SVA [20] was performed on the matrix of expression measurements, after controlling for the effects of age, sex, and a multidimensional scaling based quantification of genetic ancestry. For each probe, we constructed a linear model y ∼ m + A + C + R + SVi..n + e, where y is the log2 transformed probe intensity, m is the expected probe intensity, A is the effect of individual age, and C is the effect of individual sex, and R is the effect of genetic ancestry, SVi..n represents the effects of a matrix of surrogate variables, and e is the residual error. Models were implemented with the lm function in R. The residuals from this regression were used as the phenotype values for all subsequent analyses.
