FDR Control in Feature Screening for Ultrahigh-Dimensional Right-Censored Data via Data-Driven Threshold Selection

Identifying key factors related to patients’ survival times from a vast set of covariates is essential for optimizing treatment strategies and advancing precision medicine. Feature screening is a powerful tool for this task, where threshold selection plays a critical role in distinguishing active from inactive covariates. However, conventional methods often apply hard thresholding rules that retain a fixed number of covariates, potentially leading to suboptimal outcomes by omitting active covariates or including many irrelevant ones. Balancing the identification of active covariates while controlling the false discovery rate (FDR) remains challenging, particularly with ultrahigh-dimensional censored data. In this paper, we propose a data-driven threshold selection procedure and introduce a novel feature screening method with FDR control, specifically tailored for ultrahigh-dimensional right-censored data. Our approach leverages data-splitting and reassembling techniques to construct a series of statistics with symmetry properties. Utilizing this symmetry, we develop a threshold selection procedure and a feature screening method that accurately identifies active covariates while maintaining strict FDR control. We establish both finite-sample and asymptotic properties for the FDR control of the proposed method under mild conditions. Simulation studies and a real data application demonstrate that our method outperforms existing approaches in variable selection accuracy and FDR control.