Strategies for Building Risk Models Based on Single Nucleotide Polymorphisms (SNPs)

• Activity Number: 533
• Type: Invited
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Mitchell Henry Gail*+ and Ruth Maria Pfeiffer and Jincao Wu
• Companies: National Cancer Institute and National Cancer Institute and National Cancer Institute
• Keywords: genetic risk models ; single nucleotide polymorphisms ; discriminatory accuracy ; feature selection ; risk models

Abstract

Many applications of risk models, such as screening populations for high risk subjects, require high discriminatory accuracy, measured by the area under the receiver operating characteristic curve, AUC. Models based on SNPs that have been discovered and validated in independent genome-wide association studies (GWAS) have yielded modest AUC. One explanation is that many informative SNPs are not being used because GWAS have been too small to discover and validate them. Using data-based distributions of log relative odds per allele for SNPs, we evaluated strategies that relaxed inclusion criteria to allow many SNPs to enter risk models. We also evaluated other features of model building. In GWAS of practical size, models that included only tens or hundreds of SNPs yielded the highest AUC values in independent data, even when thousands of SNPs (many with small effects), were in fact associated with disease. The most important aspect of model building was SNP selection. The method of estimation of SNP effects following SNP selection was less important. Because of the dominant importance of SNP selection, simple model building methods performed comparably to more complex ones.