Evaluating discrimination accuracy of prognostic models with competing risk outcomes

We evaluate the discrimination accuracy of prognostic models for competing risk outcomes when multiple events are modeled simultaneously. Existing methods are mainly cause-specific assessments that focus on one event of interest and ignore model information for other events, therefore cannot fully characterize model's ability to distinguish different groups. In this work, our goal is to first introduce several new measures for overall and cause-specific evaluation that account for multiple events simultaneously, to better capture models' discrimination ability from different aspect based on varying rules of classification and concordance, and then to systematically investigate their functionalities under various scenarios via simulation studies and real data analyses. Estimators of proposed metrics are provided, along with asymptotic properties of one estimator demonstrated as an example. Our results highlight that one of the proposed cause-specific evaluation indices is an attractive alternative for distinguishing models from different families. Furthermore, all measures demonstrate sensitivity to the exclusion of important variables in prognostic models, highlighting their utility in evaluating model performance.