In recent decades, event studies have become widely used in health and social research to estimate the dynamic effects of staggered interventions. In this paper, we develop a framework grounded in experimental design principles for observational studies that encompasses common event study estimators and makes explicit how they borrow information across units and time periods. We consider two routes to identification, unconfoundedness in treatment initiation and parallel trends with limited anticipation, and construct weighted contrasts from outcome levels under the first and within-unit differences under the second. We propose a robust weighting estimator that draws on progressively larger sets of observations as stronger assumptions on treatment assignment and potential outcomes are imposed. We derive exact finite-sample decompositions of dynamic two-way fixed effects and interaction-weighted estimators, and of doubly robust estimators, into weighted contrasts of within-unit outcome differences, revealing the hypothetical experiment each approximates. These decompositions show which unit-period observations enter each contrast and how they are weighted, clarifying and supplementing the notion of forbidden comparisons. The implied weights yield diagnostics for covariate balance, sign reversal, effective sample size, information shares across observation groups, and each observation's influence on the estimate. We illustrate these diagnostics in a study of the effects of Castle Doctrine laws on homicide rates.

翻译：暂无翻译