A general framework for meta-analyzing dependent studies with overlapping subjects in association mapping
Meta-analysis of genome-wide association studies is increasingly popular and many meta-analytic methods have been recently proposed. A majority of meta-analytic methods combine information from multiple studies by assuming that studies are independent since individuals collected in one study are unlikely to be collected again by another study. However, it has become increasingly common to utilize the same control individuals among multiple studies to reduce genotyping or sequencing cost. This causes those studies that share the same individuals to be dependent, and spurious associations may arise if overlapping subjects are not taken into account in a meta-analysis. In this paper, we propose a general framework for meta-analyzing dependent studies with overlapping subjects. Given dependent studies, our approach "decouples" the studies into independent studies such that meta-analysis methods assuming independent studies can be applied. This enables many meta-analysis methods, such as the random effects model, to account for overlapping subjects. Another advantage is that one can continue to use preferred software in the analysis pipeline which may not support overlapping subjects. Using simulations and the Wellcome Trust Case Control Consortium data, we show that our decoupling approach allows both the fixed and the random effects models to account for overlapping subjects while retaining desirable false positive rate and power.