Basic demographic and phenotypic characteristics by study site for the 3,531 subjects included in the analysis are summarized in . ARIC participants from Jackson had higher BMI, were less likely to have attended college and reported lower incomes, physical activity and dietary fat intake than those from Forsyth County (all p < 0.05). The eliminated individuals (n =735, see section “Study populations”) did not significantly (all p > 0.05) different from those individuals included in the analysis in BMI, weight, waist circumference, hip circumference, and triceps skinfold thickness (data not shown), except that they had slightly higher waist-to-hip ratio (0.93 ± 0.07, p = 0.002) and thinner subscapular skinfold thickness (31.0 ± 13.4 mm, p = 0.006) than those included.
Baseline Characteristic of the Study Participants by Study Site
The distribution of estimated PEA in the African-American participants appeared right-skewed (). Overall, the median (interquartile range, IQR) PEA was 0.151 (0.115) in all participants, 0.150 (0.111) in Jackson participants, and 0.172 (0.136) in Forsyth County participants (p for the difference between the two sites = 0.004, by Wilcoxon rank-sum test). Although the PEA was higher in participants from Forsyth County, there were highly overlapping distributions of ancestry between the two ARIC study sites based on principal component analysis for detection of substructure (). We found a very high correlation between individuals' first eigenvector (axis of variation) and the estimated PEA (rs = 0.99, p < 0.0001); thus the first axis of variation may well reflect continental origins and ancestry effects.
Figure 1 Histograms of estimated proportions of European ancestry (PEA) in the ARIC African Americans. The PEA was determined using the ANCETRYMAP software for (A) 3,126 individuals in Jackson, MS, (B) 405 in Forsyth County, NC and in (C) all 3,531 subjects. Four (more ...)
Figure 2 The top two axes of variation (eigenvector) from principal component analysis in the ARIC African Americans and the HapMap YRI and CEU samples. The first axis of variation may reflect continental origins. This analysis found highly overlapping distributions (more ...)
In the ARIC participants, we observed a significant trend (p < 0.0001) towards decreased PEA from 0.164 (0.128) in the non-obese individuals (defined as BMI < 25), to 0.156 (0.119) in overweight individuals (BMI 25-29.9), and 0.143 (0.103) in the obese individuals (BMI ≥ 30). The correlations between PEA and each of the eight continuous obesity-related traits are shown in . When correcting for age, sex, and study site, we found that all traits, except waist-to-hip ratio, were significantly correlated with PEA. Additionally adjusting for SES weakened most of the associations, but did not abolish them all. That is, BMI (p = 4.3 × 10-5), weight (p = 2.3 × 10-4), hip circumference (p = 0.009) and subscapular skinfold thickness (p = 0.001) were still significantly, inversely correlated with PEA. In contrast, waist circumference adjusted for BMI was significantly, positively correlated with PEA (p = 3.6 × 10-4). The evidence of significant correlations for these five traits persisted and the strength of the correlations was similar, even after additionally correcting for other covariates, including dietary fat and physical activity during leisure time (data not shown). To minimize the potential impact of diabetes on obesity-related traits, we further examined these correlations among only non-diabetic individuals (n = 2,780) and found the strengths of correlations with BMI, weight, and hip circumference to be stronger (i.e., higher absolute value of rs). For waist circumference adjusted for BMI, the correlation appeared to be weaker in non-diabetic individuals.
Correlations between Obesity-related Traits and Proportions of European Ancestry
Although there were differences between the two study sites for some of the obesity-related traits, the site-specific correlations between the traits and PEA were similar for all traits except for waist-to-hip ratio, which was not significantly associated with PEA in either site (see Supplementary Table 2
). Therefore, in the present study, we pooled samples from two sites to increase study power for the above correlation analysis and the following admixture scans.
Admixture mapping scans were performed on each of the eight obesity-related traits (adjusted for age, sex, study site and SES). Using the top 30% of participants with the highest values as cases and the bottom 30% as controls, we found genome-wide suggestive significant evidence of associations with BMI (). The strongest association for BMI was at 27.3 Mb on chromosome 2 (2p23.3) between rs13025681 and rs7593448. The peak locus-specific LOD was 4.11 (), which meets our priori defined thresholds for suggestive significance (LOD = 4). Averaging 10 to the power of the LOD scores across all loci in the genome and taking the log-base-10 in of this average produced a genome-wide association score of 1.14, again meeting our threshold of 1 for suggestive significance accounting for multiple comparisons. The most extreme case-control statistic (Z score) in the genome is -5.09 (Bonferroni-adjusted p = 3.6 × 10-4, and was exactly at the same location as the peak locus specific LOD of 4.11.
Summary of Locus-Genome and Case-Control Statistic from the Admixture Mapping Scans of BMIa
Figure 3 Genome-wide admixture mapping scans for loci underlying BMI. (A) A suggestive significant signal (locus-specific LOD = 4.11) was detected at 2p23.3 for BMI in all subjects. Averaging the LOD scores across all loci in the genome, we obtained a genome-wide (more ...)
We also performed admixture scans on non-diabetic subjects only. The peak locus-specific LOD increased from 4.11 to 4.65 () and the genome-wide LOD score increased from 1.14 to 1.73, despite the smaller sample size. The best case-control Z score is -5.41 (Bonferroni-adjusted p = 6.3 × 10-5). For weight, a peak locus-specific LOD of 4.07 was identified at the same location with BMI at 2p23.3, with a corresponding genome-wide LOD score of 1.22 (), again reaching the thresholds for suggestive significance for a genome-wide analysis. The case-control statistic for weight at this peak is -5.11, corresponding to a Bonferroni-adjusted p = 3.2 × 10-4. We failed to find any evidence of associations with the other six obesity-related traits. However, the best locus-specific LOD scores for waist circumference and hip circumference were also at 2p23.3.
Summary of Locus-Genome Statistic from the Admixture Mapping Scans of Obesity-Related Traits except BMI
We next performed an analysis to determine whether the observed association on 2p23.3 was due to our case-control definition, by examining the locus-specific LOD score for each individual. As expected, we found that the evidence of association to BMI was contributed mostly by the subjects with the highest BMIs (those in the top 15%–45% of BMI) (). The cumulative score increases gradually after the top 15% of adjusted BMI values, reaches its maximum for the top 30%, and then drops after the top 45%. We also ranked all non-diabetic subjects according to their adjusted BMI values. Similarly, we found that the admixture association is contributed by the top 12%–50% of non-diabetic subjects. The cumulative LOD score is generally higher in non-diabetic subjects after the top 12% (), which suggests the admixture association is stronger when we exclude diabetic subjects.
Figure 4 The observed association was contributed mostly by the top 15% to 45% of subjects with the highest BMI, adjusted for age, sex, and study site and socioeconomic status. The software ANCESRRYMAP was used to calculate the LOD score of disease association (more ...)
To further examine the robustness of the admixture-generated signal on 2p23.3 and whether the signal on 2p23.3 contributes to the overall association between global PEA and BMI, we next carried out a series of linear regression analysis to assess the association between the normal-quantile transformed BMI and both global PEA and local estimates of European ancestry at the 2p23.3 peak in all individuals (). As expected, global PEA was strongly associated with transformed BMI (p = 7.3 × 10-5; Model 1 in ). Similarly, local European ancestry alone was strongly, inversely associated with transformed BMI (p = 3.2 × 10-8; Model 2). To assess how much of the association between global PEA and transformed BMI was accounted by association with ancestry at 2p23.3, we modeled transformed BMI as a function of both global and local European ancestry (Model 3). We found that the locus European ancestry at the 2p23.3 peak almost eliminated the association between BMI and individual global PEA. Conversely, after adjustment for each individual's global PEA, there was still significant evidence of residual association (p = 2.9 ×10-5) between transformed BMI and the local European ancestry (Model 3). Each additional copy of a European ancestral allele at the 2p23.3 peak was associated with a BMI decrease of 0.16 Z-score units on average (equivalent to ~0.92 kg/m2). Similarly in non-diabetic subjects, after adjustment for global PEA, the residual association with the local European ancestry remained significant (p = 2.5 × 10-6) and the size effect of the local ancestry was greater than that in all subjects.
Linear Regression Models of BMIa on Global European Ancestry and Local European Ancestry at 2p23.3 Peak