CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Model-free variable screening for ultrahigh dimensional survival data with FDR control Authors:  Chenlu Ke - Virginia Commonwealth University (United States) [presenting]
Abstract: A novel framework is proposed for variable screening for ultrahigh dimensional survival data. The contribution of each individual predictor to the survival outcome is quantified in the presence of the other candidates by kernel-based R-squared statistics. Compared with existing marginal screening methods, our proposal does not require an intermediate estimation of the survival function and relaxes the commonly imposed assumption of independent censoring. Moreover, our method can capture hidden important predictors that are marginally independent but jointly dependent on the survival outcome. We establish the sure screening property and the rank consistency property of the proposed approach in the notion of sufficiency. A knockoff procedure is also developed for controlling false discoveries. The advantages of the proposed method are demonstrated by simulation studies and an application to high-throughput gene expression data.