The Impact of Outliers on Regression Coefficients: A Sensitivity Analysis

Abstract

Empirical research analyzes real-life data that often do not conform to a normal distribution with statistical tools such as linear regressions that require the assumption of normality. The lack of conformity to a known statistical distribution requires researchers to handle outliers properly. Accounting studies typically treat outliers with a “delete-and-forget” approach, which assumes that extreme values are erroneous and results remain insensitive to the deletion of a small number of observations. Results in this study refute these assumptions by showing that the ambiguity in handling outliers and variable selection motivates researchers to explore various analytic alternatives, which in turn produce unstable regression coefficients and heightened false-positive rates.