Regarding "Testing for population subdivision and association in four case-control studies".

To the Editor: Ardlie et al. (2002) recently found no evidence for population structure in separate case-control studies of type 2 diabetes and hypertension in U.S. whites and only weak evidence of structure in a case-control study of hypertension in African Americans. These results are consistent with the theoretical results of Wacholder et al. (2000), who found that the magnitude of bias due to unrecognized population stratification is likely to be small under most plausible scenarios. To further evaluate the potential bias due to stratification for these and other conditions, we conducted a series of case-control studies for six common phenotypes in a population-based sample of U.S. adults. The study population included 444 unrelated adults (231 African Americans and 213 non-Hispanic whites) randomly selected from five U.S. communities as part of the Hypertension Genetic Epidemiology Network (HyperGEN) of the National Heart, Lung, and Blood Institute (NHLBI) Family Blood Pressure Program (Williams et al. 2000). The study was approved by the institutional review boards at each institution, and appropriate informed consent was obtained from human subjects. Phenotypes measured included: (1) obesity (BMI⩾30), (2) hypercholesterolemia (total plasma cholesterol⩾240 mg/dl or current use of medications to lower cholesterol), (3) hypertension (systolic blood pressure⩾140 mmHg, diastolic blood pressure⩾90 mmHg, or current use of medications to lower blood pressure), (4) diabetes (fasting serum glucose⩾126 mg/dl, nonfasting glucose⩾200 mg/dl, self-reported physician diagnosis of diabetes, or current use of hypoglycemic medications), (5) renal dysfunction (serum creatinine ⩾ sex-specific 90th percentile [1.4 mg/dl in men and 1.1 mg/dl in women]), and (6) cardiovascular disease (self-reported history of heart attack, stroke, or coronary artery bypass surgery). For each phenotype, those who did not meet the case definition served as control individuals. We constructed contingency tables and performed χ2 tests of association for these six phenotypes with each of 368 STR markers typed by the NHLBI Mammalian Genotyping Service at Marshfield, WI (screening set 10). Like Ardlie et al. (2002), we then computed a statistic, χs2, to test for overall differences in allele frequencies between each set of case individuals and control individuals (Pritchard and Rosenberg 1999). To simplify the analysis and ensure that expected values in contingency tables were sufficiently large (>5) for the classical χ2 test, we converted each STR marker to a biallelic marker by selecting one index allele for each marker and then collapsing all other alleles for that marker into a single alternative allele. Index alleles for each marker were selected by first choosing alleles with allele frequencies of at least 15% in both African Americans and whites and then selecting the allele that demonstrated the largest absolute difference in allele frequencies between racial groups. The prevalence of several of the phenotypes differed substantially between racial groups (table 1). In crude analysis pooling both racial groups, the percentage of markers nominally associated (P<.05) with each phenotype was higher than expected, under the null hypothesis, for diabetes (8.4%), hypertension (7.9%), renal dysfunction (7.6%), and hypercholesterolemia (5.4%) but not for cardiovascular disease (4.9%) or obesity (4.9%). The summary test for stratification incorporating all 368 markers (i.e., 368 df) was statistically significant for diabetes, renal dysfunction, and hypertension (table 1), indicating overall differences in allele frequencies between case individuals and control individuals. However, after adjustment for race or stratification by race, there was no evidence of cryptic stratification for any of the six phenotypes, with the possible exception of obesity in whites. Table 1 Summary Tests for Population Stratification by Phenotype Our results provide further evidence that hidden or unrecognized population stratification is unlikely to be a serious threat to the validity of case-control designs that appropriately account for ethnicity in either the design or analysis phase of the study (Wacholder et al. 2000; Ardlie et al. 2002). Because of the large number of markers tested, it is likely that our study was even more sensitive to subtle background genetic differences between case individuals and control individuals than that conducted by Ardlie et al. (2002), which included only 9 STR markers and 35 SNP markers. We think that other factors, such as selection bias, chance, publication bias, gene-environment interactions, and differences in linkage disequilibrium patterns across study populations, are more plausible explanations for inconsistency of results between genetic association studies.