ABSTRACT
This chapter deals with advanced topics that the expert needs to understand and figure out. Often the researcher receives data from multiple cohorts. These cohorts may come from different hospitals, different doctors, or different countries, as examples. Treatment is different from other risk factors because it can be intervened upon. Indeed, a common task for medical risk prediction models is to help the patient to decide which treatment, if any, to select. An important concern with comparative effectiveness tables is residual confounding. Unless the data underlying the models come from randomized controlled trials of the treatments being featured in the comparative effectiveness tables, one has to worry about whether the prediction models properly adjust for all important covariates. The chapter describes the learning curve paradigm, which is not at all a surprising assumption but important to keep in mind when dealing with internal validation schemes and missing values in the predictor variables.
