed at many different locations, one or
more of the choices, such as ferry or train, might be unavailable to them, and those might
vary from person to person. In this case, there is a fixed set of J alternatives, but each
individual chooses among their own Ji choices. This is called a ‘labeled’ choice set.

• Individuals each choose among their own set of Ji alternatives. However, there is no
universal choice set. Consider, for example, the choice of which shopping center to shop at.
If observations are taken in many different cities, we will observe numerous different choice
sets, but there is no well defined universal choice set. This is called an ‘unlabeled’ choice
set.

Unlabeled choice sets often arise in survey data, or ‘stated choice experiments.’ In a stated choice
experiment, an individual might be offered a set of Ji alternatives that are only differentiated by their
attributes. Configurations of features in a choice set of cars or appliances might be such a case. In
this instance, the choices are simply numbered, 1,2,…
Any of these cases can be accommodated with NLOGIT. For both cases, you will provide a
variable which gives the number of choices for each observation. This variable is then a second
; Lhs specification. The command for an unlabeled choice set, which is the simpler case, becomes

NLOGIT ; Lhs = y,nij

; ... specification of the utility functions
; ... the rest of the command $

Note that the ; Choices = list is not defined in the command, since in this case, there is no clearly
defined choice set. Nothing else need be changed. NLOGIT does all of the accounting internally. In
this case, it is simply assumed that each individual has their own choice set.
For example, one such data set might appear as follows.

y q w nij
i=1 0 q1,1 w1,1 3
>1 q2,1 w2,1 3
0 q3,1 w3,1 3

i=2 0 q1,2 w1,2 4
0 q2,2 w2,2 4
>1 q3,2 w3,2 4
0 q4,2 w4,2 4

i=3 >1 q1,3 w1,3 2
0 q2,3 w2,3 2
N20: Choice Sets and Utility Functions N-358

Note that nij is the usual group size variable for a panel in NLOGIT. The model command might be

NLOGIT ; Lhs = y,nij ; Rhs = q,w $

Notice, once again, that the command does not contain a definition of the choice set, such as
; Choices = list specification.

For the case of a universal choice set, suppose that the data set above were, instead:

Y q w nij altij
i=1 0 q1,1 w1,1 3 1 (Air)
>1 q2,1 w2,1 3 2 (Train)
0 q3,1 w3,1 3 4 (Car)

i=2 0 q1,2 w1,2 4 1 (Air)
0 q2,2 w2,2 4 2 (Train)
>1 q3,2 w3,2 4 3 (Bus)
0 q4,2 w4,2 4 4 (Car)

i=3 >1 q1,3 w1,3 2 3 (Bus)
0 q2,3 w2,3 2 4 (Car)

The specific choice identifier, when it is needed, is provided as a third Lhs variable. For this case,
the choice set would have to be defined. For example,

NLOGIT ; Lhs = y,nij,altij

; Choices = air,train,bus,car
; Rhs = q,w $

In this case, every individual is assumed to choose from a set of four alternatives, though the altij
variable indicates that some of these choices are unavailable to some individuals.
Do note that if you are not defining a universal choice set, NLOGIT simply uses the largest
number of choices for any individual in the sample to determine J for the model. As such, an
expanded set of choice specific constants is not likely to be meaningful, though you can create one
with ; Rh2 = one. Also, if you do not specify a universal choice set, the variable altij will not be
meaningful.

N20.2.2 Restricting the Choice Set

The IIA test described later in Section N21.4.1 is carried out by fitting the model to a
restricted choice set, then comparing the two sets of parameter estimates. You can restrict the choice
set used in estimation, irrespective of the IIA test, by a slight change in the command. In the
; Choices = list of alternatives specification, enclose any choices to be excluded in parentheses. For
example, in our CLOGIT application, the specification

; Choices = air,(train),(bus),car

produces the following display in the model output:

N20: Choice Sets and Utility Functions N-359

+------------------------------------------------------+
|WARNING: Bad observations were found in the sample. |
|Found 93 bad observations among 210 individuals. |
|You can use ;CheckData to get a list of these points. |
+------------------------------------------------------+
Sample proportions are marginal, not conditional.
Choices marked with * are excluded for the IIA test.
+----------------+------+---
|Choice (prop.)|Weight|IIA
+----------------+------+---
|AIR .49573| 1.000|
|TRAIN .00000| 1.000|*
|BUS .00000| 1.000|*
|CAR .50427| 1.000|
+----------------+------+---
Normal exit: 6 iterations. Status=0, F= 52.79148
-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable Choice
Log likelihood function -52.79148
Estimation based on N = 117, K = 5
Number of obs.= 210, skipped 93 obs
Restricted choice set. Excluded choices are
TRAIN BUS
--------+--------------------------------------------------------------------
| Standard Prob. 95% Confidence
MODE| Coefficient Error z |z|>Z* Interval
--------+--------------------------------------------------------------------
INVC| -.04871* .02757 -1.77 .0772 -.10274 .00532
INVT| -.01195*** .00395 -3.03 .0025 -.01969 -.00422
GC| .08576*** .02654 3.23 .0012 .03374 .13778
TTME| -.08222*** .01854 -4.43 .0000 -.11855 -.04588
A_AIR| 2.12899* 1.20531 1.77 .0773 -.23337 4.49135
A_TRAIN| 0.0 .....(Fixed Parameter).....
A_BUS| 0.0 .....(Fixed Parameter).....
--------+--------------------------------------------------------------------
Note: ***, **, * ==> Significance at 1%, 5%, 10% level.
Fixed parameter ... is constrained to equal the value or
had a nonpositive st.error because of an earlier problem.

Note that as in the IIA test, this procedure results in exclusion of some ‘bad’ observations, that is, the
ones that selected the excluded choices. Because of the model specification, the ASCs for train and
bus have been fixed at zero.
You may combine the choice based sampling estimator with the restricted choice set. All
the necessary adjustments of the weights are made internally. Thus, the specification

; Choices = air,(train),(bus),car / .14,.13,.09,.64

produces the following listing:
+----------------+------+---+
|Choice (prop.)|Weight|IIA|
+----------------+------+---+
|AIR .49573| .387| |
|TRAIN .00000| .000| * |
|BUS .00000| .000| * |
|CAR .50427| 1.739| |
+----------------+------+---+
N20: Choice Sets and Utility Functions N-360

N20.2.3 A Shorthand for Choice Sets

You may use
; Choices = number_name
To define a set of choice labels of the form name1, name2, … For example,

; Choices = 5_brand
Creates choice labels brand1, brand2, brand3, brand4, brand5. This sort of construction is likely to
be useful for unlabeled choice experiments.

N20.2.4 Large Choice Sets – A Panel Data Equivalence

The conditional logit estimator can fit a model with up to 500 choices, which is quite large.
Chamberlain’s fixed effects model for the binary logit model described in Section N9.5 can also be
used to fit a discrete choice model. The log likelihood function for this model is
Ti

∏t =1
exp[ yit β′xit ]
Lc =