Outliers May Not Be Automatically Removed

Abstract

Researchers often remove outliers when comparing groups. It is well documented that the common practice of removing outliers within groups leads to inflated type I error rates. However, it was recently argued by André that if outliers are instead removed across groups, type I error rates are not inflated. The same study discusses that removing outliers across groups is a specific case of the more general concept of hypothesis-blind removal of outliers, which is consequently recommended. In this paper, I demonstrate that, contrary to this advice, hypothesis-blind outlier removal is problematic. Specifically, it almost always invalidates confidence intervals and biases estimates if there are group differences. It moreover inflates type I error rates in certain situations, for example when variances are unequal and data nonnormal. Consequently, a data point may not be removed solely because it is deemed an outlier, whether the procedure used is hypothesis-blind or hypothesis-aware. I conclude by recommending valid alternatives.

Authors 1

1. Julian David Karch corresponding Aachen Institute of Psychology, Leiden University Methodology and Statistics Department

   Leiden University

   Affiliation as printed

   Institute of Psychology Leiden University

   Methodology and Statistics Department , Institute of Psychology , Leiden University , PO Box 9555 , 2300 RB Leiden , The Netherlands