After adjusting for age, sex, ancestry, and unmeasured confounding variables (quantified by surrogate variable analyses, see Methods and [20]), we found 1,787 gene models with significant cis-linked genetic effects on expression levels (UC log10 Bayes Factor (BF)>5; SNP to TSS distance <250 kb; Figure 2A, Figure 3A, Table S1). The distribution of t-test p-values in the replication sets, adjusted for the same covariates, for the UC best associated gene-SNP pairs were significantly enriched for small values (Figure 3B), indicating that a large fraction of cis-eQTLs are reproducible in independent sample collections. As with demographic effects, we defined replication as a p-value<0.05 and a concordant allele effect direction (Figure 3C). While the significance of association in the discovery cohort has a large effect on replication probability, the relationship between significance and replication was effectively binary (Figure 3C). Cis-eQTLs with BFs>5 were much more likely to replicate than those with BFs<5 (chi square p-value<2×10−16). However, among genes with BFs>5, replication probability was only weakly dependent upon BF (Figure 3C; logistic regression chi-squared p-value = 0.00319).
