Broken Effects? How to Reduce False Positives in Panel Regressions

This paper presents Monte Carlo simulations to assess which regression estimators can lead to elevated rates of false positives: results that appear “significant" while they are not. Structuralestimators, such as random effects, logit, and Poisson, are compared to ordinary least squares (OLS), fixed effects, and clustered standard errors. The simulations indicate that structural estimators - which are still widely used in published management papers - can lead to elevated rates of false positives in plausible empirical settings, while OLS is less prone to such errors. Though these results generally point in favor of linear regressions, we also discuss several situations in which OLS and/or fixed effects are less suitable and other estimators should be considered. We conclude the paper with guidance when to use OLS versus other estimators and which caveats to watch out for