Logit, Probit, and Tobit: Limited Dependent Variables

www.realoptionsvaluation.com ROV Technical Papers Series: Volume 23

Theory
The term limited dependent variables describes the situation where the dependent variable
In This Issue contains data that are limited in scope and range, such as binary responses (0 or 1) and
1. Learn about limited
truncated, ordered, or censored data. For instance, given a set of independent variables (e.g.,
dependent variables age, income, education level of credit card or mortgage loan holders), we can model the
probability of default using maximum likelihood estimation (MLE). The response or
2. Using Risk Simulator’s dependent variable Y is binary; that is, it can have only two possible outcomes that we
Maximum Likelihood denote as 1 and 0 (e.g., Y may represent presence/absence of a certain condition,
Models to identify which
variables affect the
defaulted/not defaulted on previous loans, success/failure of some device, answer yes/no
default behavior of on a survey, etc.). We also have a vector of independent variable regressors X, which are
individuals assumed to influence the outcome Y. A typical ordinary least squares regression approach is
invalid because the regression errors are heteroskedastic and non-normal, and the resulting
estimated probability estimates will return nonsensical values of above 1 or below 0. MLE
analysis handles these problems using an iterative optimization routine to maximize a log
likelihood function when the dependent variables are limited.
A logit, or logistic, regression is used for predicting the probability of occurrence of an
event by fitting data to a logistic curve. It is a generalized linear model used for binomial
regression, and like many forms of regression analysis, it makes use of several predictor
“How can we model situations variables that may be either numerical or categorical. MLE applied in a binary multivariate
involving dependent logistic analysis is used to model dependent variables to determine the expected probability
variables?” of success of belonging to a certain group. The estimated coefficients for the logit model are
the logarithmic odds ratios and cannot be interpreted directly as probabilities. A quick
computation is first required and the approach is simple.
The logit model is specified as Estimated Y = LN[Pi/(1 – Pi)] or, conversely, Pi =
EXP(Estimated Y)/(1 + EXP(Estimated Y)), and the coefficients βi are the log odds ratios. So,
taking the antilog or EXP(β i), we obtain the odds ratio of Pi/(1 – Pi). This means that with
an increase in a unit of βi, the log odds ratio increases by this amount. Finally, the rate of
change in the probability is dP/dX = βiPi(1 – Pi). The standard error measures how accurate
the predicted coefficients are, and the t-statistics are the ratios of each predicted coefficient
to its standard error and are used in the typical regression hypothesis test of the significance
of each estimated parameter. To estimate the probability of success of belonging to a certain
group (e.g., predicting if a smoker will develop chest complications given the amount
smoked per year), simply compute the Estimated Y value using the MLE coefficients. For
example, if the model is Y = 1.1 + 0.005 (Cigarettes) then someone smoking 100 packs per
year has an Estimated Y of 1.1 + 0.005(100) = 1.6. Next, compute the inverse antilog of the
odds ratio by doing: EXP(Estimated Y)/[1 + EXP(Estimated Y)] = EXP(1.6)/(1 + EXP(1.6))
= 0.8320. Such a person has an 83.20% chance of developing some chest complications in
his or her lifetime.
A probit model (sometimes also known as a normit model) is a popular alternative
specification for a binary response model. It employs a probit function estimated using
maximum likelihood estimation and the approach is called probit regression. The probit and
logistic regression models tend to produce very similar predictions where the parameter
estimates in a logistic regression tend to be 1.6 to 1.8 times higher than they are in a
Contact Us corresponding probit model. The choice of using a probit or logit is entirely up to
Real Options Valuation, Inc. convenience, and the main distinction is that the logistic distribution has a higher kurtosis
4101F Dublin Blvd., Ste. 425, (fatter tails) to account for extreme values. For example, suppose that house ownership is
Dublin, California 94568 U.S.A. the decision to be modeled, and this response variable is binary (home purchase or no home
admin@realoptionsvaluation.com
purchase) and depends on a series of independent variables Xi such as income, age, and so
www.realoptionsvaluation.com forth, such that Ii = β0 + β1X1 + ...+ βnXn, where the larger the value of Ii, the higher the
www.rovusa.com probability of home ownership. For each family, a critical I* threshold exists, where if it is
 exceeded, the house is purchased; otherwise, no home is purchased, and the outcome probability (P) is assumed to be
normally distributed, such that Pi = CDF(I) using a standard normal cumulative distribution function (CDF). Therefore,
use the estimated coefficients exactly like those of a regression model and using the Estimated Y value, apply a standard
normal distribution (you can use Excel’s NORMSDIST function or Risk Simulator's Distributional Analysis tool by
selecting Normal distribution and setting the mean to be 0 and standard deviation to be 1). Finally, to obtain a probit or
probability unit measure, set Ii + 5 (because whenever the probability Pi < 0.5, the estimated Ii is negative, as a result of the
normal distribution being symmetrical around a mean of zero).
The tobit model (censored tobit) is an econometric and biometric modeling method used to describe the relationship
between a non-negative dependent variable Yi and one or more independent variables Xi. A tobit model is an econometric
model in which the dependent variable is censored; that is, the dependent variable is censored because values below zero
are not observed. The tobit model assumes that there is a latent unobservable variable Y*. This variable is linearly
dependent on the Xi variables via a vector of βi coefficients that determine their interrelationships. In addition, there is a
normally distributed error term Ui to capture random influences on this relationship. The observable variable Yi is defined
to be equal to the latent variables whenever the latent variables are above zero and Yi is assumed to be zero otherwise.
That is, Yi = Y* if Y* > 0 and Yi = 0 if Y* = 0. If the relationship parameter βi is estimated by using ordinary least squares
regression of the observed Yi on Xi, the resulting regression estimators are inconsistent and yield downward-biased slope
coefficients and an upward-biased intercept. Only MLE would be consistent for a tobit model. In the tobit model, there is
an ancillary statistic called sigma, which is equivalent to the standard error of estimate in a standard ordinary least squares
regression, and the estimated coefficients are used the same way as a regression analysis.

Procedure
 Start Excel and open the example file Advanced Forecasting Model, go to the MLE worksheet, select the dataset
including the headers, and click on Risk Simulator | Forecasting | Maximum Likelihood.
 Select the dependent variable from the drop-down list (see Figure 1) and click OK to run the model and report.

Figure 1. Logit, Probit, Tobit

© Copyright 2005-2012. All rights reserved. www.realoptionsvaluation.com Page | 2