Given differences in sample sizes among these studies, we sought to define a baseline replication rate against which to compare the observed levels of reproducibility. We therefore conducted a re-sampling experiment in which, for each gene expression trait, 100 sets of 60 sex and age (+/−3 years) matched samples were selected at random and used to define replication (i.e. concordant effect direction and p<0.05). We found that simulated replication rates increase dramatically near a BF of 5 (95.5% replication at BF>5; Figure 3C) and are effectively 100% at higher thresholds. These observations suggest that power differential among the studies cannot alone explain the observed rates of replication, as there are many genes with effect sizes in one study that should be readily detected in both (let alone either) replication panels. This is further supported by the observation that concordance alone (i.e. no p-value threshold) yielded similar levels of reproducibility, as did direct comparisons of allelic coefficients (Spearman's rho of 0.663 and 0.681 for UC–UW and UC-Merck comparisons, respectively; Figure S6).
