ABSTRACT In clinical trials, multiple outcomes are often of interest. For time‐to‐event outcomes, it is the norm rather than exception that one or more nonterminal events (e.g., heart failure hospitalization) and a terminal event (e.g., cardiovascular death) are encountered, subject to competing risks in addition to independent censoring. In this setting, traditionally, time‐to‐first event analysis is used to assess treatment effect, with popular procedures such as the log‐rank test and the Cox proportional hazards regression. However, this approach fails to fully utilize available data, resulting in relatively wide confidence intervals and less powerful tests. Various methods have been proposed in recent years, some involving complex models and others having intuitive appeal but hidden conditions. Furthermore, each method was motivated to address specific scientific questions. Navigating this landscape can be challenging for users. In addition, although competing risks can often be accommodated by considering cause‐specific hazards, that is not the case for some of the recently proposed methods, which may lead to difficulty in interpretation, requirement of strong assumptions, and less generalizability. This article provides a brief summary of several recently developed methods, discussing the strengths and limitations of each approach. Then, the problem is framed in a more general setup than in the literature for a wider range of competing risks scenarios. Furthermore, new nonparametric testing and estimation procedures are proposed, and their asymptotic validity is established. The methods are illustrated in a recent large scale clinical trial.
Song Yang (Fri,) studied this question.