Online Program

Modeling Curvilinearity, Interactions, and Curvilinear Interactions in Logistic Regression: Having More Fun with Your Data (302759)

Keywords: logistic regression, statistical practice, interaction, curvilinear effect

ANOVA and regression analyses are part of the same general linear model, yet they have historically existed as two separate traditions with different procedural norms (i.e., routinely examining ANOVA analyses for interactions, but not regression). Statistical software seems to promulgate this type of traditional disparity, making examination of interactions default and routine in ANOVA-type analyses, but not in regression. Similarly, software packages rarely test for curvilinear effects without direct (and sometimes difficult) user direction. This is a plea for statisticians to routinely examine their data for interactions, curvilinear effects, and curvilinear interactions. In this presentation, I will use examples from logistic regression, but the points apply to any linear modeling situation. One important assumption we make in reporting results is that the model is appropriately specified and important terms are not left out of the model. Curvilinear effects present but not modeled violate this assumption, as do unspecified interactions. More critically, these type of effects are often the most interesting and fun types for statisticians to explore.