ABSTRACT

Variable screening has gained increasing popularity in high-dimensional survival analysis. Most existing methods for variable screening with survival data suffer from that variable importance is assessed based on marginal models which relate the timeto-event outcome to each variable separately, implying that the relevance of one variable is examined when other variables are excluded. These methods will preclude variables that only manifest their influence jointly and may retain irrelevant variables that are correlated with relevant ones. To circumvent these difficulties, we propose a new approach to directly evaluating joint variable importance and achieving nonspurious screening. We establish the sure screening properties of the proposed method and demonstrate its effectiveness through simulation studies and a real data application. A novel stability selection-based procedure is also proposed for tuning.