Conceptual

Time-Dependent Predictive Accuracy Metrics for Interval-Censored Competing-Risks Models

How to compute time-dependent discrimination and calibration metrics — the time-dependent AUC, Brier score, and expected predictive cross-entropy — for a risk-prediction model when the outcome is interval-censored and a competing event is present. Two estimators are contrasted: a model-based approach that weights every at-risk subject by its predicted risk, and an inverse-probability-of-censoring-weighting approach that reweights the subset of subjects with a known case/control status.