An important caveat is that these estimates of reproducibility are less meaningful for sequence-based studies of gene expression, which offer advantages in dynamic range and measurement accuracy [9], [10]; sequencing is also largely immune to the SNP-in-probe effect that significantly inflates false positives in our data (Figure 3C). However, the observation that age, sex, race, drug exposures, clinical covariates, and other global factors have such strong influences on expression (e.g., this study and [36], [2]) coupled with observations in other studies and different tissues that factors like cause of death are relevant [42], suggests that much of the non-reproducibility is in fact driven by systematic differences in tissue source. Such differences will likely be important to all studies of primary tissue samples, whether assayed by arrays or by sequencing. The reproducibility of future results would benefit from analysis of samples from multiple centers with as much clinical information as possible. Furthermore, our results confirm previous observations that the effects of unknown, unmeasured, or unquantified covariates can confound genetic effects with structured error sources [19], [36], [2], and that controlling for
