Feature Selection in GLM with Large Model Spaces

• Activity Number: 198
• Type: Invited
• Date/Time: Monday, August 3, 2009 : 2:00 PM to 3:50 PM
• Sponsor: SSC
• Author(s): Jiahua Chen*+ and Zehua Chen
• Companies: The University of British Columbia and National University of Singapore
• Keywords: BIC ; Consistency ; genome-wide association study ; single nucleotide polymorphism ; variable selection

Abstract

In genome-wide association studies and other investigations, the number of independent variables far exceeds the sample size where a generalized linear model is applicable. In such ``large-p-small-n" situations, the classical variable selection criteria such as BIC are found far too liberal. The extended Bayes information criterion proposed by Chen and Chen (2008), in contrast, provides effective control on false discovery rate while retaining comparable positive discovery rate. In this paper, we show that this criterion is also consistent at identifying the true significant set of variables under the generalized linear model. Its finite sample properties are investigated through simulation studies.