Reading Event-Study Plots: Beyond the Lines
Event-study plots visualize estimated treatment effects $\theta_k$ across event time $k$. But interpreting these lines requires care—especially in the presence of treatment effect heterogeneity and changing cohort composition.
What the Plot Shows (and What It Hides)
- Each point: An estimated $\theta_k$ (or bin) relative to the omitted reference period (usually $k = -1$).
- Confidence intervals: Reflect statistical uncertainty, but not bias from model misspecification or compositional shifts.
- Pre-trend check: Flat, near-zero pre-treatment coefficients ($k < 0$) support parallel trends, but sparse support or binning can mask violations.
Cohort Composition: The Hidden Axis
At each event time $k$, the set of units contributing to the estimate changes. Early post-treatment periods include all treated cohorts; later periods include only early adopters. If treatment effects differ by cohort, the plot can mislead:
- Apparent dynamics may reflect changing cohort mix, not true effect evolution.
- Support tables/figures: Always report the number of units and cohorts contributing to each $k$; plot effective sample size alongside $\theta_k$.
- Cohort-specific plots: Plot $\theta_{g,k}$ for major cohorts to reveal heterogeneity.
Common Pitfalls and Diagnostic Tools
- Sparse support at tails: Wide confidence intervals and unstable estimates. Use binning and report support.
- Over-interpreting late $k$: Later event times reflect only early adopters—do not generalize.
- Ignoring anticipation or contamination: If $k=0$ is contaminated, interpret $k \geq 1$ with caution.
- Neglecting pre-trend diagnostics: Always check for credible pre-treatment flatness.
Best Practices
- Always accompany event-study plots with support tables/figures.
- Use binning for sparse event times, but be transparent about bin definitions.
- Interpret late post-treatment effects as cohort-specific, not average.
- When possible, estimate and plot cohort-specific effects.
Summary
Event-study plots are powerful, but their interpretation is subtle. Always check support, cohort composition, and pre-trends before drawing causal conclusions. Use diagnostic tools to distinguish genuine effect dynamics from artifacts of the data structure.