Interaction-Weighted Event-Study Estimation

Why Interact Treatment Cohort With Event Time?

In a staggered-adoption design, units begin treatment in different calendar periods. Event-study analysis describes how effects evolve with time since treatment, but a single regression coefficient for a given event time can mix information across cohorts in problematic ways. In particular, conventional two-way fixed-effects event-study regressions can use already-treated cohorts as comparisons for later-treated cohorts. If effects differ across cohorts or change over event time, those comparisons can contaminate the reported coefficients.

Interaction-weighted estimation addresses this by estimating effects separately for each treatment cohort at each event time, using untreated comparison units for identification, and then combining the cohort-specific effects with stated weights. The result is both a dynamic summary and a clearly defined average for the cohorts that contribute at each event time.

A common reference period is \(e=-1\), the last period before treatment. Effects are interpreted as differences from that untreated baseline, not as absolute outcome levels. The set \(\mathcal{G}_e\) may change with event time: late event times are often observed only for early-adopting cohorts. Therefore, the weights—and the population represented by the aggregate—can also change across event times.

Identification With Untreated Comparisons

The cohort-specific effect can be identified by comparing the cohort’s outcome change from its baseline period to the relevant event period with the corresponding change among units that remain untreated over that comparison. A comparison group can be never-treated units, or units not yet treated by the event period, provided they are untreated throughout the comparison interval. Already-treated units are not valid untreated controls merely because they belong to a different cohort.

For a simple two-period contrast, let \(Y_{i,t}(0)\) be the outcome unit \(i\) would have in period \(t\) without treatment. The key parallel-trends condition is that, absent treatment, the average change for cohort \(g\) between the baseline \(g-1\) and period \(t\) would equal the average change for its untreated comparison group. Also assume no anticipation: treatment does not affect outcomes before the defined treatment date. These assumptions, along with a well-defined treatment date and available comparison observations, identify the cohort-specific effect. With covariates, the corresponding conditions may be stated conditionally on those covariates.

Here \(C_{g,t}=1\) identifies the chosen untreated comparison units for cohort \(g\) over the interval from \(g-1\) to \(t\). The formula is a difference-in-differences contrast: subtract the comparison group’s change from the treated cohort’s change. Its validity follows directly from parallel trends. Under no anticipation, the cohort’s baseline outcome is untreated; parallel trends supplies the counterfactual change it would have experienced without treatment. Subtracting that change from its observed change leaves the average treatment effect.

Regression With Cohort-by-Event-Time Interactions

An interaction-weighted event study estimates separate coefficients for cohort-by-event-time indicators, rather than imposing one common effect at each event time. A schematic specification with unit and calendar-time fixed effects is

Here \(\alpha_i\) and \(\lambda_t\) are unit and calendar-time fixed effects; \(\mathbf{1}\{\cdot\}\) is an indicator; and the \(e=-1\) cohort indicators are omitted as the reference period. The set \(\mathcal{E}_g\) includes event times supported by the observed data for cohort \(g\). In an implementation, the comparison group and estimation sample must be chosen so the identifying contrasts use never-treated or not-yet-treated units, not already-treated units as untreated controls. The regression is a way to organize the cohort-specific contrasts; it does not replace the assumptions that identify them.

The coefficient \(\beta_{g,e}\) represents the cohort-specific change at event time \(e\), relative to the omitted baseline, under the model and identification conditions. The interaction-weighted estimate is formed after estimating those coefficients:

Weights should be reported or otherwise clearly defined. A common choice is cohort-size weights among the cohorts with support at event time \(e\): \(\widehat{\omega}_{g,e}=N_{g,e}/\sum_{h\in\mathcal{G}_e}N_{h,e}\), where \(N_{g,e}\) is the number of contributing units in cohort \(g\). Equal-cohort weights are another possible target, but they answer a different question. Neither choice is universally correct; the weights specify which average effect is being summarized.

Why Already-Treated Comparisons Contaminate Event Studies

A treated cohort’s outcome after adoption combines its untreated outcome path and its treatment effect. If that cohort is used as a control for another cohort, its change may include a treatment effect that is itself evolving. The resulting contrast need not equal the later cohort’s treatment effect. Cohort-by-event-time estimation prevents this particular problem by first estimating each cohort’s effect against untreated comparisons and only then aggregating those estimates.

Support, Interpretation, and Inference

An event-time estimate should be reported only for cohorts and periods with the necessary comparison data. For example, if no untreated units remain by a late calendar period, the data do not identify the same untreated comparison contrast without additional assumptions or design choices. Likewise, a long-run aggregate may describe only early-adopting cohorts, while a near-term aggregate includes more cohorts. Changes in \(\widehat{\theta}_e\) across \(e\) can therefore reflect both changing effects and changing cohort composition.

Uncertainty for an aggregate must account for uncertainty and covariance among the cohort-specific estimates. If the weights are treated as fixed, its estimated variance is based on the variance of the weighted sum, including covariances between estimates. Inference should also respect the level at which treatment is assigned and outcomes may share shocks; clustering is commonly needed in panel designs. The point estimate’s weighting formula alone does not determine valid standard errors.

Common Mistake

Wrong reasoning: “Two-way fixed effects remove differences between groups, so any already-treated cohort is a suitable comparison group for a later cohort.” Fixed effects do not make a treated outcome untreated. If effects are heterogeneous or evolve over time, the treated comparison cohort can contribute its own treatment-effect change to the contrast.

Correction: Identify each cohort’s event-time effect using never-treated or not-yet-treated units that remain untreated over the comparison interval, under the relevant parallel-trends and no-anticipation assumptions. Then aggregate the cohort-specific estimates with explicit weights. This separates identification from the choice of summary.