Statistical Methods for Heritability Estimation and Genetic Risk Prediction with Application to Addiction Phenotypes

• Activity Number: 618
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Yue-Ming Chen*+ and Peng Wei
• Companies: and University of Texas School of Public Health
• Keywords: heritability ; genetic risk prediction ; GWAS ; linear mixed model ; penalized regression

Abstract

Recent genome-wide association studies (GWAS) research reveals possible sources of missing heritability, which may illuminate the underlying genetic architecture, and may improve the accuracy of genetic risk prediction of individuals. In the framework of quantitative genetics, we construct a linear mixed model (LMM) for two purposes: to estimate proportion of the phenotypic variance explained by genome-wide single nucleotide polymorphisms (SNPs), i.e., chip heritability, and to make genetic risk prediction at the individual level. For risk prediction, we also compare LMM to a set of penalized regression models, such as the ridge regression and the lasso, and a polygenic score method including only the top ranking marginally significant SNPs. Using extensive simulations, we found that the LMM can estimate the chip heritability well without requiring a large sample size and the risk prediction accuracy depends on the training sample size and the genetic architecture of the phenotype. We further investigate the prediction models on a GWAS dataset of addiction phenotypes. We discuss and compare how different model building strategies can influence the genetic risk prediction results.