Sparse Estimation of Genetic Relatedness to Control for Population Structure and Sample Relatedness in Genome-Wide Association Studies

• Activity Number: 310 - SPEED:Statistical Methods for GWAs, Genetics, Genomics, and Other Omics Studies, Part 2
• Type: Contributed
• Date/Time: Tuesday, July 30, 2019 : 9:25 AM to 10:10 AM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Rounak Dey* and Yaowu Liu and Zilin Li and Junwei Lu and Zheng Tracy Ke and Xihong Lin
• Companies: Harvard TH Chan School of Public Health and Harvard TH Chan School of Public Health and Harvard TH Chan School of Public Health and Harvard TH Chan School of Public Health and Harvard University and Harvard
• Keywords: GWAS; Sparse Covariance Estimation; Generalized Linear Mixed Models

Abstract

In genome-wide association studies, generalized linear mixed models are commonly used to control for relatedness among the samples by modeling the covariance of the random effects based on the genetic relationship matrix (GRM). Even though the empirically estimated GRM or its population-adjusted version can be element-wise consistent to the true GRM, they can result in biased estimation in the spectral norm when the sample size becomes large compared to the number of variants used to estimate the GRM. This can lead to miscalibration of type I errors in the association tests as the variance of the test statistic depends on the quadratic forms of the GRM. In this paper, we first investigate how this miscalibration can affect modern large-scale association studies with hundreds of thousands of samples, and propose an alternative method using a sparse estimation of the GRM to provide well-calibrated p values for large-scale association studies while controlling for population heterogeneity. Using numerical simulations, and an application on TOPMed lipid data, we demonstrate the superior performance our method in terms of controlling the type I errors in large-scale association tests.