2

Cox proportional hazards models in the presence of censoring assume that the censoring mechanism is independent of the outcome of interest (1). In the context of censoring that occurs at least in part from competing risks, this assumption is rarely tenable. For that reason, we conducted a sensitivity analysis using a discrete-time failure model for death as the censoring event. The approach used in these analyses created two models for the informative censoring mechanism as a function of covariates and then used predicted probabilities of censoring to derive adjustment weights (2, 3). Modeling the censoring mechanism. The first model used pooled regression to estimate the risk of censoring from the competing risk as a function of time and all possible baseline covariates (and interactions) that predict the competing risk. To achieve this result, in a discrete-time failure model with time as a covariate, patients censored for reasons other than competing risks were included in the dataset until the time of censoring. The time interval was from the start of follow-up until the patient had the competing risk or was censored for another reason. Time was grouped to reduce model complexity. The outcome in this model was the binary indicator for the competing risk event. This model produced predicted probabilities of censoring from the competing risk of death for each patient at each time period. A second model predicted censoring from the competing risk as a function of treatment assignment (and time) only. For each subject and for each unit of time, one can estimate an inverse probability of censoring weight as