Reproducible learning for accelerated failure time models via deep knockoffs

Selecting truly relevant variables contributing to the response is a fundamental problem in many scientific fields. One of major challenges in variable selection is to effectively control the false discovery rate (FDR). Most existing variable selection procedures in survival analysis neglect the FDR control. In this article, we fill such a gap and propose a new and flexible variable selection method with guaranteed FDR control for accelerated failure time models. The proposed method combines the strengths of deep knockoffs and the weighted M-estimation procedure and enjoys the FDR control for arbitrarily high dimension with finite samples. More importantly, our method does not require prior knowledge about the joint distribution of covariates. Extensive simulation studies confirm the generality, effectiveness, and power of the proposed method. Finally, the proposed method is used to analyze a primary biliary cirrhosis data to demonstrate its practical utility.