1 Introduction
Event study plots have become ubiquitous in applied economics. Nearly every paper using a difference-in-differences (DiD) design now includes a figure showing how the treatment effect evolves before and after the policy change. The pre-period coefficients are interpreted as a falsification test: flat pre-trends provide evidence that the parallel trends assumption holds. The post-period coefficients are interpreted as the dynamic causal path of the treatment.
But how reliable are these interpretations? A growing literature argues that event study coefficients are more fragile than their apparent precision suggests. Pre-trend coefficients may be uninformative about the most relevant violations of parallel trends [Roth, 2022]. Post-treatment coefficients in staggered adoption settings can be contaminated by heterogeneous treatment effects in ways that are not captured by the standard TWFE event study specification [Sun and Abraham, 2021]. Anticipation effects, compositional changes in the sample, and researcher choices about normalisation and binning can all distort the event study picture.
This debate matters because event study plots are now routinely used not just for validation, but to make substantive policy claims about effect timing, persistence, and fade-out interpretations that may not be warranted by the underlying identification.
2 The Case for Event Studies as Valid Causal Dynamic Estimates
Proponents of event study designs argue that, when properly implemented, they provide a powerful and uniquely transparent window into causal dynamics.
2.1 Testing and Demonstrating Parallel Trends
The standard event study specification in the TWFE framework is:
where Ti* is the treatment adoption period, k is event time (periods before/after treatment), and k = -1 is the omitted reference period. The βk for k < 0 are the pre-treatment coefficients that should be near zero if parallel trends holds; the βk for k > 0 are the dynamic treatment effects.
Proponents argue that a flat pre-trend pattern visually verified and tested with joint significance tests provides credible evidence that the identifying assumption is satisfied. Coupled with the event study's ability to show whether effects are immediate or delayed, growing or fading, the design provides much richer evidence than a single post-treatment ATT.
2.2 Dynamic Effects Are Economically Important
For many policy questions, the time path of effects matters as much as the contemporaneous effect. Does a minimum wage increase affect employment immediately, or with a lag as firms adjust? Does an environmental regulation reduce pollution continuously or only when enforcement is credible? Do job training effects persist or fade out? Event studies are uniquely equipped to answer these questions, and proponents argue that alternatives (e.g., a simple post-treatment dummy) would sacrifice important information [Freyaldenhoven et al., 2019].
3 The Sceptical View: Why Event Study Coefficients Can Mislead
3.1 Pre-Test Power and the Roth Critique
Roth [2022] identifies a fundamental problem with using pre-trend tests to validate DiD designs. Standard pre-trend tests have low power against economically meaningful violations of parallel trends: they can fail to reject the null of no pre-trend even when trends are diverging in ways that would substantially bias the treatment effect estimate.
The reason is statistical: pre-trend coefficients are noisy, especially when the number of pre-treatment periods is small relative to the noise in the outcome. A pre-trend slope of 0.3 (in the units of the estimated treatment effect) is consistent with "flat" pre-trends at conventional significance levels if standard errors are large. Yet a pre-trend of 0.3 per period continuing into the post-treatment window would dramatically inflate the treatment effect estimate.
The implication, formalised by Rambachan and Roth [2023], is that pre-trend tests should be supplemented by sensitivity analyses that ask: "How much would the result change if the true pre-trend of the counterfactual continued at the same rate?" or "at the maximum rate consistent with the confidence interval?"
3.2 Staggered Adoption: TWFE Event Studies Are Contaminated
The standard TWFE event study was designed for a single treatment date shared by all treated units. In the ubiquitous staggered adoption setting where different units are treated at different times the TWFE event study suffers from the same negative-weighting problems as the static TWFE estimator [de Chaisemartin and D'Haultfœuille, 2020, Goodman-Bacon, 2021].
Sun and Abraham [2021] show that the TWFE event study coefficient β̂k in staggered settings is a weighted average of cohort-specific event study coefficients, where the weights can be negative for some cohorts. A negative-weighted average of positive treatment effects can be negative or zero, even if the true effect is positive in every cohort at every event time.
de Chaisemartin and D'Haultfœuille [2024] document that different software implementations of event study plots (Stata's eventstudy2, R's fixest iplot, and the did package) can produce substantially different-looking event study figures for the same data, because they differ in normalisation, binning of distant event times, and how they handle the contamination from heterogeneous adoption timing.
3.3 Compositional Effects in Post-Treatment Periods
Another threat arises from compositional changes: the sample of units that have been treated for k periods changes with k. At k = 1, all treated units are one period post-treatment. At k = 5, only units treated early enough to have 5 post-treatment periods are in the k = 5 cell.
If early adopters differ systematically from late adopters in their outcomes or treatment effects, the pattern of event study coefficients across k reflects a mixture of the true time path and the changing sample composition.
Freyaldenhoven et al. [2019] propose controlling for leads of the instrument (in a IV-event study context) to absorb anticipation effects that would contaminate pre-treatment coefficients. But compositional effects in the post-treatment window have received less attention and remain a concern.
3.4 Anticipation Effects
When treated units anticipate the treatment and adjust their behaviour in advance, the pre-treatment coefficients will not be zero even if parallel trends holds for the no-anticipation potential outcomes. Researchers typically address this by defining the treatment as the announcement (rather than implementation) or by incorporating the anticipation window into the specification. But distinguishing anticipation from a violation of parallel trends is not always possible from data alone.
4 Better Alternatives and Repairs
4.1 Cohort-Specific Event Studies
The Callaway and Sant'Anna [2021] and Sun and Abraham [2021] frameworks estimate cohort-specific treatment effects ATT(g,t) and aggregate them into event-study-style plots without the contamination from heterogeneous adoption timing. The did R package (Callaway-Sant'Anna) and fixest (sunab() function for Sun-Abraham) implement these.
4.2 Honest Sensitivity Analysis
Rambachan and Roth [2023] provide a formal framework for conducting sensitivity analysis on the parallel trends assumption. Rather than using the pre-trend test as a binary pass/fail, they compute the identified set for the treatment effect under restrictions on how much the post-treatment counterfactual trend can deviate from the pre-treatment trend. The HonestDiD R package implements this.
4.3 Freyaldenhoven-Hansen-Shapiro Pre-Controls
When a "contamination" covariate Pit is available one that is affected by the anticipation of treatment but not by the treatment itself Freyaldenhoven et al. [2019] propose including leads and lags of Pit as controls to absorb the pre-trends driven by anticipation, recovering cleaner pre- and post-treatment event study coefficients.
5 Where the Evidence Points
The weight of evidence supports a nuanced position:
Event studies are indispensable. The combination of pre-trend testing and dynamic post-treatment analysis provides evidence that no static DiD estimate can offer. Removing event studies from the toolkit would throw away valuable information.
Standard TWFE event studies in staggered settings should be replaced. The contamination from heterogeneous adoption timing is real, well-documented, and easy to address with modern estimators. There is no good reason to use contaminated TWFE event studies in 2026.
Pre-trend tests should be supplemented, not replaced. Visual pre-trend inspection and joint significance tests are useful, but should be paired with formal sensitivity analysis (Rambachan-Roth) that quantifies how robust the conclusion is to pre-trend violations.
Compositional changes and anticipation need case-by-case attention. There is no universal solution; researchers must think carefully about the institutional context to determine whether these threats are relevant.
6 Conclusion
Event study designs are among the most powerful tools in applied econometrics, but their apparent precision can be deceptive. Pre-trend coefficients are often underpowered against economically meaningful violations of parallel trends; post-treatment coefficients in staggered settings can reflect contamination from heterogeneous timing rather than true causal dynamics; and compositional changes and anticipation effects can distort the picture further. None of these problems is fatal each has known solutions but the casual consumer of event study figures should be aware that flat pre-trends and smooth post-treatment trajectories are not sufficient for causal inference. The toolkit for valid dynamic DiD is richer and more demanding than a standard TWFE event study, and the best empirical work in 2026 uses it fully.
References
- Callaway, B. and Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2):200-230.
- de Chaisemartin, C. and D'Haultfœuille, X. (2020). Two-way fixed effects estimators with heterogeneous treatment effects. American Economic Review, 110(9):2964-2996.
- de Chaisemartin, C. and D'Haultfœuille, X. (2024). Difference-in-differences estimators of intertemporal treatment effects. Review of Economics and Statistics, forthcoming.
- Freyaldenhoven, S., Hansen, C., and Shapiro, J. M. (2019). Pre-event trends in the panel event-study design. American Economic Review, 109(9):3307-3338.
- Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. Journal of Econometrics, 225(2):254-277.
- Rambachan, A. and Roth, J. (2023). A more credible approach to parallel trends. Review of Economic Studies, 90(5):2555-2591.
- Roth, J. (2022). Pretest with caution: Event-study estimates after testing for parallel trends. American Economic Review: Insights, 4(3):305-322.
- Sun, L. and Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics, 225(2):175-199.