Part IV
Limited Dependent Variables
In many empirical applications, the dependent variable is not continuously observed but
is limited in its range. In contrast to variables such as age or salary (which are bounded by
definition), the observable outcomes in these models are restricted by their observability,
that is, the lack of information. Examples of limited response variables include:
• Binary response: decisions to work or purchase durable goods.
• Discrete response: survey opinions, ratings, individual choices.
• Count data: numbers of infrequent events.
• Truncated data: female labor supply.
• Censored data: duration, income, hours worked, expenditures on durables.
• Sample selection: nonrandom samples resulting from endogenous decisions.
In the following section, we focus particularly on the censored regression. The aim is
not only to demonstrate the estimation by the maximum likelihood, but to include other
aspects such as interpretation of the estimation results.
6 Models for Censored and Truncated Data
In many empirical applications, the dependent variable is not fully continuous, but it is
observed only over part of its range. In censored regression models the latent continuous
variable is observed only if it exceeds or falls below a certain threshold; otherwise, only
a bound is recorded. For instance, personal income is recorded by the tax office only if
it exceed the minimum taxable income and the lower levels of income are not necessarily
observed in the data. Similarly, in corner solution models, the outcome is observed as
zero for many observations and is continuously distributed for the rest. For example, an
individual’s expenditures on alcohol can be positive or zero, but irrespective of the dislike
for alcohol, they cannot be negative. In both cases, the standard linear regression model
is not appropriate, and the same specialized model will be employed.
6.1 Introduction to Censored Models
Similarly to the binary-choice case, we assume there is a latent variable y ∗, which cannot
be fully observed; for example an individual’s actual income or an individual’s utility of
alcohol consumption. The latent variable is assumed to follow a linear regression model,
yi∗ = x⊤ 2
i β + εi , typically under the assumption of normal errors, εi |xi ∼ N(0, σ ). When
the data are censored, the observed response yi is assumed to depend on the latent variable
yi∗ according to
yi = max{yi∗, a}, or yi = min{yi∗, a}.
30
Without loss of generality, we focus on a common case of censoring from below at zero:
yi = max{yi∗, 0} = max{x⊤
i β + εi , 0} (11)
since one can easily transform the dependent variable otherwise (e.g., by considering the
latent variable yi∗ − a, which is censored at zero if the original response is censored at
a). Hence, while yi∗ is typically assumed to be continuously distributed, the observed yi
has a point mass at zero as all values below zero are recorded as 0 and is continuously
distributed above zero.
One might be tempted to fit a linear regression model,
E(yi |xi ) = x⊤
i β,
to censored data. However, this approach is generally unsatisfactory because the con-
ditional expectation E(yi |xi ) under censoring is no longer linear in xi . In particular,
under censoring (11) of the latent model yi∗ = x⊤ 2
i β + εi , εi |xi ∼ N(0, σ ), one has
E(yi |xi ) > E(yi∗ |xi ) = x⊤
i β and the difference between the two expectations corresponds
to the amount of censoring and is generally different at different points xi . Moreover, the
predicted values of the linear regression may fall outside the allowable range (i.e., result-
ing in negative predicted values when only nonnegative outcomes are observed). Thus, a
different model that explicitly accounts for censoring is required.
6.2 The Tobit Model
A widely used model for censored data is the Tobit model (sometimes called Tobit Type
I). It is defined by a latent variable equation
yi∗ = x⊤ ∗
i β + εi , ε∗i | xi ∼ N(0, σ 2 ),
where the observations (yi∗ , xi ) are assumed to be independent, E(xi x⊤
i ) is nonsingular to
ensure identification, and the observation rule is
yi = max{0, yi∗ }.
In other words, if the latent outcome yi∗ is negative, we observe yi = 0; if yi∗ is positive,
then yi = yi∗ . Note that in this formulation, the error term is assumed to be normally
distributed so that, for any given xi , there is a positive probability of censoring:
x⊤ β x⊤ β
∗ ∗ ⊤ i
P (yi = 0|xi ) = P (yi ≤ 0|xi ) = P εi ≤ −xi β = Φ − =1−Φ i .
σ σ
6.3 Likelihood Estimation of the Tobit Model
Because the observed outcome is a mixture of a point mass at zero and a continuous
distribution above zero, the likelihood function must combine probability masses and
densities. For an individual observation, we have:
• If yi = 0 (an observation is censored), then the likelihood contribution is the condi-
tional probability of observing the outcome yi = 0 (derived in the previous section):
x⊤ β
Li (β, σ) = P (yi = 0|xi ) = P (yi∗ ≤ 0|xi ) = 1 − Φ i .
σ
31
• If yi > 0 (an observation is not censored), then the likelihood contribution is the
conditional density of the outcome yi = yi∗ :
1 yi − x⊤
i β
Li (β, σ) = fy (yi |xi ) = fεi (yi − x⊤
i β|xi ) = φ ,
σ σ
where φ(·) denotes the standard normal density and φ(t/σ)/σ is thus the density
function of εi ∼ N(0, σ 2 ).
Thus, the complete log-likelihood contribution for the observation i can be written as
h x⊤ β i
i 1 yi − x⊤ i β
Li (β, σ) = I(yi = 0) ln 1 − Φ + I(yi > 0) ln φ ,
σ σ σ
and the log-likelihood function for the sample of size n is
( )
Xn h x⊤ β i y − x⊤ β
i i i
ln Ln (β, σ) = I(yi = 0) ln 1 − Φ + I(yi > 0) − ln σ + ln φ .
i=1
σ σ
The maximum likelihood estimation proceeds by finding the values of β and σ that max-
imize ln Ln (β, σ).
6.4 Marginal Effects in the Tobit Model
Just as with the binary-choice models, the Tobit model’s coefficients do not have the same
direct interpretation as in the classical linear regression. To interpret the estimated pa-
rameters, we examine the marginal effects – the derivatives of the conditional expectation
with respect to the explanatory variables. Due to their complexity, we just state them
without derivation. In the Tobit model, several conditional expectations and marginal
effects can be of interest; in all cases, the marginal effects can be seen as products of the
regression coefficients and positive functions.
• The expectation of the latent response:
E(yi∗ |xi ) = x⊤
i β
with the marginal effect
∂E(yi∗ |xi )
= βk ;
∂xik
• The probability of not being censored:
x⊤ β
P (yi > 0|xi ) = Φ i
σ
with the marginal effect
∂P (yi > 0|xi ) ∂Φ(x⊤
i β/σ) x⊤
i β βk
= =φ ;
∂xik ∂xik σ σ
32
• The conditional expectation for uncensored observations (using the notation λ(t) =
φ(t)/[1 − Φ(t)] for inverse Mill’s ratio):
⊤
x β
φ iσ x⊤
i β
E(yi |xi , yi > 0) = x⊤
i β + σ = x⊤
i β + λ(− )
x⊤ β σ
Φ iσ
with the marginal effect
⊤
∂E(yi |xi , yi > 0) x⊤
i β xi β x⊤
i β
= βk 1 − λ(− ) + λ(− )
∂xik σ σ σ
• The overall expectation of the observed response:
x⊤ β x⊤ β
E(yi |xi ) = P (yi > 0|xi )E(yi |xi , yi > 0) = x⊤
i β · Φ i
+ σφ i
σ σ
with the marginal effect
∂E(yi |xi )
= Φ(x⊤
i β/σ)βk .
∂xik
In practice, marginal effects are typically evaluated at the sample means or averaged over
observations and their standard errors are obtained using the delta method.
6.5 Example: Female Labor Supply
An influential application of the Tobit model is the analysis of female labor supply. In
this example, we consider data very similar to the one in Section 5.3, but with a different
dependent variable:
• The dependent variable is the annual number of hours worked, which can be zero
(if a woman chooses not to work) or positive. This is the case of the corner solution
problem, since the number of hours worked cannot be negative, only the latent
utility of working can.
• As in Section 5.3, explanatory variables include non-wife household income, educa-
tion, labor market experience (and its square), age, and the number of young and
older children.
Because about 40% of the sample individuals reports zero hours of work, a simple linear
regression would yield biased estimates. Therefore, we apply the Tobit model, which is
estimated using the log-likelihood function described above. The output for an example
data set is presented in the following table.
33
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
ho ur s | Coef . Std . Err . t
−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
nwifeinc | − 8.81424 4.4591 − 1.98
educ | 80.64561 21.5832 3.74
exper | 1 3 1 . 5 6 4 3 0 17.2794 7.61
expersq | − 1.86416 .5376 − 3.47
age | − 54.40501 7.4185 − 7.33
k i d s l t 6 | − 894.02170 111.8779 − 7.99
k i d s g e 6 | − 16.21800 38.6414 − 0.42
cons | 965.30530 446.4358 2.16
−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
/ sigma | 1122.022 41.57903
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
After estimating the model, the results can be interpreted either in a limited way using
the signs of the coefficients only, or by computing the marginal effects. For example in
the case of education, the coefficient is significantly positive, indicating that all discussed
marginal effects – for example, for the expected hours of work for those working – are
positive, ceteris paribus. To evaluate the actual effect in terms of working hours, the
marginal effects for the expected hours of work for those who work – E(yi |xi , yi > 0) –
are in the following table. In the case of education, we can see that one extra year of
education increases on average the expected hours worked by the employed individuals
by 36 hours per year.
Ma r g i na l e f f e c t s o f E( y i | xi , y i > 0 )
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
| Delta −method
| dy/dx Std . Err . z P>| z |
−−−−−−−−−−+−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
nwifeinc | − 3.96878 2.00758 − 1.98 0 . 0 4 8
educ | 36.31225 9.70304 3.74 0.000
exper | 59.23938 7.83368 7.56 0.000
expersq | − .83937 .24232 − 3.46 0 . 0 0 1
age | − 24.49691 3.36249 − 7.29 0 . 0 0 0
k i d s l t 6 | − 402.55070 50.74877 − 7.93 0 . 0 0 0
k i d s g e 6 | − 7.302468 17.40427 − 0.42 0 . 6 7 5
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
34
