Outliers and Robust Inference in Archival Accounting Research

We study the nature and consequences of outliers in archival research, as well as the merits and limitations of robust regression estimators in identifying and downweighting their influence. Using simulated and actual data, we demonstrate how outliers can arise non-randomly from the data-generating process, research design choices such as scaling, and model misspecification. We find that robust regression estimators generate more precise estimates than OLS in common archival data. At the same time, these estimators can bias inferences due to their downweighting of substantial and nonrandom proportions of the data. We further demonstrate that model misspecification (e.g., a failure to account for nonlinear relations) can induce biases in robust regression estimates that are more severe than with OLS. Based on our analyses, we recommend researchers to carefully evaluate the causes and consequences of the nonrandom nature of outliers in their samples, to implement and interpret robust regression estimators with care, and to evaluate and disclose the sensitivity of robust estimators to key design choices.