Empirical logit analysis is not logistic regression

- We review the differences between empirical logit analysis and logistic regression.
- We show that the two models differ substantially when there are few observations.
- We show that models differ substantially when ceiling or floor effects are present.

Many recent psycholinguistic studies have used empirical logit analysis as a substitute for mixed-effects logistic regression. In this paper, we describe the differences between empirical logit analysis and mixed-effects logistic regression and highlight three interacting sources of bias in empirical logit analysis. We then report on two sets of simulations comparing logistic regression and empirical logit analysis. We show that relative to logistic regression, empirical logit analysis generally yields biased parameter estimates when proportions are close to 0 or 1, especially when the number of observations underlying a proportion is very low. We also show that, in some circumstances, this bias can create spurious interactions, leading to unacceptable Type I error rates. While these two models may provide similar answers to some questions, we encourage readers to interpret empirical logit parameters cautiously.