Ordinary Linear Mixed Model Approaches May Lead to Invalid Inference in Genetic Association Studies for Binary Traits

• Activity Number: 648
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Han Chen* and Chaolong Wang and Matthew Conomos and Adrienne Stilp and Zilin Li and Tamar Sofer and Adam Szpiro and Timothy Thornton and Cathy Laurie and Kenneth Rice and Xihong Lin
• Companies: and Genome Institute of Singapore and University of Washington and University of Washington and and University of Washington and University of Washington and University of Washington and University of Washington and University of Washington and Harvard School of Public Health
• Keywords: linear mixed model ; population structure ; cryptic relatedness ; homoscedasticity ; genetic association study ; type I error

Abstract

In recent years, linear mixed models (LMMs) have been widely used to control for population structure and cryptic relatedness in genetic association studies, and various LMM-based association test software programs have been developed. LMMs have been applied to genetic association studies for both quantitative and binary traits in the genetic literature. However, for binary traits, the homoscedasticity assumption of ordinary LMMs can be severely violated. This issue has often been overlooked in the field of genetic epidemiology, leading to widespread misunderstanding and misuse of ordinary LMMs in analyses. Here we show that for binary traits, heteroscedasticity can be a serious problem for ordinary LMMs and may lead to invalid inference in practice, which cannot be remedied by simply increasing the sample size. In large-scale genetic association studies, applying ordinary LMMs and ignoring heteroscedasticity may result in seriously inflated or deflated Type I Error rates for many variants, and this problem is not rectified by standard use of genomic control.