gms | German Medical Science

Introduction: In time-to-event analysis, the time to occurrence of a specific event is of interest. However, patients may be subject to multiple potential events, and the occurrence of the event of interest, e.g. disease-specific death, may be prevented by another (competing) event, e.g. death from another cause. In randomized clinical trials with such competing risks, treatment effects on the event of interest are commonly assessed using the Cox model for cause-specific hazards [1] or the Fine and Gray model for subdistribution hazards [2].

In recent years, several alternative approaches for analyzing competing risks have been proposed, such as Temporal Process Regression (TPR) [3], Inverse Probability of Censoring Weighting (IPCW) [4], Pseudo Values (PV) [5], and Multiple Imputation (MI) [6]. However, they seem to be seldom used in practice, and comprehensive comparisons are rare.

Methods: We conduct a simulation study [7] following the ADEMP framework, a structured approach to planning simulation studies introduced by Morris et al. [8], to compare the TPR, IPCW, PV and MI approaches as well as the cause-specific Cox model and Fine and Gray model in terms of type I error, power and treatment effect estimation. The simulation study targets the treatment effect on the event of interest in a univariable RCT setting with one competing event and varying assumptions for the data-generating model (proportional cause-specific hazards and subdistribution hazards), effect sizes, sample sizes, and degrees of censoring.

Results: First results for univariable simulation under proportional cause-specific hazards indicate that CSH, TPR, and IPCW generally hold the nominal significance level, while SDH, PV, and MI exceed the nominal significance level in scenarios with non-zero effect on the competing event. This becomes even more evident as the sample size increases, whereas the degree of censoring shows no difference with respect to the type I error. Furthermore, the simulation results suggest that the statistical power depends on the specific scenarios, in particular on the combination of effects on the event of interest and the competing event.

Discussion: The performance of the investigated approaches in different scenarios of the simulation study will be discussed. In particular, the influence of the combination of treatment effects on the event of interest and the competing event will be addressed.

Conclusion: The summarized evidence obtained from the simulation study will be used to develop recommendations on which of the considered approaches should be used in the situations studied.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.