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

Principal component analysis (PCA) is widely used to control for population structure in genome-wide association studies (GWAS). Top principal components (PCs) typically reflect population structure, but challenges arise in deciding how many PCs are needed and ensuring that PCs do not capture other artifacts such as regions with atypical linkage disequilibrium (LD). In response to the latter, many groups suggest performing LD pruning or excluding known high LD regions prior to PCA. However, these suggestions are not universally implemented and the implications for GWAS are not fully understood, especially in the context of admixed populations. In this paper, we investigate the impact of pre-processing and the number of PCs included in GWAS models in African American samples from the Women’s Health Initiative SNP Health Association Resource and two Trans-Omics for Precision Medicine Whole Genome Sequencing Project contributing studies (Jackson Heart Study and Genetic Epidemiology of Chronic Obstructive Pulmonary Disease Study). In all three samples, we find the first PC is highly correlated with genome-wide ancestry whereas later PCs often capture local genomic features. The pattern of which, and how many, genetic variants are highly correlated with individual PCs differs from what has been observed in prior studies focused on European populations and leads to distinct downstream consequences: adjusting for such PCs yields biased effect size estimates and elevated rates of spurious associations due to the phenomenon of collider bias. Excluding high LD regions identified in previous studies does not resolve these issues. LD pruning proves more effective, but the optimal choice of thresholds varies across datasets. Altogether, our work highlights unique issues that arise when using PCA to control for ancestral heterogeneity in admixed populations and demonstrates the importance of careful pre-processing and diagnostics to ensure that PCs capturing multiple local genomic features are not included in GWAS models.Author summary: Principal component analysis (PCA) is a widely used technique in human genetics research. One of its most frequent applications is in the context of genetic association studies, wherein researchers use PCA to infer, and then adjust for, the genetic ancestry of study participants. Although a powerful approach, prior work has shown that PCA sometimes captures other features or data quality issues, and pre-processing steps have been suggested to address these concerns. However, the utility and downstream implications of this recommended pre-processing are not fully understood, nor are these steps universally implemented. Moreover, the vast majority of prior work in this area was conducted in studies that exclusively included individuals of European ancestry. Here, we revisit this work in the context of admixed populations—populations with diverse, mixed ancestry that have been largely underrepresented in genetics research to date. We demonstrate the unique concerns that can arise in this context and illustrate the detrimental effects that including principal components in genetic association study models can have when not implemented carefully. Altogether, we hope our work serves as a reminder of the care that must be taken—including careful pre-processing, diagnostics, and modeling choices—when implementing PCA in admixed populations and beyond.