Chapter 4

Binary Choice Models for Panel Data

Théophile T. Azomahou
University Clermont Auvergne, CNRS, CERDI
Maastricht University, School of Business and Economics
Email: [email protected]

Pooled Estimator
Random Effects Models
Fixed Effects Models
Dynamic Binary Choice Models
Théophile T. Azomahou (CERDI) Février 20-28, 2020 1 / 14
 Introduction

1. Introduction
The structural model for a possibly unbalanced panel of data would be written

yit∗ = x0it β + εit , i = 1, . . . , n, t = 1, . . . , Ti ,

yit = 1 if yit∗ > 0, and 0 otherwise, (1)

The second line of this definition is often written

yit = 1(x0it β + εit > 0)

to indicate a variable that equals one when the condition in parentheses is true
and zero when it is not.
A more promising approach is an effects
model,

yit∗ = x0it β + vit + ui , i = 1, . . . , n, t = 1, . . . , Ti ,

yit = 1 if yit∗ > 0, and 0 otherwise, (2)

where, ui is the unobserved, individual specific heterogeneity.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 2 / 14

 Introduction

We distinguish between “random” and “fixed” effects models by the relationship

between ui and xit .

The assumption that ui is unrelated to xit , so that the conditional

distribution f (ui | xit ) is not dependent on xit , produces the random effects
model. Note that this places a restriction on the distribution of the
heterogeneity.

If that distribution is unrestricted, so that ui and xit may be correlated, then

we have what is called the fixed effects model.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 3 / 14

 The Pooled Estimator

2. The Pooled Estimator

The pooled estimator results if we simply ignore the heterogeneity, ui and fit the
model as if the cross-section specification applies.
If the fixed effects model is appropriate, then all the results for omitted
variables apply. The pooled MLE that ignores fixed effects will be
inconsistent. Observe that since the estimator is ML, not least squares,
converting the data to deviations from group means is not a
solution—converting the binary dependent variable to deviations will
produce a continuous variable with unknown properties.
The random effects case is more benign. From (17-39), the marginal
probability implied by the model is

Prob(yit = 1 | xit ) = Prob(vit + ui > −x0it β)

1/2 
= F x0it β/ 1 + σu2


= F (x0it δ).

Théophile T. Azomahou (CERDI) Février 20-28, 2020 4 / 14

 The Pooled Estimator

The implication is that based on the marginal distributions, we can consistently

estimate δ (but not β or σu separately) by pooled MLE.

The implication, which is absent in the linear case is that ignoring the
random effects in a pooled model produces an attenuated (inconsistent —
downward biased) estimate of β; the scale factor that produces δ is
1/(1 + σu2 )1/2 which is between zero and one.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 5 / 14

 The Random Effects Models

3. The Random Effects Models

The random effects model specifies
εit = vit + ui ,

where vit and ui are independent random variables with

E [vit | X] = 0; Cov[vit , vjs | X] = Var[vit | X] = 1, if i = j and t = s; 0 otherwise

E [ui | X] = 0; Cov[ui , uj | X] = Var[ui | X] = σu2 , if i = j; 0 otherwise,
Cov[vit , uj | X] = 0 for all i, t, j,

and X indicates all the exogenous data in the sample, xit for all i and t. Then,

E [εit | X] = 0,
Var[εit | X] = σv2 + σu2 = 1 + σu2 ,

and σu2
Corr[εit , εis | X] = ρ = .
1 + σu2
The new free parameter is σu2 = ρ/(1 − ρ).
Théophile T. Azomahou (CERDI) Février 20-28, 2020 6 / 14
 The Fixed Effects Models

4. The Fixed Effects Models

The fixed effects model is

yit∗ = αi dit + x0it β + εit , i = 1, . . . , n, t = 1, . . . , Ti ,

(3)
yit = 1 if yit∗ > 0, and 0 otherwise,

where dit is a dummy variable that takes the value one for individual i and zero
otherwise. For convenience, we have redefined xit to be the nonconstant variables
in the model.

The parameters to be estimated are the K elements of β and the n individual

constant terms, possibly a huge number of parameters, (n + K ) and could be in
the thousands in a typical application.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 7 / 14

 The Fixed Effects Models

Estimation

The log-likelihood function for the fixed effects model is

Ti
n X
X
ln L = ln P(yit | αi + x0it β), (4)
i=1 t=1

where P(.) is the probability of the observed outcome, for example,

Φ[qit (αi + x0it β)] for the probit model or Λ[qit (αi + x0it β)] for the logit model,
where qit = 2yit − 1.

What follows can be extended to any index function model, but for the present,
we’ll confine our attention to symmetric distributions such as the normal and
logistic, so that the probability can be conveniently written as
Prob(Yit = yit | xit ) = P[qit (αi + x0it β)]. It will be convenient to let
zit = αi + x0it β so Prob(Yit = yit | xit ) = P(qit zit ).

Théophile T. Azomahou (CERDI) Février 20-28, 2020 8 / 14

 The Fixed Effects Models

Virtues and shortcomings of the fixed effects model

The problems with the fixed effects estimator are statistical, not practical.
The estimator relies on Ti increasing for the constant terms to be
consistent—in essence, each αi is estimated with Ti observations. But, in
this setting, not only is Ti fixed, it is likely to be quite small.

The estimator of β is a function of the estimators of α, which means that

the MLE of β is not consistent either. This is the incidental parameters
problem.

Hsiao (1986) found that for Ti = 2, the bias in the MLE of β is 100
percent, which is extremely pessimistic. Heckman and MaCurdy found in a
Monte Carlo study that in samples of n = 100 and T = 8, the bias appeared
to be on the order of 10 percent, which is substantive, but certainly less
severe than Hsiao’s results suggest.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 9 / 14

 Dynamic Binary Choice Models

5. Dynamic Binary Choice Models

A random or fixed effects model that explicitly allows for lagged effects would be

yit = 1(x0it β + αi + γyi,t−1 + εit > 0)

Lagged effects, or persistence, in a binary choice setting can arise from three
sources:

serial correlation in εit

the heterogeneity, αi
true state dependence through the term γyi,t−1

A related problem is that with a relatively short panel, the initial conditions, yi0 ,
have a crucial impact on the entire path of outcomes. Modeling dynamic effects
and initial conditions in binary choice models is more complex than in the linear
model.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 10 / 14

 Dynamic Binary Choice Models

The correlation between αi and yi,t−1 in the dynamic binary choice model makes
yi,t−1 endogenous. Thus, the estimators we have examined thus far will not be
consistent. Two familiar alternative approaches that have appeared in recent
applications are due to Heckman (1981) and Wooldridge (2005), both of which
build on the random effects specification.

Heckman’s approach

Heckman’s approach provides a separate equation for the initial condition,

Prob(yi1 = 1 | xi1 , zi , αi ) = Φ(x0i1 δ + z0i τ + θαi )

Prob(yit = 1 | xit , yi,t−1 , αi ) = Φ(x0it β + γyi,t−1 + αi ), t = 2, . . . , Ti ,

where zi is a set of “instruments” observed at the first period that are not
contained in xit .

Théophile T. Azomahou (CERDI) Février 20-28, 2020 11 / 14

 Dynamic Binary Choice Models

The conditional log-likelihood is

n
( Ti
)
X Y
ln L | α = ln Φ [(2yi1 − 1)(x0i1 δ + z0i τ + θαi )] Φ [(2yit − 1)(ξ)]
i=1 t=2
Xn
= ln Li | αi
i=1

with ξ = x0i1 β + γyi,t−1 + αi . Stewart and Arulampalam (2007) suggest a useful

shortcut for formulating the Heckman model. Let Dit = 1 in period 1 and 0 in
every other period and let Cit = 1 − Dit . Then, the two parts may be combined in
n
X Ti
Y
ln L | α = ln {Φ[(2yit − 1) hCit (ζ) + Dit (x0it δ + z0i τ ) + (1 + λDit )αi i]}
i=1 t=1

where ζ = x0i1 β
+ γyi,t−1 In this form, the model can be viewed as a random
parameters (random constant term) model in which there is heteroscedasticity in
the random part of the constant term.

Théophile T. Azomahou (CERDI) Février 20-28, 2020 12 / 14

 Dynamic Binary Choice Models

Wooldridge’s approach

Wooldridge’s approach suggests a model for the random effect conditioned on the
initial value. Thus,

αi | yi1 , zi ∼ N α0 + ηyi1 + z0i τ , σα2

 

Assembling the parts, Wooldridge’s model is a bit simpler than Heckman’s;

Prob(yit = 1 | xit , yi1 , ui )
= Φ[(2yit − 1)(α0 + x0it β + γyi,t−1 + ηyi1 + z0i τ + ui )], t = 2, . . . , Ti

Théophile T. Azomahou (CERDI) Février 20-28, 2020 13 / 14

 Dynamic Binary Choice Models

6. Application using STATA

The effect of informality on social protection

xtprobit, xtlogit, xtologit, xtoprobit

Théophile T. Azomahou (CERDI) Février 20-28, 2020 14 / 14