Keywords: Anderson-Gill’s model, Bayesian posterior distribution, Metropolis-Hastings, Non-informative censoring, Piece-wise constant hazard, Recurrent events
This paper introduces the sensitivity analysis for the robustness of primary conclusion drawn from the data analysis in clinical trials, where the endpoint of interest is the time to event (including time to the first event or time to recurrent events). The non-informative censoring is a key assumption that most statistical methods for the time to event endpoint rely on. However, this assumption could be hardly verified using the observed data, which could affect the validity of the primary analysis results. This paper addresses the evaluation of sensitivity to missing data by the tipping point sensitivity analysis. The tipping-point sensitivity analysis requires no assumption on the missing data mechanism and is realized via assessing the clinical possibility of the “margin point” that reverses the primary conclusion. As a particular case, the method and algorithm can be used for the control-based jump to reference analysis, which assumes the hazards of missing event in the treatment group are similar to the hazards in the control group and is of interest in some clinical trials. Simulations are performed to demonstrate and evaluate the proposed methods through the R package developed by the authors. A data example from a cardiovascular outcome study is provided for illustration of the methods.