Mixed logistic regression in genome-wide association studies.

Area of Science:

• Genetics
• Biostatistics
• Population Genetics

Background:

• Mixed linear models (MLM) are commonly used for genome-wide association studies (GWAS) but can be inappropriate for binary traits.
• Existing score tests for mixed logistic regression (MLR) do not provide effect size estimation.
• Population structure is a key consideration in GWAS.

Purpose of the Study:

• To develop and evaluate computationally efficient methods for estimating variant effects in mixed logistic regression models for GWAS.
• To assess the performance of proposed methods against traditional MLM and logistic regression using simulated and real data.
• To introduce a stratified QQ-plot for improved diagnosis of p-value inflation/deflation in GWAS.

Main Methods:

• Proposed two novel, computationally efficient methods for estimating variant effects.
• Evaluated methods using simulated genomic data and real GWAS data from West African populations.
• Compared performance against mixed linear models (MLM) and standard logistic regression.

Main Results:

• Mixed linear models (MLM) are inappropriate for binary traits when disease prevalence differs between population strata.
• Mixed logistic regression (MLR) demonstrated superior performance across all evaluated scenarios.
• The proposed methods provided well-estimated variant effects with moderate bias for large effect sizes.
• A stratified QQ-plot was introduced to enhance the detection of p-value inflation or deflation.

Conclusions:

• The two proposed methods are implemented in the R package milorGWAS, available on CRAN.
• These methods are computationally scalable to at least 10,000 individuals.
• The computational strategies can be extended to other models, such as mixed Cox models for survival analysis.