NHANES III was a nationally representative sample of the non-institutionalized civilian U.S. population collected using stratified multistage probability sampling. NHANES participants underwent a physical examination, phlebotomy, and a household interview [20]. This study was limited to non-diabetic patients (aged 20 or older) with 8–23 hours of fasting prior to blood sampling. Blood from NHANES III Phase II (1991–1994) participants aged 12 or older were used to generate Epstein-Barr transformed lymphocyte cell lines for DNA extraction. Mortality data (death within a mean of 13.5 years of follow-up) were merged from the NHANES III mortality-linked data file. Race-ethnic group was assigned based on self-report. The survey asked each subject to categorize his/her race as “white,” “black,” or “other” and his/her ethnicity as “Mexican-American,” “other Hispanic,” or “not Hispanic.” Of 3,894 individuals with complete data for analysis, we excluded 149 who were not of NHB, MA or NHW race-ethnicity and 704 with diabetes (293 NHW, 167 NHB and 244 MA), leaving 901 NHB, 909 MA, and 1,231 NHW individuals in the analysis. Written informed consent was obtained from all subjects and this study was approved by the National Center for Health Statistics (NCHS) Ethics Review Board.

Individuals with diabetes were excluded to avoid the confounding effects of treatment on HbA_1c. We defined diabetes as a fasting plasma glucose ≥ 7.0 mmol/L, report of a diagnosis of diabetes or use of hypoglycemic medications. HbA_1c levels were measured using HPLC (Bio-Rad DIAMAT glycosylated hemoglobin analyzer system) [21].

Genotyping was performed using Sequenom iPLEX. We genotyped 11 SNPs at ten loci shown among white non-diabetic individuals in MAGIC to have genome-wide significant association with HbA_1c.[7] We used SNP rs282606 as a proxy for ATP11A rs7998202 (CEU r² = 1.0), SNP rs10830956 as a proxy for MTNR1B rs1387153 (CEU r² = 1.0), and rs2022003 as a proxy for SPTA1 rs2779116 (CEU r² = 0.927) [r² for ASW and MEX populations not available]. The minimum call rate for genotyping was 95%. Allele frequencies of all SNPs were in Hardy Weinberg Equilibrium (HWE) based on National Center for Health Statistics standards (HWE rejected if p < 0.01 in ≥ 2 or more race-ethnic groups). We compared NHANES observed allele frequencies with those available from HapMap (http://hapmap.ncbi.nlm.nih.gov/, Release 27, Phases II and III, NCBI build 36), comparing NHW with CEU (Utah residents with Northern and Western European ancestry from the CEPH collection), NHB with ASW (African ancestry in Southwest USA), and MA with MEX (Mexican ancestry in Los Angeles, California).

We calculated a genotype risk score to test the collective association with HbA_1c of 11 SNPs at 10 loci (2 uncorrelated SNPs at ANK1). We assumed that each SNP was associated with HbA_1c based on previous association results in whites, despite potential ancestral differences in NHB or MA in linkage disequilibrium (LD) patterns [22]. Since we did not know the effect size of the MAGIC SNPs in non-white populations, we did not apply SNP-specific weights to account for SNP-specific differences in effect on HbA_1c, but simply summed the presence of 0, 1, or 2 risk alleles carried by individuals at each SNP. In addition to the 11-SNP GRS, we also performed a secondary analysis using an eight SNP “non-glycemic” risk score by excluding the three glycemic loci (G6PC2, GCK, MTNR1B) for score calculation.

We stratified the analysis by race-ethnicity (NHB, MA, and NHW) and to estimate rates and proportions within groups used weights to account for sampling probabilities using methods previously described [23]. P-values for differences across race-ethnic groups were calculated using Satterthwaite adjusted- F statistics for continuous variables and chi-square tests for categorical variables. To estimate the significance of differences in allele frequencies across groups we used Fisher’s Exact tests.

To investigate the relationship between SNPs and HbA_1c level we used linear regression and an additive genetic model adjusted for age and sex. We included one SNP at a time in the models for individual SNP associations with HbA_1c, with genotypes coded as 0, 1 or 2 depending on the number of HbA_1c-raising alleles present. To study the collective effect of the 11 SNPs on HbA_1c we used linear regression adjusted for age and sex, totaled the number of risk alleles at all 11 SNPs to calculate a risk score, and tested associations of a per-risk-allele increase in genotype risk score with HbA_1c. We calculated the adjusted model R² with and without the genotype risk score for each group to determine the percent variance in HbA_1c explained by genetic effects. The same procedure was carried out for the 8 SNP “non-glycemic” risk score, as well as for genetic associations with mortality (percent dead as of 13.5 years post-baseline exam). To determine if a significant genetic risk score x ethnicity interaction effect on HbA_1c exists, we also applied the following linear regression model on the whole sample: Hba_1c level (outcome) = sex, age, genetic risk score, ethnicity, genetic risk score x ethnicity interaction. For tests of association with mortality we used logistic regression to estimate the odds of mortality with per-risk-allele increase in HbA_1c. For analysis of mortality, Cox models yielded similar results to logistic regression, so Cox model results are not shown. We also applied the following logistic regression model on the whole sample: mortality (outcome) = sex, age, GRS, ethnicity, GRS x ethnicity interaction. For the analyses we used SUDAAN (version 10.0) [24] and SAS (version 9.2, SAS Institute Inc, Cary, NC). We considered p values less than 0.05 to indicate statistical significance, based on one test per previously established SNP at each locus for each hypothesis (SNP is associated with HbA1c; SNP is associated with mortality).

To evaluate inter-ethnic differences in LD near the SNPs, we examined 500 kb around each SNP (HapMap Release 27, Build 36, phases II and III) for four populations (CEU, YRI, ASW, and MEX). Using Haploview version 4.2, [25] we counted the number of “Gabriel” LD regions (based on confidence intervals) [26] in that region for each population. We investigated natural selection around the ten loci using Haplotter [27] and HapMap Phase II data. Standardized Integrated Haplotype Score (iHS) (a statistic based on differential LD around positively selected alleles that compares haplotype length with ancestral allele versus derived allele to detect positive selection) [27], Fay and Wu’s H + statistic (a measure used to scan a region for allele frequencies that are skewed from the neutral model) [28] and the Fixation Index (F_ST) (a statistic using allele frequencies to measure genetic divergence between subpopulations) [29] were obtained through Haplotter SNP queries spanning 2 Mb regions at each locus.

NHW individuals were older, had lower BMI and lower mean HbA_1c than did NHB and MA individuals (global p values all <0.0001, Table 1).

Risk allele frequencies across the 11 loci varied widely within the three race-ethnic groups ( Additional file 1 Table S1). Six out of 11 HbA_1c–associated SNPs had risk allele frequencies that differed significantly across race-ethnic groups (Fisher’s p <0.0002). At five of these six loci, risk allele frequency of NHB was most divergent, including SNPs near ANK1 (two uncorrelated SNPs), MTNR1B, ATP11A/TUBGCP3 and TMPRSS6. At the SNP near SPTA1, risk allele frequency differed most in MA. The HbA_1c -raising allele was the minor (less frequent) allele in all three ethnic groups for SNPs near SPTA1, GCK, MTNR1B, FN3K, and TMPRSS6. The HbA_1c –raising allele was the major (more frequent) allele at SNPs near ABCB11, HFE, ANKI (rs6474359) and HK1. At two loci, ATP11A, and ANK1 (rs4737009), the HbA_1c –raising allele was the minor allele in NHW and MA, but the major allele in NHB ( Additional file 1: Table S1). Risk allele frequencies observed in this study and those available from HapMap were generally similar, although at some loci minor dissimilarity with HapMap was observed in NHB and MA cohorts (Figure 1; Additional file 1: Table S2).

Though single-SNP associations are underpowered ( Additional file 1: Table S3), we did observe that in NHW, eight of the 11 SNPs in NHW were consistent with Soranzo et al. (2010) in having a positive risk effect on HbA_1c levels, with three of the SNPs used in our analysis (rs282606 [ATP11A], rs10830956 [MTNR1B], and rs2022003 [SPTA1]) serving as proxies for those in the MAGIC study. Beta coefficients were negative for three, three and four of the 11 SNPs in NHW, NHB and MA groups, respectively, but corresponding SNPs did not generate significant associations (Table 2). Three out of 11 HbA_1c-associated SNPs had nominally significant (p < 0.05) associations with HbA_1c levels in at least one of the three race-ethnic groups, but altogether only four of the 33 possible associations (11 SNPs x three race-ethnic groups) were significant (p < 0.05). No significant associations were observed in NHB. Two HbA_1c-SNPs produced a significant association only in NHW (both SNPs at ANK1), and one produced a significant association in both NHW and MA (rs855791 near TMPRSS6).

The mean 11-SNP genotype scores (actual scores ranged from 1–18) were 11.0 (± 0.09 [SE]) in NHB, 10.7 (± 0.08) in MA and 10.4 (± 0.07) in NHW, (p value for global difference across race-ethnicity < 0.0001, Table 3). Median genetic risk scores (unweighted) were 11.0 (SD =2.2), 11.0 (SD =2.3) and 11.0 (SD =2.0) in NHW, NHB and MA, respectively, with distributions of genetic risk scores negatively skewed toward a lower score in all three ethnic groups. The per-risk allele increase in the score was significantly associated with HbA_1c levels in NHW and MA, but not NHB. When comparing the top and bottom 10% of the genotype score distribution for each race-ethnic group, the smallest difference in HbA_1c was observed in NHW (NHW: 0.49%; NHB: 0.56%; MA: 0.54%). The genotype score explained very little of the variance in HbA_1c levels in NHB (0.0005%) compared with NHW (0.0016%) and MA (0.0121%). Variance explained in NHW is comparable to the previously published value [7]. We observed no significant genetic risk score x ethnicity interaction on HbA_1c level (p = 0.68).

The mean “non-glycemic” 8-SNP genotype scores (actual scores ranged from 4–15) were 8.80 (± 0.06[SE]) in NHB, 8.72 (± 0.06) in MA and 8.41(± 0.06) in NHW, (p value for global difference across race-ethnicity < 0.0001) ( Additional file 1: Table S4). The per-risk allele increase in the score was significantly associated with HbA_1c levels in NHW, but not in NHB and MA.

Mortality rates differed between race-ethnic groups (Table 4) with a higher mortality rate observed in NHB (19.4%) compared with NHW (12.8%) and MA (14.5%). The 11-SNP genotype score was not associated with mortality in any race-ethnic group. We observed no significant genetic risk score x ethnicity interaction on mortality (p=0.62). Power calculations for the mortality analysis are provided in Additional file 1: Table S5.

There were consistently fewer LD regions in the CEU population compared to YRI at every locus (YRI:CEU): SPTA1 42:25; ABCB11/G6PC2 49:27; HFE 31:17; GCK 28:21; ANK1 (rs41668351) 35:29; ANK1 (rs41749562) 30:25; HK1 45:38; MTNR1B 39:22; FN3K 42:27; TMPRSS6 78:52; ATP11A/TUBGCP3 49:31 ( Additional file 1: Table 6). ASW, which represents a population with African ancestry in the southwestern United States, only had higher numbers of LD regions compared to CEU in two out of 11 regions, possibly due to lower coverage of ASW compared to CEU (and YRI) in HapMap Release 27.

Fay and Wu’s H + was highly skewed at two loci (HK1 and ATP11A) in CEU ( Additional file 1: Table S7). Integrated haplotype scores (iHS) were not highly negative or positive at these SNPs, as would be characteristic for regions undergoing recent natural selection. F_ST, a measure of the amount of allelic fixation due to drift, was greater than 15% at ANK1 and ATP11A in both CEU and YRI, suggesting population differentiation at these loci [29]. Haplotter queries by gene did not reveal evidence of natural selection directly at the genes queried, but evidence of natural selection was observed within a 2 Mb region of ABC11/G6PC2 and TMPRSS6 for CEU and YRI, respectively.