For the meta-analysis, to create genotypic data for the same SNPs for all datasets, we imputed data for each sample for HapMap II SNPs that were not genotyped in that sample, using MACH 1.031 (autosomal SNPs) or IMPUTE32 (X chromosome). For each dataset, imputation was based on SNPs that passed QC for both cases and controls. MACH and IMPUTE are two of several available methods with similar accuracy.33 Using a Hidden Markov Model algorithm with phased CEU HapMap haplotypes as training data, a non-integer “allele dosage” is assigned to each individual for each SNP based on weighted probabilities of possible genotypes. For each SNP, an r2 value estimates concordance with actual genotypes (and thus the predicted concordance with the association tests they would produce). A low r2 predicts greater variance in the concordance of genotypes and of test statistics. This uncertainty is taken into account in the meta-analysis procedure. SNPs have been excluded from analysis if MAF was less than 1% in any dataset or if imputation r2 was less than 0.3. This threshold was used in four previous large
