However, we noticed that two of the phenotypes, rheumatoid arthritis and type 1 diabetes, show significant deflation of test statistics beyond the 95% confidence interval (λ = 0.965 for rheumatoid arthritis, λ = 0.946 for type 1 diabetes). This is not unexpected, considering that a substantial fraction of the phenotypic variance in these autoimmune diseases is explained by the HLA loci, leading to inaccurate estimation of variance parameters under the null hypothesis when the HLA effect is not accounted for. In fact, the set of genome-wide significant SNPs (P < 7.2 × 10−8; ref. 32) in this region account for 47% and 60% of the phenotypic variance of rheumatoid arthritis and type 1 diabetes, respectively6. We re-estimated the variance parameters by conditioning on the 57 and 134 SNPs within the extended human MHC region36 that explain more than 1% of phenotypic variance of rheumatoid arthritis and type 1 diabetes, respectively (as described in the Supplementary Note). As a result, the genomic control λ increased to 0.989 for rheumatoid arthritis and 0.991 for type 1 diabetes. We performed this conditioning procedure
