Logit models analyze the relationship between a dichotomous dependent variable and some predictor variables. They are presented from the interactions of the dichotomous parole outcome with the four dichotomous predictor variable of the type of commitment offense, age, prior record, and alcohol or drug dependency. When the predictors are categorical variables, the maximum likelihood estimates for the parameters of these models and the tests for goodness of fit can be calculated by fitting equivalent log linear models. A six-step general stepwise procedure for fitting linear logit models with categorical predictors via equivalent log linear models provides for detecting nonrelevant predictor variables and collapsing individual data tables over them. The procedure applied to the van Alstyne and Gottfredson study on parole outcomes demonstrates that the previously reported discrepancies between the models emerging from their construction and validation samples can be removed by collapsing the table over the nonrelevant variable of age. The results emphasize the importance of properly screening the set of possible predictor variables and detecting the relevant ones. Tables and four references are provided.