Association tests in the Phase 1 and Phase 2 GWAS datasets were performed using linear or logistic association models embedded in generalized estimating equations (GEE) to correct for correlations among related individuals (18). The Phase 3 sample containing unrelated individuals only was evaluated using linear and logistic models. All models were adjusted for age, sex, and the first three PCs of ancestry. The data were analyzed separately within population groups and the results were combined by meta-analysis using the inverse variance method implemented in the computer program METAL (19). In Phase 2, Phase 1 SNPs with p-values<10−4 in either population were tested for association in the SAGE dataset in the specific population group(s) and statistical model(s) yielding the specified result in Phase 1. Tests were restricted in this manner to minimize their number and because it is unlikely that a SNP not meeting the cut-off for follow-up would attain a noteworthy significance level when combining results from the discovery and replication samples. Phase 1 and 2 results were combined within population groups by meta-analysis. In Phase 3, we selected for
