Semi-Parametric Methods for Competing Risks Data with Applications in Organ Transplantation.

Abstract

Competing risks data arise naturally in many biomedical settings, and it is often of interest to compare outcomes between subgroups of subjects. The second chapter proposes a measure to contrast group specific cumulative incidence functions (CIF). One such application involves evaluating organ procurement organizations (OPO) with respect to the cumulative incidence of kidney transplantation, with the competing risks being (i) death on the wait list and (ii) removal from the wait-list. The proposed method assumes proportional cause-specific hazards, which are estimated through Cox models stratified by OPO. The effect measure compares the average CIF of an OPO to the average CIF that would have resulted if that particular OPO had cause-specific hazard functions equal to those of the national average.

The third chapter proposes a measure, based on direct standardization, which contrasts two average cumulative incidence functions. In the context of evaluating a particular OPO, the contrast would be between (i) national average CIF (ii) what the national average would equal if all patients were subject to the practices of the OPO of interest. The proposed methods are nonparametric in the sense that no models are assumed for the cause-specific hazards or the subdistribution function. Observed event counts are weighted using Inverse Probability of Treatment Weighting and Inverse Probability of Censoring Weighting.

The fourth chapter develops a multiple imputation method for competing risks data. For individuals who experienced a competing risk not-of-interest, we impute censoring times in order to create censoring-complete data. The subdistribution hazard regression model developed by Fine and Gray (1999) can then be applied to the censoring-complete data, without the need to use inverse weighting. For each of the proposed methods, large sample properties are derived and the finite-sample properties are evaluated using simulations. We apply each method to national kidney transplantation data from the Scientific Registry of Transplant Recipients.