Improving the Robustness of Variable Selection and Predictive Performance of Lasso and Elastic-Net Regularized Generalized Linear Models and Cox Proportional Hazard Models

• Activity Number: 240
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Feng Hong*+ and Viswanath Devanarayan
• Companies: AbbVie and AbbVie
• Keywords: Predictive ; Prognostic ; Variable Selection ; Signatures ; Genomics

Abstract

In diagnostic and drug development applications, high-dimensional data from genomics, proteomics, imaging, etc., are generated for deriving signatures that predict patient phenotypes such as disease status/progression, drug efficacy and safety. Various statistical algorithms are utilized to identify an optimal subset of biomarkers, that when applied to an appropriate model, predicts the desired phenotype. Both the composition and predictive performance of such biomarker signatures are critical. Recent algorithms proposed by Friedman et al (2010) and Simon et al (2011) for the regularization of generalized linear and cox regression models via cyclical coordinate descent are extremely useful as they are very fast and can handle different phenotypes (multinomial, counts, continuous, time-to-event). However the variable selection results tend to be unstable and affect the composition of the biomarker signature. In this paper, we propose a Monte-Carlo approach with a cross-validation wrapper to improve the robustness and stability of the variable selection results and predictive performance evaluation. We illustrate the improvements via real datasets and simulations.