Probit, Logit and Tobit

Models
S.Madheswaran
Institute for Social and Economic Change
Bangalore

1
Logit and Probit Models
 Another criticism of the linear probability model is
that the model assumes that the probability that Yi =
1 is linearly related to the explanatory variables
 However, the relation may be nonlinear
 For example, increasing the income of the very poor or the very
rich will probably have little effect on whether they buy an
automobile, but it could have a nonzero effect on other income
groups
 Logit and probit models are nonlinear and provide
predicted probabilities between 0 and 1

2
Logit and Probit Models

3
Logit and Probit Models
 Suppose our underlying dummy dependent
variable depends on an unobserved utility
index, Y*
 If Y is discrete—taking on the values 0 or 1 if
someone buys a car, for instance
 Can imagine a continuous variable Y* that reflects
a person’s desire to buy the car
 Y* would vary continuously with some explanatory
variable like income
4
Logit and Probit Models
 Written formally as

 If the utility index is “high enough,” a person

will buy a car

 If the utility index is not “high enough,” a

person will not buy a car

5
Logit and Probit Models

 The basic problem is selecting F—the cumulative density

function for the error term
 This is where where the two models differ
6
Logit and Probit Models
 Interested in estimating the β’s in the model
 Typically done using a maximum likelihood
estimator (MLE)
 Each outcome Yi has the density function ƒ(Yi)
= PiYi (1 − Pi)1 − Yi
 Each Yi takes on either the value of 0 or 1 with
probability ƒ(0) = (1 − Pi) and ƒ(1) = Pi

7
Logit and Probit Models
 The likelihood function is

8
Logit Model
 For the logit model we specify

 Prob(Yi = 1) → 0 as β0 + β1X1i → −∞
 Prob(Yi = 1) → 1 as β0 + β1X1i → ∞
 Thus, probabilities from the logit model will be
between 0 and 1

9
Logit Model
 A complication arises in interpreting the estimated
β’s
 With a linear probability model, a β estimate measures the
ceteris paribus effect of a change in the explanatory
variable on the probability Y equals 1
 In the logit model

The derivative is
nonlinear and
depends on the
value of X.

10
Probit Model
 In the probit model, we assume the error in the
utility index model is normally distributed
 εi ~ N(0,σ2)

 Where F is the standard normal cumulative

density function (c.d.f.)

11
Probit Model
 The c.d.f. of the logit and the probit look quite
similar
 Calculating the derivative is moderately complicated

 Where ƒ is the density function of the normal distribution

12
Probit Model
 The derivative is nonlinear
 Often evaluated at the mean of the explanatory variables
 Common to estimate the derivative as the probability
Y = 1 when the dummy variable is 1 minus the
probability Y = 1 when the dummy variable is 0
 Calculate how the predicted probability changes when the
dummy variable switches from 0 to 1

13
Which is Better? Logit or Probit?
 From an empirical standpoint logits and probits typically
yield similar estimates of the relevant derivatives
 Because the cumulative distribution functions for the two models
differ slightly only in the tails of their respective distributions
 The derivatives are different only if there are enough
observations in the tail of the distribution
 While the derivatives are usually similar, the parameter
estimates associated with the two models are not
 Multiplying the logit estimates by 0.625 makes the logit estimates
comparable to the probit estimates

14
Censored Regression Model
 Often the dependent variable is constrained
(or censored)
 Takes on a positive value for some observations
and zero for other observations
 Represents non-continuous data as there is a large
cluster of observations at zero
 Using OLS leads to biased estimates of the
parameters
15
Censored Regression Model
 Examples include data sets containing
information on
 The number of hours people worked last week
along with their age
 Some people will have worked a positive number of
hours
 Others (such as retirees) will not have worked at all
and will report working zero hours
 Families’ expenditures on new automobile
purchases during a particular year
16
Censored Regression Model
 For the probit and logit we defined a latent
variable Y*i = βXi + ui with

 If Yi is not a binary variable but rather is

observed as Y*i if Y*i > 0 and is not observed
for Y*i ≤ 0, then
u is assumed to follow the
normal distribution with
mean 0 and variance σ2. 17
Censored Regression Model
 Called the Tobit model or the censored regression
model
 To estimate this model, specify the likelihood
function for this problem and generate the maximum
likelihood estimator
 The (log) likelihood for the Tobit model is

18
Heckman Two-Step Estimator
 As an alternative to estimation of the Tobit
model using maximum likelihood methods,
James Heckman has developed a two-step
estimation procedure
 Yields consistent estimates of the parameters
 Suppose the model takes the form

19
Heckman Two-Step Estimator
 The mean value of Y (if it is greater than zero)
may be written as

 It can be shown that

 Where
Called the inverse
Mills ratio or
the hazard rate.
20
Heckman Two-Step Estimator
 Regressing the positive values of Yi on Xi
would lead to omitted variable bias
 If we could get an estimate of λ we could run
ordinary least squares on X and λ

21
Heckman Two-Step Estimator
 Heckman proposes
 Defining I as a dummy variable taking on the value 1 for
the positive values of Y and 0 otherwise
 Ii = 1 if Yi > 0; 0 otherwise
 Estimate λ by estimating a probit model of Ii on X
 Since the probit model specifies Prob(Y = 1) = F(βXi), we can
get estimates of β by estimating the probit model
 Can use these estimates to form

 Using the positive values of Y, run OLS on X and the

estimated λ—will yield consistent estimates of β 22