Population Structure, Admixture, and Aging-Related Phenotypes in African American Adults: The Cardiovascular Health Study

Abstract

U.S. populations are genetically admixed, but surprisingly little empirical data exists documenting the impact of such heterogeneity on type I and type II error in genetic-association studies of unrelated individuals. By applying several complementary analytical techniques, we characterize genetic background heterogeneity among 810 self-identified African American subjects sampled as part of a multisite cohort study of cardiovascular disease in older adults. On the basis of the typing of 24 ancestry-informative biallelic single-nucleotide–polymorphism markers, there was evidence of substantial population substructure and admixture. We used an allele-sharing–based clustering algorithm to infer evidence for four genetically distinct subpopulations. Using multivariable regression models, we demonstrate the complex interplay of genetic and socioeconomic factors on quantitative phenotypes related to cardiovascular disease and aging. Blood glucose level correlated with individual African ancestry, whereas body mass index was associated more strongly with genetic similarity. Blood pressure, HDL cholesterol level, C-reactive protein level, and carotid wall thickness were not associated with genetic background. Blood pressure and HDL cholesterol level varied by geographic site, whereas C-reactive protein level differed by occupation. Both ancestry and genetic similarity predicted the number and quality of years lived during follow-up, but socioeconomic factors largely accounted for these associations. When the 24 genetic markers were tested individually, there were an excess number of marker-trait associations, most of which were attenuated by adjustment for genetic ancestry. We conclude that the genetic demography underlying older individuals who self identify as African American is complex, and that controlling for both genetic admixture and socioeconomic characteristics will be required in assessing genetic associations with chronic-disease–related traits in African Americans. Complementary methods that identify discrete subgroups on the basis of genetic similarity may help to further characterize the complex biodemographic structure of human populations.

Introduction

Genetic-association studies are often performed in population samples of unrelated individuals to identify susceptibility loci for complex human traits. If subjects are sampled from two or more subpopulations for which the frequencies of marker alleles and traits differ, spurious associations may arise due to confounding by population substructure (Pritchard et al. ^2000b; Schork et al. ²⁰⁰¹; Risch et al. ²⁰⁰²). On the other hand, the increased extent of linkage disequilibrium between markers on the same chromosome, created by population admixture, may actually facilitate genome mapping of complex trait genes when exploited appropriately in the design of a study (Chakraborty and Weiss 1988; McKeigue ¹⁹⁹⁸).

Prior studies assessing population stratification have primarily considered the impact of population subdivision and ancestral admixture proportions. Additional population genetic factors, however, may contribute to genetic background heterogeneity (Schork et al. 2001). Variation in allele frequencies as a result of genealogical differences between people in a sample may occur even in the absence of overt admixture. In addition, time-dependent population shifts, due to environmental or socioeconomic factors that influence migration or mating patterns, might create genetic heterogeneity across different age groups. These demographic movements may be especially relevant for studies of older adults in assessing complex diseases related to aging, as well as interindividual variation in life span or longevity (Yashin et al. 1999).

Biodemographic factors contributing to population heterogeneity and substructure are particularly important for genetic-association studies involving African Americans, among whom admixture with whites and Native Americans varies by geographic region (Parra et al. 1998, 2001; Pfaff et al. ²⁰⁰¹; Smith et al. ²⁰⁰⁴). Among older adults in the United States, African Americans have a higher prevalence of cardiovascular disease (CVD) risk factors (Hutchinson et al. 1997; Kuller et al. ¹⁹⁹⁸; Sundquist et al. ²⁰⁰¹) and also greater clustering of CVD risk factors (Sharma et al. 2004), compared with non-Hispanic whites.

In light of the potential for confounding due to population stratification (Kittles et al. 2002; Freedman et al. ²⁰⁰⁴), as well as the opportunity for efficient genetic mapping of complex diseases by admixture linkage disequilibrium, we empirically evaluated the influence of population stratification on several common chronic-disease and aging-related phenotypes in a multicenter African American cohort. Our results show that there is substantial population admixture and substructure among the African American population, and that controlling for genetic ancestry not only may reduce false-positive associations but also may uncover a true association previously obscured by stratification. Our findings also show that controlling for social economic status, in addition to population stratification, is necessary in assessing genetic associations with chronic-disease–related traits in African American subjects.

Methods

Study Subjects

Study subjects were self-identified African American men and women aged ⩾65 years old who participated in the Cardiovascular Health Study (CHS) (Fried et al. 1991). CHS participants were recruited from lists of Medicare beneficiaries in four U.S. communities: Winston-Salem, NC; Pittsburgh, PA; Washington County, MD; and Sacramento, CA. The original CHS cohort, recruited from 1989 to 1990, included 246 African American participants. A second cohort of 678 African American participants was recruited from 1992 to 1993. Of 924 total African American participants, 810 are included in the present study. The reason for exclusion was either refusal of consent for genetic testing (n=62) or lack of an available DNA sample (n=52). All procedures were conducted under institutionally approved protocols for study of human subjects, and all subjects provided written informed consent.

Data Collection and Definition of Phenotypes Related to Vascular Disease and Aging

Data collection methods in the CHS have been described elsewhere (Fried et al. 1991). The baseline evaluation included demographic, lifestyle, and medical histories; physical examination; and fasting blood collection (Cushman et al. 1995). Quantitative phenotypes and CVD risk factors that were considered include systolic blood pressure (mm Hg), BMI (kg/m²), fasting blood glucose (mg/dL), HDL cholesterol level (mg/dL), and C-reactive protein (CRP) level (mg/liter). Carotid wall thickness, a quantitative measure of subclinical vascular disease, was defined as the mean maximal intimal-medial thickness of the near and far walls on both the left and right arteries, as determined by high-resolution ultrasonography (O’Leary et al. ¹⁹⁹¹). The outcome “years of life” (YOL) was defined as the number of years a participant was alive during 10 years of follow-up, and “years of healthy life” (YHL) was defined as the number of years the person reported being in excellent, very good, or good health during the 10 years of follow-up. This outcome was derived from standard information on self-rated health status (excellent/very good/good/fair/poor) collected at baseline and every 6 mo during follow-up (Diehr et al. 1998).

Selection of Ancestry-Informative Markers (AIMs) and Genotype Analysis

Twenty-four biallelic SNP markers (table 1) were chosen on the basis of known allele-frequency differences (δ values) between African, European, and Native American populations. A subset of these markers has been characterized and published by Mark Shriver and colleagues (Hoggart et al. 2003; Shriver et al. ²⁰⁰³). Additional markers were selected by identifying markers from dbSNP that had been typed in both European Americans and African Americans and that had a high allele-frequency difference between those populations (δ>0.5). The ancestral allele frequencies were then confirmed by genotyping the markers in populations collected from Sub-Saharan Africa (Nigeria, Central African Republic, and Sierra Leone [n=481]), Europe (Ireland, England, Germany, and Spain [n=243]), and Native American populations indigenous to the United States and Mexico (Maya, Pima, Cheyenne, and Pueblo [n=148]). The ancestral DNA samples were kindly provided by Dr. Mark Shriver. Detailed information regarding the markers characterized by Shriver and colleagues can be found at the dbSNP Web site under the submitter handle “PSU-ANTH,” or, for the newly identified markers, under the submitter handle “HapMap-UCSF-WU-FP-TDI.”

Table 1.

Marker Chromosomal Locations, Ancestral Population Allele Frequencies, and Allele Frequencies and Tests for Hardy-Weinberg Proportions in CHS African American Participants

			Ancestral PopulationAllele 1 Frequencies			Marker FST Values betweenAncestral Populationsa			Findings for CHS African Americans
Marker	Location	Alleles 1/2	African	European	NativeAmerican	African/European	African/ NativeAmerican	European/ NativeAmerican	Allele 1Frequency	Hardy-WeinbergEquilibriumb
rs2814778	1q23.2	A/G	.003	.994	.991	.982	.976	.000	.256	.034
rs930072	5p13	C/T	.960	.096	.447	.749	.315	.156	.731	.045
rs7349	10p11.22	C/T	.039	.873	.956	.701	.841	.022	.214	.0004
rs723632	1q32.3	G/C	.100	.919	.674	.671	.347	.093	.277	.822
rs722098	21q21.1	G/A	.902	.177	.717	.529	.055	.295	.702	.038
rs146026	13q13.1	C/T	.256	.917	.826	.450	.327	.018	.377	.132
rs6003	1q31.3	G/A	.702	.083	.031	.402	.485	.013	.570	.674
rs1985080	7p14.3	G/A	.100	.643	.966	.316	.753	.166	.224	.418
rs518116	9q33.3	G/A	.131	.669	.581	.302	.221	.008	.245	.052
rs3287	2p16.2	G/A	.730	.196	.205	.287	.277	.000	.590	.057
rs1989486	19q13.42	C/T	.045	.578	.404	.331	.185	.030	.219	.120
rs7041	4q13.3	T/G	.928	.413	.451	.300	.266	.001	.815	.223
rs994174	10q23.1	G/A	.758	.246	.264	.262	.244	.000	.667	.062
rs1800498	11q23.2	T/C	.138	.648	.088	.273	.006	.337	.258	.968
rs2816	17p13.1	T/C	.003	.494	.075	.323	.035	.216	.151	.490
rs2891	17p13.2	G/A	.021	.507	.425	.304	.235	.007	.122	.280
rs3188520	20q11.22	G/C	.828	.349	.439	.237	.163	.008	.747	.156
rss1042602	11q14.3	A/C	.004	.467	.053	.298	.022	.223	.090	.012
rs326946	11q23.1	G/T	.609	.167	.067	.206	.328	.024	.482	.305
rs2077863	18p11.21	C/G	.511	.925	.926	.212	.213	.000	.660	.861
rs3188519	4q28.2	C/T	.758	.369	.318	.154	.195	.003	.623	.216
rs594689	11q13.1	A/G	.094	.467	.130	.172	.003	.136	.193	.102
rs2228478	16q24.3	G/A	.508	.136	.043	.158	.271	.027	.393	.458
rs584059	3q23	C/A	.494	.140	.467	.145	.001	.126	.419	.917

^aMarker F_ST values (inverse of the variance of the estimated ancestral contributions) were calculated in accordance with Pfaff et al. (2004).

^bP values <.05 for the test of Hardy-Weinberg equilibrium are shown in bold italics.

The 24 AIMs were distantly spaced throughout the genome so that they offer independent association about genetic background/ancestry. The average distance between adjacent markers on the same chromosome was 26 Mb (range 1–60 Mb). The mean δ value between African and European populations was 0.56 (range 0.36–0.99). The mean allele-frequency differential between African and Native American populations was 0.44 (range 0.03–0.99). The mean allele-frequency difference between European and Native American populations was 0.19 (range 0.001–0.56). Marker F_ST values were calculated as the inverse of the variance of the estimated ancestral contributions, in accordance with Pfaff et al. (2004), and are shown in table 1.

Genotyping assays were performed on blood drawn from 810 CHS African American participants who gave informed consent to DNA preparation and testing. Genotyping was performed using the AcycloPrime-FP (Perkin Elmer) method (Chen et al. 1999) under standard conditions: 5 μl PCR volume with Platinum Taq buffer, 2.5 mM MgCl₂, 2.4–4.0 ng of genomic DNA, 50 μM dNTPs, 0.1 μM of primers, and 0.1 U of Platinum Taq (Invitrogen). Cycling conditions were 95°C for 2 min, followed by 35 cycles at 92°C for 10 s, 58°C for 20 s, and 68°C for 30 s, with a final extension at 68°C for 10 min. PCR products were purified enzymatically, and genotyping extension reactions were performed in accordance with kit directions. The primer sequences for PCR and genotyping extension reactions and any changes to standard conditions are presented in table A1 (online only).

Characterization of Population Structure, Admixture, and Genetic Background Similarity

Exact tests for Hardy-Weinberg equilibrium and linkage disequilibrium and Wright’s hierarchical F statistics (Wright 1951) as estimators of allele-frequency variation under a pure-drift model (Weir and Cockerham 1984) were computed using Genetic Data Analysis, version 1.1 (see Genetic Data Analysis Web site).

Group admixture proportions were estimated from the average coalescent times for a pair of alleles taken from within and between populations by use of the program ADMIX, version 2.0 (Dupanloup and Bertorelle 2001). SEs for the group admixture coefficients were calculated on the basis of 1,000 bootstraps. The proportion of African, European, and Native American ancestry for each individual was estimated by a maximum-likelihood method (Chakraborty et al. 1986) by use of the program IAE3 (Bonilla et al. 2004), kindly provided by Mark Shriver. This program also gives 1-SD support intervals for the estimated ancestral proportions.

We used two Bayesian Markov Chain–Monte Carlo methods to provide complementary information on genetic differentiation between and among populations under nonequilibrium conditions. Population structure and evidence for allelic association between linked markers caused by correlation in ancestry (i.e., “admixture linkage disequilibrium”) were evaluated by estimating the average recombination rate by use of the program STRUCTURE 2.1 (Pritchard et al. ^2000a; Falush et al. ²⁰⁰³), with a burn-in of 50,000 iterations and 1,000,000 iterations. By relaxing the requirement for Hardy-Weinberg equilibrium within geographic subpopulations and by allowing for recent migration, local inbreeding coefficients were estimated using the program BayesAss 1.2 of Wilson and Rannala (2003), which was run for a total of 3,000,000 iterations, including an initial burn-in of 1,000,000 iterations.

A genetic-clustering algorithm based on pairwise, weighted allele-frequency sharing was used to assess genetic background similarity (Schork 2001; Schork et al. ²⁰⁰¹). Allele-sharing matrices were constructed in accordance with the method of Lynch and Ritland (1999), as implemented in the program IDENTIX (Belkhir et al. 2002). The resulting similarity matrices were used in an agglomerative hierarchical cluster analysis with complete linkage, under the assumption of the existence of 2–15 genetically similar groups of individuals within the total sample. To determine the most likely number of groups in the sample, we assigned each individual to his or her most likely genetic subgroup on the basis of his or her allele-frequency profile, and we assessed phenotypic differences across the groups by performing standard ANOVA and nonparametric ANOVA or the Kruskall-Wallis test (Lehmann and D’Abrera ¹⁹⁹⁸). To identify any genetic “outliers” whose genetic background is extremely different from the remaining cohort, we applied the multilocus genotype-based permutation test of Curtis et al. (2002), which was run for 10,000 iterations, with a significance threshold of P ⩽ 1 × 10⁻⁶.

Tests of Associations between Quantitative Traits and Biodemographic Variables

Associations between quantitative traits and biodemographic predictor variables (estimates of individual ancestry, genetic-cluster membership, socioeconomic status, clinic site, or individual AIM genotypes) were assessed by multiple linear regression, by use of the statistical package Stata 8.0. Levels of blood glucose, HDL cholesterol, and CRP were log transformed to reduce skewness and kurtosis. Individual marker genotypes were coded 0, 1, and 2, under the assumption of an additive genetic model. An individual's percentage of African ancestry was coded as a continuous variable, by use of his or her proportion of African ancestry estimated by maximum likelihood. Each clinic site was represented by an indicator variable, and the largest clinic (North Carolina) was omitted from the regression model as the reference group. Similarly, the four genetic-similarity clusters were coded as indicator variables, with cluster 1 as reference. We created categorical variables for education, income, and occupation type as proxies for socioeconomic status (SES). A three-level ordinal categorical variable for education was created by dividing the cohort on the basis of education level (from none to grade 9; high school or general equivalency diploma; or college, vocational, graduate, or professional training). Similarly, a three-level ordinal categorical variable was created on the basis of annual income levels <$8,000; $8,000–$35,000; and >$35,000. For type of occupation, we created three nonordered categories on the basis of a response card that indicated lifetime occupation: professional/technical/managerial/administrative positions and sales/clerical service were classified as “white-collar” occupations; craftsman/machine operator/laborer and farming/forestry work were grouped together as “blue-collar” occupations; and housewife, other occupation, or refusal to answer were combined into the category of “other” occupations.

Covariate-adjusted P values for associations between quantitative traits and population characteristics (clinic site, genetic similarity cluster, individual ancestry, or SES defined by education, income, or occupation) were determined by likelihood-ratio tests. The log likelihood of a “full” regression model containing the variable(s) for a particular characteristic was compared with a reduced model without the characteristic. We adjusted the nominal 5% significance level by the number of traits analyzed (n=8) and used a P value threshold of <.00625 for significance. Mean-adjusted trait values (and 95% CIs) for different levels of predictor variables were calculated from the linear-regression coefficients and SEs, with any additional covariates set to their respective mean values. All analyses were adjusted for age at baseline and for sex. Additionally, we adjusted some analyses for other clinical covariates known to be important for the particular quantitative trait (Hutchinson et al. 1997; Kuller et al. ¹⁹⁹⁸; Sundquist et al. ²⁰⁰¹; Sharma et al. ²⁰⁰⁴). Thus, systolic blood pressure was adjusted for treated hypertension; blood glucose level was adjusted for baseline diabetes; CRP level was adjusted for BMI, smoking, and diabetes; and carotid wall thickness was adjusted for smoking, hypertension, diabetes, BMI, and HDL cholesterol level. Both YOL and YHL were adjusted for hypertension, diabetes, current smoking, BMI, coronary heart disease, cancer, and self-reported health status.

Results

Population Substructure and Admixture in African American Cohort

Of the 24 AIMs tested, 6 (including 4 of the 5 markers having the largest allele-frequency differential between Africans and Europeans) deviated significantly from Hardy-Weinberg proportions (table 1). There was increased homozygosity, both overall (F_IT=0.034; 95% CI 0.016–0.052) and within the four regional subpopulations (F_IS=0.033; 95% CI 0.015–0.050). Even though the markers were unlinked or widely spaced throughout the genome, 170 (60%) of 285 pairwise combinations showed significant allelic association. Together, the excess homozygosity and association between unlinked markers suggest substantial population substructure and admixture in the CHS African American cohort due to continuous gene flow or nonrandom mating. The program STRUCTURE 2.1 showed there was a greater likelihood that the cohort descended from two ancestral populations (log likelihood −20,725) than three (−20,767) or four (−20,823) ancestral populations or than a single homogeneous population (−21,363). Under a linkage model with two ancestral populations, the presence of significant admixture linkage disequilibrium was confirmed (Falush et al. 2003).

The mean proportions (± SEs) of African, European, and Native American ancestry, estimated for the cohort as a whole, were 76.4 ± 0.6%, 20.9 ± 1.2%, and 2.7 ± 1.6%, respectively. We also estimated individual ancestry by using maximum-likelihood, but the mean SEs were much larger (15.6% for African, 17.9% for European, and 20.9% for Native American); these presumably reflect both the wide interindividual variation in degree of admixture and the lack of precision in distinguishing European from Native American ancestry by the current set of 24 markers. Individual African ancestral proportions estimated by use of STRUCTURE under a two-population admixture model were virtually identical to those estimated by maximum likelihood under a three-population model (correlation coefficient 0.98; P<.0001). We also conducted a principal-components analysis and compared the scores that individuals received for the principal components with ancestral proportions calculated by use of the maximum-likelihood model (table 2). We found a very high correlation between an individual’s score on the first principal component and estimated African ancestry (correlation coefficient 0.97; P<.0001). For European and Native American ancestry, the correlations with the first principal component were weaker. The second and third principal components were also weakly correlated with percentage European versus percentage Native American ancestry.

Table 2.

Correlations between Individual Ancestry Estimated by Maximum Likelihood and Principal-Component Analysis

	Ancestral Proportions Estimated by Maximum Likelihood
	African		European		Native American
PrincipalComponent	CorrelationCoefficienta	P Valueb	CorrelationCoefficienta	P Valueb	CorrelationCoefficienta	P Valueb
1st	.9718	.0000	−.8523	.0000	−.2884	.0000
2nd	−.0085	1.0000	−.0656	1.0000	.1018	.3458
3rd	−.0589	1.0000	−.1619	.0027	.3076	.0000
4th	.0048	1.0000	.1034	.3102	−.1477	.0101

^aPearson correlation coefficient for the comparison of ancestry with principal component.

^bBonferroni-corrected P value.

Genetic Differentiation among Geographic Subpopulations

The coancestry coefficient estimator of F_ST was 0.0013 (95% CI 0.0003–0.0026), suggesting a small but significant amount of genetic differentiation among the four regions of the United States from which the CHS participants were sampled. Mean age- and sex-adjusted individual ancestry estimates differed across the four CHS clinic sites (P=.005). Group admixture estimates by clinic site are shown in table 3. Exclusion of the Maryland African American residents did not appreciably alter the variation in admixture (P=.01) but did attenuate the allele-frequency differences among the three larger population samples (F_ST=0.0007; 95% CI −0.00009 to 0.0017). Potential local inbreeding effects for the North Carolina, California, Maryland, and Pennsylvania African American populations were estimated at 0.025 ± 0.014, 0.092 ± 0.065, −0.006 ± 0.037, and −0.035 ± 0.079, respectively. Together, these results suggest that local population differences may play a role in shaping the overall genetic heterogeneity and structure of the entire CHS African American cohort.

Table 3.

Group Admixture Estimates by CHS Clinic Site or by Clusters Inferred on the Basis of Genetic Similarity^[Note]

		Estimated Ancestral Proportions(% ± SE)
BiodemographicCharacteristic	N	African	European	Native American
Clinical centera:
Winston-Salem	299	79.1 ± .9	17.0 ± 1.7	3.9 ± 2.1
Sacramento	214	74.4 ± 1.0	20.6 ± 2.0	4.9 ± 2.4
Pittsburgh	285	75.2 ± .9	23.9 ± 1.8	.9 ± 2.2
Genetic similarity:
Cluster 1	467	86.4 ± .7	12.1 ± 1.3	1.6 ± 1.7
Cluster 2	32	76.1 ± 2.3	25.0 ± 5.0	.0 ± 6.0
Cluster 3	74	41.4 ± 1.6	37.1 ± 3.6	21.5 ± 4.2
Cluster 4	236	67.2 ± 1.0	33.2 ± 2.1	.0 ± 2.6

Note.— Group ancestral proportions and SEs were estimated separately for each subpopulation by use of the program ADMIX 2.0 (Dupanloup and Bertorelle 2001), as described in the “Methods” section.

^aThe small number of subjects in the Maryland sample (n=11) were excluded from this analysis.

Population Subdivision Due to Genetic Background Similarity

As a complementary approach to population-structure assessment, discrete clusters of genetically similar individuals were identified through the use of pairwise, allele-frequency–weighted, identity-by-state, allele-sharing matrices. The most likely number of genetically similar clusters of individuals within the total sample was determined by testing allele frequency and phenotypic differences within the total sample of 810 individuals. As shown in figure 1, the most significant differences in male BMI involved the assumption of four groups of individuals (P<.0001). These and other data (N.J.S., unpublished data) suggest that, within the cohort, there are likely four genetically distinct subgroups. The distribution of these four subpopulations, identified on the basis of genetic background similarity, did not differ among the four geographic subregions (P=.20) but did differ with respect to individual admixture proportions (P<.001) and group admixture estimates (table 3). These results suggest that the empirically determined clusters actually reflect the differences in degrees of admixture among the study subjects.

Figure 1.

Differences in BMI for males across groups, determined by the use of cluster analyses involving allele-similarity matrices. The solid line represents the P values associated with standard ANOVA, and the dotted line represents the P values associated with nonparametric ANOVA.

Two participants, one from North Carolina and the other from California, had multilocus genetic backgrounds that were extremely different from the remaining cohort. These two genetic “outliers” were confirmed to have 0% and 3% African ancestry, respectively, and 100% and 97% European ancestry, respectively. Both belonged to genetic similarity cluster 4. These two individuals were excluded from further analyses. Additional investigation did not reveal any evidence that either individual had been inadvertently misclassified within the CHS data set.

Relationships among Population Genetic, Demographic, and SES Variables

The mean age at baseline of the CHS African American participants was 73 years (range 65–93 years). Individual African ancestry averaged 74% among the 65–69-year-olds (n=269), 72% among the 70–74-year-olds (n=279), 74% among the 75–79-year-olds (n=160), and 78% among the participants aged ⩾80 years (n=102). These differences were not significant (P=.10). Education, income, and occupation type all differed strongly by individual admixture proportions and genetic background similarity (all P values <.001). Moreover, there were differences in education (P<.001), income (P<.001), and occupation (P=.02) among clinic sites. These data highlight the complex interrelationships that exist between genetic or ancestral background and current social, economic, and environmental conditions in human populations.

Associations between Population-Structure Characteristics and Chronic-Disease Phenotypes

To address the influence of different forms of population structure on traits related to CVD and aging, we performed direct tests of association between several quantitative phenotypes and estimates of individual ancestry, genetic similarity, SES, and clinic site (tables 4, 5, 6, and 7). Each biodemographic characteristic was examined separately and in a multivariable model that was simultaneously adjusted for other characteristics. Higher fasting blood glucose levels were associated most strongly with African ancestry. Mean glucose levels, adjusted for age, sex, and baseline diabetes status, were 19 mg/dL higher (95% CI 5–33) among subjects with 100% African ancestry compared with those with 0% African ancestry. The glucose-ancestry association was altered only minimally by additional adjustment for SES and clinic site. Systolic blood pressure and HDL cholesterol levels were higher among African Americans sampled from the Sacramento area than among those sampled from Winston-Salem or Pittsburgh. In multivariable-adjusted models, genetic background or SES did not seem to account appreciably for these geographic differences (tables 4 and 5). CRP levels were influenced by type of occupation (table 6). Blue-collar workers had 23% higher (range 5%–41% higher) CRP levels relative to those of white-collar workers. Carotid arterial wall thickness did not vary significantly by any of the biodemographic indicators in table 6.

Table 4.

Blood Glucose and Systolic Blood Pressure by Clinic Site, Genetic Background Similarity, Admixture, and SES^[Note]

		Mean Blood Glucose (95% CI)[mg/dl]		Mean Systolic BloodPressure (95% CI)[mm Hg]
BiodemographicCharacteristic	N	Minimally Adjusted	Fully Adjusted	Minimally Adjusted	Fully Adjusted
Clinical centera:
Winston-Salem	299	114 (111–117)	114 (111–117)	141 (138–143)	141 (138–143)
Sacramento	213	109 (106–112)	110 (106–114)	146 (143–149)	145 (142–148)
Pittsburgh	285	114 (111–116)	113 (110–116)	140 (138–143)	141 (138–143)
Genetic similarityb:
Cluster 1	467	114 (111–116)	114 (112–117)	141 (139–143)	141 (139–144)
Cluster 2	32	105 (97–114)	105 (97–114)	144 (137–152)	144 (136–152)
Cluster 3	74	107 (101–113)	108 (102–114)	142 (136–147)	141 (136–147)
Cluster 4	234	112 (109–115)	111 (108–115)	142 (140–145)	142 (139–145)
Genetic ancestryc:
0% African (estimated)		102 (96–108)	104 (98–111)	139 (133–145)	138 (132–144)
100% African (estimated)		116 (113–119)	116 (112–119)	143 (140–145)	143 (141–146)
Educationd:
None–grade 9	237	116 (112–119)	114 (111–117)	140 (138–143)	140 (139–143)
High school	294	110 (107–113)	112 (111–114)	142 (140–143)	142 (140–143)
Professional/vocational	274	111 (108–114)	111 (108–114)	143 (141–145)	143 (140–145)
Annual incomee:
<$8,000	285	115 (112–118)	114 (111–117)	143 (140–145)	143 (140–145)
$8,000–$35,000	398	111 (109–114)	112 (110–114)	142 (140–143)	141 (140–143)
>$35,000	77	109 (103–115)	110 (105–114)	141 (137–144)	140 (136–144)
Occupation typef:
White collar	300	111 (108–114)	112 (109–115)	142 (139–144)	142 (139–144)
Blue collar	232	111 (108–115)	110 (107–114)	139 (136–142)	140 (137–143)
Housewife/other	275	114 (111–118)	115 (111–118)	144 (141–146)	144 (141–146)

Note.— Likelihood-ratio tests of association were performed by multiple linear regression of each phenotypic trait on biodemographic characteristics. Minimally adjusted models were adjusted for age, sex, and any clinically relevant covariates, as described in the “Methods” section. Fully adjusted models additionally contained variables for remaining biodemographic characteristics. In the footnotes below, P values in bold italics are less than the nominal significance level of 5% adjusted for the number of traits assessed (n=8; P<.00625).

^aFor measurements by clinical center, the minimally adjusted P value was .05 and the fully adjusted P value was .25 for blood glucose, and the minimally adjusted P value was .005 and the fully adjusted P value was .04 for systolic blood pressure.

^bFor measurements by genetic similarity, the minimally adjusted P value was .06 and the fully adjusted P value was .09 for blood glucose, and the minimally adjusted P value was .86 and the fully adjusted P value was .90 for systolic blood pressure.

^cFor measurements by genetic ancestry, the minimally adjusted P value was .002 and the fully adjusted P value was .01 for blood glucose, and the minimally adjusted P value was .29 and the fully adjusted P value was .23 for systolic blood pressure.

^dFor measurements by education, the minimally adjusted P value was .08 and the fully adjusted P value was .26 for blood glucose, and the minimally adjusted P value was .16 and the fully adjusted P value was .24 for systolic blood pressure.

^eFor measurements by annual income, the minimally adjusted P value was .04 and the fully adjusted P value was .20 for blood glucose, and the minimally adjusted P value was .43 and the fully adjusted P value was .32 for systolic blood pressure.

^fFor measurements by occupation type, the minimally adjusted P value was .30 and the fully adjusted P value was .21 for blood glucose, and the minimally adjusted P value was .14 and the fully adjusted P value was .16 for systolic blood pressure.

Table 5.

HDL Cholesterol and BMI by Clinic Site, Genetic Background Similarity, Admixture, and SES^[Note]

		Mean HDL Cholesterol (95% CI)[mg/dl]		Mean BMI (95% CI)[kg/m2]
BiodemographicCharacteristic	N	Minimally Adjusted	Fully Adjusted	Minimally Adjusted	Fully Adjusted
Clinical centera:
Winston-Salem	299	55 (53–56)	55 (53–56)	28.3 (27.7–28.9)	28.3 (27.7–29.0)
Sacramento	213	59 (57–61)	59 (57–61)	29.3 (27.9–29.3)	28.9 (28.1–29.6)
Pittsburgh	285	55 (53–56)	55 (53–57)	28.5 (27.9–29.2)	28.5 (27.9–29.1)
Genetic similarityb:
Cluster 1	467	56 (55–57)	56 (55–57)	28.7 (28.2–29.2)	28.7 (28.2–29.2)
Cluster 2	32	56 (51–61)	55 (51–60)	26.4 (24.5–28.2)	26.5 (24.4–28.4)
Cluster 3	74	54 (51–57)	53 (50–56)	29.4 (28.2–30.6)	29.7 (28.4–31.0)
Cluster 4	234	57 (55–59)	56 (55–58)	28.1 (27.4–28.8)	28.2 (27.5–28.9)
Genetic ancestryc:
0% African (estimated)		55 (52–59)	54 (51–58)	27.8 (26.4–29.1)	28.0 (26.5–29.5)
100% African (estimated)		56 (54–58)	56 (55–58)	28.8 (28.2–29.4)	28.7 (28.1–29.4)
Educationd:
None–grade 9	237	55 (54–57)	56 (54–57)	28.6 (28.0–29.2)	28.6 (28.0–29.3)
High school	294	56 (55–57)	56 (55–57)	28.5 (28.1–28.9)	28.5 (28.1–28.9)
Professional/vocational	274	56 (55–58)	56 (54–58)	28.4 (27.8–28.9)	28.4 (27.8–29.0)
Annual incomee:
<$8,000	285	55 (53–56)	55 (53–56)	28.9 (28.3–29.5)	28.9 (28.3–29.6)
$8,000– $35,000	398	56 (55–57)	56 (55–57)	28.5 (28.0–28.9)	28.4 (28.0–28.9)
>$35,000	77	58 (56–61)	58 (55–60)	28.0 (27.1–28.9)	28.0 (27.0–28.9)
Occupation typef:
White collar	300	58 (56–59)	58 (56–59)	28.0 (27.4–28.6)	28.0 (27.4–28.6)
Blue collar	232	54 (52–56)	54 (53–56)	28.7 (28.0–29.4)	28.7 (28.0–29.5)
Housewife/other	275	55 (53–57)	55 (53–57)	28.8 (28.2–29.5)	28.8 (28.2–29.5)

Note.— Likelihood-ratio tests of association were performed by multiple linear regression of each phenotypic trait on biodemographic characteristics. Minimally adjusted models were adjusted for age, sex, and any clinically relevant covariates, as described in the “Methods” section. Fully adjusted models additionally contained variables for remaining biodemographic characteristics. In the footnotes below, P values in bold italics are less than the nominal significance level of 5% adjusted for the number of traits assessed (n=8; P<.00625).

^aFor measurements by clinical center, the minimally adjusted P value was .0005 and the fully adjusted P value was .005 for HDL cholesterol, and the minimally adjusted P value was .61 and the fully adjusted P value was .57 for BMI.

^bFor measurements by genetic similarity, the minimally adjusted P value was .45 and the fully adjusted P value was .28 for HDL cholesterol, and the minimally adjusted P value was .03 and the fully adjusted P value was .02 for BMI.

^cFor measurements by genetic ancestry, the minimally adjusted P value was .84 and the fully adjusted P value was .43 for HDL cholesterol, and the minimally adjusted P value was .28 and the fully adjusted P value was .46 for BMI.

^dFor measurements by education, the minimally adjusted P value was .47 and the fully adjusted P value was .90 for HDL cholesterol, and the minimally adjusted P value was .57 and the fully adjusted P value was .64 for BMI.

^eFor measurements by annual income, the minimally adjusted P value was .05 and the fully adjusted P value was .10 for HDL cholesterol, and the minimally adjusted P value was .20 and the fully adjusted P value was .16 for BMI.

^fFor measurements by occupation type, the minimally adjusted P value was .008 and the fully adjusted P value was .02 for HDL cholesterol, and the minimally adjusted P value was .13 and the fully adjusted P value was .14 for BMI.

Table 6.

CRP Levels and Carotid Wall Thickness by Clinic Site, Genetic Background Similarity, Admixture, and SES^[Note]

		Mean CRP Levels (95% CI)[mg/liter]		Mean Carotid WallThickness (95% CI)[mm]
BiodemographicCharacteristic	N	Minimally Adjusted	Fully Adjusted	Minimally Adjusted	Fully Adjusted
Clinical centera:
Winston-Salem	299	2.82 (2.51–3.17)	2.81 (2.49–3.17)	1.12 (1.09–1.14)	1.12 (1.10–1.15)
Sacramento	213	2.43 (2.11–2.78)	2.54 (2.20–2.93)	1.12 (1.10–1.15)	1.13 (1.10–1.16)
Pittsburgh	285	2.57 (2.28–2.88)	2.58 (2.29–2.90)	1.12 (1.10–1.15)	1.12 (1.10–1.15)
Genetic similarityb:
Cluster 1	467	2.66 (2.42–2.91)	2.66 (2.41–2.92)	1.13 (1.11–1.15)	1.13 (1.11–1.15)
Cluster 2	32	2.67 (1.88–3.80)	2.57 (1.80–3.67)	1.09 (1.01–1.16)	1.08 (1.00–1.15)
Cluster 3	74	2.35 (1.86–2.96)	2.41 (1.89–3.08)	1.09 (1.04–1.14)	1.09 (1.04–1.14)
Cluster 4	234	2.58 (2.27–2.94)	2.66 (2.33–3.03)	1.13 (1.10–1.15)	1.13 (1.10–1.16)
Genetic ancestryc:
0% African (estimated)		2.12 (1.63–2.76)	2.37 (1.79–3.14)	1.09 (1.03–1.14)	1.09 (1.03–1.15)
100% African (estimated)		2.80 (2.50–3.14)	2.72 (2.42–3.07)	1.14 (1.11–1.16)	1.13 (1.11–1.16)
Educationd:
None–grade 9	237	2.80 (2.48–3.14)	2.75 (2.44–3.11)	1.14 (1.11–1.16)	1.14 (1.11–1.16)
High school	294	2.62 (2.44–2.81)	2.62 (2.44–2.81)	1.12 (1.11–1.14)	1.12 (1.11–1.14)
Professional/vocational	274	2.45 (2.20–2.74)	2.49 (2.23–2.79)	1.11 (1.09–1.13)	1.11 (1.09–1.14)
Annual incomee:
<$8,000	285	2.81 (2.50–3.15)	2.77 (2.46–3.11)	1.13 (1.10–1.15)	1.13 (1.10–1.15)
$8,000– $35,000	398	2.55 (2.36–3.77)	2.57 (2.37–3.79)	1.12 (1.11–1.14)	1.13 (1.11–1.14)
>$35,000	77	2.48 (2.25–2.75)	2.39 (2.00–2.86)	1.12 (1.08–1.15)	1.12 (1.08–1.16)
Occupation typef:
White collar	300	2.52 (2.25–2.83)	2.54 (2.26–2.86)	1.12 (1.10–1.15)	1.12 (1.10–1.15)
Blue collar	232	3.15 (2.75–3.62)	3.12 (2.70–3.59)	1.12 (1.09–1.15)	1.12 (1.09–1.15)
Housewife/other	275	2.30 (2.03–2.60)	2.31 (2.03–2.62)	1.12 (1.10–1.15)	1.12 (1.09–1.15)

Note.— Likelihood-ratio tests of association were performed by multiple linear regression of each phenotypic trait on biodemographic characteristics. Minimally adjusted models were adjusted for age, sex, and any clinically relevant covariates, as described in the “Methods” section. Fully adjusted models additionally contained variables for remaining biodemographic characteristics. In the footnotes below, the P value in bold italics is less than the nominal significance level of 5% adjusted for the number of traits assessed (n=8; P<.00625).

^aFor measurements by clinical center, the minimally adjusted P value was .23 and the fully adjusted P value was .49 for CRP level, and the minimally adjusted P value was .95 and the fully adjusted P value was .94 for carotid wall thickness.

^bFor measurements by genetic similarity, the minimally adjusted P value was .80 and the fully adjusted P value was .91 for CRP level, and the minimally adjusted P value was .44 and the fully adjusted P value was .32 for carotid wall thickness.

^cFor measurements by genetic ancestry, the minimally adjusted P value was .11 and the fully adjusted P value was .46 for CRP level, and the minimally adjusted P value was .19 and the fully adjusted P value was .31 for carotid wall thickness.

^dFor measurements by education, the minimally adjusted P value was .16 and the fully adjusted P value was .31 for CRP level, and the minimally adjusted P value was .20 and the fully adjusted P value was .26 for carotid wall thickness.

^eFor measurements by annual income, the minimally adjusted P value was .13 and the fully adjusted P value was .27 for CRP level, and the minimally adjusted P value was .70 and the fully adjusted P value was .93 for carotid wall thickness.

^fFor measurements by occupation type, the minimally adjusted P value was .004 and the fully adjusted P value was .01 for CRP level, and the minimally adjusted P value was .99 and the fully adjusted P value was .97 for carotid wall thickness.

Table 7.

YOL and YHL by Clinic Site, Genetic Background Similarity, Admixture, and SES^[Note]

		Mean YOL (95% CI)[years]		Mean YHL (95% CI)[years]
BiodemographicCharacteristic	N	Minimally Adjusted	Fully Adjusted	Minimally Adjusted	Fully Adjusted
Clinical centera:
Winston-Salem	299	8.13 (7.83–8.42)	8.11 (7.80–8.42)	5.07 (4.76–5.38)	5.10 (4.77–5.43)
Sacramento	213	8.54 (8.19–8.90)	8.47 (8.10–8.83)	5.63 (5.26–6.01)	5.45 (5.06–5.84)
Pittsburgh	285	8.25 (7.95–8.55)	8.26 (7.95–8.56)	4.95 (4.63–5.27)	4.93 (4.61–5.26)
Genetic similarityb:
Cluster 1	467	7.98 (7.74–8.21)	7.97 (7.72–8.21)	4.93 (4.68–5.18)	4.94 (4.68–5.20)
Cluster 2	32	7.99 (7.09–8.89)	7.96 (7.04–8.88)	5.05 (4.09–6.00)	5.07 (4.10–6.05)
Cluster 3	74	8.69 (8.10–9.28)	8.64 (8.01–9.26)	5.90 (5.27–6.53)	5.79 (5.13–6.45
Cluster 4	234	8.76 (8.43–9.09)	8.71 (8.37–9.05)	5.48 (5.13–5.83)	5.34 (4.98–5.69)
Genetic ancestryc:
0% African (estimated)		9.15 (8.48–9.82)	9.02 (8.30–9.75)	6.08 (5.37–6.80)	5.59 (4.82–6.36)
100% African (estimated)		7.97 (7.68–8.26)	7.98 (7.67–8.28)	4.86 (4.55–5.17)	4.98 (4.65–5.30)
Educationd:
None–grade 9	237	8.16 (7.85–8.46)	8.28 (7.97–8.59)	4.69 (4.37–5.01)	4.77 (4.44–5.10)
High school	294	8.26 (8.09–8.44)	8.27 (8.09–8.44)	5.15 (4.96–5.33)	5.15 (4.96–5.34)
Professional/vocational	274	8.37 (8.09–8.65)	8.26 (7.97–8.55)	5.60 (5.30–5.90)	5.53 (5.22–5.84)
Annual incomee:
<$8,000	285	8.13 (7.83–8.42)	8.25 (7.95–8.55)	4.83 (4.51–5.14)	4.94 (4.62–5.27)
$8,000– $35,000	398	8.30 (8.10–8.50)	8.25 (8.05–8.45)	5.27 (5.05–5.48)	5.22 (5.01–5.44)
>$35,000	77	8.48 (8.04–8.92)	8.25 (7.79–8.70)	5.70 (5.24–6.17)	5.50 (5.02–5.99)
Occupation typef:
White collar	300	8.41 (8.11–8.71)	8.32 (8.02–8.62)	5.63 (5.32–5.95)	5.48 (5.11–5.85)
Blue collar	232	8.07 (7.71–8.42)	8.13 (7.77–8.49)	4.60 (4.23–4.98)	4.75 (4.33–5.17)
Housewife/other	275	8.30 (7.98–8.62)	8.34 (8.02–8.66)	5.17 (4.83–5.51)	5.09 (4.72–5.47)

Note.— Likelihood-ratio tests of association were performed by multiple linear regression of each phenotypic trait on biodemographic characteristics. Minimally adjusted models were adjusted for age, sex, and any clinically relevant covariates, as described in the “Methods” section. Fully adjusted models additionally contained variables for remaining biodemographic characteristics. In the footnotes below, P values in bold italics are less than the nominal significance level of 5% adjusted for the number of traits assessed (n=8; P<.00625).

^aFor measurements by clinical center, the minimally adjusted P value was .32 and the fully adjusted P value was .36 for YOL, and the minimally adjusted P value was .04 and the fully adjusted P value was .14 for YHL.

^bFor measurements by genetic similarity, the minimally adjusted P value was .0007 and the fully adjusted P value was .003 for YOL, and the minimally adjusted P value was .008 and the fully adjusted P value was .08 for YHL.

^cFor measurements by genetic ancestry, the minimally adjusted P value was .008 and the fully adjusted P value was .03 for YOL, and the minimally adjusted P value was .01 and the fully adjusted P value was .23 for YHL.

^dFor measurements by education, the minimally adjusted P value was .37 and the fully adjusted P value was .93 for YOL, and the minimally adjusted P value was .003 and the fully adjusted P value was .004 for YHL.

^eFor measurements by annual income, the minimally adjusted P value was .27 and the fully adjusted P value was .99 for YOL, and the minimally adjusted P value was .01 and the fully adjusted P value was .12 for YHL.

^fFor measurements by occupation type, the minimally adjusted P value was .36 and the fully adjusted P value was .66 for YOL, and the minimally adjusted P value was .0003 and the fully adjusted P value was .002 for YHL.

BMI appeared to be influenced more by genetic similarity than by ancestral proportions (table 5). Moreover, the associations differed by sex (P value for sex-genetic similarity cluster interaction on BMI was .03). The mean age-, clinic-, and SES-adjusted BMI was highest among men in genetic similarity cluster 3 (28.7 kg/m²; 95% CI 27.1–30.3) and was lowest among men in genetic similarity cluster 2 (24.6 kg/m²; 95% CI 22.6–26.5). In contrast, income level remained the only significant predictor of BMI among women, after multivariable adjustment (P=.03). The mean-adjusted BMI was 30.3 kg/m² (95% CI 29.6–31.1) for women in the lowest income group, compared with 28.3 kg/m² (95% CI 26.8–29.8) for women in the highest income group.

African ancestry and genetic similarity, even when adjusted for baseline age, BMI, sex, self-rated health status, smoking, hypertension, diabetes, coronary heart disease, and cancer, were associated with both YOL and YHL during follow-up (table 7). For the YHL outcome, SES adjustment attenuated these associations. Moreover, when all of the biodemographic covariates in table 7 were included together simultaneously, income, education, and occupation (P=.005), rather than individual ancestry (P=.50) or genetic similarity (P=.12), remained the only significant predictor of YHL. In contrast, genetic background similarity (P=.02) remained the only significant biodemographic predictor of the outcome YOL in a multivariable model; individuals belonging to genetic similarity cluster 4 lived, on average, an additional 9 mo (95% CI 3–14), compared with genetic similarity cluster 1.

Associations of Phenotypic Traits with Individual AIMs

Tests of association for each AIM with each phenotypic trait are shown in table 8. In general, trait-associated markers tended to be among those with the highest African/European allele-frequency differential. Under the null hypothesis, ∼1/24 markers would be expected by chance to be associated with any single trait (with the significance threshold of P<.05). By comparison, blood glucose was associated with five markers. Adjustment for individual ancestry attenuated each of the five marker–blood glucose associations (fig. 2). These findings strongly suggest false-positive associations due to population stratification. Note also in figure 2 that an additional marker (rs722098) showed no association before adjustment but was associated with glucose level (P=.002) only after correction for African ancestry.

Table 8.

Tests of Association between Each AIM and the Cardiovascular Risk Factors and Longevity Outcomes^[Note]

	P Values for
Markera	BloodGlucose	SystolicBlood Pressure	HDLCholesterol	BMI	CRP Levels	Carotid WallThickness	YOL	YHL
rs2814778	.37	.46	.65	.89	.004	.55	.01	.23
rs930072	.005	.18	.49	.16	.05	.02	.03	.02
rs7349	.04	.28	.75	.18	.97	.91	.10	.01
rs723632	.07	.94	.25	.05	.95	.85	.65	.40
rs722098	.56	.58	.16	.28	.29	.87	.05	.09
rs146026	.04	.21	.02	.93	.54	.02	.41	.58
rs6003	.12	.33	.18	.95	.07	.82	.43	.57
rs1985080	.91	.92	.74	.86	.81	.56	.94	.78
rs518116	.01	.28	.68	.76	.60	.38	.08	.07
rs3287	.01	.49	.15	.73	.50	.69	.49	.25
rs1989486	.52	.43	.08	.32	.32	.99	.16	.99
rs7041	.07	.23	.72	.38	.66	.10	.06	.22
rs994174	.45	.18	.44	.76	.30	.57	.33	.18
rs1800498	.67	.56	.15	.71	.86	.75	.03	.15
rs2816	.77	.36	.38	.59	.15	.21	.82	.54
rs2891	.15	.41	.59	.18	.28	.98	.19	.16
rs3188520	.77	.62	.49	.07	.19	.10	.61	.13
rss1042602	.96	.04	.13	.95	.62	.72	.22	.18
rs326946	.10	.67	.87	.21	.23	.64	.27	.10
rs2077863	.23	.40	.99	.38	.63	.35	.17	.91
rs3188519	.92	.32	.96	.16	.84	.08	.73	.48
rs594689	.24	.62	.48	.20	.44	.07	.29	.93
rs2228478	.45	.37	.26	.03	.04	.81	.01	.21
rs584059	.43	.04	.41	.11	.27	.45	.71	.86

Note.— Likelihood-ratio tests of association were performed using multiple linear regression, adjusted for age, sex, and any clinically relevant covariates, as described in the “Methods” section. P values ⩽.05 are indicated in bold italics.

^aMarkers are listed in decreasing order of African/European F_ST values.

Figure 2.

Associations of markers with glucose, before and after adjustment for African ancestry proportions

Discussion

Our results suggest a complex relationship between aging-related traits, genetic background heterogeneity, and population structure among older African American adults. Overall, there was evidence of substantial population subdivision and genetic admixture, as demonstrated by decreased marker heterozygosity and excess allelic association of unlinked markers. Remarkably, ∼60% of pairs of markers in this data set were associated, despite the fact that the markers were randomly scattered throughout the genome. The association between markers that are not in physical proximity indicates that the rate of spurious associations without adjustment for population stratification is likely to be particularly high in this population, since association cannot be a reliable measure of genomic position. However, the rate of association between these markers is much higher than the rate expected with random markers, since there is a linear relationship between the strength of allelic association and marker informativeness in admixed populations (Chakraborty and Weiss 1988).

All of the phenotypic traits under study are known to be influenced by environmental factors, some of which are related to SES. In our analyses, SES adjustment seemed to weaken the association of genetic background with some traits but not others. However, education, income, and occupation type likely represent only crude proxies for current SES (Kaufman et al. 1997), given the retirement status of the participants. Moreover, clinical disease is common among older cohorts such as the CHS cohort (Kuller et al. 1998). Although we additionally adjusted our analyses for known clinical confounders, residual nonrandom associations between health care access or adequacy of treatment and social characteristics may persist. Therefore, we cannot exclude residual confounding by environmental determinants as a possible explanation for the observed genetic ancestry-trait associations (Risch et al. 2002; Kittles and Weiss ²⁰⁰³). Ultimately, proof that association between genetic ancestry and a particular phenotype is due to genetic etiology lies in the identification of a specific genomic region or regions that account for the association. This would require a whole-genome admixture mapping survey, which depends on the typing of hundreds to thousands of markers (Smith et al. 2004; McKeigue ²⁰⁰⁵).

The observed association of increased blood glucose levels with African ancestry is consistent with reports of higher fasting glucose and increased prevalence of diabetes in older African Americans (Haffner et al. 1996). A similar association between insulin resistance and African ancestry, independent of SES, was recently reported in a study of children residing in the southern United States, which included individuals whose self-reported race/ethnicity was white and African American (Gower et al. 2003). Our analysis included only individuals who self reported as African American and was also adjusted for education and income levels.

The sex-dependent associations we observed for BMI are noteworthy in light of the higher prevalence of female obesity but lower levels of male obesity among older African Americans compared with older whites (Hutchinson et al. 1997; Kuller et al. ¹⁹⁹⁸; Sundquist et al. ²⁰⁰¹). The propensity for weight gain and obesity in black women has been associated with lower SES and higher physical inactivity (Fernandez et al. 2003). Our findings suggest a possible contribution of either genetic background or other distinct environmental factors correlated with genetic background to the lower rates of obesity in African American men.

SES has a major impact on all-cause mortality within and among age, sex, and race strata (Lin et al. 2003). In the CHS African American cohort, longevity appeared to be influenced by various indicators of population and social structure. The association with SES indicators was particularly strong for YHL, a measure that incorporates both length and quality of life. For YOL, which more objectively quantifies survival time or all-cause mortality, the association with genetic background appeared stronger but was attenuated somewhat by SES adjustment. Genetic factors may influence mortality in older adults, particularly at very advanced ages (Perls et al. 2002). It is important to recognize, however, that genetic similarity or shared ancestry are likely correlated with a range of social, cultural, and/or environmental variables that influence disease occurrence and mortality yet remain unmeasured or not adequately accounted for in our analysis (Risch et al. 2002; Kittles and Weiss ²⁰⁰³). The substantial effect of SES on the genetic associations with longevity highlights an important principle: excess type I error can occur in admixed populations even as a result of environmental factors (Risch et al. 2002; Cardon and Palmer ²⁰⁰³). In this case, SES is associated with genetic ancestry, leading to confounding in tests for individual markers.

Our findings for CHS strongly suggest that controlling for population structure/admixture will be required in large, multicenter genetic-association studies that assess common chronic-disease–related traits in African American population samples. Individual ancestry can be estimated in African American samples by typing a reasonable number of markers that are highly differentiated in allele frequency across parental populations. Conditioning on admixture proportions in a multilocus analysis can control for confounding due to population stratification (McKeigue et al. ²⁰⁰⁰; Pfaff et al. ²⁰⁰¹; Hoggart et al. ²⁰⁰³). As illustrated in figure 2, controlling for genetic ancestry should not only reduce false-positive associations but may also uncover a true association previously obscured by stratification. On the basis of dynamic relationships among various genetic and environmental determinants of disease susceptibility, additional multilocus methods—such as those that detect genetic similarity, cryptic relatedness, or rates of migration under nonequilibrium conditions—may help to characterize the complex genetic demography of an epidemiologic sample (Overall and Nichols 2001; Schork ²⁰⁰¹; Schork et al. ²⁰⁰¹; Curtis et al. ²⁰⁰²; Wilson and Rannala ²⁰⁰³) and thereby provide additional information about the genetic architecture of common diseases of aging in heterogeneous outbred populations.

The limited number of markers we used may have resulted in imprecise estimates of individual ancestry or genetic background similarity. However, several different statistical methods of differentiating individuals, including the Bayesian algorithm in the program STRUCTURE and the results of the principal-components method, demonstrated a very high degree of correlation with our estimates of African ancestry from the maximum-likelihood model. This is not unexpected, since the markers were chosen primarily on the basis of frequency differential between African and European parental populations.

Since the markers we used had less ability to distinguish Native American ancestry from European ancestry, the correlations were less robust for Native American or European ancestry estimated by maximum likelihood, STRUCTURE, and principal-components analysis. This is also reflected by the wide CIs associated with our estimates of Native American ancestry. Our results are not inconsistent with previous studies, such as those of Parra et al. (1998) and Smith et al. (2004), who estimated the Native American ancestry of African American populations at 1%–2%. Interestingly, there appeared to be a somewhat higher proportion of Native American ancestry among individuals within genetic similarity cluster 3. Typing more markers informative for Native American ancestry will be necessary to confirm these findings, which might lead to greater precision in controlling for admixture in association studies.

Our results also are in agreement with other studies showing ∼20% European admixture among African Americans, with somewhat higher contributions of European ancestry in northern or western U.S. populations (Chakraborty et al. 1986; Parra et al. ¹⁹⁹⁸; McKeigue et al. ²⁰⁰⁰; Pfaff et al. ²⁰⁰¹; Hoggart et al. ²⁰⁰³). Whether genetic heterogeneity among the African parental source populations has contributed to local variations in admixture among modern African American populations remains uncertain. Despite earlier studies suggesting genetic heterogeneity within continental Africa (reviewed by Tishkoff and Williams ²⁰⁰²; Kittles and Weiss ²⁰⁰³), markers such as the ones we typed, which have large frequency differences between European and African populations, appear to have much smaller variations within continental Africa (Collins-Schramm et al. ²⁰⁰²).

Our analysis excluded the individuals who self reported as white in the CHS cohort. Since other studies report that the proportion of African ancestry among U.S. non-Hispanic whites is <5%, the exclusion of self-identified white CHS participants from our study sample is unlikely to have impacted our findings in a substantial way. Our study does not address the question of population stratification among European Americans. Since allele-frequency differences between European subpopulations are likely smaller than those between European and African ancestral populations, a larger of number of markers will be required to assess the impact of stratification within non-Hispanic white populations.

Individuals who self identify as African American are culturally, socially, and genetically heterogeneous. SES and related factors, such as access to health care, play a major role in healthy aging. In some instances, these nongenetic factors may account for all or part of the association between a phenotype and ancestry. Nevertheless, from a public health and epidemiologic standpoint, an objective assessment of genetic background may provide additional information relevant to potential nongenetic confounders and predictors of disease risk as well as insight into genetic contributions. These considerations highlight the need for further investigation of the various SES and biodemographic factors that influence life span or quality of life in older adults.

In summary, there is evidence of substantial substructure and admixture among the CHS African American population. In addition, our analyses have shown that nongenetic factors may, in fact, confound genetic associations among populations with recent admixture and population substructure. Therefore, both controlling for population admixture by use of genetic markers and controlling for sociodemographic measures will be required in assessing genetic associations with complex chronic-disease traits in African American subjects.

Appendix A

Table A1.

PCR Assay Conditions and Oligonucleotide Primers for the 24 AIMs^[Note]

						PCR Primer (5′→3′)
Marker	Alternate Namea	Conditionsb	Tm(°C)	Strandc	Dye Terminator Mixd	Forward	Reverse	Genotyping Primer (5′→3′)
rs2814778	FY-NULL	.2 μM primers, 100 μM dNTPs, 5% DMSO	58	Reverse	CT	GCATAGGGATAAGGGACTCT	AACCTCAAAACAGGAAGACC	GCCCTCATTAGTCCTTGGCTCTTA
rs930072		Standard	58	Reverse	GA	TGCCACTGTCTCTTAGATTACAT	CAAAGAAAGTCACTTCAAATCTC	CTTCTGGCCCCTTTATAGACTAGCTT
rs7349		Standard	58	Reverse	GA	TAACATTTATACTTGCCTTGGAC	ATTGTTCCCAACACTGTTTATAG	GGAAATGAGAGTTGTATGGTTAGGCT
rs723632		Standard	58	Forward	GC	ATATTTGAACCTTGTGCTTCAG	TAAGAGAGTTTATGATGGGTGTC	CCCATCCCCAGGCCTCTTTA
rs722098		Standard	58	Reverse	CT	TAATGAACACAGCCATCTAGTCT	AAATATTCAGCACATCCAAAAT	GAAATATTCAGCACATCCAAAATTTAAATCCTTA
rs146026		Standard	58	Reverse	GA	TGAGCACTGGCTTTATTATTATT	CAACTGCTTTAAAAAGAGAGAGA	TGAAACACACAAAATGCATTACAATCGTA
rs6003	F13B	Standard	58	Reverse	CT	ATGAAATCGCCAATAATAACAT	AGAGAGACATTGAACCACTTCTT	CCCTGAAGCGCAACCATAA
rs1985080		Standard	58	Reverse	CT	TTTTTAAAATGGTATCAAACACC	GTCATCCCTATACCTTCTGACTT	GTGGATTGTACAGTGAGCCACTTA
rs518116		Standard	58	Reverse	CT	TTACATCATGAGAATCACTTTTTC	GTCATCTACTACACAGGGAGGTA	TTCAGTTTGGGCAAATTAGTTTGCAC
rs3287	WI-16857	5% DMSO	57	Reverse	CT	TATAATCCATCCTCCAACACA	AAGAGCAGATTTCTCCACATTAT	TCTTGTGAATTGGGAAGACCACTAG
rs1989486		Standard	58	Reverse	GA	AAATGGACACAAAAATAAAGATG	GTTTAAAATTGATTGCTGTTCTC	CCCTTGATGAACAAGGCTAACCT
rs7041		5% DMSO	58	Forward	GT	AATGGCTATTATTTTGCATTAGA	TAAAAGATTCTGCCATGTTAAGT	GCGACTAAAAGCAAAATTGCCTGA
rs994174		Standard	58	Reverse	CT	ATTGTGAAGTACAGGTGTGGTAT	TTTGGAGTCTCAGTGTTCTTATC	GTACTGCTTCTAATAGTACTCAGTGCCA
rs1800498	DRD2Taq_D	.4 μM primers, 3.5mM MgCl2	58	Reverse	GA	TATAAGCATCAAGTGTTTGGAAC	GAGAAGAAAACAGAACAAGATCA	GCCCCAGGTTCCCTAGTC
rs2816	WI-7423	Standard	58	Forward	CT	CTCATAGCTGCTGCTTCAAT	ACAGCATGCCTTTATTTCAC	CATTGCTGGGCTGTGTTCC
rs2891	WI-14867	Standard	58	Forward	GA	CAAAGTGTTGGAAATGTCATC	GAACAACTGATGAACTGAATTTT	ATGGGGCTGCAGACACTC
rs3188520		Standard	61	Reverse	GC	CTCTGACTGAGAAACTGAACAAT	AGGGTACCTAGCACCTTGTATT	AAAACAGGCAATCCTCCTAAGTCT
rs1042602	TYR-192	Standard	58	Reverse	GT	CTACCTCACTTTAGCAAAGCATA	ACCGCAACAAGAAGAGTCTAT	GCAAAATCAATGTCTCTCCAGATTTCA
rs326946		Standard	58	Reverse	CA	AACCCCTGAGATAGAACATAGAT	GGTAAGAAATGACTCAGAAGACA	TTTTTGCAAAGTGGCAGGATTTAATAATAATAAT
rs2077863		Standard	58	Reverse	GC	AGGTCCAGAAATGGCTCTAT	GACTGATAGTTGTTCCTGCTG	GCATCTATCCTAGGTTTGTGGATAGGA
rs3188519		Standard	58	Forward	CT	TAGGTCTCAGGTAGCAAGAGTC	CAGATGATTCATACACACCCTAT	CCAGCCACCAGTCACCC
rs594689	D11S429	5% DMSO	58	Reverse	CT	GATCTGAAAGGCATTGTAGG	CTTTGTAGGGTAGGAGAAATTG	GGGGACGGGCAAGGAG
rs2228478	MC1R-314	Standard	58	Reverse	CT	CAAGAACTTCAACCTCTTTCTC	AGATCATTTAGTCCATCCTCTTT	GCGCTCACCAGGAGCA
rs584059		Standard	58	Forward	CA	TCATCAGGATCCTCCAATACT	CTCCACTTCTCTGGAATTTTT	CAACCCAGGACCCACAGAAG

Note.— T_m = the annealing temperature used in PCR cycles.

^aAdditional information on markers with an alternate name can be found in dbSNP under the handle PSU-ANTH.

^bStandard PCR conditions were 5 μl PCR volume with 2.4 ng genomic DNA, Platinum Taq buffer, 2.5 mM MgCl₂, 50 μM dNTPs, 0.1μM PCR primers, 0.1 unit Platinum Taq, and 1 μl of extra water to counteract evaporation.

^cThe strand of DNA for which the genotyping primer will give results.

^dThe mix of dye terminators used in the genotyping reaction (based on the strand used for the genotyping primer).