Statistical power and utility of meta-analysis methods for cross-phenotype genome-wide association studies

Abstract

Advances in recent genome wide association studies (GWAS) suggest that pleiotropic effects on human complex traits are widespread. A number of classic and recent meta-analysis methods have been used to identify genetic loci with pleiotropic effects, but the overall performance of these methods is not well understood. In this work, we use extensive simulations and case studies of GWAS datasets to investigate the power and type-I error rates of ten meta-analysis methods. We specifically focus on three conditions commonly encountered in the studies of multiple traits: (1) extensive heterogeneity of genetic effects; (2) characterization of trait-specific association; and (3) inflated correlation of GWAS due to overlapping samples. Although the statistical power is highly variable under distinct study conditions, we found the superior power of several methods under diverse heterogeneity. In particular, classic fixed-effects model showed surprisingly good performance when a variant is associated with more than a half of study traits. As the number of traits with null effects increases, ASSET performed the best along with competitive specificity and sensitivity. With opposite directional effects, CPASSOC featured the first-rate power. However, caution is advised when using CPASSOC for studying genetically correlated traits with overlapping samples. We conclude with a discussion of unresolved issues and directions for future research.

Fig 1. Power (K = 8, all effects in the same direction).

In each simulation, a total of 10,000 summary association statistics were generated for eight traits. The numbers of subjects were 1,000 cases and 1,000 controls. The minor allele frequency (MAF) of each causal variant was 0.1. The three panels from the top to the bottom represent when the effect of the variant was drawn from a normal distribution, a bimodal normal distribution, and a uniform distribution. In each graph, the Y-axis denotes the power of each method at an alpha level of 0.05 while the X-axis shows the number of truly associated traits. All associated traits shared the effects in the same direction. Ten meta-analysis methods were separated into three groups each based on: (1) the fixed-effects model (blue hue), (2) the random-effects model (red hue), and (3) the p-value-based model (green hue).

Fig 2. Power (K = 8, some effects in the opposite direction).

The same simulation was conducted as previously except that the direction of some effects was opposite. The three panels (A), (B), and (C) from the top to the bottom represent when 25% of associated traits carry effects in the opposite direction and effect sizes were drawn from a normal distribution, a bimodal normal distribution, and a uniform distribution, respectively. In each graph, the Y-axis denotes the power of each method at an alpha level of 0.05 while the X-axis shows the number of truly associated traits. The panels (D), (E), and (F) summarize the power of the ten methods when 50% of effects are in opposite direction.

Fig 3. Meta analysis results of GWAS data with shared controls.

Three control sample overlapping scenarios (0%, 50%, 100% overlapping) were set up to be used in the association study of both disorders. Total of four methods (ASSET1, ASSET2, CPASSOC, WICS) that are able to control overlapping samples between studies were compared with standard fixed-effect method (FEMA) for their performance in this analysis. The Y-axis denotes number of genome–wide significant loci.