Econometrics II

Lecture 2: Discrete Choice Models

Måns Söderbom

4 April 2011

University of Gothenburg. [email protected]. www.economics.gu.se/soderbom. www.soderbom.net

1. Introduction

Linear regression is primarily designed for modelling a continuous, quantitative variable - e.g. economic

growth, the log of value-added or output, the log of earnings etc.

Many economic phenomena of interest, however, concern variables that are not continuous or perhaps

not even quantitative.

What characteristics (e.g. parental) a¤ect the likelihood that an individual obtains a higher degree?

What determines labour force participation (employed vs not employed)?

What factors drive the incidence of civil war?

Today we will discuss binary choice models. These are central models in applied econometrics.

Obviously binary choice models are useful when our outcome variable of interest is binary - a common

situation in applied work. Moreover, the binary choice model is often used as an ingredient in other

models. For example:

In propensity score matching models (to be covered in lecture 3), we identify the average

treatment e¤ect by comparing outcomes of treated and non-treated individuals who, a priori, have

similar probabilities of being treated. The probability of being treated is typically modelled using

probit.

In Heckman’s selection model, we use probit in the …rst stage to predict the likelihood that

someone is included (selected) in the sample. We then control for the likelihood of being selected

when estimating our equation of interest (e.g. a wage equation)

The binary choice model is also a good starting point if we want to study more complicated models.

Later on in the course we will thus cover extensions of the binary choice model, such as models for

multinomial or ordered response, and models combining continuous and discrete outcomes (e.g. corner

response models).

Useful references for this lecture:

2
 Greene, W (2008). Econometric Analysis, 6th edition.

Angrist, Joshua and Jörn-Stefen Pischke (2009). Mostly Harmless Econometrics. An Empiricist’s

Companion. Chapter 3.4.2

In addition, for my empirical examples I will draw on material presented in the following paper:.

Kingdon, G. (1996) ’The quality and e¢ ciency of private and public education: a case-study of urban

India,’Oxford Bulletin of Economics and Statistics 58: 57-81

2. Binary Response

Whenever the variable that we want to model is binary, it is natural to think in terms of probabilities,

e.g.

’What is the probability that an individual with such and such characteristics owns a car?’

’If some variable X changes by one unit, what is the e¤ect on the probability of owning a car?’

When the dependent variable y is binary, it is typically equal to one for all observations in the data for

which the event of interest has happened (’success’) and zero for the remaining observations (’failure’).

Provided we have a random sample, the sample mean of this binary variable is an unbiased estimate

of the unconditional probability that the event happens. That is, letting y denote our binary dependent

variable, we have
P
i yi
Pr (y = 1) = E (y) = ;
N

where N is the number of observations in the sample.

Estimating the unconditional probability is trivial, but usually not the most interesting thing we can

do with the data. Suppose we want to analyze what factors ’determine’changes in the probability that

y equals one. Can we use the classical linear regression framework to this end?

3
3. The Regression Approach

Consider the linear regression model

y = 1 + 2 x2 + ::: + K xK +u

= x + u; (3.1)

where is a K 1 vector of parameters, x is a N K matrix of explanatory variables, and u is a residual.

Assume that the residual is uncorrelated with the regressors, i.e. endogeneity is not a problem. This

allows us to use OLS to estimate the parameters of interest.

To interpret the results, note that if we take expectations on both sides of the equation above we

obtain

E (yjx; ) = x :

Now, just like the unconditional probability that y equals one is equal to the unconditional expected

value of y, i.e. E (y) = Pr (y = 1), the conditional probability that y equals one is equal to the

conditional expected value of y:

Pr (y = 1jx) = E (yjx; ) ;

Pr (y = 1jx) = x : (3.2)

Because probabilities must sum to one, it must also be that

Pr (y = 0jx) = 1 x :

Equation (3.2) is a binary response model. In this particular model the probability of success

(i.e. y = 1) is a linear function of the explanatory variables in the vector x. This is why using

OLS with a binary dependent variable is called the linear probability model (LPM).

4
 Notice that in the LPM the parameter j measures the change in the probability of ’success’, resulting

from a change in the variable xj , holding other factors …xed:

Pr (y = 1jx) = j xj :

This can be interpreted as a partial e¤ect on the probability of ’success’.

EXAMPLE: Modelling the probability of going to a private, unaided school (PUA) in India.1 See

appendix, Table 1a.

3.1. Shortcomings of the Linear Probability Model

Clearly the LPM is straightforward to estimate, however there are some important shortcomings.

One undesirable property of the LPM is that, if we plug in certain combinations of values for the

independent variables into (3.2), we can get predictions either less than zero or greater than one. Of

course a probability by de…nition falls within the (0,1) interval, so predictions outside this range are

hard to interpret. This is not an unusual result; for instance, based on the above LPM results, there

are 61 observations for which the predicted probability is larger than one and 81 observations for

which the predicted probability is less than zero. That is, 16 per cent of the predictions fall outside

the (0,1) interval in this application (see Figure 1 in the appendix, and the summary statistics for

the predictions reported below the table).

Angrist and Pischke (p.103): "...[linear regression] may generate …tted values outside the LDV

boundaries. This fact bothers some researchers and has generated a lot of bad press for the linear

probability model."

A related problem is that, conceptually, it does not make sense to say that a probability is linearly

related to a continuous independent variable for all possible values. If it were, then continually

increasing this explanatory variable would eventually drive P (y = 1jx) above one or below zero.

1 The data for this example are taken from the study by Kingdon (1996).

5
For example, the model above predicts that an increase in parental wealth by 1 unit increases the

probability of going to a PUA school by about 1 percentage point. This may seem reasonable for

families with average levels of wealth, however in very rich or very poor families the wealth e¤ect

is probably smaller. In fact, when taken to the extreme our model implies that a hundred-fold

increase in wealth increases the probability of going to a PUA by more than 1 which, of course, is

impossible (the wealth variable ranges from 0.072 to 82 in the data, so such an comparison is not

unrealistic).

A third problem with the LPM - arguably less serious than those above - is that the residual is

heteroskedastic by de…nition. Why is this? Because y takes the value of 1 or 0, the residuals in

equation (3.1) can take only two values, conditional on x: 1 x and x. Further, the respective

probabilities of these events are x and 1 x. Hence,

2
var (ujx) = Pr (y = 1jx) [1 x ]

2
+ Pr (y = 0jx) [ x ]

2 2
= x [1 x ] + (1 x )[ x ]

= x [1 x ];

which clearly varies with the explanatory variables x. The OLS estimator is still unbiased, but the

conventional formula for estimating the standard errors, and hence the t-values, will be wrong. The

easiest way of solving this problem is to obtain estimates of the standard errors that are robust to

heteroskedasticity.

EXAMPLE continued: Appendix - LPM with robust standard errors, Table 1b; compare to LPM

with non-robust standard errors (Table 1a).

A fourth and related problem is that, because the residual can only take two values, it cannot be

normally distributed. The problem of non-normality means that OLS point estimates are unbiased

6
 but its violation does mean that inference in small samples cannot be based on the usual suite of

normality-based distributions such as the t test.

Summarizing:

The LPM can be useful as a …rst step in the analysis of binary choices, but awkward issues arise if

we want to argue that we are modelling a probability.

As we shall see next, probit and logit solve these particular problems. Nowadays, these are just as

easy to implement as LPM/OLS - but they are less straightforward to interpret.

However, LPM remains a reasonably popular modelling framework because certain econometric

problems are easier to address within the LPM framework than with probits and logits.

If, for whatever reason, we use the LPM, it is important to recognize that it tends to give better

estimates of the partial e¤ects on the response probability near the centre of the distribution of

x than at extreme values (i.e. close to 0 and 1). The LPM graph in the appendix illustrates this

(Figure 1).

3.2. Logit and Probit Models for Binary Response

The two main problems with the LPM were: nonsense predictions are possible (there is nothing to bind

the value of Y to the (0,1) range); and linearity doesn’t make much sense conceptually.

To address these problems we abandon the LPM and thus the OLS approach to estimating binary

response models. Consider instead a class of binary response models of the form

Pr (y = 1jx) = F( 1 + 2 x2 + ::: + K xK )

Pr (y = 1jx) = F (x ) ; (3.3)

where F is a function taking on values strictly between zero and one: 0 < F (z) < 1, for all real numbers

z. The model (3.3) is often referred to in general terms as an index model, because Pr (y = 1jx) is a

7
function of the vector x only through the index

x = 1 + 2 x2 + ::: + k xk ;

which is simply a scalar. Notice that 0 < F (x ) < 1 ensures that the estimated response probabilities

are strictly between zero and one, which thus addresses the main worries of using LPM. F is usually a

cumulative density function (cdf), monotonically increasing in the index z (i.e. x ), with

Pr (y = 1jx) ! 1 as x ! 1

Pr (y = 1jx) ! 0 as x ! 1:

It follows that F must be a non-linear function, and hence we cannot use OLS.

Various non-linear functions for F have been suggested in the literature. By far the most common

ones are the logistic distribution, yielding the logit model, and the standard normal distribution, yielding

the probit model.

The logit model:

exp (x )
Pr (y = 1jx) = = (x ) ;
1 + exp (x )

which is between zero and one for all values of x (recall that x is a scalar). This is the cumulative

distribution function (CDF) for a logistic variable.

The probit model:

Z x
F (x ) = (x ) (v) dv;
1

where
1 v2
(v) = p exp ;
2 2

is the standard normal density. This choice of F also ensures that the probability of ’success’is strictly

between zero and one for all values of the parameters and the explanatory variables.

8
 EXAMPLE: See graphs in Figure 2, appendix.

The logit and probit functions are both increasing in x . Both functions increase relatively quickly

at x = 0, while the e¤ect on F at extreme values of x tends to zero. The latter result ensures that the

partial e¤ects of changes in explanatory variables are not constant, a concern we had with the LPM.

Also notice that the standard normal CDF has a shape very similar to of the logistic CDF, suggesting

that it doesn’t much matter which one of the two we choose to use in our analysis. I will come back to

this point later.

Interpretation of probit and logit estimates is less straightforward than what we are used to for

linear regression. Note in particular that the marginal e¤ects - the e¤ects on the response probability

Pr (y = 1jx) resulting from a change in one of the explanatory variables - cannot be read o¤ the parameter

vector directly. Let’s look at how to compute marginal e¤ects in a few cases.

3.2.1. Case I: The explanatory variable is continuous.

In linear models the marginal e¤ect of a unit change in some explanatory variable on the dependent

variable is simply the associated coe¢ cient on the relevant explanatory variable.

However, for logit and probit models obtaining measures of the marginal e¤ect is more complicated

(which should come as no surprise, as these models are non-linear). When xj is a continuous

variable, its partial e¤ect on Pr (y = 1jx) is obtained from the partial derivative:

@ Pr (y = 1jx) @F (x )
=
@xj @xj

= f (x ) j;

where
dF (z)
f (z)
dz

is the probability density function associated with F .

Because the density function is non-negative, the partial e¤ect of xj will always have the same

9
 sign as j.

Notice that the partial e¤ect depends on f (x ); i.e. for di¤erent values of x1 ; x2 ; :::; xk the partial

e¤ect will be di¤erent. Hence, one has to take a stand on how to evaluate the marginal e¤ects.

– One possibility is to evaluate marginal e¤ects at the sample mean values of x1 ; x2 ; :::; xk .

This is what Stata command ’mfx compute’does (unless you tell it otherwise). This command

also provides standard errors for the marginal e¤ects - more on this below.

– Alternatively, you could compute marginal e¤ects for each observation in the sample and

average them - this gives you the average marginal e¤ect.

– Or, you could evaluate them anywhere you like, depending on what kind of argument you

want to make (e.g. suppose income is an explanatory variable, and suppose you want to say

something about the e¤ect among low-income people - then it makes sense to evaluate the

marginal e¤ect at a low income level.

Can you see at what values of x the partial (or marginal) e¤ect will be relatively small/large? See

graphs of the standard normal and the logistic CDFs in handout.

EXAMPLE: Suppose we use the Indian data introduced above to estimate a probit modelling the

probability that a child goes to a private unaided school as a function of the child’s ability, measured by

the score on the Raven’s test. For simplicity, abstract from other explanatory variables. Our model is

thus:

Pr (pua = 1jsraven) = ( 0 + 1 sraven) :

The probit results are

coef. t-value

0 -1.82 12.84

1 0.050 11.76

Since the coe¢ cient on sraven is positive, we know that the marginal e¤ect must be positive. Treating

10
sraven as a continuous variable, it follows that the marginal e¤ect is equal to

@ Pr (pua = 1jsraven)
= ( 0 + 1 sraven) 1
@sraven

= ( 1:82 + 0:05 sraven) 0:05;

where (:) is the standard normal density function:

1
(z) = p exp z 2 =2 :
2

We see straight away that the marginal e¤ect depends on the level of sraven. We see from the summary

statistics that the mean value of sraven is about 31, so let’s evaluate the marginal e¤ect at sraven = 31:

@ Pr (pua = 1jsraven = 31)

@sraven
1 2
= p exp ( 1:82 + 0:05 31) =2 0:05
2
= 0:019;

Evaluated at the mean of sraven, we see that the results imply that an increase in sraven by one unit

raises the probability of going to a private school by about two percentage points. At lower levels of

sraven, the marginal e¤ect is smaller:

@ Pr (pua = 1jsraven = 15)

@sraven
1 2
= p exp ( 1:82 + 0:05 20) =2 0:05
2
= 0:011:

Of course, the fact that the marginal e¤ect is smaller at lower levels re‡ects the non-linearity of the probit

model (again: see graphs in Figure 2 in handout).

STUDENT EXERCISE: Now consider logit:

11
 Pr (pua = 1jsraven) = ( 0 + 1 sraven) ;

exp (z)
(z) = :
1 + exp (z)

The logit results are

coef. t-value

0 -3.07 12.00

1 0.084 11.20

Task: Calculate and interpret the marginal e¤ect. Compare the result to the probit marginal e¤ect.

3.2.2. Case II: The explanatory variable is discrete.

If xj is a discrete variable then we should not rely on calculus in evaluating the e¤ect on the response

probability. To keep things simple, suppose x2 is binary. In this case the partial e¤ect from changing x2

from zero to one, holding all other variables …xed, is

F( 1 + 2 1 + ::: + K xK ) F( 1 + 2 0 + ::: + K xK ) :

Again this depends on all the values of the other explanatory variables and the values of all the other

coe¢ cients.

Again, knowing the sign of 2 is su¢ cient for determining whether the e¤ect is positive or not, but

to …nd the magnitude of the e¤ect we have to use the formula above.

The Stata command ’mfx compute’can spot dummy explanatory variables. In such a case it will use

the above formula for estimating the partial e¤ect.

3.2.3. Case III: Non-linear explanatory variables.

Suppose the model is

2
Pr (y = 1jx) = F 1 + 2 x2 + 2 x3 + 22 x2 ;

12
where x22 is a continuous variable. What is the marginal e¤ect of x2 on the response probability?

4. Latent Regression - Index Function Models

As we have seen, the probit and logit models resolve some of the problems with the LPM model. The

key, really, is the speci…cation

Pr (y = 1jx) = F (x ) ;

where F is the cdf for either the standard normal or the logistic distribution, because with any of these

models we have a functional form that is easier to defend than the linear model.

The traditional way of introducing probits and logits in econometrics, however, is not as a response

to a functional form problem. Instead, probits and logits are traditionally viewed as models suitable for

estimating parameters of interest when the dependent variable is not fully observed. Let’s have a look at

this perspective.

Let y be a continuous variable that we do not observe - a latent variable - and assume y is

determined by the model

y = 1 + 2 x2 + ::: + K xK +e

= x + e; (4.1)

where e is a residual, assumed uncorrelated with x (i.e. x is not endogenous). While we do not observe

y , we do observe the discrete choice made by the individual, according to the following choice rule:

y = 1 if y > 0

y = 0 if y 0:

Why is y unobserved? Think about y as representing net utility of, say, buying a car. The individual

13
undertakes a cost-bene…t analysis and decides to purchase the car if the net utility is positive. We do not

observe (because we cannot measure) the ’amount’of net utility; all we observe is the actual outcome of

whether or not the individual does buy a car. (If we had data on y we could estimate the model (5.4)

with OLS as usual.)

Now, we want to model the probability that a ’positive’choice is made (e.g. buying, as distinct from

not buying, a car). Assuming that e follows a logistic distribution2 ,

exp ( e)
(e) = 2 (density),
(1 + exp( e))
exp (e)
(e) = (CDF),
1 + exp(e)

it follows that

Pr (y = 1jx) = Pr (y > 0jx)

= Pr (x + e > 0jx)

= Pr (e > x )

= 1 ( x ) (integrate)

= (x ) (exploit symmetry):

Notice that the last step here exploits the fact that the logistic distribution is symmetric, so that F (z) =

1 F ( z) for all z. This equation is exactly the binary response model (3.3) for the logit model. This

is how the binary response model can be derived from an underlying latent variable model.

We can follow the same route to derive the probit model. Assume e follows a standard normal

distribution

2 Note exp(e) exp( e)

that symmetry of the probability density function implies (e) = (1+exp(e))2
= (1+exp( e))2
. In some expositions
exp(e) exp( e)
you see (1+exp(e))2
, in others (1+exp( e))2
. Don’t let this confuse you.

14
 Pr (y = 1jx) = Pr (y > 0jx)

= Pr (x + e > 0jx)

= Pr (e > x )

x
= 1 N (integrate)

= (x ) ,

where again we exploit symmetry and use = 1 implied by the standard normal distribution. This is

the binary response model (3.3) for the probit model.3

5. Estimation and Inference in Binary Choice Models

To estimate the LPM we can use OLS. Because of the non-linear nature of the probit and logit models

(see graphs), however, linear estimators are not applicable for these. Instead we rely on Maximum

Likelihood (ML) estimation. The principle of ML is very general and not con…ned to probit and logit

models. Before turning the details of how ML is used to estimate probits and logits, here is an informal

recap of ML.

5.1. Maximum Likelihood: Recap

Suppose that, in the population, there is a variable w which is distributed according to some

distribution f (w; ), where is a vector of unknown parameters.

Suppose we have a random sample fw1 ; w2 ; :::; wN g drawn from the population distribution f (w; )

where is unknown.

Our objective is to estimate . Our sample is more likely to have come from a population charac-

3 The assumption that = 1 may appear restrictive. In fact, this is a necessary normalisation, because we cannot
estimate by means of a binary response model.

15
terized by one particular set of parameter values, say ~, than from another set of parameter values,

say .

ML
The maximum likelihood estimate (MLE) of is simply the particular vector ^ that gives the

greatest likelihood (or, if you prefer, probability) of observing the sample fw1 ; w2 ; :::; wN g.

Random sampling (an assumption) implies that w1 ; w2 ; :::; wN are independent of each other,

hence the likelihood of observing fw1 ; w2 ; :::; wN g (i.e. the sample) is simply

L ( ; w1 ; w2 ; :::; wN ) = f (w1 ; ) f (w2 ; ) ::: f (wN ; ) ;

or, in more compact notation,

N
Y
L ( ; w1 ; w2 ; :::; wN ) = f (wi ; ) :
i=1

i.e. the product of the individual likelihoods. The equation just de…ned is a function of : for

some values of the resulting L will be relatively high while for other values of it will be low.

This is why we refer to equations of this form as likelihood functions.

The value of that gives the maximum value of the likelihood function is the maximum likelihood

estimate of .

For computational reasons it is much more convenient to work with the log-likelihood function:

N
X
ln L ( ; w1 ; w2 ; :::; wN ) = ln f (wi ; ) :
i=1

ML
The value of that gives the maximum value of the log likelihood function is the ^ :

16
5.2. Maximum likelihood estimation of logit and probit models

We now return to the logit and probit models. How can ML be used to estimate the parameters of

interest in these models, i.e. ? Assume that we have random sample of size N . The ML estimate of
ML
is the particular vector ^ that gives the greatest likelihood of observing the sample fy1 ; y2 ; :::; yN g,

conditional on the explanatory variables x.

By assumption, the probability of observing yi = 1 is F (x ) while the probability of observing yi = 0

is 1 F (x ) : It follows that the probability of observing the entire sample is

Y Y
L (yjx; ) = [1 F (xi )] F (xi ) ;
yi =0 yi =1

We can rewrite this as

N
Y y (1 yi )
L (yjx; ) = F (xi ) i [1 F (xi )] ;
i=1

because when y = 1 we get F (xi ) and when y = 0 we get [1 F (xi )].

The log likelihood for the sample is

N
X
ln L (yjx; ) = fyi ln F (xi ) + (1 yi ) ln [1 F (xi )]g :
i=1

The MLE of maximizes this log likelihood function.

If F is the logistic CDF then we obtain the logit likelihood:

N
X
ln L (yjx; ) = fyi ln (xi ) + (1 yi ) ln [1 (xi )]g
i=1
N
X exp (xi ) 1
ln L (yjx; ) = yi ln + (1 yi ) ln ;
i=1
1 + exp (xi ) 1 + exp (xi )

17
 which simpli…es to

N
X
ln L (yjx; ) = fyi [xi ln (1 + exp (xi ))] (1 yi ) ln (1 + exp (xi ))g :
i=1

If F is the standard normal CDF we get the probit estimator:

N
X
ln L (yjx; ) = fyi ln (xi ) + (1 yi ) ln [1 (xi )]g :
i=1

How maximize the log likelihood function?

Sample log likelihood:

N
X
ln L (yjx; ) = fyi ln F (xi ) + (1 yi ) ln [1 F (xi )]g :
i=1

Because the objective is to maximize the log likelihood function with respect to the parameters in

the vector , it must be that, at the maximum, the following K …rst order conditions will hold:

N
X f (xi ) f (xi )
yi + (1 yi ) xi = 0:
i=1
F (xi ) [1 F (xi )]

1x1 1xK

Greene refers to these as likelihood equations.

It is typically not possible to solve analytically for here. Instead, to obtain parameter estimates,

we rely on some sophisticated iterative ’trial and error’ technique. There are lots of algorithms

that can be used, but we will not study these here. The most common ones are based on …rst and

sometimes second derivatives of the log likelihood function. Think of a blind man walking up a

hill, and whose only knowledge of the hill comes from what passes under his feet. Provided the hill

is strictly concave, the man should have no trouble …nding the top. Fortunately the log likelihood

18
 functions for logit and probit are concave, but this is not always the case for other models.

EXAMPLE: Appendix, Tables 2-3.

5.3. Hypothesis Tests

5.3.1. Inference based on the log likelihood function

We have already discussed how the ML estimates of the parameters are those that maximize the

likelihood of observing the sample. It must then be that all other parameter values - which, by

de…nition, are not the ML estimates - will result in a lower (worse) log likelihood value.

Now let’s revisit our Indian dataset and investigate what happens to the log likelihood value if we

change the value of the coe¢ cient on sraven. See Figure 3 in the handout.

As expected, values of b_raven not equal to 0.03 produce lower log likelihood values.

Is it important how much the log L falls as a result of moving b_sraven a given distance away

from the ML estimate of 0.03? Yes, very important, because this, essentially, is the general basis

for our inference. Think about the log likelihood ratio test.

The log likelihood ratio test is de…ned as two times the di¤erence in two log likelihood values:

LR = 2 (ln LR ln LU ) ;

where ln LU is the log likelihood value for the unrestricted model and ln LR is the log likelihood

value for the restricted model. LR follows a chi-squared distribution with q degrees of freedom

under H0 , where q is the number of restrictions.

Suppose now I want to test the following null hypothesis:

H0 : b_sraven = 0.

19
Looking at Figure 3, I see that

ln LR ' 356;

(i.e. this is the log likelihood value associated with b_sraven = 0) and I know from the regression

output (or from the graph) that

ln LU = 340:4:

Hence

LR = 2 ( 340:4 + 356) = 31:2:

To test H0 at the 5% level we use as our critical value the 95th percentile in the 2
q distribution.

With q = 1 (because there is only one restriction here) the critical value is 3.84, so I …rmly reject

the null hypothesis at the 5% level. If you want a speci…c p-value, we can type

chiprob(1; 31:2)

in Stata which is equal to 0.00000002. We can thus reject the null at any conventional level of

signi…cance.

Key point: It is the large fall in the log L resulting from imposing b_sraven = 0 that enables us

to reject the null hypothesis. Had the log likelihood function been ‡atter in b_sraven, we might

not have been able to reject the null hypothesis.

The log likelihood ratio is often used to test whether a sub-set of the explanatory variables can be

omitted from the model. Again the idea is that, because ML maximizes the log likelihood function,

dropping variables will lead to a lower log likelihood value (this is similar to the result that the

R-squared falls when variables are dropped from an OLS regression). The question is whether the

fall in the log likelihood is large enough to conclude that the dropped variables are important. The

20
 likelihood ratio statistic:

LR = 2 (ln LR ln LU ) ;

where ln LU is the log likelihood value for the unrestricted model, e.g.

F( 0 + 1 x1 + 2 x2 + 3 x3 + 4 x4 ) ;

and ln LR is the value for the restricted model, e.g.

F( 0 + 1 x1 ) :

So estimate these two models and compare the two log likelihood values. We can obtain p-values

directly in Stata by using the command

chiprob(q; LR).

What’s q in this case?

Example in appendix, Table 8.

In Table 2 in the appendix, how should we interpret the information in ’LR chi2(9)’?

5.3.2. Standard errors for parameters

1
In linear models the conventional covariance matrix is given by 2
X 0X which is straightforward

to estimate. In non-linear models, however, such as the probit and the logit, deriving formulas for the

covariance matrix, and hence the standard errors, is more complicated. The conventional estimator for

the covariance matrix is based on the inverse of the negative Hessian:

ML
V ar ^ = H 1
; (5.1)

21
where the Hessian is the matrix of second order derivatives of the log likelihood function:

ML
@ 2 ln L yjxi ; ^
H( )= 0 :
@ @

Note that provided the log likelihood function is concave, the second derivative is negative which ensures

that the variance is positive.

This is (somewhat) intuitive. Note that the second derivative of the log likelihood function with
ML
respect to (evaluated at ^ ) measures the curvature of the log likelihood function - and so

variance formula (5.1) says that the more curvature, the lower is the variance.

Recall that how big is the quantitative fall in the log L as a result of imposing other parameter

values than the ML estimate, is central for our inference if we use the log likelihood ratio test.

Clearly curvature plays a central role here: with little curvature you have to move the parameter

value for b_sraven a long way away from its ML estimate before the LR test rejects, but with a

lot of curvature you will not have to move far. So you see the variance and the LR test are very

closely related.

Sometimes you see ’robust’ standard errors reported - these are obtained from the ’sandwich’

formula:
ML
h ML
i 1 h ML
ih ML
i 1
V ar ^ = H ^ V ar s ^ H ^ ;

where
ML
ML @ ln L yjx; ^
s ^ ML
@^

is the gradient vector, or score vector.

5.3.3. Standard errors for marginal e¤ects

Once we have estimated the variance matrix, we can calculate standard errors by taking the square

root of the diagonal elements of the covariance matrix, and subsequently obtain t-values and con-

22
…dence intervals in the usual ways.

We can also calculate standard errors for the marginal e¤ects (recall these are non-linear functions

of the parameters). The Stata mfx command does this for us using the delta method, which
ML @ Pr(y=1jx)
involves transforming the standard errors of ^ into standard errors of @xj by means of a

Taylor series approximation. Here is how it works:

@ Pr(y=1jx)
– Our goal is to estimate the standard error of the marginal e¤ect @xj : De…ne

@ Pr (y = 1jx)
h( );
@xj

making it explicit that the marginal e¤ect is a function of the parameters . We have obtained
ML ML
an estimate of ; denoted ^ . We have also estimated the covariance matrix V ar ^ .

We now need to obtain V ar ^ M

j
L
.

M LE
– Now de…ne ^ M
j
LE
= h ^ , and then take a Taylor series approximation of ^ M
j
LE
=
M LE
h ^ around the true value :

XK
@h ^ M LE
^M
j
LE
' j + i i :
i=1
@ i

In matrix notation

^ M LE ' ^ M LE ; (5.2)

where 2 3
@h1 @h1 @h1
6 @ 1 @ 2 ::: @ K 7
6 7
6 7
6 @h2 @h2
::: @h2 7
6 @ 1 @ 2 @ K 7
=6
6
7
7
6 :: 7
6 7
6 7
4 5
@hJ @hJ
@ 1 ::: @ K

is a J K matrix of derivatives. Post-multiply (5.2) by the transpose of (5.2), and take

23
 expectations, and you get the variance matrix for the marginal e¤ects:

ML
V ar ^ M LE = V ar ^ 0
:

5.4. Speci…cation Tests for Binary Choice Models

A lot has been written about the problems posed by heteroskedasticity for the probit and logit models.

You often hear statements to the e¤ect that probit and logit estimates are inconsistent in the presence

of heteroskedasticity. Greene (2006, p. 787) argues that this is a serious problem "because the probit

model is most often used with microeconomic data, which are frequently heteroscedastic". What is the

nature of the problem? Consider the following illustration:4

Start from a latent variable model with one explanatory variable xi1 :

yi = 0 + 1 xi1 + ui : (5.3)

Suppose the residual ui is heteroskedastic. Consider the following - admittedly very special and arguably

peculiar - form of heteroskedasticity:

ui N ormal 0; x2i1 :

Recall that we do not observe yi - all we observe is the binary dependent variable:

yi = 1 if yi > 0

yi = 0 if yi 0:

Thus,

yi = 1 if 0 + 1 xi1 + ui > 0:

4 This is taken from Section 15.7.4 in Wooldridge (2002) "Econometric Analysis of Cross Section and Panel Data".

24
What is the probability that y = 1? We have

Pr (yi = 1jxi ) = Pr (yi > 0jxi )

= Pr ( 0 + 1 xi1
+ ui > 0jxi )
q
= Pr 0 + 1 xi1 + x2i1 ei > 0jxi ;

where ei follows a standard normal distribution (i.e. with mean zero and variance equal to one). Hence,

1
Pr (yi = 1jxi1 ) = Pr ei > ( + 1 xi1 )
xi1 0
1
= 1 ( + 1 xi1 ) (integrate)
xi1 0
1
= ( + 1 xi1 ) , (symmetry)
xi1 0
1
= 0 + 1 :
xi1

We now see how the presence of heteroskedasticity radically has altered the functional form of the

probit model. Given that the underlying latent model is

yi = 0 + 1 xi1 + ui ;

we might be tempted to specify the probit model as

Pr (yi = 1jxi ) = ( 0 + 1 xi1 ) ;

but this would not be the correct speci…cation. This is quite important. Think about the partial e¤ect

of xi1 . The correct speci…cation is

1
Pr (yi = 1jxi ) = 0 + 1 ;
xi1

25
hence the correct marginal e¤ect is

!
2
@ Pr (yi = 1jxi1 ) 1 1
= 0 + 1 0 :
@xi1 xi1 xi1

Remarkably, the sign of the marginal e¤ect is the opposite of that of 0 - i.e. the constant in the latent

variable model - and does not depend on the sign of 1 - the slope coe¢ cient on xi1 in the latent variable

model. It follows that if 0 and 1 are both positive, the marginal e¤ect of xi1 on the probability of

’success’has the opposite sign to the marginal e¤ect of xi1 on the latent dependent variable yi .

Of course the latter result is driven by the speci…c form of heteroskedasticity considered here, and

should not be viewed as a general result. The main point is that if the residual in the latent variable model

is heteroskedastic this alters the functional form. Exactly how depends on the form of heteroskedasticity.

Now, suppose you were to specify (incorrectly) the probit as

Pr (yi = 1jxi ) = ( 0 + 1 xi1 ) :

Do you think your coe¢ cient on x1 would be a good estimate of the coe¢ cient 1 in the latent variable

model

yi = 0 + 1 xi1 + ui ?

Answer: no. And this is an example of how the presence of heteroskedasticity leads to "inconsistent

estimates" of the parameters in the latent variable model.

How can we proceed if we believe heteroskedasticity is a problem? One possibility is to use Stata’s

hetprob command, which estimates a generalized probit model:

y = 1 + 2 x2 + ::: + K xK +e

y = x + e; (5.4)

26
where
2 2
e = [exp (z )] ;

where z is a vector of variables (not including a constant - since not identi…ed) thought to a¤ect the

variance of e, and is the corresponding vector of coe¢ cients. We obtain

Pr (y = 1jx; z) = Pr (y > 0jx; z)

= Pr (x + e > 0jx; z)

= Pr (x + exp (z ) u > 0jx; z) ;

where u follows a standard normal distribution (a normalization). Hence

x
Pr (y = 1jx; z) = Pr u >
exp (z )
x
Pr (y = 1jx; z) = 1 N (integrate)
exp (z )
x
Pr (y = 1jx; z) = .
exp (z )

Of course, if a variable xk is included in both x and z, the marginal e¤ect is somewhat more involved:

@ Pr (y = 1jx; z) x k (x ) k
= :
@xk exp (z ) exp (z )

This shows that the sign of the marginal e¤ect is not necessarily the same as the sign of k.

EXAMPLE: Heteroskedasticity in school choice probit. Appendix.

5.5. Measuring Goodness of Fit

In linear models where the dependent variable is continuous, we often rely on the R-squared as a measure

of the goodness of …t of the model. If for some reason we use linear regression in a binary choice setting

(i.e. LPM here), you will obviously get an estimate of the R-squared. However, you should probably not

27
pay too much attention to this statistic. Why?

Recall:
var (^
y)
R2 = ;
var (y)

where y^ denotes the predictions from the regression. But remember the main problem with LPM is

that linearity is an unattractive feature of the model - both conceptually and in the sense that nonsense

probability predictions may result. Consequently, we should not take the predictions of the LPM too

seriously and so any measures of how ’good’these predictions are, is of limited interest.

The two most common alternative measures of goodness of …t for binary choice models are the percent

correctly predicted, and the pseudo R-squared.

Percent correctly predicted. To obtain the percent correctly predicted we begin by computing the

estimated probability that yi equals one for each observation in the sample. For the probit model, for

instance, this is given by

M LE
Est: Pr (y = 1jx) = x^ ;

M LE
where Est: denotes ’estimated’. We then say that the predicted outcome of yi is one if x^ >

0:50 and zero otherwise. The percentages of times the predicted yi matches the actual yi is the per cent

correctly speci…ed. Note the di¤erence between predicted outcome (which is binary, 0 or 1) and predicted

probability (any number between 0 and 1).

The per cent correctly predicted is a useful measure in this context, but we need to be careful. Consider

a case where out of 200 observations, 180 have yi = 0. If, say, 150 of these are predicted to be zero we

obtain 75% correct predictions, even if our model fails to predict any of the observations for which y = 1

correctly. This is not an uncommon outcome in practice. For this reason, it is a good idea to report the

percentages (or frequencies) correctly predicted for each of the two outcomes.

Appendix, Table 7.

Note: Hard to say a priori what makes up a ’satisfactory’percentage of correct predictions.

28
Pseudo R-squared. Various pseudo R-squared measures for binary response models have been devel-

oped. The most common one is

~2 = 1 ln Lur
R ;
ln Lr

where ln Lur is the value of the log likelihood at the ML estimates (the ’unrestricted’model) and ln Lr

is the log likelihood value for a ’restricted’model in which the only ’explanatory’variable is a constant.

What is the logic of using this formula? Notice that if our explanatory variables have no explanatory

~ 2 = 0.
power at all, then ln Lr = ln Lur (why?). In this case we get R

In contrast, if our model is doing very well indeed in predicting the actual observations of y, then the

~ 2 will tend to
log likelihood value (of the unrestricted model) will approach zero from below, and hence R

one. Why?

Recall that the log likelihood function is

N
X
ln L (yjxi ; ) = fyi ln F (xi ) + (1 yi ) ln [1 F (xi )]g :
i=1

A very good model will be such that F (xi ) will be very close to one for all observations for which yi = 1

and very close to zero for all observations for which yi = 0. To illustrate the point, suppose F (xi ) is

exactly one for all observations for which yi = 1 and exactly zero for all observations for which yi = 0 -

i.e. the model predicts the dependent variable perfectly. In that extreme case, we have

N
X
ln L (yjxi ; ) = fyi ln 1 + (1 yi ) ln [1 0]g
i=1
N
X
= fyi 0 + (1 yi ) 0g
i=1

= 0;

29
and so

~2 0
R = 1
ln Lr

= 1:

~ 2 uses the same information as that underlying the log likelihood ratio test.
Notice that R

~ 2 is consistent with the LR test.

See Table 2 in the appendix. Verify that the reported R

30
PhD Programme: Econometrics II
Department of Economics, University of Gothenburg
Appendix Lecture 2
Måns Söderbom

Binary Choice Models

Application: School Choice in India

The data used below were kindly provided by Dr Geeta Kingdon. These data have been used
in

Kingdon, G. (1996) ‘The quality and efficiency of private and public education: a case-study
of urban India,’ Oxford Bulletin of Economics and Statistics 58: 57-81.

See Table 1 in the paper for details on how variables are defined.

Key variables and summary statistics:

Contains data from kingdon96.dta

obs: 902
vars: 9
size: 36,080 (99.7% of memory free)
-------------------------------------------------------------------------------
storage display value
variable name type format label variable label
-------------------------------------------------------------------------------
numsib float %9.0g Number of siblings
sraven float %9.0g Raven ability score
wealth float %9.0g Index of household asset value
male float %9.0g Gender dummy: male=1, female=0
lowcaste float %9.0g Low caste? yes=1,no=0
muslim float %9.0g Muslim? yes=1,no=0
medyrs float %9.0g Mother's education in years
sikhchr float %9.0g Sikh or Christian? yes=1,no=0
stype float %9.0g School type: 0=govt, 1=private
aided, 2=private unaided
-------------------------------------------------------------------------------

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
puaind | 902 .3991131 .4899878 0 1
numsib | 902 3.988914 1.705215 1 11
sraven | 902 30.52661 11.22551 3 57
wealth | 902 24.25723 21.08854 .072 82
male | 902 .5321508 .4992421 0 1
-------------+--------------------------------------------------------
lowcaste | 902 .1330377 .3398039 0 1
muslim | 902 .2184035 .4133919 0 1
medyrs | 902 8.665188 4.954049 0 20
medyrsq | 902 99.60089 79.36289 0 400
sikhchr | 902 .0310421 .1735278 0 1

1
Now consider results from OLS, probit and logit using the Stata code in Box 1.

Box 1: Stata code for estimation of binary choice models

#delimit;

use kingdon96.dta, clear;

describe;
summarize;

tabstat numsib sraven wealth male ,

by(stype) s(mean p50 sd);

tabstat lowcaste muslim medyrs sikhchr,

by(stype) s(mean p50 sd);

ge puaind=stype==2;
replace puaind=. if stype==.;

ge medyrsq = medyrs^2;

sum puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr;

/*** LPM, probit, logit ***/

reg puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr;
reg puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr, robust;
predict yhat; /* obtain predicted probability */
predict xb, xb;
label var xb "xb (index)";
scatter puaind yhat xb, symbol(+ o) jitter(2) l1title("Linear prediction &
actual outcome");

probit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr;
predict phat, p; /* obtain predicted probability */

ge phat_d=phat>.5; /* predicted outcome */

tab phat_d puaind; /* compare predicted & acutal outcomes */

test sraven wealth; /* Wald test, joint significance */

mfx compute;

logit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr;
predict lhat, p; /* obtain predicted probability */

sum yhat phat lhat;

count if yhat<0; /* number of negative predicted probabilities, LPM
*/
count if yhat>1; /* number of predicted probabilities in excess of
one, LPM */

exit;

2
Table 1a. LINEAR PROBABILITY MODEL

> regress puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq sikhchr;

Source | SS df MS Number of obs = 902

-------------+------------------------------ F( 9, 892) = 85.25
Model | 100.026088 9 11.1140097 Prob > F = 0.0000
Residual | 116.293203 892 .130373546 R-squared = 0.4624
-------------+------------------------------ Adj R-squared = 0.4570
Total | 216.31929 901 .240088003 Root MSE = .36107

------------------------------------------------------------------------------
puaind | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0223168 .0082608 -2.70 0.007 -.0385297 -.0061038
sraven | .0075825 .0012103 6.27 0.000 .0052072 .0099578
wealth | .0101314 .0007259 13.96 0.000 .0087067 .0115561
male | .1732116 .0245567 7.05 0.000 .1250159 .2214072
lowcaste | -.1412188 .0392124 -3.60 0.000 -.2181782 -.0642594
muslim | -.1387535 .0321586 -4.31 0.000 -.2018689 -.0756381
medyrs | -.0245589 .0078772 -3.12 0.002 -.0400188 -.0090989
medyrsq | .0016972 .0005077 3.34 0.001 .0007008 .0026937
sikhchr | .220197 .0702409 3.13 0.002 .0823403 .3580538
_cons | .0047471 .0624763 0.08 0.939 -.1178706 .1273648
------------------------------------------------------------------------------

Figure 1: Predictions based on LPM shown in Table 1a

1.5
Linear prediction & actual outcome
1
.5
0
-.5

-.5 0 .5 1 1.5
xb (index)

puaind Fitted values

Note: The linear prediction is denoted yhat, and puaind is the actual binary dependent
variable. The puaind variable has been “jittered” to facilitate interpretation.

3
. sum yhat;

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
yhat | 902 .3991131 .3331918 -.2283615 1.499489

. count if yhat>1;
61

. count if yhat<0;
81

Table 1b. LINEAR PROBABILITY MODEL WITH STANDARD ERRORS ROBUST TO HETEROSKEDASTICITY

> regress puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq sikhchr,
> robust ;

Regression with robust standard errors Number of obs = 902

F( 9, 892) = 189.05
Prob > F = 0.0000
R-squared = 0.4624
Root MSE = .36107

------------------------------------------------------------------------------
| Robust
puaind | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0223168 .0084851 -2.63 0.009 -.0389699 -.0056636
sraven | .0075825 .00126 6.02 0.000 .0051096 .0100554
wealth | .0101314 .0006368 15.91 0.000 .0088817 .0113811
male | .1732116 .0241983 7.16 0.000 .1257193 .2207039
lowcaste | -.1412188 .0374195 -3.77 0.000 -.2146593 -.0677782
muslim | -.1387535 .0317488 -4.37 0.000 -.2010645 -.0764425
medyrs | -.0245589 .007757 -3.17 0.002 -.039783 -.0093347
medyrsq | .0016972 .000503 3.37 0.001 .00071 .0026845
sikhchr | .220197 .0775071 2.84 0.005 .0680796 .3723145
_cons | .0047471 .0618058 0.08 0.939 -.1165547 .1260488
------------------------------------------------------------------------------

4
Figure 2: The Logit and Probit CDFs

Logit model: G(bx)=exp(bx)/(1+exp(bx))

1
0.9
0.8
0.7
0.6
Pr(y=1|x)

0.5
0.4
0.3
0.2
0.1
0
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

bx

Probit model: G(bx)=PHI(bx)

1
0.9
0.8
0.7
0.6
Pr(y=1|x)

0.5
0.4
0.3
0.2
0.1
0
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

bx

5
Table 2. PROBIT MODEL

> probit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq sikhchr;

Iteration 0: log likelihood = -606.73067

Iteration 1: log likelihood = -373.10677
Iteration 2: log likelihood = -343.27331
Iteration 3: log likelihood = -340.47774
Iteration 4: log likelihood = -340.43889
Iteration 5: log likelihood = -340.43888

Probit estimates Number of obs = 902

LR chi2(9) = 532.58
Prob > chi2 = 0.0000
Log likelihood = -340.43888 Pseudo R2 = 0.4389

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0998298 .0382835 -2.61 0.009 -.1748641 -.0247956
sraven | .0301986 .0054653 5.53 0.000 .0194869 .0409103
wealth | .0461453 .0043108 10.70 0.000 .0376963 .0545943
male | .8575159 .1199153 7.15 0.000 .6224862 1.092546
lowcaste | -.5496526 .1865875 -2.95 0.003 -.9153575 -.1839478
muslim | -.7229197 .1530685 -4.72 0.000 -1.022929 -.4229109
medyrs | -.1260082 .0373075 -3.38 0.001 -.1991296 -.0528868
medyrsq | .0079365 .0024278 3.27 0.001 .0031781 .0126948
sikhchr | .8875504 .3272338 2.71 0.007 .246184 1.528917
_cons | -1.882662 .287822 -6.54 0.000 -2.446783 -1.318541
------------------------------------------------------------------------------

. /* marginal effects using mfx compute */

> mfx compute;

Marginal effects after probit

y = Pr(puaind) (predict)
= .38659838
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
numsib | -.0382063 .01465 -2.61 0.009 -.066911 -.009502 3.98891
sraven | .0115574 .00208 5.54 0.000 .007472 .015643 30.5266
wealth | .0176605 .00172 10.27 0.000 .014289 .021032 24.2572
male*| .3167018 .04152 7.63 0.000 .235328 .398076 .532151
lowcaste*| -.1926379 .05762 -3.34 0.001 -.305563 -.079713 .133038
muslim*| -.2513949 .04612 -5.45 0.000 -.341793 -.160997 .218404
medyrs | -.0482251 .01433 -3.37 0.001 -.076308 -.020142 8.66519
medyrsq | .0030374 .00093 3.26 0.001 .001211 .004864 99.6009
sikhchr*| .3401752 .11123 3.06 0.002 .122159 .558191 .031042
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

6
Table 3. LOGIT MODEL

> logit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq sikhchr;

Iteration 0: log likelihood = -606.73067

Iteration 1: log likelihood = -373.59867
Iteration 2: log likelihood = -342.99987
Iteration 3: log likelihood = -338.7387
Iteration 4: log likelihood = -338.60417
Iteration 5: log likelihood = -338.604

Logit estimates Number of obs = 902

LR chi2(9) = 536.25
Prob > chi2 = 0.0000
Log likelihood = -338.604 Pseudo R2 = 0.4419

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.1763513 .0702654 -2.51 0.012 -.3140689 -.0386336
sraven | .0566268 .0099437 5.69 0.000 .0371376 .0761161
wealth | .0812388 .0079356 10.24 0.000 .0656852 .0967923
male | 1.573975 .2206061 7.13 0.000 1.141595 2.006355
lowcaste | -1.097276 .35503 -3.09 0.002 -1.793122 -.4014304
muslim | -1.30201 .2766134 -4.71 0.000 -1.844162 -.7598576
medyrs | -.2326064 .0666913 -3.49 0.000 -.363319 -.1018938
medyrsq | .0142063 .0043246 3.28 0.001 .0057302 .0226824
sikhchr | 1.672074 .5791094 2.89 0.004 .5370403 2.807107
_cons | -3.388287 .5269294 -6.43 0.000 -4.421049 -2.355524
------------------------------------------------------------------------------

. /* marginal effects using mfx compute */

> mfx compute;

Marginal effects after logit

y = Pr(puaind) (predict)
= .36959733
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
numsib | -.041089 .01632 -2.52 0.012 -.073077 -.009101 3.98891
sraven | .0131938 .00229 5.76 0.000 .008701 .017686 30.5266
wealth | .0189282 .00196 9.65 0.000 .015082 .022774 24.2572
male*| .3480658 .04426 7.86 0.000 .261309 .434823 .532151
lowcaste*| -.2195702 .05702 -3.85 0.000 -.331332 -.107809 .133038
muslim*| -.26307 .04577 -5.75 0.000 -.352781 -.173359 .218404
medyrs | -.0541962 .01563 -3.47 0.001 -.084826 -.023567 8.66519
medyrsq | .00331 .00101 3.27 0.001 .001324 .005296 99.6009
sikhchr*| .3900823 .10997 3.55 0.000 .174542 .605622 .031042
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

Table 4. PREDICTED PROBABILITIES: PROBIT AND LOGIT

. sum phat lhat;

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
phat | 902 .4023303 .3436492 .0009482 .999997
lhat | 902 .3991131 .3473132 .0037879 .9996895

Note: phat = predicted probability based on probit model; lhat = predicted probability
based on logit model.

7
Table 5. SIMPLE PROBIT MODEL: PUA = PHI(SRAVEN)

Probit estimates Number of obs = 902

LR chi2(1) = 152.66
Prob > chi2 = 0.0000
Log likelihood = -530.40283 Pseudo R2 = 0.1258

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
sraven | .0499438 .0042474 11.76 0.000 .0416191 .0582685
_cons | -1.816083 .1414252 -12.84 0.000 -2.093271 -1.538894
------------------------------------------------------------------------------

Table 6: The baseline probit model (same as Table 2)

Probit estimates Number of obs = 902

LR chi2(9) = 532.58
Prob > chi2 = 0.0000
Log likelihood = -340.43888 Pseudo R2 = 0.4389

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0998298 .0382835 -2.61 0.009 -.1748641 -.0247956
sraven | .0301986 .0054653 5.53 0.000 .0194869 .0409103
wealth | .0461453 .0043108 10.70 0.000 .0376963 .0545943
male | .8575159 .1199153 7.15 0.000 .6224862 1.092546
lowcaste | -.5496526 .1865875 -2.95 0.003 -.9153575 -.1839478
muslim | -.7229197 .1530685 -4.72 0.000 -1.022929 -.4229109
medyrs | -.1260082 .0373075 -3.38 0.001 -.1991296 -.0528868
medyrsq | .0079365 .0024278 3.27 0.001 .0031781 .0126948
sikhchr | .8875504 .3272338 2.71 0.007 .246184 1.528917
_cons | -1.882662 .287822 -6.54 0.000 -2.446783 -1.318541
------------------------------------------------------------------------------

. test sraven wealth;

( 1) sraven = 0
( 2) wealth = 0

chi2( 2) = 150.74
Prob > chi2 = 0.0000

Now vary the coefficient on sraven around the ML estimate of 0.03 – see Figure 1.

8
Figure 3: The log-likelihood as function of b_sraven

-340
-350
log likelihood
-370 -360
-380

-.02 0 .02 .04 .06 .08

b_sraven

As expected, values of b_raven not equal to 0.03 produce a lower log likelihood value. Is
it important how much the log L falls as a result of moving b_sraven away from the ML
estimate of 0.03?

Predictions:

. predict phat, p;
. ge phat_d=phat>.5;
. table phat_d pua;

Table 7: Frequencies of correct predictions

----------------------
| puaind
phat_d | 0 1
----------+-----------
0 | 491 87
1 | 51 273
----------------------

9
Illustration of LR test

Estimate restricted model without sraven and wealth. Compare the resulting log likelihood
value to that obtained in the unrestricted model (Table 2):

Table 8: Restricted probit: sraven and wealth omitted

Probit estimates Number of obs = 902

LR chi2(7) = 311.29
Prob > chi2 = 0.0000
Log likelihood = -451.08324 Pseudo R2 = 0.2565

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.1092096 .0343527 -3.18 0.001 -.1765395 -.0418796
male | .6587308 .1010961 6.52 0.000 .4605861 .8568756
lowcaste | -.5847738 .1689565 -3.46 0.001 -.9159225 -.253625
muslim | -.6309888 .1307152 -4.83 0.000 -.8871859 -.3747918
medyrs | -.0969399 .0336861 -2.88 0.004 -.1629634 -.0309163
medyrsq | .0127564 .0021721 5.87 0.000 .0084992 .0170135
sikhchr | .8240591 .3017656 2.73 0.006 .2326093 1.415509
_cons | -.4723225 .2159322 -2.19 0.029 -.8955419 -.0491031
------------------------------------------------------------------------------

. display 2*(-340.43888 - -451.08324 )

221.28872

. disp chiprob(2,221.29)
8.861e-49

(=0.0000000000…)

10
Heteroskedasticity in the school choice probit

Consider the benchmark probit model reported in Table 2 above:

Probit estimates Number of obs = 902

LR chi2(9) = 532.58
Prob > chi2 = 0.0000
Log likelihood = -340.43888 Pseudo R2 = 0.4389

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0998298 .0382835 -2.61 0.009 -.1748641 -.0247956
sraven | .0301986 .0054653 5.53 0.000 .0194869 .0409103
wealth | .0461453 .0043108 10.70 0.000 .0376963 .0545943
male | .8575159 .1199153 7.15 0.000 .6224862 1.092546
lowcaste | -.5496526 .1865875 -2.95 0.003 -.9153575 -.1839478
muslim | -.7229197 .1530685 -4.72 0.000 -1.022929 -.4229109
medyrs | -.1260082 .0373075 -3.38 0.001 -.1991296 -.0528868
medyrsq | .0079365 .0024278 3.27 0.001 .0031781 .0126948
sikhchr | .8875504 .3272338 2.71 0.007 .246184 1.528917
_cons | -1.882662 .287822 -6.54 0.000 -2.446783 -1.318541
------------------------------------------------------------------------------

Now relax the assumption that the error term is homoskedastic, by writing the variance of the
error term as [exp(g*sraven)]^2, where g is a parameter to be estimated (note: if g=0 we’re
back to the homoskedastic model). I can obtain results for this generalized model by using the
hetprob command in Stata:

hetprob puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr, het(sraven);

Heteroskedastic probit model Number of obs = 902

Zero outcomes = 542
Nonzero outcomes = 360

Wald chi2(9) = 36.56

Log likelihood = -336.3521 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
puaind |
numsib | -.0672515 .0246799 -2.72 0.006 -.1156232 -.0188798
sraven | .0229242 .0041345 5.54 0.000 .0148208 .0310276
wealth | .0270485 .0057435 4.71 0.000 .0157914 .0383056
male | .5055253 .1190219 4.25 0.000 .2722466 .7388039
lowcaste | -.3304765 .127884 -2.58 0.010 -.5811245 -.0798284
muslim | -.410586 .1214911 -3.38 0.001 -.6487041 -.1724678
medyrs | -.0691932 .0268416 -2.58 0.010 -.1218016 -.0165847
medyrsq | .0041276 .0017081 2.42 0.016 .0007798 .0074754
sikhchr | .4802327 .2136518 2.25 0.025 .0614829 .8989825
_cons | -1.272155 .2532967 -5.02 0.000 -1.768607 -.7757022
-------------+----------------------------------------------------------------
lnsigma2 |
sraven | -.0171179 .0059309 -2.89 0.004 -.0287422 -.0054935
------------------------------------------------------------------------------
Likelihood-ratio test of lnsigma2=0: chi2(1) = 8.17 Prob > chi2 = 0.0043

Clearly there is evidence here that the variance of the error term falls with sraven.
Alternatively, this can be interpreted as indicating that the functional form of the baseline
probit is wrong. Now consider adding sraven squared to the baseline model, on the
grounds that this is a generalization of the baseline probit. Results:

11
. ge sraven2=sraven^2;

. probit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq sikhchr
sraven2;

Probit regression Number of obs = 902

LR chi2(10) = 539.44
Prob > chi2 = 0.0000
Log likelihood = -337.00859 Pseudo R2 = 0.4445

------------------------------------------------------------------------------
puaind | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
numsib | -.0976174 .0383671 -2.54 0.011 -.1728155 -.0224192
sraven | -.0364237 .0257712 -1.41 0.158 -.0869344 .0140869
wealth | .0467619 .0044055 10.61 0.000 .0381273 .0553964
male | .8598127 .1210329 7.10 0.000 .6225926 1.097033
lowcaste | -.5200105 .1866679 -2.79 0.005 -.8858729 -.1541481
muslim | -.7285526 .1536653 -4.74 0.000 -1.029731 -.4273742
medyrs | -.1232646 .0372906 -3.31 0.001 -.1963527 -.0501765
medyrsq | .00768 .0024345 3.15 0.002 .0029084 .0124516
sikhchr | .9275063 .3282048 2.83 0.005 .2842368 1.570776
sraven2 | .0011059 .0004214 2.62 0.009 .00028 .0019317
_cons | -1.0303 .4258538 -2.42 0.016 -1.864958 -.195642
------------------------------------------------------------------------------

Clearly the squared term is quite significant.

Based on the three models shown above, the following graph illustrates how the predicted
probability of going to a private unaided school varies with sraven, holding all other
explanatory factors constant.
1
.8
.6
.4
.2

0 20 40 60
(mean) sraven

Baseline homoskedastic probit Heteroskedastic probit

Homoskedastic probit, quadratic sraven

12
It seems the heteroskedastic probit and the homoscedastic probit with sraven squared
included tell a similar story: the likelihood that y=1 is relatively insensitive to changes to
sraven at low levels, but more sensitive to changes to sraven at high levels than what is
implied by the benchmark model.

Box 2: Stata code generating the graph on the previous page

probit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr;
estimates store base;
hetprob puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr, het(sraven);
estimates store het;

ge sraven2=sraven^2;
probit puaind numsib sraven wealth male lowcaste muslim medyrs medyrsq
sikhchr sraven2;
estimates store quad;

collapse _all;
ge junk=50;
expand junk;
replace sraven=sraven+(_n-25);
replace sraven2=sraven^2;

estimates restore base;

predict p1, p;
estimates restore het;
predict p2, p;
estimates restore quad;
predict p3, p;

label var p1 "Baseline homoskedastic probit";

label var p2 "Heteroskedastic probit";
label var p3 "Homoskedastic probit, quadratic sraven";

scatter p1 p2 p3 sraven, s(+ d o);

exit;

13