Penalized Variable Selection for Multi-center Competing Risks Data

TL;DR

This paper develops penalized variable selection methods for multi-center competing risks data, specifically in kidney transplant studies, accounting for center effects and improving variable selection and estimation accuracy.

Contribution

It introduces novel penalization strategies for stratified and marginal PSH models to effectively select variables while handling intra-center correlations.

Findings

Simulations show good performance and efficiency of the proposed methods.

Application to UNOS data demonstrates practical utility.

Methods outperform existing approaches in variable selection accuracy.

Abstract

We consider variable selection in competing risks regression for multi-center data. Our research is motivated by deceased donor kidney transplants, from which recipients would experience graft failure, death with functioning graft (DWFG), or graft survival. The occurrence of DWFG precludes graft failure from happening and therefore is a competing risk. Data within a transplant center may be correlated due to a latent center effect, such as varying patient populations, surgical techniques, and patient management. The proportional subdistribution hazard (PSH) model has been frequently used in the regression analysis of competing risks data. Two of its extensions, the stratified and the marginal PSH models, can be applied to multi-center data to account for the center effect. In this paper, we propose penalization strategies for the two models, primarily to select important variables and estimate their effects whereas correlations within centers serve as a nuisance. Simulations demonstrate good performance and computational efficiency for the proposed methods. It is further assessed using an analysis of data from the United Network of Organ Sharing.