allele frequency) to replace the missing value for that SNP. We used these ethnic-specific per allele summary odds ratios and the total number of risk alleles among control subjects to estimate the distribution of relative risks conveyed by all risk alleles. To avoid making the reference group carriers of zero risk alleles (a group which does not exist) we centered the distribution on the mean number of risk alleles observed in the control population (18.5). The log relative risk for each subject was calculated as logRR = (RA−18.5)×log(ORi) (where RA is equal to the subject's total risk alleles and log(ORi) is the log of the ethnic specific per allele odds ratio. A spline function was used to capture the shape of the distributions of log OR for display purposes. Two variants in KCNQ1 were included in the risk modeling because both were significantly associated with T2D when co-modeled (results were similar when only the most significant of the two, rs2237897, was included). The variant in FTO was excluded from risk modeling procedures, as we found (as have others) that it is not a risk factor for diabetes independent of its effect on obesity.
