# Statistical Advisor, Probit/Logit Regression for Categorical Dependent Variable

Use GENERALIZED LINEAR MODEL (GLZ) or NONLINEAR ESTIMATION. These chapters discuss techniques to perform logit and probit regression, using a maximum likelihood criterion. The GLZ chapter also discusses techniques for stepwise and best subset regression, for continuous and/or categorical predictors (ANCOVA-like designs).

Logit and probit regression are used for analyzing the relationship between one or more independent (predictor) variables with a categorical dependent (criterion) variable at two levels. For example, one may want to predict success or failure in a test from various continuous variables, such as length preparation time for the test. Here, success vs. failure represents a categorical dependent variable at two levels while preparation time would be a continuous independent variable.

There are numerous advantages of logit/probit regression over linear multiple regression (see the manual). In most general terms, these regression methods imply that the dependent variable is actually the result of a transformation of an underlying variable, which is not restricted in range. For example, the probit model assumes that the actual underlying dependent variable is measured in terms of z-values for the normal curve; if one transforms those values to probabilities (under the normal curve) then the predictions for the dependent variable will always fall between 0 and 1. Thus, we are actually predicting probabilities from the independent variables.