We performed cis-eQTL analysis in each of the eQTL discovery datasets by calculating Spearman correlations within each cohort, followed by a sample size-weighted z-score meta-analysis approach, as described previously14. We opted for this approach to minimize the influence of potential heterogeneity between cohorts and showed that it performs comparably to FastQTL/QTLTools87 (Supplementary Note and Supplementary Fig. 35). To correct for multiple testing, we used an approach similar to FastQTL/QTLTools87, where we used 1,000 permutations of the sample labels to fit a β-distribution per gene and, after adjustment using this distribution, calculated the q-values88 over the top association per gene to determine significance (Supplementary Note). Genes with q-value < 0.05 were deemed significant. We limited these analyses to 19,373 protein-coding genes and to SNPs located within 1 Mb of the TSS, with MAF > 1% and Hardy–Weinberg P > 0.0001. The RNA-seq data were corrected for up to 20 technical covariates, dataset indicator variables and four multidimensional scaling components derived from the genotype data using OLS. In addition, we evaluated the impact of regressing out increasing numbers of PCs and defined
