Restrictions on Candidate Predictors

Abstract

A major challenge in prediction modeling is that we may have more candidate predictors available for the analysis than we would like to include for further analysis, in particular if our data set is relatively small. A small sample size leads to problems as discussed in Chap. 5, such as limited power to test effects of potential predictors, and too extreme predictions when predictions are based on the standard regression coefficients (overfitting). We discuss some procedures to increase the robustness and validity of a prediction model, including restriction of the number of candidate predictors based on subject knowledge, considering distributions of predictors, combining similar variables, and averaging the effects of similar variables. We provide a detailed description of a case study of modeling similar effects of aspects of family history for robust prediction of the presence of a genetic mutation.