Unified Methods for Variable Selection in Large-Scale Genomic Studies with Censored Survival Outcomes

• Activity Number: 219 - SLDS 2017 Student Paper Awards Session
• Type: Topic Contributed
• Date/Time: Monday, July 31, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Lauren Spirko* and Karthik Devarajan
• Companies: Temple University and Fox Chase Cancer Center
• Keywords: non-proportional hazards ; survival analysis ; Kullback-Leibler divergence ; high-dimensional data ; gene expression ; proportional odds model

Abstract

A major goal in genomic studies is to identify genes with a prognostic impact on time-to-event outcomes. With rapid development in genomic technologies, the scientific community is able to compile data sets with tens of thousands of such variables. Methods based on univariate Cox regression are often used to select genes related to survival outcomes. However, the Cox model assumes proportional hazards, which is unlikely to hold for each gene and could lead to an incorrect estimation of the effects.

In this paper, we propose variable selection methods that accommodate various forms of non-proportional hazards. First, we develop methods to test for gene effects in the proportional odds (PO) and Yang-Prentice models based on Kullback-Leibler divergence. We then propose pseudo-R^2 measures for the PO and proportional log-odds (PLO) models. We evaluate the performance of our methods using extensive simulation studies and large-scale genomic datasets in ovarian and oral cancer, and compare their performance with that of existing methods.