flat even at extremely stringent thresholds (Figure 3C). Additionally, the resampling experiments demonstrated that, in direct contrast with a winner's curse prediction, effect sizes would need to be increasingly more severely over-estimated at higher thresholds (3-fold or more) to result in the observed rates of replication (Figure S8). Second, the definition of replication (concordance and p-value<0.05) is relatively loose when applied to eQTLs with a BF>5 (typical linear regression p-values<5×10−8) and should accommodate substantial drops in effect sizes for both replication panels but especially for the larger Merck dataset. This is further supported by the observation that concordance alone yielded similar rates of replication (Figure S6). We conclude that statistical power and winner's curse cannot explain the observed rates of non-replication for eQTLs with BF>5.
