Simulations to evaluate the summary-based method were performed in 6,000 unrelated METSIM GWAS individuals. 100 genes and the SNPs in the surrounding 1MB were randomly selected for testing. For each gene, normally distributed gene expression was simulated as E = Xβ + ε; where X is a matrix of the desired number of causal genotypes, sampled randomly from the locus; β is a vector of normally distributed effect-sizes for each causal variant; and ε is a vector of normally distributed noise to achieve a cis-h2g of 0.17 (corresponding to the mean observed in our significant gene sets). 1,000 individuals with SNPs and simulated expression were then withheld for training the predictors. For the remaining 5,000 individuals, normally distributed noise was applied to the expression to generate a heritable phenotype where expression explained 0.10/180 or 0.20/180 of the phenotypic variance (the former corresponding to the average effect-sizes for associated genes observed in a large GWAS of height56 and the latter to high-effect loci). Association between SNP and phenotype was estimated in the 5,000 individuals (standard Z-score), and the phenotype generation repeated
