Primary Submission Category: Difference in Differences, Synthetic Control, Methods for Panel and Longitudinal Data
When can we get away with using the two-way fixed effects regression?
Authors: Apoorva Lal,
Presenting Author: Apoorva Lal*
Two-way fixed effects (TWFE) regression is a fundamental tool in causal inference with panel data, traditionally used to estimate average treatment effects on the treated (ATT). Recent econometric work has highlighted potential biases in TWFE estimates when treatment effects are heterogeneous across time and adoption cohorts, leading to the development of several robust but higher-variance alternatives. This creates a bias-variance tradeoff: while TWFE may yield biased estimates under heterogeneity, alternative estimators often have substantially higher variance and computational complexity, thereby posing challenges in adoption from applied researchers. We develop simple specification tests to diagnose when TWFE is likely to yield reliable estimates versus when more complex estimators are needed. Our approach uses F-tests on cohort-time interactions in a fully saturated specification to detect patterns of effect heterogeneity that may yield problematic contamination bias that biases TWFE. Through simulation studies, we show these tests maintain correct size under homogeneous effects while having good power to detect heterogeneity, which validates their use in specification selection. The tests are computationally simple and readily implementable in standard software. By helping researchers identify cases where TWFE is adequate, our diagnostic framework facilitates principled choices between estimation strategies in practice.