For each gene, weighted least-squares linear regression was performed using limma to yield coefficients for the effect on gene expression of each variable on the right-hand side: gene expression~diagnosis+covariates Then, for each gene, the SCZ disease status coefficient was statistically tested for being non-zero, implying an estimated effect for SCZ, above and beyond any other effect from the covariates. This test produces a t-statistic (then moderated in a Bayesian fashion) and corresponding P value. P values were then adjusted for multiple hypothesis testing using false discovery rate (FDR) estimation, and the differentially expressed genes were determined as those with an estimated FDR ≤ 5%. FDR was calculated by the limma package, which uses Benjamini-Hochberg from p.adjust() function in R. Significance was also assessed by permuting case-control status for 1,000 experiments. Of these experiments, the average number of significant genes at FDR ≤ 0.05 was 4.3, well below 693 found in our sample. If 5% of the 693 were false, the threshold established of 34.7 genes is exceed in 9 out of 1000 experiments, slightly less than 1%. Differential expression of gene isoforms was performed analogously.
