Difference-in-Differences with Time-Varying Covariates in the Parallel Trends Assumption
Abstract
In this paper, we study difference-in-differences identification and estimation strategies where the parallel trends assumption holds after conditioning on time-varying covariates and/or time-invariant covariates. Our first main contribution is to point out a number of weaknesses of commonly used two-way fixed effects (TWFE) regressions in this context. In addition to issues related to multiple periods and variation in treatment timing that have been emphasized in the literature, we show that, even in the case with only two time periods, TWFE regressions are not generally robust to (i) paths of untreated potential outcomes depending on the level of time-varying covariates (as opposed to only the change in the covariates over time), (ii) paths of untreated potential outcomes depending on time-invariant covariates, and (iii) violations of linearity conditions for outcomes over time and/or the propensity score. Even in cases where none of the previous three issues hold, we show that TWFE regressions can suffer from negative weighting and weight-reversal issues. Thus, TWFE regressions can deliver misleading estimates of causal effect parameters in a number of empirically relevant cases. Second, we extend these arguments to the case of multiple periods and variation in treatment timing. Third, we provide simple diagnostics for assessing the extent of misspecification bias arising due to TWFE regressions. Finally, we propose alternative (and simple) estimation strategies that can circumvent these issues with two-way fixed regressions.
- Publication:
-
arXiv e-prints
- Pub Date:
- February 2022
- DOI:
- 10.48550/arXiv.2202.02903
- arXiv:
- arXiv:2202.02903
- Bibcode:
- 2022arXiv220202903C
- Keywords:
-
- Economics - Econometrics
- E-Print:
- 47 pages. Substantially expands arguments about interpreting TWFE regressions that include covariates. Results on covariates that could be affected by the treatment are moved to "Difference-in-differences with bad controls" (Caetano, Callaway, Payne, and Rodrigues, 2023)