We next quantified the role of several additional factors that may generate spurious associations. Most strikingly, failure to control for unknown or unmeasured confounding variables by surrogate variable analysis (SVA) produced a large decrease in the number of significant (BF>5) cis-eQTL signals (1,787 vs. 873; Figure 4A; McNemar's chi-squared test p-value<2×10−16), similar to a recent study of gene expression within twins [23]. Not only did SVA produce a larger number of significant cis-eQTL associations, but these associations were also significantly more likely to replicate (McNemar's Chi-squared test p-value≪2×10−16; Figure 4B). While it has been shown that unknown or unquantified confounders can lead to unreliable genetic predictions [19], [36], [2], our data show that such factors, if unaccounted for, dramatically decrease the number of eQTL signals and their reproducibility across multiple independent collections of primary human tissues.
