Regency EF
Comparison of Methods to Analyze Clustered Time-to-Event Data with Competing Risks (304079)
Denise A Esserman, Yale University
Erich A Greene, Yale University
*Wenhan A Lu, Yale University
Can A Meng, Yale University
Guanqun A Meng, Yale University
*Zehua Pan, Yale University
Peter A Peduzzi, Yale University
Yuxuan A Wang, Yale University
Keywords: time-to-event; cluster; competing risk
With growing use of pragmatic clinical trials to increase generalizability of results, reduce costs, and emulate real world scenarios, the complexity of their design and analysis has also increased. We explored the impact of these design complexities on model convergence, bias and coverage for a cluster randomized clinical trial with a time-to-event outcome and a competing risk of death. We conducted simulations using two methods to generate the clustered data. We varied the event rate, competing risk rate, censoring rate, and amount of clustering. We investigated six models: generalized linear model with a Poisson link function and its extension to allow for clustering; Cox proportional hazards model and its extension to allow for clustering via a sandwich variance estimator; Fine and Grey competing risk model and its extension to allow for clustering (developed by Zhou et al). We found that only under extreme scenarios did we encounter model convergence issues; the Zhou model performed the best under most scenarios; and the models that ignored clustering had worse bias and coverage compared to those that ignored the competing risk.