Plasma glucose levels were measured by hexokinase (COBAS MIRA; Roche Diagnostics, Montclair, NJ). We used HOMA as a surrogate measure of β-cell function (HOMA-B; calculated as [20 × fasting insulin {μIU/ml}/FG {mmol/l} − 3.5]) (14).

We defined IFG as having a fasting plasma glucose ≥5.6 mmol/l, but <7.0 mmol/l and diabetes as use of antidiabetes medications (including insulin) or a fasting plasma glucose ≥7.0 mmol/l (4).

We genotyped 16 SNPs that were confirmed as associated with FG levels among nondiabetic people of European ancestry in the MAGIC GWAS (10). We used SNP rs573225 as a proxy (r² = 1.0 in CEU) for SNP rs560887 in G6PC2 and SNP rs11558471 as a proxy (r² = 1.0 in CEU) for SNP rs13266634 in SLC30A8. Genotyping was performed with the use of iPLEX (Sequenom) (15). The minimum call rate was 98.1%, and all SNPs were in Hardy-Weinberg equilibrium (HWE) according to the National Center for Health Statistics criterion (if P < 0.01 in two or more race/ethnicity groups, HWE is rejected).

We constructed a weighted genetic risk score (GRS) to examine the combined effect of 16 SNPs on FG levels, HOMA-B, and risk for IFG. We used the β-coefficients from the published European ancestry GWASs of these FG-associated SNPs (10). We multiplied these β-coefficients by 0, 1, or 2, according to the number of risk alleles carried by each individual, and then summed them. To facilitate the comparison with the simple GRS (summing the number of risk alleles), we divided the score by 0.948 (twice the sum of the β-coefficients) and multiplied by 32 (total number of alleles) (16). Although a number of SNPs did not show significant association with FG or HOMA-B in the NHANES III, we assumed, on the basis of GWAS results, that each SNP is independently associated with FG for computation of a weighted GRS. This assumption might not be appropriate if an index SNP is correlated with the causal SNP in the discovery population but not so in other racial/ethnic groups due to differences in linkage disequilibrium patterns (17). We used an additive genetic model for each SNP and applied a linear weighting of 0, 1, or 2 to genotypes containing 0, 1, or 2 risk alleles (16). We fit the weighted GRS as a continuous variable and categorized it into quintiles in multivariate analyses. In presenting the results, we rounded the weighted GRS to the whole number.

We calculated weighted allele frequencies and their 95% CIs, stratified by race/ethnicity (non-Hispanic white, non-Hispanic black, and Mexican American) using the NHANES sample weights. We carried out HWE tests for each genetic variant for each racial/ethnic group. To test for differences in allele frequencies, we pooled all racial/ethnic groups and used χ² to test the differences across the racial/ethnic groups. We used linear regression modeling to evaluate the relationship between 16 SNPs and FG levels and HOMA-B, adjusted for age and sex. We examined the adjusted marginal effect on FG by including one SNP at a time in the model for each racial/ethnic group. SNPs were coded as 0, 1, or 2 for those who were noncarriers, heterozygous, or homozygous for the risk (FG raising) alleles, respectively. We modeled the GRS in a similar fashion by testing associations of a per–risk allele increase with FG. For the GRS models, we calculated adjusted R², with and without the GRS to determine the explained variance, by adding the GRS in the models for each racial/ethnic group. We tested for homogeneity of effects of these SNPs in marginal and GRS models across race/ethnicity by including the interaction terms of SNP or GRS with race/ethnicity in the regression models using the pooled dataset and reported the associated P value.

We used age- and sex-adjusted logistic regression to evaluate the association of the GRS with prevalence of IFG for each racial/ethnic group. We modeled the GRS to test the association of a per–risk allele increase with odds of IFG. We also categorized the GRS into quintiles to examine the trend of effect on risk of IFG. We tested for homogeneity of the GRS effect on IFG across race/ethnicity by including the interaction term using the pooled dataset. Statistical tests, using linear or logistic regression modeling, were based on Satterthwaite adjusted-F statistics. All tests were two sided, and a P value of <0.05 at each locus or the GRS was considered significant. To avoid nonresponse bias among NHANES III participants for whom DNA were available, the sample weights were recalculated for the NHANES III DNA dataset by using previously described methods (18). We used SUDAAN (version 10.0) and SAS (version 9.2; SAS Institute, Cary, NC) statistical software for our analyses to account for the complex sampling design.

All allele frequencies (16 SNPs) were significantly different across racial/ethnic groups (P < 0.05). Risk allele frequencies varied from 17.1% (rs4607517, GCK) to 86.1% (rs10885122, ADRA2) among non-Hispanic whites, from 7.9% (rs10830963, MTNR1B) to 93.7% (rs7944584, MADD) among non-Hispanic blacks, and from 18.0% (rs7903146 (TCF7L2) to 86.5% (rs11920090, SLC2A2) among Mexican Americans (Table 2) (online appendix Table 1, available at http://care.diabetesjournals.org/cgi/content/full/dc10-0898/DC1).

We observed three SNPs (rs573225, rs780094, and rs11558471) that were significantly associated with FG levels among non-Hispanic whites, one SNP (rs10830963) among non-Hispanic blacks, and three variants (rs573225, rs4607517, and rs10830963) among Mexican Americans (online appendix Tables 2 and 3). However, the majority of β-coefficients (37 of 48) showed positive associations in the expected direction with FG levels among all three racial/ethnic groups. We found no evidence of heterogeneity effects of these SNPs on FG levels across racial/ethnic groups, except for rs11885122 for which non-Hispanic whites had lower FG and non-Hispanic blacks and Mexican Americans had higher FG levels (P = 0.009). Generally, similar patterns of associations of these 16 SNPs were seen for HOMA-B (online appendix Table 2).

The number of risk alleles at 16 SNPs ranged from 10 to 27 (median 19), 12 to 26 (median 20), and 10 to 26 (median 18) among non-Hispanic whites, non-Hispanic blacks, and Mexican Americans, respectively. For the weighted GRS, the corresponding numbers were 8–27 (median 17), 10–24 (median 18), and 9–26 (median 17), respectively. The GRS was normally distributed in all three racial/ethnic groups (Kolmogorov-Smirnov P < 0.001) (online appendix Fig. 1). FG increased significantly across the quintile of GRS for all racial/ethnic groups. Each additional risk allele in the score was associated with a 0.022, 0.036, and 0.033 mmol/l increase in FG among non-Hispanic whites, non-Hispanic blacks, and Mexican Americans, respectively (Table 3). There was no evidence of a different effect of the GRS on FG levels or HOMA-B across race/ethnicity (P = 0.291 and 0.637 for testing heterogeneity of the GRS on FG levels and HOMA-B across race/ethnicity, respectively). In a model including age and sex, the GRS explained more of the total variation in FG in Mexican Americans and in non-Hispanic blacks than in non-Hispanic whites (Table 3). Generally, similar patterns of associations of these 16 SNPs were seen for HOMA-B (Table 3).

The prevalence of IFG, adjusted for age and sex, was lowest among non-Hispanic blacks and highest among Mexican Americans. The adjusted odds ratios for IFG were highest among Mexican Americans, comparing people in the highest with those in the lowest quintiles of the GRS for all racial/ethnic groups (Table 4) (online appendix Fig. 2). There was no evidence of a different effect of the GRS on risk for IFG across race/ethnicity (P = 0.365 for testing heterogeneity across race/ethnicity). Each additional allele was associated with a significant increased risk of IFG for all racial/ethnic groups.