Interacting Biases, Non-Normal Return Distributions and the Performance of Tests for Long-Horizon Event Studies

Abstract

We report simulations, using actual stock return data, of statistical tests of long-horizon buy-and-hold stock returns. We use benchmark portfolios purged of new-listings and rebalancing biases, and find that many proposed tests are misspecified, due in part to skewness. The use of a single control firm instead of a portfolio still produces misspecification, particularly in large samples. We document an inverse relation between skewness bias and sample size, and also document an overlapping-horizons bias. Both biases worsen as the holding period lengthens. Due to interacting biases, tests can be well-specified in one testing scenario but not another, seeming similar, one. A two groups test applied to winsorized abnormal returns exhibits correct specification and considerable power most often in among the tests simulated.