Logit Regression: Best Model Selection

Abstract

The typical method of analyzing categorical variables is to use the chi-square statistic. However, with more than two categorical variables, simultaneous examination of main and interaction effects is not feasible. The logit regression technique permits analysis of categorical variables, the modeling of main and interaction effects, control of Type I error, and distribution freer assumptions. This study investigated parsimonious model fit related to the selection of the best set of categorical predictor variables. Findings indicated that the various variable selection criteria (L^2, z, log-odds ratio, R^2L, model variance, and L\.C2) provided different results. Order of variable entry also produced significantly different results. The use of a Tabu search procedure and L\.C2 criteria is recommended to determine the best set of categorical independent predictor variables in logit regression.