Sample overlap in genome-wide association meta-analysis: practical comparisons across significance thresholds

In the realm of genome-wide association studies (GWAS), meta-analysis is a commonly employed technique to consolidate evidence from multiple studies without necessitating individual-level data. However, the increasing number of studies and their expanding sample sizes have led to a phenomenon known as sample overlap. This occurs when a single participant is included in more than one study, thereby introducing dependence across studies and breaching the independence assumptions that form the basis of standard meta-analysis methods.

In a Canadian initiative called HostSeq, 15 clinical and epidemiological studies were combined, resulting in DNA samples from around 10,000 Canadians who were infected by the SARS-CoV-2 virus. Through this process, the researchers observed instances of duplicate samples across different participating studies. This observation posed a significant practical question: what implications does sample overlap have on association results, and under what circumstances should it be considered a cause for concern?

To address this question, the researchers sought to compare five different meta-analysis methods and assess their performance in comparison to a reference mega-analysis in which the true degree of sample overlap was known. They discovered that when the degree of sample overlap is relatively modest, its impact on association results remains minimal, even at the stringent genome-wide significance threshold of 5*10^-8.

To substantiate their findings, the researchers conducted extensive simulation studies and applied the methods to overlapping cohorts within the HostSeq initiative. Through these rigorous evaluations, they were able to demonstrate the negligible impact of sample overlap on association results when the degree of overlap is relatively small.