A0652
Title: Sufficient feature screening for ultrahigh-dimensional right-censored data with high censoring rate
Authors: Wenbo Wu - University of Texas at San Antonio (United States) [presenting]
Abstract: Feature screening for ultrahigh-dimensional right-censored data poses significant challenges, especially when censoring rates are high. Most existing marginal screening methods focus on identifying predictors with strong marginal associations with the response variable. However, these approaches often fail to detect active predictors that exhibit strong conditional dependence on the response, and their performance declines as censoring rates increase. Two distance correlation-based sufficient variable screening frameworks are proposed that effectively identify active predictors with both marginal and conditional associations to the response variable. These model-free screening methods leverage censoring information via Buckley-James imputation, achieving superior performance even at high censoring rates. The sure screening property and rank consistency are established for all proposed procedures. The effectiveness and advantages of the methods are demonstrated through extensive simulation studies and analysis of a real-world dataset.
