From Improvised to Standardized: Testing Interaction Effects in Logit, Probit, and Other Nonlinear Models

Abstract

Management theories are routinely tested as moderation hypotheses about whether a firm enters, adopts, or fails, or whether a candidate is hired. In the nonlinear models such outcomes call for, like logit and probit, the coefficient on a product term is not the interaction effect: it can differ in magnitude, significance, and sign. Across 205 articles in five leading journals, only 5% test the interaction effect from a model suited to the outcome; nearly half retreat to a linear probability model. I introduce a standardized remedy: the average interaction effect, computed by ginteff in Stata and R for binary, ordered, multinomial, and fractional models. Reanalyses of three published studies change or sharpen conclusions; simulations show each common practice failing while the average interaction effect holds.