Logistic or linear? Estimating causal effects of experimental treatments on binary outcomes using regression analysis.

When the outcome is binary, psychologists often use nonlinear modeling strategies such as logit or probit. These are neither optimal nor justified when objective to estimate causal effects of experimental treatments. Researchers need take extra steps convert and probit coefficients into interpretable quantities, they do, these quantities remain difficult understand. Odds ratios, for instance, described obscure in many textbooks (e.g., Gelman & Hill, 2006, p. 83). I draw on econometric theory established statistical findings demonstrate that linear regression generally best strategy treatments binary outcomes. Linear directly terms probabilities and, interaction fixed included, safer. review Neyman-Rubin model, which prove analytically yields unbiased estimates treatment Then, run simulations analyze existing data 24,191 students from 56 middle schools (Paluck, Shepherd, Aronow, 2013) illustrate effectiveness regression. Based grounds, recommend (PsycInfo Database Record (c) 2021 APA, all rights reserved).