Sample overlap in genome-wide association meta-analysis: practical comparisons across significance thresholds
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Meta-analysis is widely used in genome-wide association studies (GWAS) to combine evidence across studies without requiring individual-level data. However, as both the number of studies included in meta-analyses and their sample sizes continue to grow, the same participant may be included in more than one study. Such sample overlap induces dependence across studies and violates the independence assumptions underlying standard meta-analysis methods. HostSeq is a Canadian initiative in which 15 clinical and epidemiological studies were assembled with DNA samples from approximately 10,000 Canadians infected by the SARS-CoV-2 virus. We observed duplicate samples across different participating studies. This observation raises an important practical question: what is the impact of sample overlap on association results, and when should such overlap be a concern? Motivated by this real problem, we sought to compare five meta-analysis methods and evaluate their performance against a reference mega-analysis in which the true sample overlap is known. We show that when the degree of sample overlap is relatively small, its impact on association results is minimal, even at the genome-wide significance threshold of α = 5 × 10 −8 . We demonstrate these findings through extensive simulation studies and an application to overlapping cohorts within HostSeq.
Author summary
Combining results from multiple genetic studies can help researchers identify genetic variants associated with diseases without requiring access to individual participants’ data. However, some individuals may be involved in more than one study, creating dependence between study results. Ignoring this dependence may lead to misleading conclusions about genetic associations. Motivated by overlapping participants identified in the Canadian HostSeq initiative, we investigated how sample overlap affects genetic association results and compared existing statistical methods that account for it. Using simulations and HostSeq genetic data, we found that small sample overlaps may have little impact on association results, but ignoring larger overlaps can lead to reporting associations when none exist. This problem can become more pronounced under the stringent significance thresholds commonly used in genetic studies. We also found that existing methods for correcting sample overlap have different strengths and limitations. Our findings provide practical guidance on when sample overlap may require correction and which statistical methods researchers should consider.
