Laplace Error Penalty-Based Variable Selection and Its Application in GWAS

• Activity Number: 348
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: International Chinese Statistical Association
• Author(s): Shaoli Wang*+ and Xueqin Wang and Xiaolin Sang
• Companies: Shanghai University of Finance and Economics and Sun Yat-Sen University and Sun Yat-Sen University
• Keywords: Variable selection ; Laplace error penalty ; Block coordinate gradient descent algorithm ; Oracle properties ; GWAS

Abstract

In this talk we propose a new penalty function for selecting variables which are influential to response. This penalty function, called Laplace Error Penalty (LEP), is singular at the origin, bounded, and infinitely differentiable except for the origin. The LEP method is originally derived from a noncontiguous penalty function known as L0 (which results in the same estimations as subset selection), is an improvement on LASSO (L1 penalty) and SCAD. With an extra parameter k, LEP is homotopic to L0 penalty, and its performance varies according to different values of the parameter k. The LEP method possesses the Oracle properties: it performs the same as if the true model were already given. A modified block coordinate gradient descent (BCGD) algorithm is introduced for fast calculation in processing large scale data sets. Finally, simulations and real data analysis are performed.