A Distribution Theory for Predictive Correlation Coefficient in High-Dimensional Models and Its Implications for Genome-Wide Association

• Activity Number: 248
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Ju-Hyun Park*+ and Nilanjan Chatterjee
• Companies: National Cancer Institute and National Cancer Institute
• Keywords: genome wide assoication ; predictive model ; effect-szie distribution

Abstract

Although recent genome-wide association studies have successfully identified thousands of common susceptibility loci across a variety of complex traits, several studies have shown that susceptibility loci which reach stringent genome-wide significance level currently have poor predictive power. Noting that there are many potentially associated common susceptibility SNPs that could have been detected by lowering the significance level, we study the behavior of the predictive power of polygenic models as a function of the sample size of a training dataset and threshold for SNP selection. We show that such relationship relies heavily on the distribution of effect sizes for all underlying susceptibility SNPs. We use above theory and empirically derived estimates of effect-size distribution for several complex traits to project sample size requirement for future studies for improving performance of predictive models.