In each cohort, association testing was based on an additive logistic regression model using PLINK.55 As covariates we used a subset of the first 20 principal components (PCs), derived within each cohort. By default, we included the first 4 PCs and thereafter every PC that was nominally significantly associated (p<0.05) with case-control status. We conducted a meta-analysis of the results using a standard error inverse-weighted fixed effects model. For chrX, gene dosages in males were scored 0 or 2, in females, 0/1/2, then association analysis was conducted separately for males and females and meta-analysed. We summarized the associations as number of independently associated index SNPs. Index SNPs were LD independent and had r2 < 0.1 within 3 Mb windows. We recorded the left and rightmost variant with r2 <0.1 to an index SNP to define an associated clump. To define loci, we added a 50kb window on each side of the LD clump and combined overlapping LD-clumps into a single locus. Due to the strong signal and high LD in the MHC region, only one SNP was kept from the extended MHC region (chr6:25–35Mb).
