MULTIPLE IMPUTATION
IN MPLUS
EMPLOYEE DATA
• Data
set containing scores from 480 employees on eight
work-related variables
• Variables:
• Age, gender, job tenure, IQ, psychological well-being, job
satisfaction, job performance, and turnover intentions
• 33% of the cases have missing well-being scores, and 33% have
missing satisfaction scores
ANALYSIS EXAMPLE
• Multiple regression model that predicts job performance from
psychological well-being and job satisfaction
jobperf = B0 + B1(wbeing) + B2(jobsat) + ε
B1
Well-Being Performance ε
B2
Satisfaction
MPLUS COMMANDS FOR
DIAGNOSTIC ANALYSIS
• TITLE (optional)
• DATA (same as ML analysis)
• VARIABLE (same as ML analysis)
• ANALYSIS
• MODEL
• OUTPUT
• PLOT
DATA COMMAND
• Full file path (data in different directory than input file)
DATA:
file = ‘c:\Data\employee.dat’;
• Omitted file path (data in same directory as input file)
DATA:
file = employee.dat;
VARIABLE COMMAND
• The VARIABLE command serves the same function as in an ML analysis
• The USEVARIABLES list should include all variables that will be part of
the imputation process
VARIABLE:
! Information about the contents of the data file;
names = id age tenure female wbeing jobsat jobperf turnover iq;
usevariables = age tenure wbeing jobsat jobperf iq;
missing = all (-99);
ANALYSIS COMMAND
• The following commands apply to the preliminary MCMC
analysis that generates graphical diagnostics
ANALYSIS:
! Bayesian estimation;
estimator = bayes;
! Random number seed for MCMC algorithm;
bseed = 48932;
! Convergence criterion (.05 = PSR < 1.05);
bconvergence = .05;
BCONVERGENCE
SUBCOMMAND
• Mplus generates imputed data sets only after the MCMC
algorithm has converged (this feature prevents imputations
from pathological chains)
• Setting
BCONVERGENCE = .05 (the default) specifies that
convergence is achieved once the PSR drops below 1.05
• Smaller
values give a more conservative picture of
convergence but require more computational time
MODEL COMMAND
• In the diagnostic run, MODEL should specify an analysis that estimates
all means, variances, and covariances (i.e., a saturated model)
MODEL:
[jobperf tenure wbeing jobsat turnover iq]; ! Means;
jobperf tenure wbeing jobsat turnover iq; ! Variances;
jobperf tenure wbeing jobsat turnover iq with
jobperf tenure wbeing jobsat turnover iq; ! Covariances;
OUTPUT COMMAND
• The TECH8 option prints the PSR statistic (updated after
every 100th iteration) to the output file
OUTPUT:
! Tech8 gives the PSR statistic;
tech8;
PLOT COMMAND
• The PLOT command generates graphical diagnostics
• This
command is only operational with the ESTIMATOR =
BAYES option (typically used only for the diagnostic run)
PLOT:
! Graphical convergence diagnostics;
type = plot2;
MPLUS DIAGNOSTIC PROGRAM
DATA:
file = employee.dat;
VARIABLE:
names = id age tenure female wbeing jobsat jobperf turnover iq;
usevariables = age tenure wbeing jobsat jobperf iq;
missing = all (-99);
ANALYSIS:
estimator = bayes;
bseed = 48932;
bconvergence = .05;
MODEL:
[age tenure wbeing jobsat jobperf iq];
age tenure wbeing jobsat jobperf iq;
age tenure wbeing jobsat jobperf iq with
age tenure wbeing jobsat jobperf iq;
OUTPUT:
tech8;
PLOT:
type = plot2;
PSR OUTPUT
(TECH8 OPTION)
• The PSR dropped below 1.05 (i.e., the MCMC algorithm
converged) by the 300th iteration
• A between-imputation interval of 200-300 is appropriate
TECHNICAL 8 OUTPUT
TECHNICAL 8 OUTPUT FOR BAYES ESTIMATION
POTENTIAL PARAMETER WITH
ITERATION SCALE REDUCTION HIGHEST PSR
100 1.134 13
200 1.058 15
POSTERIOR DISTRIBUTION OF
THE JOB SATISFACTION MEAN
TRACE PLOT OF THE JOB
SATISFACTION MEAN
TRACE PLOT OF THE JOB
SATISFACTION VARIANCE
TRACE PLOT OF THE
WELL-BEING MEAN
TRACE PLOT OF THE
WELL-BEING VARIANCE
TRACE PLOT OF THE SATISFACTION -
WELL-BEING COVARIANCE
MPLUS COMMANDS FOR
IMPUTATION ANALYSIS
• TITLE (optional)
• DATA (same as ML analysis)
• VARIABLE (same as ML analysis)
• ANALYSIS
• DATA IMPUTATION
• OUTPUT
ANALYSIS COMMAND
• Thefollowing commands apply to the final MCMC run that
generates the imputed data sets
ANALYSIS:
! Saturated imputation model;
type = basic;
! Random number seed for MCMC algorithm;
bseed = 48932;
! Convergence criterion (.05 = PSR < 1.05);
bconvergence = .05;
DATA IMPUTATION
COMMAND
• Thefollowing commands apply to the final MCMC run that
generates the imputed data sets
DATA IMPUTATION:
! Incomplete variables to be imputed;
! The USEVARIABLES list specifies other variables in the model;
impute = wbeing jobsat;
! Number of imputed data sets;
ndatasets = 50;
! File name prefix for imputed data sets;
save = employeeimp*.dat;
! Between-imputation interval;
thin = 300;
MPLUS IMPUTATION PROGRAM
DATA:
file = employee.dat;
VARIABLE:
names = id age tenure female wbeing jobsat jobperf turnover iq;
usevariables = age tenure wbeing jobsat jobperf iq;
missing = all (-99);
ANALYSIS:
type = basic;
bseed = 48932;
bconvergence = .05;
DATA IMPUTATION:
impute = wbeing jobsat;
ndatasets = 50;
save = employeeimp*.dat;
thin = 300;
OUTPUT:
tech8;
IMPUTATION OUTPUT FILE
• Mplus lists the variable order in the imputed data near the bottom of the
output file
• Use this variable list for all subsequent analyses
SAVEDATA INFORMATION
Order of variables
AGE
TENURE
WBEING
JOBSAT
JOBPERF
IQ
IMPUTED DATA FORMAT
• Mplus saves each imputed data set to a separate file
• The file names use the prefix specified in the SAVE command
(e.g., employeeimp1.dat, employeeimp2.dat, etc.)
• The
program also generates a list file that contains the file
names of the imputed data sets (e.g., employeeimplist.dat)
• The list file serves as input data for all subsequent analyses
THE IMPLIST FILE
• The imputation program
generates a text file (e.g.,
employeimplist.dat) that
contains the imputed data
set names
• The implist file serves as
input data for all subsequent
analyses
ANALYZING IMPUTED DATA
• Mplus fully automates the analysis and pooling phases
• Analyzing
imputed data sets requires a small change to the
DATA command, but the remaining commands are identical to
a complete-data analysis
• Theanalyses simplify a bit (e.g., no need to list incomplete
predictors, no need to use the auxiliary command)
DATA COMMAND
• Omitted file path (implist file in the same directory as input file)
• TYPE = imputation specifies multiply imputed data sets and
invokes pooling rules
DATA:
file = employeeimplist.dat; ! List of imputation file names;
type = imputation; ! Imputation data;
MPLUS ANALYSIS PROGRAM
DATA:
file = employeeimplist.dat;
type = imputation;
VARIABLE:
names = age tenure wbeing jobsat jobperf iq;
usevariables = wbeing jobsat jobperf;
ANALYSIS:
estimator = ml;
MODEL:
jobperf on wbeing (b1);
jobperf on jobsat (b2);
MODEL TEST:
b1 = 0;
b2 = 0;
OUTPUT:
standardized;
DESCRIPTIVES
SAMPLE STATISTICS
NOTE: These are average results over 50 data sets.
SAMPLE STATISTICS
Means
JOBPERF WBEING JOBSAT
________ ________ ________
1 6.021 6.296 5.949
Covariances
JOBPERF WBEING JOBSAT
________ ________ ________
JOBPERF 1.570
WBEING 0.676 1.380
JOBSAT 0.263 0.448 1.404
DESCRIPTIVES, CONTINUED
Correlations
JOBPERF WBEING JOBSAT
________ ________ ________
JOBPERF 1.000
WBEING 0.459 1.000
JOBSAT 0.177 0.322 1.000
WALD TEST
(MODEL TEST COMMAND)
• The Wald statistic (a chi-square with 2 degrees of freedom) is akin
to the omnibus F test in OLS regression
Wald Test of Parameter Constraints
Value 136.976
Degrees of Freedom 2
P-Value 0.0000
• The significant chi-square, χ2(2)= 136.976, indicates that the set of
predictors explain significant variation in the dependent variable
UNSTANDARDIZED
ESTIMATES
MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value
JOBPERF ON
WBEING 0.479 0.055 8.783 0.000
JOBSAT 0.035 0.056 0.622 0.534
Intercepts
JOBPERF 2.797 0.402 6.959 0.000
Residual Variances
JOBPERF 1.236 0.088 14.052 0.000
INTERPRETATIONS
• Interpret
and report MI estimates in the same way as a
complete-data analysis
• Controlling
for job satisfaction, a one-point increase in
psychological well-being results in a .479 increase in job
performance, on average
• Controlling for psychological well-being, a one-point increase
job satisfaction in results in a .035 increase in job
performance, on average
STANDARDIZED ESTIMATES
(STANDARDIZED OPTION)
STANDARDIZED MODEL RESULTS
STDYX Standardization
Two-Tailed
Estimate S.E. Est./S.E. P-Value
JOBPERF ON
WBEING 0.449 0.048 9.400 0.000
JOBSAT 0.033 0.053 0.622 0.534
STANDARDIZED ESTIMATES,
CONTINUED
R-SQUARE
Observed Two-Tailed
Variable Estimate S.E. Est./S.E. P-Value
JOBPERF 0.213 0.041 5.250 0.000
INTERPRETATIONS
• The STDYX standardization gives beta weights
• Controlling for job satisfaction, a one standard deviation increase in
psychological well-being results in a .449 standard deviation increase
in job performance, on average
• Controlling for psychological well-being, a one standard deviation
increase job satisfaction in results in a .033 standard deviation
increase in job performance, on average
• Together, the two predictors explain 21.3% of the variance in job
performance ratings
