Validation of prediction models for patient outcomes and individualized treatment effect

In this chapter we discuss statistical methods to evaluate the performance of treatment effect models. We first explain the principles of prediction model validation and highlight common metrics of model performance. Subsequently, we explain additional issues of relevance when assessing the accuracy of treatment benefit predictions. To this purpose, we start out from a setting where RCT data are available for model validation, and distinguish between the evaluation of prognostic outcomes, and estimates of absolute treatment benefit. We discuss various measures of model performance, including discrimination, and calibration. In the context of data from multiple sources, we introduce internal-external cross-validation as a framework to assess the generalizability of model predictions. Finally, we touch upon recent developments with respect to the importance and evaluation of model fairness.