Exploring the performance of the GPSC method under several levels of outliers

Abstract

The aim of this study is to evaluate the performance of the Groupwise Principal Sensitivity Components method which is a robust iterative procedure for fitting linear regression models with fixed group effects. A simulation study is carried out to assess its ability to detect multiple outliers located in the response variable (vertical outliers) or in the explanatory and response variable (high leverage outliers). Several levels of outliers are considered ranging from 5% to 45% within selected groups. The results suggest that the GPSC method is able to avoid the masking effect under low or moderate level of outliers -approximately below to 30%. Additionally, in almost all cases the GPSC method reports lower levels of false outlier detection under high leverage outliers.