- The paper shows that TWFE estimators produce biased dynamic treatment effects when treatment effects are heterogeneous across cohorts.
- It proposes an interaction-weighted estimator that uses cohort shares to isolate period-specific treatment effects.
- Empirical illustrations on hospitalization data demonstrate that the new estimator yields more interpretable and reliable causal estimates.
Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects
The paper under consideration presents a methodological critique and an alternative approach to estimating dynamic treatment effects in econometric event studies. In the context of staggered treatment adoption, where treatment timing varies across units, the authors argue that conventional two-way fixed effects (TWFE) regressions with leads and lags can produce biased estimates if treatment effects are heterogeneous across groups. This is particularly relevant in the estimation of dynamic treatment effects, which are often of interest in empirical research to understand the temporal impact of an intervention.
Key Findings and Contributions
The authors identify a critical issue arising from the contamination of TWFE estimates in the presence of treatment effects heterogeneity. Specifically, they show that the coefficient on a given time relative to treatment can inadvertently capture effects from other periods due to the underlying cohort-specific heterogeneity. This complicates the causal interpretation of regression coefficients on dynamic indicators, especially when heterogeneity is not accounted for in the estimation process.
To address the limitations of TWFE regressions under heterogeneity, the authors propose an alternative estimator that constructs weights based on cohort shares of units exposed to the treatment. This method avoids the contamination problem by ensuring that each coefficient on a relative time indicator consistently reflects the treatment effect for that period relative to a baseline. The estimator leverages a decomposition strategy that provides a clear representation of the weights applied to cohort-specific treatment effects in TWFE regressions, thereby informing researchers about the potential biases inherent in conventional approaches.
Numerical Results and Empirical Illustrations
The authors illustrate the pitfalls of TWFE estimators and the advantages of their proposed method through an empirical application using data on the economic consequences of hospitalization. By focusing on outcomes such as out-of-pocket medical expenditure and labor earnings, they demonstrate that conventional TWFE estimates can deviate significantly from the true weighted average effect due to contamination by lagged or lead effects. Their interaction-weighted estimator, in contrast, provides estimates that are more interpretable as they are convex averages of the underlying cohort-specific effects.
Implications and Future Directions
The practical implications of this research are substantial for empirical economists and other researchers relying on TWFE models to estimate dynamic treatment effects. The findings highlight the necessity for greater scrutiny in interpreting dynamic TWFE estimates, particularly when heterogeneity in treatment effects across cohorts is likely.
Theoretically, this paper contributes to an ongoing dialogue in econometrics concerning the limitations of standard regression techniques under complex treatment assignments. By developing an alternative estimator that mitigates biases associated with heterogeneous treatment effects, the authors pave the way for more robust and reliable estimation techniques.
Looking forward, this work suggests several avenues for further research. One potential direction could involve extending the proposed methodology to account for additional complexities such as simultaneous exogeneity or the incorporation of instrumental variables within the event paper framework. Additionally, the development and dissemination of software packages, as noted by the authors, would facilitate broader adoption and ease of application in empirical studies.
Conclusion
In summary, this paper presents a rigorous critique of TWFE models for dynamic treatment effect estimation and proposes a robust alternative that enhances the reliability of causal inference in event studies. By acknowledging and addressing treatment effects heterogeneity, the authors offer valuable insights and tools that can significantly improve empirical practices across disciplines reliant on econometric evaluation of policy interventions.