ABSTRACT
This chapter focuses on meta-analysis for prediction models in the medical field but generalizes easily to other fields. Prediction models may inform patients and physicians by providing individualized risk predictions. Meta-analysis of prediction models requires data from studies that predict the same outcome using the same set of predictors. A specific challenge is the assessment of heterogeneity in absolute risk predictions, where the hope is to obtain a single “global model” that would be valid for all similar studies. Two sources of between-study heterogeneity have to be checked: (1) differences in distributions of the predictors and (2) differences in prediction models. We focus on binary outcomes, which can be modeled by random-effect logistic models allowing random variation in intercept and regression parameters. In the case of substantial random variation, it is possible to obtain an “average model” that might, however, be far off for a particular patient population. We describe different ways of checking the effect of the observed random variation between studies. We also address how the meta-analysis can be used to obtain a patient-specific prediction in a population not included in the meta-analysis.
