Are Robust Estimators Really Necessary?

Abstract

Although there is substantial literature on robust estimation, most scientists continue to employ traditional methods. They remain skeptical about the practical benefit of employing robust techniques and doubt the realism of the long-tailed error distributions commonly employed by their proponents in Monte Carlo studies. In this article a method of comparing the performance of estimators of location is developed and applied to a series of historical data sets in the physical sciences and to a collection of modern analytical-chemistry data sets. Both sets of results suggest that either severely trimmed means or modern robust estimators are required for optimal efficiency.