Abstract:
|
In situations where randomized control trials (RCT) cannot be conceptualized (e.g. large observational studies with survival endpoints), various propensity score (PS) techniques are popularly used to control for pretreatment confoundings in baseline characteristics. However, this inferential setup of time-to-event outcomes can be further plagued with a myriad of complexities, such as heavy censoring, high imbalance between comparator groups, and clustered nature of the data. We develop a 2-step approach that addresses the complexities observed in the United Network of Organ Sharing (UNOS) database, which monitors the survival behavior of patients undergoing kidney transplantations. First, we estimate the propensity scores for multiple groups with the generalized boosted model to adjust for large number of confounders. Next, we incorporate the estimated propensity scores into a semi-parametric Cox-proportional hazards modeling framework, adjusted for multicenter clustering (via a parametric frailty specification) and excess censoring (via cure-rates), through an inverse propensity treatment weighting (IPTW).
|