Quantifying and Estimating the Predictive Accuracy for Censored Time-to-Event Data with Competing Risks

摘要

This paper focuses on quantifying and estimating the predictive accuracy ofprognostic models for time-to-event outcomes with competing events. We considerthe time-dependent discrimination and calibration metrics, including thereceiver operating characteristics curve and the Brier score, in the context ofcompeting risks. To address censoring, we propose a unified nonparametricestimation framework for both discrimination and calibration measures, byweighting the censored subjects with the conditional probability of the eventof interest given the observed data. We demonstrate through simulations thatthe proposed estimator is unbiased, efficient and robust against modelmisspecification in comparison to other methods published in the literature. Inaddition, the proposed method can be extended to time-dependent predictiveaccuracy metrics constructed from a general class of loss functions. We applythe methodology to a data set from the African American Study of Kidney Diseaseand Hypertension to evaluate the predictive accuracy of a prognostic risk scorein predicting end-stage renal disease (ESRD), accounting for the competing riskof pre-ESRD death.