Systematic comparison of approaches to analyze clustered competing risks data

Abstract

Abstract Background In many clinical trials the study interest lies in the comparison of a treatment to a control group regarding a time to event endpoint like time to myocardial infarction, time to relapse, or time to a specific cause of death. Thereby, an event can occur before the primary event of interest that alters the risk for or prohibits observing the latter, i.e. a competing event. Furthermore, multi-center studies are often conducted. Hence, a cluster structure might be observed. However, commonly only the aspect of competing events or the aspect of the cluster structure is modelled within primary analysis, although both are given within the study design. Methods to adequately analyze data in such a design were recently described but were not systematically compared yet. Methods Within this work we provide a systematic comparison of four approaches for the analysis of competing events where a cluster structure is present based on a real life data set and a simulation study. The considered methods are the commonly applied cause-specific Cox proportional hazards model with a frailty, the Fine and Gray model for considering competing risks, and extensions of the latter model by Katsahian et al. and Zhou et al. Results Based on our simulation results, the model by Katsahian et al. showed the best performance in bias, square root of mean squared error, and power in nearly all scenarios. In contrast to the other three models this approach allows both unbiased effect estimation and prognosis. Conclusion The provided comparison and simulations help to guide applied researchers to choose an adequate method for the analysis of competing events where a cluster structure is present. Based on our simulation results the approach by Katsahian et al. can be recommended.