We used several conservative criteria to include newly discovered SNPs in our database. First, to avoid the winner's curse bias, we used the OR and p-value from the replication stages of the discovery GWAS. Second, when several replication stages from the same GWAS were available, the OR from the stage with largest sample size was recorded. Only when no replication stages were available did we use the OR from the GWAS stage. Third, SNPs associated uniquely in sex-specific analyses were excluded. Fourth, ORs coming from allelic tests and additive models were prioritized over genotypic tests and other genetic models. Fifth, the genome-wide significance level for a newly discovered SNP to be included in our analysis was set at P<5×10−7, unless imputed SNPs were used in the GWAS, in which we toughened up the threshold to P<5×10−8. Sixth, for genomic regions with several genome-wide significant SNPs (SNPs less than 200 Kb from each other), we included in the study the SNP with lowest p-value. Finally, disease-associated SNPs from the MHC region and HLA alleles were not included in the study. In
