Adjusting for principal components can induce collider bias in genome-wide association studies

Area of Science:

• Genetics
• Population Genetics
• Statistical Genomics

Background:

• Principal component analysis (PCA) is a common method to account for population structure in genome-wide association studies (GWAS).
• Deciding the optimal number of principal components (PCs) and ensuring they don't capture confounding factors like linkage disequilibrium (LD) are significant challenges.
• Pre-processing steps like LD pruning or excluding high LD regions are suggested but not universally applied, with unclear implications for admixed populations.

Purpose of the Study:

• To investigate the impact of PCA pre-processing strategies and the number of PCs included on GWAS results in African American admixed populations.
• To identify potential biases and spurious associations arising from PCA in admixed samples.
• To provide recommendations for robust PCA implementation in GWAS for diverse populations.

Main Methods:

• Analysis of African American samples from the Women's Health Initiative and two Trans-Omics for Precision Medicine studies.
• Evaluation of different pre-processing techniques, including LD pruning and exclusion of high LD regions, prior to PCA.
• Assessment of the correlation between PCs and genome-wide ancestry versus local genomic features.
• Examination of the downstream consequences of including various PCs in GWAS models, focusing on effect size estimation and spurious associations.

Main Results:

• The first PC strongly correlated with genome-wide ancestry, while subsequent PCs captured local genomic features in all three admixed samples.
• The pattern of variant correlations with PCs differed from European populations, leading to biased effect size estimates and inflated false positive rates (collider bias).
• Excluding high LD regions did not resolve these issues; LD pruning was more effective, but optimal thresholds varied by dataset.

Conclusions:

• PCA in admixed populations presents unique challenges not fully addressed by standard pre-processing methods.
• Later PCs can capture local genomic features, leading to collider bias and spurious associations in GWAS.
• Careful pre-processing and diagnostic checks are essential to prevent the inclusion of confounding PCs in GWAS models for admixed ancestries.