ABSTRACT
This chapter discusses some more recently proposed criteria for model comparison that have no "single model" analogue. They include risk stratification tables, the net reclassification improvement (NRI), and continuous net reclassification improvement (cNRI). These latter criteria were developed to quantify the improvement in a model from adding a novel predictive marker. Examples of serum markers include CA-125 for predicting ovarian cancer incidence, and PSA (prostate-specific antigen) to predict the presence of prevalent prostate cancer and help decide if a biopsy is warranted. There has also been interest in determining how much the performance of risk prediction models could be improved by including information on single nucleotide polymorphisms (SNPs) identified in genome wide association studies (GWAS). Although individual disease-associated SNPs may confer only modest risk for most diseases, the combined effect of all identified disease-associated SNPs, summarized in a "polygenic risk score", might provide substantial information.
