Robust Model Selection via Weighted Cross-Validation

• Activity Number: 18
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Marianthi Markatou*+ and Claudio Agostinelli
• Companies: SUNY, Buffalo and Ca'Foscari University
• Keywords: cross-validation ; computational efficiency ; model selection ; robustness ; weighted likelihood

Abstract

Model selection criteria such as AIC, BIC, Mallow's Cp and Shao's (1993) cross-validation have been proposed in the literature along with their robust versions (Ronchetti (1985), Ronchetti & Staudte (1994) and Ronchetti, Field and Blanchard (1997)). We propose a weighted cross-validation model selection method based on the weighted likelihood approach (Markatou, Basu and Lindsay (1998)). We study both, its asymptotic properties and its finite sample behavior. The main advantage of our method is that it is computationally faster than other competing methods therefore, it is suitable for the analysis of very large data sets. This method differs from robust cross-validation in that a) it downweighs only those observations that do not fit well the full model, b) it is fully efficient in the absence of contamination, and c) it is computationally faster than many of the robust model selection procedures proposed in the literature. We present a detailed simulation study and the results of the analysis of a study that aims to understand the significance of normal alanine aminotransferase (ALT) levels in patients with HIV/HCV co-infection.