Cis and trans variance components were estimated using the REML algorithm implemented in GCTA19. As in previous studies, estimates were allowed to converge outside the expected 0–1 bound on variance to achieve unbiased mean estimates across all genes24. Standard error across gene sets was estimated by dividing the observed standard deviation by the square root of the number of genes that converged (this will lead to underestimation due to correlated genes, but is presented for completeness). Genome-wide h2g for the four traits in the GWAS cohort was estimated with GCTA from a single relatedness matrix constructed over all post-QC SNPs in the strictly unrelated individuals. For estimating expression-wide h2GE, each predicted expression value was standardized to mean=0 and variance=1, and sample covariance across these values used to define the relatedness matrix. The hGE2 was then estimated from this component with GCTA, with P-values for difference from zero computed using a likelihood ratio test. 20 principal components (PCs) were always included as fixed-effects to account for ancestry. Genetic correlation between traits in the GWAS cohort was estimated from all post-QC SNPs
