Econometrics II

Ordered & Multinomial Outcomes. Tobit regression.

Måns Söderbom

2 May 2011

Department of Economics, University of Gothenburg. Email: mans.soderbom@economics.gu.se. Web:

www.economics.gu.se/soderbom, www.soderbom.net.
1. Introduction

In this lecture we consider the following econometric models:

Ordered response models (e.g. modelling the rating of the corporate payment default risk, which

varies from, say, A (best) to D (worst))

Multinomial response models (e.g. whether an individual is unemployed, wage-employed or self-

employed)

Corner solution models and censored regression models (e.g. modelling household health expendi-

ture: the dependent variable is non-negative, continuous above zero and has a lot of observations

at zero)

These models are designed for situations in which the dependent variable is not strictly continuous and

not binary. They can be viewed as extensions of the nonlinear binary choice models studied in Lecture

2 (probit & logit).

References:

Greene 23.10 (ordered response); 23.11 (multinomial response); 24.2-3 (truncation & censored data).

You might also …nd the following sections in Wooldridge (2002) "Cross Section and Panel Data"

useful: 15.9-15.10; 16.1-5; 16.6.3-4; 16.7; 17.3 (these references in Wooldridge are optional).

2. Ordered Response Models

What’s the meaning of ordered response? Consider credit rating on a scale from zero to six, for

instance, and suppose this is the variable that we want to model (i.e. this is the dependent variable).

Clearly, this is a variable that has ordinal meaning: six is better than …ve, which is better than four etc.

The standard way of modelling ordered response variables is by means of ordered probit or ordered

logit. These two models are very similar. I will discuss the ordered probit, but everything below carries

over to the logit if we replace the normal CDF (:) by the logistic CDF (:).

2
 Can you think of reasons why OLS may not be suitable for modelling an ordered response variable?

Could binary choice models (LPM, probit, logit) potentially be used?

2.1. Ordered Probit

Let y be an ordered response taking on the values f0; 1; 2; :::; Jg: We derive the ordered probit from a

latent variable model (cf. probit binary choice)

y = 1 x1 + ::: + k xk +"

= x0 + "; (2.1)

where e is a normally distributed variable with the variance normalized to one. Notice that this model

does not contain a constant. Why will become clear in a moment.

Next de…ne J cut-o¤ points (or threshold parameters) as follows:

1 < 2 < ::: J:

We do not observe the latent variable, but we do observe choices according to the following:

y = 0 if y 1

y = 1 if 1 <y 2

y = 2 if 2 <y 3

(:::)

y = J if J <y :

Think of the cut-o¤ points as intercept shifters. This is how Stata speci…es the model. Greene does it

slightly di¤erently, including an intercept in the x vector and (e¤ectively) normalizes 1 = 0.

3
 Suppose y can take three values: 0, 1 or 2. We then have

y = 0 if x0 + " 1

y = 1 if 1 < x0 + " 2

y = 2 if 2 < x0 + ":

We can now de…ne the probabilities of observing y = 0; 1; 2. For the smallest and the largest value, the

resulting expressions are very similar to what we have seen for the binary probit:

Pr (y = 0jx) = Pr (x0 + " 1)

= Pr (e 1 x0 )

= ( 1 x0 ) ;

= 1 (x0 1)

Pr (y = 2jx) = Pr (x0 + " > 2)

= Pr (" > 2 x0 )

= 1 ( 2 x0 )

= (x0 2) :

For the intermediate category, we get:

Pr (y = 1jx) = Pr ( 1 < x0 + " 2)

= Pr (" > 1 x0 ; " 2 x0 )

= [1 ( 1 x0 )] (x0 2) ;

= 1 (1 (x0 1 )) (x0 2) ;

= (x0 1) (x0 2) ;

4
or equivalently

Pr (y = 1jx) = ( 2 x0 ) ( 1 x0 )

(remember: (a) = 1 ( a), because the normal distribution is symmetric - keep this in mind when

studying ordered probits or you might get lost in the algebra). In the general case where there are several

intermediate categories, all the associated probabilities will be of this form; see Greene, p.832. Notice

that the probabilities sum to one.

2.2. Interpretation

When discussing binary choice models we paid a lot of attention to marginal e¤ects - i.e. the partial

e¤ects of a small change in explanatory variable xj on the probability that we have a positive outcome.

For ordered models, we can clearly compute marginal e¤ects on the predicted probabilities along the

same principles. It is not obvious (to me, anyway) that this the most useful way of interpreting the

results is, however. Let’s have a look the marginal e¤ects and then discuss.

2.2.1. Partial e¤ects on predicted probabilities

When discussing marginal e¤ects for binary choice models, we focussed on the e¤ects on the probability

that y (the binary dependent variable) is equal to one. We ignored discussing e¤ects on the probability

that y is equal to zero, as these will always be equal to minus one times the partial e¤ect on the probability

that y is equal to one.

Since we now have more than two outcomes, interpretation of partial e¤ects on probabilities becomes

somewhat more awkward. Sticking to the example in which we have three possible outcomes, we obtain:

@ Pr (y = 2jx)
= (x0 2) k;
@xk

5
for the highest category (note: analogous to the expression for binary probit).1 Moreover,

@ Pr (y = 1jx)
= [ (x0 1) (x0 2 )] k;
@xk

for the intermediate category, and

@ Pr (y = 0jx)
= (x0 1) k;
@xk

for the lowest category, assuming that xk is a continuous variable enter the index model linearly (if xk is

discrete - typically binary - you just compute the discrete change in the predicted probabilities associated

with changing xk by one unit, for example from 0 to 1). We observe:

The partial e¤ect of xk on the predicted probability of the highest outcome has the same sign as

k.

The partial e¤ect of xk on the predicted probability of the lowest outcome has the opposite sign to

The sign of the partial e¤ect of xk on predicted probabilities of intermediate outcomes cannot, in

general, be inferred from the sign of k. This is because there are two o¤setting e¤ects - suppose

k > 0, then the intermediate category becomes more likely if you increase xk because the the

probability of the lowest category falls, but it also becomes less likely because the the probability of

the highest category increases (illustrate this in a graph). Typically, partial e¤ects for intermediate

probabilities are quantitatively small and often statistically insigni…cant. Don’t let this confuse you!

Discussion - how best interpret results from ordered probit (or logit)?

1 Remember that (a) = ( a) - i.e. I could just as well have written

@ Pr (y = 2jx)
= ( 2 x) k;
@xk
for instance - cf. Greene’s exposition on p.833.

6
Clearly one option here is to look at the estimated -parameters, emphasizing the underlying latent

variable equation with which we started. Note that we don’t identify the standard deviation of

" separately. Note also that consistent estimation of the -parameters requires the model to be

correctly speci…ed - e.g. homoskedasticity and normality need to hold, if we are using ordered

probit. Such assumptions are testable using, for example, the methods introduced for binary choice

models. You don’t often see this done in applied work however.

Another option might be to look at the e¤ect on the expected value of the ordered response

variable, e.g.

@E (yjx; ) @ Pr (y = 0jx) @ Pr (y = 1jx) @ Pr (y = 2jx)

= 0+ 1+ 2;
@xk @xk @xk @xk

in our example with three possible outcomes. This may make a lot of sense if y is a numerical

variable - basically, if you are prepared to compute mean values of y and interpret them. For

example, suppose you’ve done a survey measuring consumer satisfaction where 1="very unhappy",

2="somewhat unhappy", 3="neither happy nor unhappy", 4="somewhat happy", and 5="very

happy", then most people would be prepared to look a the sample mean even though strictly the

underlying variable is qualitative, thinking that 3.5 (for example) means something (consumers are

on average a little bit happy?). In such a case you could look at partial e¤ects on the conditional

mean.

Alternatively, you might want investigate the e¤ect on the probability of observing categories j; j +

1; :::; J. In my consumer satisfaction example, it would be straightforward to compute the partial

e¤ect on the probability that a consumer is "somewhat happy" or "very happy", for example.

Thus, it all boils down to presentation and interpretation here, and exactly what your quantity

of interest is depends on the context. We can use the Stata command ’mfx compute’ to obtain

estimates of the partial e¤ects on the predicted probabilities, but for more elaborate partial e¤ects

you may have to do some coding tailored to the context.

7
 EXAMPLE: Incidence of corruption in Kenyan …rms. Section 1 in the appendix.

3. Multinomial response: Multinomial logit

Suppose now the dependent variable is such that more than two outcomes are possible, where the outcomes

cannot be ordered in any natural way. For example, suppose we are modelling occupational status

based on household data, where the possible outcomes are self-employed (SE), wage-employed (WE) or

unemployed (UE). Alternatively, suppose we are modelling the transportation mode for commuting to

work: bus, train, car,...

Binary probit and logit models are ill suited for modelling data of this kind. Of course, in principle we

could bunch two or more categories and so construct a binary outcome variable from the raw data (e.g.

if we don’t care if employed individuals are self-employed or wage-employees, we may decide to construct

a binary variable indicating whether someone is unemployed or employed). But in doing so, we throw

away potentially interesting information. And OLS is obviously not a good model in this context.

However, the logit model for binary choice can be extended to model more than two outcomes.

Suppose there are J possible outcomes in the data. The dependent variable y can then take J values, e.g.

0,1,...,J-1. So if we are modelling, say, occupational status, and this is either SE, WE or UE, we have

J = 3. There is no natural ordering of these outcomes, and so what number goes with what category is

arbitrary (but, as we shall see, it matters for the interpretation of the results). Suppose we decide on the

following:

y = 0 if individual is UE,

y = 1 if individual is WE,

y = 2 if individual is SE.

8
We write the conditional probability that an individual belongs to category j = 0; 1; 2 as

Pr (yi = jjxi ) ;

where xi is a vector of explanatory variables. Reasonable restrictions on these probabilities are:

that each of them is bounded in the (0,1) interval,

that they sum to unity (one).

One way of imposing these restrictions is to write the probabilities in logit form:

exp (xi 1 )
Pr (yi = 1jxi ) = ;
1 + exp (xi 1 ) + exp (xi 2)

exp (xi 2 )
Pr (yi = 2jxi ) = ;
1 + exp (xi 1 ) + exp (xi 2)

Pr (yi = 0jxi ) = 1 Pr (yi = 1jxi ) Pr (yi = 2jxi )

1
= :
1 + exp (xi 1 ) + exp (xi 2)

The main di¤erence compared to what we have seen so far, is that there are now two parameter vectors,

1 and 2 (in the general case with J possible responses, there are J 1 parameter vectors). This makes

interpretation of the coe¢ cients more di¢ cult than for binary choice models.

The easiest case to think about is where 1k and 2k have the same sign. If 1k and 2k are positive

(negative) then it is clear that an increase in the variable xk makes it less (more) likely that the

individual belongs to category 0.

But what about the e¤ects on Pr (yi = 1jxi ) and Pr (yi = 2jxi )? This is much trickier than what

we are used to. We know that, for sure, the sum of Pr (yi = 1jxi ) and Pr (yi = 2jxi ) will increase,

but how this total increase is allocated between these two probabilities is not obvious. To …nd out,

9
 we need to look at the marginal e¤ects. We have

@ Pr (yi = 1jxi ) 1
= 1k exp (xi 1 ) [1 + exp (xi 1) + exp (xi 2 )]
@xik
2
exp (xi 1 ) [1 + exp (xi 1) + exp (xi 2 )]

( 1k exp (xi 1) + 2k exp (xi 2 )) ;

which can be written as

@ Pr (yi = 1jxi )
= 1k Pr (yi = 1jxi )
@xik
1
Pr (yi = 1jxi ) [1 + exp (xi 1) + exp (xi 2 )]

( 1k exp (xi 1) + 2k exp (xi 2 )) ;

or
@ Pr (yi = 1jxi ) 1kexp (xi 1 ) + 2k exp (xi 2 )
= Pr (yi = 1jxi ) 1k : (3.1)
@xik 1 + exp (xi 1 ) + exp (xi 2 )

Similarly, for j = 2:

@ Pr (yi = 2jxi ) 1k exp (xi 1 ) + 2k exp (xi 2 )

= Pr (yi = 2jxi ) 2k ;
@xik 1 + exp (xi 1 ) + exp (xi 2 )

while for the base category j = 0:

@ Pr (yi = 0jxi ) 1k exp (xi 1 ) + 2k exp (xi 2 )

= Pr (yi = 0jxi ) :
@xik 1 + exp (xi 1 ) + exp (xi 2 )

Of course it’s virtually impossible to remember, or indeed interpret, these expressions. The point is that

whether the probability that y falls into, say, category 1 rises or falls as a result of varying xik , depends

not only on the parameter estimate 1k , but also on 2k . As you can see from (3.1), the marginal e¤ect
@ Pr(yi =1jxi )
@xik may in fact be negative even if 1k is positive, and vice versa. Why might that happen?

10
 EXAMPLE: Appendix, Section 2. Occupational outcomes amongst Kenyan manufacturing workers.

3.0.2. Independence of irrelevant alternatives (IIA)

Reference: Greene, 23.11.3.

The multinomial logit is very convenient for modelling an unordered discrete variable that can take

on more than two values. One important limitation of the model is that the ratio of any two probabilities

j and m depends only on the parameter vectors j and m, and the explanatory variables xi :

Pr (yi = 1jxi ) exp (xi 1)

=
Pr (yi = 2jxi ) exp (xi 2)

= exp (xi ( 1 2 )) :

It follows that the inclusion or exclusion of other categories must be irrelevant to the ratio of the two

probabilities that y = 1 and y = 2. This is potentially restrictive, in a behavioral sense.

Example: Individuals can commute to work by three transportation means: blue bus, red bus, or

train. Individuals choose one of these alternatives, and the econometrician estimates a multinomial logit

modelling this decision, and obtains an estimate of

Pr (yi = red busjxi )

:
Pr (yi = trainjxi )

Suppose the bus company were to remove blue bus from the set of options, so that individuals can choose

only between red bus and train. If the econometrician were to estimate the multinomial logit on data

generated under this regime, do you think the above probability ratio would be the same as before?

If not, this suggests the multinomial logit modelling the choice between blue bus, red bus and train is

mis-speci…ed: the presence of a blue bus alternative is not irrelevant for the above probability ratio, and

thus for individuals’decisions more generally.

Some authors (e.g. Greene; Stata manuals) claim we can test the IIA assumption for the multinomial

logit by means of a Hausman test. The basic idea is as follows:

11
1. Estimate the full model. For example, with red bus, train and blue bus being the possible outcomes,

and with red bus de…ned as the benchmark category. Retain the coe¢ cient estimates.

2. Omit one category and re-estimate the model - e.g. exclude blue bus, and model the binary decision

to go by train as distinct from red bus.

3. Compare the coe¢ cients from (1) and (2) above using the usual Hausman formula. Under the null

that IIA holds, the coe¢ cients should not be signi…cantly di¤erent from each other.

This procedure does not make sense to me.

First, you don’t really have data generated in the alternative regime (with blue bus not being an

option) and so how can you hope to shed light on the behavioral e¤ect of removing blue bus from

the set of options?

Second, obviously sample means of ratios such as

Pr (yi = red bus) Nred bus =N

=
Pr (yi = train) Ntrain =N

don’t depend on blue bus outcomes. So if you estimate a multinomial logit with only a constant

included in the speci…cation, the estimated constant in the speci…cation train speci…cation (with

red bus as the reference outcome) will not change if you omit blue bus outcomes when estimating

(i.e. step (2) above). Conceptually, a similar issue will hold if you have explanatory variables in the

model, at least if you have a ‡exible functional form in your xi indices (e.g. mutually exclusive

dummy variables)

Third, from what I have seen the Hausman test for the IIA does not work well in practice (not very

surprising).

While testing for IIA in the context of a multinomial logit appears problematic, it may more sense

in a di¤erent setting. For example, it will work …ne for conditional logit models, i.e. models where

12
 choices are made based on observable attributes of each alternative (e.g. ticket prices for blue bus,

red bus and train may vary). So, what I have said above applies speci…cally for the multinomial

logit.

Note that there are lots of other econometric models that can be used to model multinomial response

models - notably multinomial probit, conditional logit, nested logit etc. These will not be discussed here.

EXAMPLE: Hausman test for IIA based on multinomial logit gives you nonsense - appendix, Section

3.

4. Tobit Estimation of Corner Solution Models

Reference: Greene. 24.2-3.

We now consider econometric issues that arise when the dependent variable is bounded but continuous

within the bounds. We focus …rst on corner solution models, and then turn to the censored regression

model (duration data is often censored) and truncated regression.

In general, a corner solution response variable is bounded such that

lo yi hi;

where lo denotes the lower bound (limit) and hi the higher bound, and where these bounds are the result

of real economic constraints.

By far the most common case is lo = 0 and hi = 1, i.e. there is a lower limit at zero and no upper

limit. The dependent variable takes the value zero for a nontrivial fraction of the population, and

is roughly continuously distributed over positive values. You will often …nd this in micro data, e.g.

household expenditure on education, health, alcohol,... .

You can thus think of this type of variable as a hybrid between a continuous variable (for which

the linear model is appropriate) and a binary variable (for which one would typically use a binary

13
 choice model). Indeed, as we shall see, the econometric model designed to model corner solution

variables looks like a hybrid between OLS and the probit model. In what follows we focus on the

case where lo = 0, hi = 1, however generalizing beyond this case is reasonably straightforward.

Let y be a variable that is equal to zero for some non-zero proportion of the population, and that is

continuous and positive if it is not equal to zero. As usual, we want to model y as a function of a set of

variables x1 ; x2 ; :::; xk - or in matrix notation:

x= 1 x1 x2 ::: xk :

4.1. OLS

We have seen how for binary choice models OLS can be a useful starting point (yielding the linear

probability model), even though the dependent variable is not continuous. We now have a variable which

is ’closer’to being a continuous variable - it’s discrete in the sense that it is either in the corner (equal

to zero) or not (in which case it’s continuous).

OLS is a useful starting point for modelling corner solution variables:

y = x + u:

We’ve seen that there are a number of reasons why we may not prefer to estimate binary choice models

using OLS. For similar reasons OLS may not be an ideal estimator for corner response models:

Based on OLS estimates we can get negative predictions, which doesn’t make sense since the

dependent variable is non-negative (if we are modelling household expenditure on education, for

instance, negative predicted values do not make sense).

Conceptually, the idea that a corner solution variable is linearly related to a continuous independent

variable for all possible values is a bit suspect. It seems more likely that for observations close to

the corner (close to zero), changes in some continuous explanatory variable (say x1 ) has a smaller

14
 e¤ect on the outcome than for observations far away from the corner. So if we are interested in

understanding how y depends on x1 among low values of y, linearity is not attractive.

A third (and less serious) problem is that the residual u is likely to be heteroskedastic - but we can

deal with this by simply correcting the standard errors.

A fourth and related problem is that, because the distribution of y has a ’spike’at zero, the residual

cannot be normally distributed. This means that OLS point estimates are unbiased, but inference

in small samples cannot be based on the usual suite of normality-based distributions such as the t

test.

So you see all of this is very similar to the problems identi…ed with the linear probability model.

4.2. Tobit

To …x these problems we follow a similar path as for binary choice models. We start, however, from the

latent variable model, written as

y = x + u; (4.1)

2
where the residual u is assumed normally distributed with a constant variance u, and uncorrelated

with x: As usual, the latent variable y is unobserved - we observe

8 9
>
> >
>
< y if y > 0 =
y= ; (4.2)
>
> >
: 0 if y 0 >
;

which can be written equivalently as

y = max (y ; 0) :

Two things should be noted here.

First, y satis…es the classical linear model assumptions, so had y been observed the obvious choice

of estimator would have been OLS.

15
 Second, it is often helpful to think of y as a variable that is bounded below for economic reasons,

and y as a variable that re‡ects the ’desired’value if there were no constraints. Actual household

expenditure on health is one example - this is bounded below at zero. In such a case y could

be interpreted as desired expenditure, in which case y < 0 would re‡ect a desire to sell o¤ ones

personal (or family’s) health. This may not be as far-fetched as it sounds - if you’re very healthy

and very poor, for instance, perhaps you wouldn’t mind feeling a little less healthy if you got paid

for it (getting paid here, of course, would be the same as having negative health expenditure).

We said above that a corner solution variable is a kind of hybrid: both discrete and continuous. The

discrete part is due to the piling up of observations at zero. The probability that y is equal to zero can

be written

Pr (y = 0jx) = Pr (y 0) ;

= Pr (x + u 0) ;

= Pr (u x )

x
= (integrate; normal distribution)
u
x
Pr (y = 0jx) = 1 (by symmetry),
u

exactly like the probit model. In contrast, if y > 0 then it is continuous:

y = x + u:

It follows that the conditional density of y is equal to

1[y(i)>0]
1[y(i)=0] yi xi
f (yjx; ; u ) = [1 (xi = u )] ;
u

16
where 1[a] is a dummy variable equal to one if a is true. Thus the contribution of observation i to the

sample log likelihood is

yi xi
ln Li = 1[y(i)=0] ln [1 (xi = u )] + 1[y(i)>0] ln ;
u

and the sample log likelihood is

N
X
ln L ( ; u) = ln Li :
i=1

Estimation is done by means of maximum likelihood.

4.2.1. Interpreting the tobit model

Suppose the model can be written according to the equations (4.1)-(4.2), and suppose we have obtained

estimates of the parameter vector . How do we interpret these parameters?

We see straight away from the latent variable model that j is interpretable as the partial (marginal)

e¤ects of xj on the latent variable y , i.e.

@E (y jx)
= j;
@xj

if xj is a continuous variable, and

E (y jxj = 1) E (y jxj = 0) = j

if xj is a dummy variable (of course if xj enters the model nonlinearly these expressions need to be

modi…ed accordingly). I have omitted i-subscripts for simplicity. If that’s what we want to know, then

we are home: all we need is an estimate of the relevant parameter j.

Typically, however, we are interested in the partial e¤ect of xj on the expected actual outcome y;

rather than on the latent variable. Think about the health example above. We are probably primarily

interested in the partial e¤ects of xj (perhaps household size) on expected actual - rather than desired

17
- health expenditure, e.g. @E (yjx) =@xj if xj is continuous. In fact there are two di¤erent potentially

interesting marginal e¤ects, namely

@E (yjx)
; (Unconditional on y)
@xj

and
@E (yjx; y > 0)
: (Conditional on y>0)
@xj

We need to be clear on which of these we are interested in. Now let’s see what these marginal e¤ects

look like.

The marginal e¤ects on expected y, conditional on y positive. We want to derive

@E (yjx; y > 0)
:
@xj

Recall that the model can be written

y = max (y ; 0) ;

y = max (x + u; 0)

(see (4.1)-(4.2)). We begin by writing down E (yjx; y > 0):

E (yjy > 0; x) = E (x + ujy > 0; x) ;

E (yjy > 0; x) = x + E (ujy > 0; x) ;

E (yjy > 0; x) = x + E (uju > x )

Because of the truncation (y is always positive, or, equivalently, u is always larger than x ), dealing

with the second term is not as easy as it may seem. We begin by taking on board the following result for

normally distributed variables:

18
 A useful result. If z follows a normal distribution with mean zero, and variance equal to one (i.e.

a standard normal distribution), then

(c)
E (zjz > c) = ; (4.3)
1 (c)

where c is a constant (i.e. the lower bound here), denotes the standard normal probability density,

and is the standard normal cumulative density.

The residual u is not, in general, standard normal because the variance is not necessarily equal to one,

but by judiciously dividing and multiplying through with its standard deviation u we can transform u

to become standard normal:

E (yjy > 0; x) = x + uE (u= u ju= u > x = u) :

That is, (u= u) is now standard normal, and so we can apply the above ’useful result’, i.e. eq (4.3), and

write:
( x = u)
E (uju > x )= u ;
1 ( x = u)

and thus
( x = u)
E (yjy > 0; x) = x + u :
1 ( x = u)

With slightly cleaner notation,

(x = u)
E (yjy > 0; x) = x + u ;
(x = u)

which is often written as

E (yjy > 0; x) = x + u (x = u) ; (4.4)

19
where the function is de…ned as
(z)
(z) = :
(z)

in general, and known as the inverse Mills ratio function.

Have a look at the inverse Mills ratio function in Section 4 in the appendix, Figure 1.

Equation (4.4) shows that the expected value of y, given that y is not zero, is equal to x plus a

term u (x = u) which is strictly positive (how do we know that?).

We can now obtain the marginal e¤ect:

@E (yjy > 0; x) @ (x = u)
= j + u ;
@xj @xj
0
= j + u j= u ;

0
= j 1+ ;

0
where denotes the partial derivative of with respect to (x = u) (note: I am assuming here that xj

is continuous and not functionally related to any other variable - i.e. it enters the model linearly - this

means I can use calculus, and that I don’t have to worry about higher-order terms). It is tedious but

fairly easy to show that

0
(z) = (z) [z + (z)]

in general, hence
@E (yjy > 0; x)
= j f1 (x = u ) [x = u + (x = u )]g :
@xj

This shows that the partial e¤ect of xj on E (yjy > 0; x) is not determined just by j. In fact, it depends

on all parameters in the model as well as on the values of all explanatory variables x, and the

standard deviation of the residual. The term in f g is often referred to as the adjustment factor, and

it can be shown that this is always larger than zero and smaller than one (why is this useful to know?).

20
 It should be clear that, just as in the case for probits and logits, we need to evaluate the marginal

e¤ects at speci…c values of the explanatory variables. This should come as no surprise, since one of the

reasons we may prefer tobit to OLS is that we have reasons to believe the marginal e¤ects may di¤er

according to how close to the corner (zero) a given observation is (see above). In Stata we can use the mfx

compute command to compute marginal e¤ects without too much e¤ort. How this is done will be clearer

in a moment, but …rst I want to go over the second type of marginal e¤ect that I might be interested in.

The marginal e¤ects on expected y, unconditional on the value of y Recall:

y = max (y ; 0) ;

y = max (x + u; 0) :

I now need to derive

@E (yjx)
:
@xj

We write E (yjx) as follows:

E (yjx) = ( x = u) E (yjy = 0; x) + (x = u) E (yjy > 0; x) ;

= ( x = u) 0+ (x = u) E (yjy > 0; x) ;

= (x = u) E (yjy > 0; x) ;

i.e. the probability that y is positive times the expected value of y given that y is indeed positive. Using

the product rule for di¤erentiation,

@E (yjx) @E (yjy > 0; x) j

= (x = u) + (x = u) E (yjy > 0; x) ;
@xj @xj u

21
and we know from the previous sub-section that

@E (yjy > 0; x)
= j f1 (x = u ) [x = u + (x = u )]g ;
@xj

and

E (yjy > 0; x) = x + u (x = u) :

Hence

@E (yjx)
= (x = u) j f1 (x = u ) [x = u + (x = u )]g
@xj
j
+ (x = u) [x + u (x = u )] ;
u

which looks complicated but the good news is that several of the terms cancel out, so that:

@E (yjx)
= j (x = u)
@xj

(try to prove this). This has a straightforward interpretation: the marginal e¤ect of xj on the expected

value of y, conditional on the vector x, is simply the parameter j times the probability that y is larger

than zero. Of course, this probability is smaller than one, so it follows immediately that the marginal

e¤ect is strictly smaller than the parameter j.

Now consider the example in section 4 in the appendix, on investment in plant and machinery among

Ghanaian manufacturing …rms.

22
PhD Programme: Econometrics II
Department of Economics, University of Gothenburg
Appendix: Ordered & Multinomial Outcomes. Tobit regression.
Måns Söderbom
 

1. Ordered probit: Incidence of corruption among Kenyan manufacturing firms

In the following example we consider a model of corruption in the Kenyan manufacturing sector. 1 Our
dataset consists of 155 firms observed in year 2000.

Our basic latent model of corruption is

⎛ profit ⎞
corrupti* = α1 ln K i + α 2 ⎜ ⎟ + si + towni + ei ,
⎝ K ⎠i

where
corrupt = incidence of corruption in the process of getting connected to public services
K = Value of the firm's capital stock
profit = Total profit
s = sector effect (food, wood, textile; metal is the omitted base category)
town = location effect (Nairobi, Mombasa, Nakuru; Eldoret – which is the most remote town – is the
omitted base category)
u = a residual, assumed homoskedastic and normally distributed with variance normalized to one.

Incidence of corruption is not directly observed. Instead we have subjective data, collected through
interviews with the firm's management, on the prevalence of corruption. Specifically, each firm was
asked the following question:

"Do firms like yours typically need to make extra, unofficial payments to get connected to public
services (e.g. electricity, telephone etc)?"

Answers were coded using the following scale:

 
N/A  Always  Usually  Frequently  Sometimes  Seldom   Never 
0  1  2  3  4  5  6 
 
Observations for which the answer is N/A or missing have been deleted from the data. Notice that this
variable, denoted obribe, is ordered so that high values indicate relatively low levels of corruption.

Given the data available, it makes sense to estimate the model using either ordered probit or ordered
logit.

                                                            
1
 These data was collected by a team from the CSAE in 2000 – for details on the survey and the data, see 
Söderbom, Måns “Constraints and Opportunities in Kenyan Manufacturing: Report on the Kenyan 
Manufacturing Enterprise Survey 2000,” 2001, CSAE Report REP/2001‐03. Oxford: Centre for the Study of 
African Economies, Department of Economics, University of Oxford. Available at 
http://www.economics.ox.ac.uk/CSAEadmin/reports. 

1 
 
Summary statistics for these variables are as follows:
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
obribe | 155 3.154839 1.852138 1 6
lk | 155 15.67499 3.197098 7.258711 22.38821
profk | 155 -.3647645 2.449862 -15.73723 11.3445
wood | 155 .2 .4012966 0 1
textile | 155 .2903226 .4553826 0 1
-------------+--------------------------------------------------------
metal | 155 .2516129 .4353465 0 1
nairobi | 155 .5096774 .5015268 0 1
mombasa | 155 .2645161 .442505 0 1
nakuru | 155 .1032258 .3052398 0 1

Table 1. Ordered probit results

. oprobit obribe1 lk profk sec2-sec4 nairobi mombasa nakuru

Iteration 0: log likelihood = -257.79967

Iteration 1: log likelihood = -248.35111
Iteration 2: log likelihood = -248.34599
Iteration 3: log likelihood = -248.34599

Ordered probit estimates Number of obs = 155

LR chi2(8) = 18.91
Prob > chi2 = 0.0154
Log likelihood = -248.34599 Pseudo R2 = 0.0367

------------------------------------------------------------------------------
obribe1 | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lk | -.0809392 .0307831 -2.63 0.009 -.141273 -.0206054
profk | -.0569773 .0377651 -1.51 0.131 -.1309955 .0170409
wood | -.543739 .2698032 -2.02 0.044 -1.072543 -.0149345
textile | .1068028 .2405553 0.44 0.657 -.3646768 .5782825
metal | -.3959804 .251102 -1.58 0.115 -.8881313 .0961706
nairobi | .0740607 .2836262 0.26 0.794 -.4818364 .6299578
mombasa | -.1443718 .3005436 -0.48 0.631 -.7334265 .4446829
nakuru | -.0242636 .3644382 -0.07 0.947 -.7385494 .6900222
-------------+----------------------------------------------------------------
_cut1 | -2.065609 .5583871 (Ancillary parameters)
_cut2 | -1.539941 .5510676
_cut3 | -1.309679 .5479021
_cut4 | -.665663 .543653
_cut5 | -.5036779 .5442082
------------------------------------------------------------------------------

2 
 
Marginal effects:
. mfx compute, predict(outcome(1));

Marginal effects after oprobit

y = Pr(obribe1==1) (predict, outcome(1))
= .26194813
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
lk | .0263548 .01007 2.62 0.009 .006619 .04609 15.675
profk | .0185525 .01232 1.51 0.132 -.005599 .042704 -.364765
sec2*| .1920139 .10092 1.90 0.057 -.005785 .389813 .2
sec3*| -.0342657 .07595 -0.45 0.652 -.183127 .114596 .290323
sec4*| .1359843 .09025 1.51 0.132 -.040909 .312877 .251613
nairobi*| -.0241228 .09275 -0.26 0.795 -.205901 .157656 .509677
mombasa*| .0479847 .10129 0.47 0.636 -.150547 .246517 .264516
nakuru*| .0079487 .12011 0.07 0.947 -.22747 .243368 .103226
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

. mfx compute, predict(outcome(3));

Marginal effects after oprobit

y = Pr(obribe1==3) (predict, outcome(3))
= .09165806
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
lk | -.0000255 .00073 -0.03 0.972 -.001459 .001408 15.675
profk | -.0000179 .00051 -0.03 0.972 -.001027 .000991 -.364765
sec2*| -.0078598 .00883 -0.89 0.374 -.025173 .009453 .2
sec3*| -.0001844 .00132 -0.14 0.889 -.002777 .002408 .290323
sec4*| -.0035893 .00571 -0.63 0.529 -.014771 .007593 .251613
nairobi*| .0000281 .00068 0.04 0.967 -.001302 .001358 .509677
mombasa*| -.0004917 .00236 -0.21 0.835 -.005126 .004143 .264516
nakuru*| -.0000289 .00079 -0.04 0.971 -.00158 .001522 .103226
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

. mfx compute, predict(outcome(6));

Marginal effects after oprobit

y = Pr(obribe1==6) (predict, outcome(6))
= .17759222
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
lk | -.0210592 .00817 -2.58 0.010 -.03708 -.005038 15.675
profk | -.0148246 .0099 -1.50 0.134 -.034227 .004578 -.364765
sec2*| -.1203152 .051 -2.36 0.018 -.220279 -.020351 .2
sec3*| .0283602 .06519 0.44 0.664 -.099409 .156129 .290323
sec4*| -.0936442 .05426 -1.73 0.084 -.199983 .012695 .251613
nairobi*| .0192561 .07374 0.26 0.794 -.12528 .163792 .509677
mombasa*| -.0363774 .07328 -0.50 0.620 -.179994 .107239 .264516
nakuru*| -.0062568 .09313 -0.07 0.946 -.188796 .176283 .103226
------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

Note: The sign of the marginal effects referring to the highest outcome are the
same as the sign of the estimated parameter beta(j), and the sign of the marginal
effects referring to the lowest outcome are the opposite to the sign of the
estimated parameter beta(j). For intermediate outcome categories, the signs of the
marginal effects are ambiguous and often close to zero (e.g. outcome 3 above). Why
is this?
 
   

3 
 
2. Multinomial Logit

In the following example we consider a model of occupational choice within the Kenyan
manufacturing sector (see footnote 1 for a reference for the data). We have data on 950 individuals and
we want to investigate if education, gender and parental background determine occupation.

We distinguish between four classes of jobs:

• management
• administration and supervision
• sales and support staff
• production workers

Sample proportions for these four categories are as follows:

 
. tabulate job

job | Freq. Percent Cum.

------------+-----------------------------------
Prod | 545 57.37 57.37
Manag | 91 9.58 66.95
Admin | 270 28.42 95.37
Support | 44 4.63 100.00
------------+-----------------------------------
Total | 950 100.00

The explanatory variables are  
 
years of education:   educ  
gender:     male  
parental background:   f_prof, m_prof (father/mother professional), f_se, m_se 
(father/mother self‐employed or trader)  
 
Summary statistics for these variables are as follows: 
 
. sum educ male f_prof f_se m_prof m_se;

Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
educ | 950 9.933684 2.86228 0 17
male | 950 .8136842 .3895664 0 1
f_prof | 950 .1347368 .3416221 0 1
f_se | 950 .1231579 .3287915 0 1
m_prof | 950 .0578947 .2336673 0 1
-------------+--------------------------------------------------------
m_se | 950 .1315789 .3382105 0 1

4 
 
A breakdown by occupation is a useful first step to see if there are any broad patterns in the data:  
. tabstat educ male f_prof f_se m_prof m_se, by(job);

Summary statistics: mean

by categories of: job

job | educ male f_prof f_se m_prof m_se

--------+------------------------------------------------------------
Prod | 8.946789 .8825688 .0715596 .1229358 .0348624 .1559633
Manag | 12.82418 .8021978 .3406593 .1208791 .1318681 .0879121
Admin | 10.75926 .7037037 .1814815 .1074074 .0666667 .1
Support | 11.11364 .6590909 .2045455 .2272727 .1363636 .1136364
--------+------------------------------------------------------------
Total | 9.933684 .8136842 .1347368 .1231579 .0578947 .1315789
---------------------------------------------------------------------
 
 
The multinomial logit seems a suitable model for modelling occupational choice with these data 
(notice in particular that there is no natural ordering of the dependent variable).  
 
I begin by coding the job variable from 0 to 3:  
 
job: 0 = prod; 1 = manag; 2 = admin; 3 = support 
 
So I will obtain three vectors of parameter estimates. Because I have set job = 0 for production 
workers, this will be the base category (I can alter this by using the basecategory( ) option). 
 

5 
 
Results: 
 
. mlogit job educ male f_prof f_se m_prof m_se;

Multinomial logistic regression Number of obs = 950

LR chi2(18) = 289.97
Prob > chi2 = 0.0000
Log likelihood = -846.16161 Pseudo R2 = 0.1463

------------------------------------------------------------------------------
job | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Manag |
educ | .738846 .0755869 9.77 0.000 .5906984 .8869935
male | .0277387 .3383262 0.08 0.935 -.6353685 .690846
f_prof | 1.135737 .3373116 3.37 0.001 .4746187 1.796856
f_se | .1189543 .4074929 0.29 0.770 -.679717 .9176256
m_prof | .3806786 .4661837 0.82 0.414 -.5330247 1.294382
m_se | -.6073577 .4413568 -1.38 0.169 -1.472401 .2576856
_cons | -10.25324 .9913425 -10.34 0.000 -12.19623 -8.310244
-------------+----------------------------------------------------------------
Admin |
educ | .2421636 .0333887 7.25 0.000 .1767229 .3076042
male | -.9075081 .2018354 -4.50 0.000 -1.303098 -.511918
f_prof | .5696015 .2570499 2.22 0.027 .065793 1.07341
f_se | -.0884656 .2616688 -0.34 0.735 -.601327 .4243958
m_prof | -.0135092 .3751632 -0.04 0.971 -.7488156 .7217972
m_se | -.5700617 .256966 -2.22 0.027 -1.073706 -.0664175
_cons | -2.350944 .3941898 -5.96 0.000 -3.123542 -1.578346
-------------+----------------------------------------------------------------
Support |
educ | .2805316 .0723475 3.88 0.000 .1387331 .4223302
male | -.9905816 .3642871 -2.72 0.007 -1.704571 -.276592
f_prof | .6547286 .4707312 1.39 0.164 -.2678877 1.577345
f_se | .8717071 .4237441 2.06 0.040 .0411839 1.70223
m_prof | .7996763 .5500412 1.45 0.146 -.2783846 1.877737
m_se | -.5924061 .5213599 -1.14 0.256 -1.614253 .4294405
_cons | -4.777905 .8675103 -5.51 0.000 -6.478193 -3.077616
------------------------------------------------------------------------------
(Outcome job==Prod is the comparison group)

Marginal effects

. mfx compute, predict(outcome(1)) nose;

Marginal effects after mlogit

y = Pr(job==1) (predict, outcome(1))
= .03792548
-------------------------------------------------------------------------------
variable | dy/dx X
---------------------------------+---------------------------------------------
educ | .0236809 9.93368
male*| .0136163 .813684
f_prof*| .0448291 .134737
f_se*| .0033605 .123158
m_prof*| .0139252 .057895
m_se*| -.0134499 .131579
-------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

6 
 
. mfx compute, predict(outcome(2)) nose;

Marginal effects after mlogit

y = Pr(job==2) (predict, outcome(2))
= .30591218
-------------------------------------------------------------------------------
variable | dy/dx X
---------------------------------+---------------------------------------------
educ | .0390723 9.93368
male*| -.1879454 .813684
f_prof*| .0950757 .134737
f_se*| -.0354302 .123158
m_prof*| -.0225145 .057895
m_se*| -.0999013 .131579
-------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

. mfx compute, predict(outcome(3)) nose;

Marginal effects after mlogit

y = Pr(job==3) (predict, outcome(3))
= .04398022
-------------------------------------------------------------------------------
variable | dy/dx X
---------------------------------+---------------------------------------------
educ | .0073048 9.93368
male*| -.0322613 .813684
f_prof*| .0184086 .134737
f_se*| .0519705 .123158
m_prof*| .0458739 .057895
m_se*| -.015106 .131579
-------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

. mfx compute, predict(outcome(0)) nose;

Marginal effects after mlogit

y = Pr(job==0) (predict, outcome(0))
= .61218213
-------------------------------------------------------------------------------
variable | dy/dx X
---------------------------------+---------------------------------------------
educ | -.0700579 9.93368
male*| .2065904 .813684
f_prof*| -.1583134 .134737
f_se*| -.0199008 .123158
m_prof*| -.0372846 .057895
m_se*| .1284572 .131579
-------------------------------------------------------------------------------
(*) dy/dx is for discrete change of dummy variable from 0 to 1

7 
 
Predicted job probabilities

Education = PRIMARY, SECONDARY, and UNIVERSITY

1. PRIMARY
. list pp1 pp2 pp3 pp0;

+-------------------------------------------+
| pp1 pp2 pp3 pp0 |
|-------------------------------------------|
1. | .0108399 .2284537 .0304956 .7302108 |
+-------------------------------------------+

2. SECONDARY
. list sp1 sp2 sp3 sp0;

+-------------------------------------------+
| sp1 sp2 sp3 sp0 |
|-------------------------------------------|
1. | .1274372 .3683361 .0573239 .4469028 |
+-------------------------------------------+

3. UNIVERSITY

. list up1 up2 up3 up0;

+-----------------------------------------+
| up1 up2 up3 up0 |
|-----------------------------------------|
1. | .6057396 .240109 .0435665 .110585 |
+-----------------------------------------+

Note: 1 = manag; 2 = admin; 3 = support; 0 = prod

How these probabilities were calculated:

/* first collapse the data: this gives a new data set consisting of one
observations and the sample means of the variables */

. collapse educ male f_prof f_se m_prof m_se;

/* now vary education: since 1985 the Kenyan education system has involved 8 years
for primary education, 4 years for secondary, and 4 years for university */

/* first do primary */
. replace educ = 8;
(1 real change made)

/* get the predicted probability that the ‘individual’ is a manager */

> predict pp1, outcome(1);
(option p assumed; predicted probability)

/* get the predicted probability that the ‘individual’ is admin */

. predict pp2, outcome(2);
(option p assumed; predicted probability)

. predict pp3, outcome(3);

(option p assumed; predicted probability)

. predict pp0, outcome(0);

(option p assumed; predicted probability)

/* now do secondary */
. replace educ=12;
(1 real change made)

8 
 
>
> predict sp1, outcome(1);
(option p assumed; predicted probability)

. predict sp2, outcome(2);

(option p assumed; predicted probability)

. predict sp3, outcome(3);

(option p assumed; predicted probability)

. predict sp0, outcome(0);

(option p assumed; predicted probability)

/* finally do university */
. replace educ=16;
(1 real change made)

> predict up1, outcome(1);

(option p assumed; predicted probability)

. predict up2, outcome(2);

(option p assumed; predicted probability)

. predict up3, outcome(3);

(option p assumed; predicted probability)

. predict up0, outcome(0);

(option p assumed; predicted probability)

9 
 
3. Illustration: The Hausman test for IIA in multinomial logit is totally useless

. use http://www.stata-press.com/data/r9/sysdsn3, clear

(Health insurance data)

.
. /* The results shown in the manual */
. mlogit insure male age

Iteration 0: log likelihood = -555.85446

Iteration 1: log likelihood = -551.32973
Iteration 2: log likelihood = -551.32802

Multinomial logistic regression Number of obs = 615

LR chi2(4) = 9.05
Prob > chi2 = 0.0598
Log likelihood = -551.32802 Pseudo R2 = 0.0081

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
male | .5095747 .1977893 2.58 0.010 .1219148 .8972345
age | -.0100251 .0060181 -1.67 0.096 -.0218204 .0017702
_cons | .2633838 .2787574 0.94 0.345 -.2829708 .8097383
-------------+----------------------------------------------------------------
Uninsure |
male | .4748547 .3618446 1.31 0.189 -.2343477 1.184057
age | -.0051925 .0113821 -0.46 0.648 -.027501 .017116
_cons | -1.756843 .5309591 -3.31 0.001 -2.797504 -.7161824
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

. estimates store allcats

.
. mlogit insure male age if insure !="Uninsure":insure

Iteration 0: log likelihood = -394.8693

Iteration 1: log likelihood = -390.4871
Iteration 2: log likelihood = -390.48643

Multinomial logistic regression Number of obs = 570

LR chi2(2) = 8.77
Prob > chi2 = 0.0125
Log likelihood = -390.48643 Pseudo R2 = 0.0111

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
male | .5144003 .1981735 2.60 0.009 .1259875 .9028132
age | -.0101521 .0060049 -1.69 0.091 -.0219214 .0016173
_cons | .2678043 .2775562 0.96 0.335 -.2761959 .8118046
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

10 
 
. hausman . allcats, alleqs constant

---- Coefficients ----

| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| . allcats Difference S.E.
-------------+----------------------------------------------------------------
male | .5144003 .5095747 .0048256 .012334
age | -.0101521 -.0100251 -.0001269 .
_cons | .2678043 .2633838 .0044205 .
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from mlogit
B = inconsistent under Ha, efficient under Ho; obtained from mlogit

Test: Ho: difference in coefficients not systematic

chi2(3) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 0.08
Prob>chi2 = 0.9944
(V_b-V_B is not positive definite)

.
. /* confirm that IIA test is nonsense in model with male dummy only */
. mlogit insure male

Iteration 0: log likelihood = -556.59502

Iteration 1: log likelihood = -553.40794
Iteration 2: log likelihood = -553.40712

Multinomial logistic regression Number of obs = 616

LR chi2(2) = 6.38
Prob > chi2 = 0.0413
Log likelihood = -553.40712 Pseudo R2 = 0.0057

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
male | .477311 .1959282 2.44 0.015 .0932987 .8613234
_cons | -.1772065 .0968274 -1.83 0.067 -.3669847 .0125718
-------------+----------------------------------------------------------------
Uninsure |
male | .46019 .3593228 1.28 0.200 -.2440698 1.16445
_cons | -1.989585 .1884768 -10.56 0.000 -2.358993 -1.620177
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

. estimates store allcats

.
. mlogit insure male if insure !="Uninsure":insure

Iteration 0: log likelihood = -395.53394

Iteration 1: log likelihood = -392.53619
Iteration 2: log likelihood = -392.53611

Multinomial logistic regression Number of obs = 571

LR chi2(1) = 6.00
Prob > chi2 = 0.0143
Log likelihood = -392.53611 Pseudo R2 = 0.0076

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
male | .477311 .1959283 2.44 0.015 .0932987 .8613234
_cons | -.1772065 .0968274 -1.83 0.067 -.3669847 .0125718
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

11 
 
.
. hausman . allcats, alleqs constant

---- Coefficients ----

| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| . allcats Difference S.E.
-------------+----------------------------------------------------------------
male | .477311 .477311 2.63e-13 .000109
_cons | -.1772065 -.1772065 -3.66e-15 .
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from mlogit
B = inconsistent under Ha, efficient under Ho; obtained from mlogit

Test: Ho: difference in coefficients not systematic

chi2(2) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 0.00
Prob>chi2 = 1.0000
(V_b-V_B is not positive definite)

.
. /* confirm that IIA test is nonsense in model with constant only */
. mlogit insure

Iteration 0: log likelihood = -556.59502

Multinomial logistic regression Number of obs = 616

LR chi2(0) = 0.00
Prob > chi2 = .
Log likelihood = -556.59502 Pseudo R2 = 0.0000

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
_cons | -.0595623 .0837345 -0.71 0.477 -.2236789 .1045544
-------------+----------------------------------------------------------------
Uninsure |
_cons | -1.876917 .1600737 -11.73 0.000 -2.190656 -1.563179
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

. estimates store allcats

.
. mlogit insure if insure !="Uninsure":insure

Iteration 0: log likelihood = -395.53394

Multinomial logistic regression Number of obs = 571

LR chi2(0) = 0.00
Prob > chi2 = .
Log likelihood = -395.53394 Pseudo R2 = 0.0000

------------------------------------------------------------------------------
insure | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Prepaid |
_cons | -.0595623 .0837345 -0.71 0.477 -.2236789 .1045544
------------------------------------------------------------------------------
(insure==Indemnity is the base outcome)

12 
 
. hausman . allcats, alleqs constant

---- Coefficients ----

| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| . allcats Difference S.E.
-------------+----------------------------------------------------------------
_cons | -.0595623 -.0595623 7.69e-15 .
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from mlogit
B = inconsistent under Ha, efficient under Ho; obtained from mlogit

Test: Ho: difference in coefficients not systematic

chi2(1) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= -0.00 chi2<0 ==> model fitted on these
data fails to meet the asymptotic
assumptions of the Hausman test;
see suest for a generalized test

13 
 
4. Tobit

Figure 1. The inverse Mills ratio function

4 3
Inverse Mills Ratio
2 1
0

-4 -2 0 2 4
z

Illustration: Modelling investment among Ghanaian manufacturing firms

In the following example we consider a model of company investment within the Ghanaian
manufacturing sector. 2 Our dataset consists of 1,202 observations on firms over the 1991-99 period (in
fact, there is a panel dimension in the data, but we will ignore this for now).

Our simple model of investment is

⎛I ⎞
⎜ ⎟ = max{0, α 0 + α1 ln TFPit + α 2 ln K i ,t −1 + uit },
⎝ K ⎠ it

where
I = Gross investment in fixed capital (plant & machinery)
K = Value of the capital stock
TFP = Total factor productivity, defined as ln(output) – 0.3ln(K) – 0.7ln(L), where L is
employment
u = a residual, assumed homoskedastic and normally distributed.

There is evidence physical capital is 'irreversible' in African manufacturing, i.e. selling off fixed
capital is difficult due to the lack of a market for second hand capital goods (Bigsten et al., 2005). We
can thus view investment as a corner response variable: investment is bounded below at zero.

Summary statistics for these variables are as follows:

                                                            
2
 This is an extension of the dataset used by Söderbom and Teal (2004). 

14 
 
Table 1. Summary statistics
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
invrate | 1202 .0629597 .1477861 0 1
invdum | 1202 .4550749 .4981849 0 1
tfp | 1202 10.20903 1.108122 5.049412 14.7326
lk_1 | 1202 16.06473 3.104121 9.555573 23.51505

Note: invrate = (I/K); invdum = 1 if invrate>0, = 0 if invrate=0; tfp = ln(TFP);

lk_1 = ln[K(t-1)]

Table 2. OLS results

. reg invrate tfp lk_1;

Source | SS df MS Number of obs = 1202

-------------+------------------------------ F( 2, 1199) = 4.54
Model | .197262981 2 .098631491 Prob > F = 0.0108
Residual | 26.0334412 1199 .021712628 R-squared = 0.0075
-------------+------------------------------ Adj R-squared = 0.0059
Total | 26.2307042 1201 .02184072 Root MSE = .14735

------------------------------------------------------------------------------
invrate | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
tfp | .0114908 .0038443 2.99 0.003 .0039484 .0190331
lk_1 | .0002798 .0013724 0.20 0.838 -.0024127 .0029723
_cons | -.058845 .0440225 -1.34 0.182 -.1452148 .0275248
------------------------------------------------------------------------------

Table 3. Tobit results

. tobit invrate tfp lk_1, ll(0);

Tobit estimates Number of obs = 1202

LR chi2(2) = 44.34
Prob > chi2 = 0.0000
Log likelihood = -398.5866 Pseudo R2 = 0.0527

------------------------------------------------------------------------------
invrate | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
tfp | .0344135 .0077922 4.42 0.000 .0191257 .0497012
lk_1 | .0123672 .0027384 4.52 0.000 .0069947 .0177397
_cons | -.6158372 .0913444 -6.74 0.000 -.7950496 -.4366247
-------------+----------------------------------------------------------------
_se | .2540915 .0083427 (Ancillary parameter)
------------------------------------------------------------------------------

Obs. summary: 655 left-censored observations at invrate<=0

547 uncensored observations

Marginal effects based on tobit

. mfx compute, predict(e(0,.));

Marginal effects after tobit

y = E(invrate|invrate>0) (predict, e(0,.))
= .18058807
------------------------------------------------------------------------------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ] X
---------+--------------------------------------------------------------------
tfp | .0106934 .00241 4.44 0.000 .005969 .015418 10.209
lk_1 | .0038429 .00084 4.56 0.000 .00219 .005496 16.0647
------------------------------------------------------------------------------

15 
 
Figure 2. Predicted investment rates as a function of log TFP
.2 .15
Investment rate
.1 .05

9 9.5 10 10.5 11
log TFP
Est imat ed E(y|y>0,x), tobit Es timat ed E(y|x), t obit
Est imat ed E(y|x), OLS

Note: Evaluated at the sample mean of lk_1.

16 
 
References

Bigsten, Arne, Paul Collier, Stefan Dercon, Marcel Fafchamps, Bernard Gauthier, Jan Willem
Gunning, Remco Oostendorp, Catherine Pattillo, Måns Söderbom, and Francis Teal (2005).
“Adjustment Costs, Irreversibility and Investment Patterns in African Manufacturing,” The B.E.
Journals in Economic Analysis & Policy: Contributions to Economic Analysis & Policy 4:1, Article
12, pp. 1-27.

Söderbom, Måns, and Francis Teal (2004). “Size and Efficiency in African Manufacturing Firms:
Evidence from Firm-Level Panel Data,” Journal of Development Economics 73, pp. 369-394.

17 
 
 18