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

Abstract

This paper assesses different approaches to testing the cross-sectional valuation effects of an event for firms with different characteristics. Standard cross-sectional return regressions typically reject at the 1% significance level more than 25% of the time in non-event periods, suggesting that the bar for rejecting in these tests is far too low. Clustering standard errors does little to reduce excess rejection. Using the time-series distribution of cross-sectional OLS coefficients from regressions in pre-event windows to conduct inference addresses the excess rejection problem but typically results in low-power tests. We propose an alternative approach using a time-series of cross-sectional GLS regressions, using principal component analysis of pre-event returns to estimate the covariance matrix, and show that this approach offers substantial improvements in power.

Keywords: event studies, inference, standard errors, clustering