Outliers vs Robustness in Nonparametric Methods of Regression

Abstract

The article addresses the question of how robust methods of regression are against outliers in a given data set. In the first part, we presented the selected methods used to detect outliers. Then, we tested the robustness of three nonparametric methods of regression: PPR, POLYMARS, and RANDOM FORESTS. The analysis was conducted applying simulation procedures to the data sets where outliers were detected. Contrary to a relatively common conviction about the robustness of nonparametric regression, the study revealed that the models built on the basis of complete data sets represent a significantly lower predictive capability than models based on the sets from which outliers were discarded.

Keywords

• outliers
• robustness
• nonparametric regression methods