ABSTRACT
This chapter assumes the case where an independent test dataset is available for the comparison of models. When comparing rival models, one of them might be more expensive or more invasive, either of which would suggest that a small drop in accuracy might outweigh the cost or pain of running the other model. In the setting with a right-censored time-to-event outcome, we must choose one or several prediction time horizon values. At each prediction time horizon, we can compare the predictive performance of rival models. The tools for model comparison begin with those described in the roadmap for individual model evaluation. There are two settings where updating a subject's prediction commonly occurs. One is where the subject remains free of the event during follow-up and would like a revised estimate of his prediction. In the second setting, additional information is available at the time when the revised prediction is requested.
