A Robust Approach to Model Selection in Regression

DOI: 10.5176/2251-1911_CMCGS15.32

Authors: Shawn X. Liu

Abstract:  A new robust model selection procedure for regression, called trimmed mean cross-validation, has been proposed to overcome the non-robustness of the usual cross-validation method. The new procedure of model selection is based on the idea that carries out the cross-validation method “almost as usual” by using the partial residuals which is “error free” rather than the whole set of residuals which may be affected by the contaminated data in order to achieve robustness. The simulation results in this paper indicate that the trimmed mean cross-validation method works quite well when the data set is uncontaminated and it works extremely well when the data set is contaminated while in this case, the usual cross-validation method is unreliable.

Keywords: cross-validation; model selection; robustness; trimmed mean cross-validation