Past is Prologue: Inference from the Cross Section of Returns Around an Event

Abstract

The possibility of contemporaneous events hampers inference about differential effects of a quasi-experimental event across firms. We show that this possibility consistently causes high false positive rates in tests of differences in short-term event-driven stock returns-nearly 50% at the 1% significance level in some cases. Clustering standard errors (e.g., by industry) is an inadequate solution. Researchers should instead use the distribution of pre-event return relationships to test for significance. We introduce a novel GLS-based variant of this testing strategy, show that it increases power substantially over OLS-based variants, and provide a Stata module that implements both.

Keywords: event studies, inference, standard errors, clustering