Search

Statistical heterogeneity refers to differences among results of studies combined in a meta-analysis beyond that expected by chance. On the one hand, excessive heterogeneity can diminish power to discover genetic signals; on the other, moderate heterogeneity can reveal important biological differences among studies. Given its double-edged nature, this thesis dissects heterogeneity in genetic association meta-analyses from three vantage points. First, a novel multi-variant statistic, M is proposed to detect genome-wide (systematic) heterogeneity patterns in genetic association meta-analyses. This was motivated by the limited availability of appropriate methodology to measure the impact of heterogeneity across genetic signals, since traditional metrics (Q, I2 and T2) measure heterogeneity at individual variants. Second, given that meta-analyses comprising small numbers of studies typically report imprecise summary effect estimates; GWAS-derived empirical heterogeneity priors are used to improve precision in estimation of average genetic effects and heterogeneity in smaller meta-analyses (e.g. ≤ 10 studies). Third, a critical evaluation of the Han-Eskin random-effects model shows how it can identify small effect heterogeneous loci overlooked by traditional fixed and random-effects methods. This work draws attention to the existence of genome-wide heterogeneity patterns, to reveal systematic differences among the ascertainment criteria of participating studies in a meta-analysis of coronary disease (CAD) risk. Furthermore, simulation studies with the Han-Eskin random-effects model revealed inflated genetic signals at small effect loci when heterogeneity levels were high. However, it did reveal an additional CAD risk variant overlooked by traditional meta-analysis methods. We therefore recommend a holistic approach to exploring heterogeneity in meta-analyses which assesses heterogeneity of genetic effects both at individual variants with traditional statistics and across multiple genetic signals with the M statistic. Furthermore, it is critically important to review forest plots for small effect loci identified using the Han-Eskin random-effects model amidst moderate-to-high heterogeneity (I2 ≥ 40%).