Examining Population Stratification via Individual Ancestry Estimates versus Self-Reported Race

Most common complex phenotypes, such as cancer, show varying incidence rates, course of disease, and genetic susceptibility by race and ancestry (1, 2). For example, the world age-adjusted incidence rate of prostate cancer in Nigeria, with very little European admixture, is 23.3 per 100,000 men, whereas in Sweden, a relatively isolated population, the incidence rate is 90.9 per 100,000 men (3). For comparison, in the United States where individuals are mixtures of ancestral populations, the age-adjusted incidence rate of prostate cancer is 274.3 per 100,000 African American men and is 171.2 per 100,000 in Caucasian men (4). Although not adjusted for smoking status, sex-specific, age-adjusted lung cancer, incidence rates also differ significantly around the world. Specifically for males, the rate per 100,000 is 1.1 in Nigeria, 21.1 in Sweden, 120.7 for African Americans in the United States, and 82.3 for Caucasians in the United States (3, 4).

Studies of disease risk associated with candidate susceptibility genes are potentially vulnerable to bias due to population stratification. In order for important bias from population stratification to exist, the following must be true: (a) the frequency of the genotype of interest varies substantially by race and ancestry, (b) the disease rate varies substantially by race and ancestry, and (c) the disease rates and genotype frequencies vary together which occurs when the genotype is related to a true risk factor or is the true risk factor with a high attributable risk (5). Methods have therefore been developed to assess population stratification in case-control studies (6-12). Some of these methods involve using DNA markers to estimate genetic ancestry at the individual level, thereby allowing studies of its association with various complex traits (9, 13-16). Ancestry-informative markers generally show large allele frequency differences between ancestral populations (7).

Estimating individual ancestry requires genotyping an additional set of DNA markers for each individual in the study population. Whether this is necessary, or whether doing stratified analyses by self-reported race is adequate to control for population stratification, is unresolved (5). For example, it has been found that commonly used racial labels were insufficient and inaccurate representations of ancestral clusters and that genotype profiles, defined by the distribution of variants in drug-metabolizing genes (CYP1A2, CYP2C19, CYP2D6, NAT2, GSTM1, and DIA4), differed significantly among ancestral clusters (17). Individuals of mixed race are becoming increasingly more common in the United States (18). Most study subjects simply cannot report what percent of their genome originated from Europe, West Africa, or Native American ancestry via questionnaire. Populations in the United States are generally formed from more recent admixture, which causes interindividual differences in genetic ancestry to become more pronounced (19, 20). The West African contribution to ancestry for African Americans in the United States is on average 80%, but it can range from 20% to 100% and ∼30% of United States Caucasian, non-Hispanics have >90% European ancestry (7, 21). African Americans also have significant admixture from European and Native American ancestral populations, whereas United States Caucasian, non-Hispanics have significant admixture from West African and Native American populations (22).

To assess the utility of using individual genetic ancestry estimates to better understand population stratification in a standard epidemiologic case-control study, we genotyped early-onset lung cancer cases (i.e., diagnosed before age 50) and population-based controls for a panel of ancestry informative markers. We then estimated individual ancestry from these markers using two different methods and used these estimates to assess population stratification within this case-control sample. We used the glutathione S-transferase μ (GSTM1) locus, a candidate gene for lung cancer risk (23-27), as an example of how using individual ancestry estimates versus self-reported race can affect estimates of disease risk associated with genotype in groups of individuals.

Cases with early-onset lung cancer were identified between September 15, 1990 and November 30, 2003 through the metropolitan Detroit Cancer Surveillance System, a participant in the National Cancer Institute's Surveillance, Epidemiology and End Results program. This study was approved by the local institutional review board and all subjects provided written informed consent. Case eligibility criteria included an incident primary, malignant cancer of the lung or bronchus, <50 years of age at diagnosis, and a resident of the Detroit tri-county area (Wayne, Macomb, and Oakland) at the time of diagnosis. Population-based controls were ascertained concurrently with the cases via random digit dialing and were frequency matched to cases by race, sex, 5-year age group, and county of residence. Over 98% of the eligible, successfully contacted controls agreed to participate. Seven hundred forty-six (N_total = 746) cases and population-based controls with available extracted normal DNA via a blood or a cheek swab sample and who self-reported their race as Caucasian, non-Hispanic, or African American were used in this analysis (n_cases = 252 and n_controls = 494).

Each individual was genotyped for the lung cancer candidate gene, GSTM1 (null or present; refs. 28, 29). In addition, all individuals were genotyped for the U.S. Federal Bureau of Investigation CODIS Core short tandem repeat (STR) set of 13 loci for analysis of individual ancestry (30). A list of these 13 loci with the chromosomal location and the number of alleles are shown in Table 1. The 13 CODIS loci were tested for Hardy-Weinberg equilibrium and linkage disequilibrium and were found to not violate Hardy-Weinberg equilibrium within loci or show linkage disequilibrium between loci (data not shown; P = 0.08-0.75 for tests of Hardy-Weinberg equilibrium and linkage disequilibrium). The average of German and Polish parental frequencies were used to represent European (31, 32) and the average of Rwandan and Nigerian parental frequencies to represent West African (32, 33), for the maximum likelihood estimations (MLE). Detroit, MI was originally settled by the Polish and Germans, with African ancestral populations settling in over time (34), making these parental populations appropriate for this study population for estimation of individual ancestry.

We estimated individual ancestry using two methods: (a) MLE (16, 22) and (b) Bayesian clustering techniques as implemented in the STRUCTURE 2.1 program (9, 35). For the first method, using the contemporary published allele frequencies mentioned above as the parental frequencies, the individual maximum likelihood ancestral proportions for the two parental populations, European and West African, were calculated for all early-onset lung cancer cases and population-based controls, using each individual's CODIS Core STR loci genotypes.

Considering a population that was formed by admixture between two genetically distinct ancestral populations (this can be easily extended to any number of populations), the frequency of the kth allele at the gth locus in the admixed population, A, is

where the two ancestral contributions, m_j, i = 1 and 2 sum to 1.0, and the δ coefficients (i.e., allele frequencies differences between parental populations) are defined as δ_g1k = p_g1k − p_g2k. The constraint

Equation B applies to all alleles at all loci. Estimates of individual admixture were obtained by treating each individual as a sample of size one because the same likelihood applies to samples of any size.

Maximum likelihood estimates for the ancestral contributions were obtained from the log-likelihood function by setting the partial derivatives, with respect to m_j,

equal to zero, and solving simultaneously for m̂₁, using the Newton-Rhapson method (36). The MLE of m̂₂ equals 1−m̂₁ (37).

For the second method, individual ancestry for two “clusters” (i.e., ancestral European and West African populations), using each individual's CODIS Core STR loci genotypes was calculated. The STRUCTURE method assigns each individual to clusters by calculating a posterior probability that an individual belongs to a cluster, given the observed marker genotypes (i.e., the CODIS STR genotypes). The number of clusters can either be inferred by the program or can be given as an initial variable. In this case, we set the number of clusters to two to compare with the MLE estimates. In the presence of admixture and hence correlated allele frequencies, the STRUCTURE method also estimates the proportion of an individual's genome that derives from each of the two cluster subpopulations.

Composite delta (δ_c) was calculated for each of the 13 CODIS loci for the European and West African ancestral combination. Composite δ was calculated as half the sum across all loci pairs of the allele frequencies in two different populations when there are multiple alleles at a locus. Spearman correlation coefficients were calculated for MLE individual ancestry compared with STRUCTURE individual ancestry. Only European ancestry estimates were used for further analyses because the West African estimates were equal to one minus the European estimates. Median European MLE and STRUCTURE ancestry were compared within self-reported racial group by case-control status using a t test; the frequency of the GSTM1 null genotype was also compared within self-reported racial group by case-control status using a χ² test. Histograms of individual European ancestry for both MLE and STRUCTURE estimates by self-reported race were generated. To assess differences in ancestry between cases and controls related to the GSTM1 null risk genotype, histograms were generated to compare the frequency of the risk genotype by case-control status within European MLE or STRUCTURE ancestral group, stratified by self-reported race. Unconditional logistic regression models were used to estimate odds ratios and 95% confidence intervals to measure the association between early-onset lung cancer and the GSTM1 null genotype. Potential confounders, including gender, age at diagnosis for cases or age at interview for controls (continuous), family history of lung cancer, and pack-years of smoking (continuous) were included in all models. To test the effects of self-reported race and individual ancestry on genetic risk, models were additionally adjusted for self-reported race or individual European MLE or STRUCTURE ancestry and were compared with the general model using the likelihood ratio test. Additionally, models were compared using the Akaike Information Criterion that adjusts the −2 log-likelihood for the model by twice the number of estimated variables in the model (38). All statistical analyses were done using SAS version 9.1 (39).

A total of 555 self-reported Caucasian, non-Hispanics and a total of 191 self-reported African Americans were available for analysis. The 13 CODIS STR loci allele frequencies varied between the ancestral groups used in this study and were also highly multiallelic markers (Table 1). The δ_c values for the majority of the CODIS loci were >0.2 making them appropriate for ancestry estimation analysis (40, 41). Because the ancestral allele frequencies were used in the MLE, it was clear which of the two MLE estimates correlated with which ancestral group; however, this correlation was less clear with the STRUCTURE results. Spearman correlation coefficients showed that the European individual MLE ancestry estimates were highly positively correlated (+0.80) with the cluster 2 estimates from STRUCTURE, whereas the West African individual MLE ancestry estimates were highly positively correlated (+0.80) with the cluster 1 estimates from STRUCTURE. Therefore, we denoted the cluster 2 STRUCTURE estimates as European and the cluster 1 STRUCTURE estimates as West African, to compare with the MLE results in subsequent analyses.

There were no significant differences in median individual European ancestry estimates within self-reported racial group by case-control group or GSTM1 null genotype frequency (Table 2). However, the distribution of European ancestry values was significantly different by self-reported race, whether using MLE or STRUCTURE estimates (Fig. 1A and B). The GSTM1 null genotype frequency in Caucasian, non-Hispanic controls was 48.3%, whereas in African American controls it was 28.8% (Table 2).

To further investigate the effects of self-reported race or individual European ancestry on case-control status and the GSTM1 “risk” genotype (i.e., the GSTM1 null genotype), histograms of the GSTM1 null genotype by case-control status were generated by ancestry, stratified by self-reported race (Figs. 2 and 3). Although the ancestral distributions differed by estimation method, the distribution of the GSTM1 null genotype showed similar patterns within self-reported racial group. Among individuals who self-reported as African American, the GSTM1 null genotype frequency showed greater variability by ancestry and case-control status, regardless of ancestry estimation method, compared with Caucasian, non-Hispanics. Risk of early-onset lung cancer associated with the GSTM1 genotype, although not significant, increased when adjusting for individual European ancestry compared with the model adjusting by self-reported race. Adjusting for self-reported race did not significantly affect the risk estimate, whereas adjusting for individual ancestry did (LRT P for race adjusted model = 0.74; LRT P for individual ancestry adjusted models = 0.001 or <0.0001; Table 3). Using the Akaike Information Criterion value to compare genetic risk models, the models adjusted for individual ancestry (MLE or STRUCTURE) clearly did better than the model adjusted for self-reported race. This information provides evidence that individual ancestry may confound disease/candidate gene associations and provides a better measure of ancestral background than self-reported race.

In epidemiologic studies, racial differences are commonly investigated by doing analyses stratified by self-reported race. Self-reported race, however, is not always a reliable measure of one's ancestral make-up. To investigate the relationship between self-reported race and ancestry, we used genetic markers to estimate individual ancestry and population structure. Using study participants from a case-control study of early-onset lung cancer, we also examined the effects of this population structure on the distribution of the risk genotype for a lung cancer candidate gene. Individual European ancestry did not correlate completely with self-reported race in a Metropolitan Detroit population. Moreover, there was significant overlap by individual ancestry between the Caucasian, non-Hispanics and African Americans in this study sample. We observed that within self-reported racial group, the frequency of a lung cancer susceptibility genotype varied by European ancestry and case-control status. Models adjusting for individual European ancestry, regardless of the estimation method, better explained genetic risk associated with early-onset lung cancer risk, compared with a model adjusting for self-reported race. Given that incidence rates of early-onset lung cancer vary worldwide (4, 42), the distribution of the risk genotype in cases and controls varied by ancestry within self-reported racial group, and estimates of risk of disease associated with the risk genotype were significantly affected when adjusting by ancestry and not by self-reported race, we conclude that population stratification could be a significant issue in this Detroit population.

Genetic ancestry estimation is not commonly used in studies of complex diseases, because of the difficulty and expense of genotyping additional markers. However, many studies use race as an eligibility criterion. Genetic ancestry proportions seem to not only vary between groups of individuals who would self-identify to the same racial group but also among individuals within a group (22). Assortative mating, patterns of linkage and linkage disequilibrium among loci, and random genetic drift can all contribute to variability in ancestry among individuals (43). Allele frequencies have been shown to vary substantially across populations that have mixed ancestry from different continents (44) and within the same continent (45). Even if common variants are shared among racial groups, the frequencies can often differ substantially (46). Although the error caused by population stratification seems stronger for rare variants compared with common ones, the greatest bias and type I error caused by population stratification is when there is a hidden subpopulation of at least 50% in the study population of interest and the allele frequencies differ by at least 25% (47). Therefore, it has been recommended that information on ethnic origin be collected in the greatest detail possible (48).

Few studies have analyzed empirical (i.e., nonsimulated data) data to compare self-reported race with individual ancestry estimates to assess population substructure. A study by Wacholder et al. (5) examined NAT2 and incidence of bladder and breast cancers in relation to ancestry. They concluded that genetic markers of ancestry were unlikely to create a better proxy than self-reported race for cultural practices that could strongly affect cancer risk. In simulation studies, it has been shown that errors that occur from using self-reported race instead of genetic ancestry would be more problematic in large studies searching for susceptibility loci with small effects (19), as the adverse effects of this stratification seem to increase with increasing sample size (49, 50). Although not a case-control design, a recent study by Wilson et al., observed that frequency of risk genotypes in six drug metabolizing genes, including GSTM1, varied by ancestral group and that self-reported race was an insufficient and inaccurate representation of these ancestral clusters (17). The conclusions from the study by Wilson et al. and from our study suggest that using individual ancestry information could enhance the validity of epidemiologic studies and improve precision of estimated effects.

The present study suffers from limitations that are common to all studies involving ancestry estimation. The precise estimation of the ancestral proportions is highly dependent on four factors: (a) choice of parental populations, (b) choice of markers for ancestry estimation, (c) precise estimation of the parental allele frequencies, and (d) choice of method for ancestry estimation. Studies show that human populations worldwide can be subdivided based on parental/ancestral population combinations from five continents: Africa, Europe and the Middle East, Asia, Pacific Islands, and America (Native American; ref. 51). Groups of Caucasian and African American individuals in the United States today, like those used in this study, have been shown to have a combination of parental/ancestral genes from West Africa and Europe (7, 16, 21). We chose to estimate these two ancestral proportions for each individual. However, to choose proper exact parental populations, with available allele frequencies for the ancestry markers, a well-described history of the immigration and migration of the study population is needed and is not always readily available. The settlement history of Detroit is well described and is available through the Center for Michigan Studies (34). Detroit was first settled by the Polish followed by the Germans and still has a large population of individuals who identify themselves as having Polish or German ancestry (34, 52). It is estimated that 20% to 30% of African Americans in the United States originated from Nigeria and it is believed to be the most homogeneous group in Western Africa (53), making it a rational choice for the estimation of African ancestry. Rwandans were also used because this country is in the most populated area in sub-Saharan Africa having the same linguistic affiliation as many other groups in Africa, indicating recent mixture of this group with other groups in Africa (33).

The choice of markers for ancestry estimation depends on the marker's informativeness for ancestry, which has generally been thought to depend solely on allele frequency differences between parental/ancestral populations or δ (7, 40, 41, 54). However, recent studies show that informativeness for ancestry can also depend on other population genetic events (55), such as which population is the contributor of genes and which population is the acceptor of genes (56). Because there currently is no “standard” set of markers available for ancestry estimation, we chose to use the Federal Bureau of Investigation CODIS Core set of 13 STR loci. This set of markers is readily available in an easy to use, reasonably priced laboratory kit. These markers show considerable allele frequency variation among racial and ancestral groups from around the world (57-59). In particular, for the two ancestral groups used in this analysis, the majority of the 13 markers had composite δ values of ≥0.2. In addition, the CODIS loci were unlinked to each other (59) and were unlinked to the candidate gene of interest in this study, GSTM1, which is a key assumption for ancestry analysis (8, 10).

If the size of the parental population is small, then the precision of the estimation of the allele frequencies is poor (22). Estimation of the allele frequencies from the parental populations used in this study, however, was based on large parental sample sizes.

Controversy about the best method to use to estimate individual ancestry still exists. Therefore, individual ancestry was estimated using both MLE and STRUCTURE to compare the ability of each estimation technique to assess population structure. From the MLE estimates, it was clear which estimates were the European and West African, because ancestral allele frequencies were specified to perform the MLE calculations. Although the STRUCTURE estimates eventually showed similar population structure results compared with MLE estimates, interpretation of the individual cluster proportion estimates in terms of which ancestral population each cluster was describing was difficult, because prior ancestral allele frequency information is not used in the estimation algorithm. However, the STRUCTURE estimates did give a better fit to the data for the modeling of genetic risk compared with the MLE estimates, possibly because it did not rely on prior allele frequency information for the estimation of ancestry.

In summary, this study is one of the first to evaluate the association of individual genetic ancestry with self-reported race in a case-control study of cancer. We found that individual European ancestry did not completely correlate with self-reported race and that there was significant overlap by individual ancestry between the Caucasian, non-Hispanics and African Americans. We also observed that the frequency of a risk genotype, GSTM1 null, varied substantially within self-reported racial group by individual ancestry and case-control status, thereby affecting models of disease risk. We conclude that individual ancestry may confound associations between disease status and a candidate gene risk genotype and could have a direct effect on accuracy of risk estimation for early-onset cancer in this Detroit population.

We thank Thomas Dyer, Ph.D. for his computational support.