Influential Observations and Inference in Accounting Research

Accounting studies often encounter observations with extreme values that can influence coefficient estimates and inferences. Two widely used approaches to address influential observations in accounting studies are winsorization and truncation. While expedient, both depend on researcher-selected cutoffs, applied on a variable-by-variable basis, which, unfortunately, can alter legitimate data points. We compare the efficacy of winsorization, truncation, influence diagnostics (Cook's Distance), and robust regression at identifying influential observations. Replication of three published accounting studies shows that the choice impacts estimates and inferences. Simulation evidence shows that winsorization and truncation are ineffective at identifying influential observations. While influence diagnostics and robust regression both outperform winsorization and truncation, overall, robust regression outperforms the other methods. Since robust regression is a theoretically appealing and easily implementable approach based on a model's residuals, we recommend that future accounting studies consider using robust regression, or at least report sensitivity tests using robust regression.