🤖 AI Summary
This paper addresses the ambiguity in causal interpretation of two-way fixed effects (TWFE) estimators in multi-period panel data. Methodologically, the authors demonstrate that the TWFE coefficient is numerically equivalent to a weighted average of all pairwise first-difference (FD) estimators across time periods and derive, for the first time, an explicit sample-level weighting decomposition framework. Building on this, they propose a relaxed “change-based” parallel trends assumption—replacing the conventional level-based strong assumption—and develop a directly implementable bias-corrected estimation strategy. The contribution is threefold: (i) it enhances transparency regarding the information sources underlying the TWFE estimate; (ii) it precisely characterizes the heterogeneous treatment effect structure actually identified by TWFE; and (iii) it provides a more robust and interpretable foundation for causal inference in staggered adoption settings.
📝 Abstract
In any multiperiod panel, a two-way fixed effects (TWFE) regression is numerically equivalent to a first-difference (FD) regression that pools all possible between-period gaps. Building on this observation, this paper develops numerical and causal interpretations of the TWFE coefficient. At the sample level, the TWFE coefficient is a weighted average of FD coefficients with different between-period gaps. This decomposition improves transparency by revealing the sources of variation that the TWFE coefficient captures. At the population level, causal interpretation of the TWFE coefficient relies on a common trends assumption for any between-period gap, conditional on changes, not levels, of time-varying covariates. I propose a simple modification to the TWFE approach that naturally eases this assumption.
