All arrays were pre-processed using robust multi-array average (RMA) normalization51 and log2 transformation in Affymetrix Power Tools version 1.14-3. We applied RMA normalization in two different ways: one in which all expression values, regardless of tissue of origin, were used together and one in which RMA normalization was applied separately on a tissue-by-tissue basis. We found no evidence for any method-dependent bias or heterogeneity in our test results, either in the estimated eQTL signal strengths (β regression coefficients) or the P values for association (Supplementary Fig. 3). For all subsequent analysis, we used RMA normalization applied once across all tissues.
