ARIC is a longitudinal study of 15,792 adults aged 45–64 years at enrollment in 1987–1989 in four communities: Forsyth County, NC; Jackson, MS (African Americans only); the northwestern suburbs of Minneapolis, MN; and Washington County, MD. Participants attended three subsequent examinations approximately every 3 years (1990–1992; 1993–1995; and 1996–1999).
CARDIA is a longitudinal study investigating 5,115 African American and white men and women aged 18–30 years at enrollment in 1985–1986 in four communities: Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA. Participant recruitment was approximately balanced on age, sex, race, and education status at each community. Six subsequent examinations were conducted (1987–1988, 1990–1991, 1992–1993, 1995–1996, 2000–2001, and 2005–2006).
The offspring cohort of the Framingham Heart Study began in 1971–1975, enrolling 5,124 offspring and spouses of the offspring of the Framingham Heart Study’s original cohort. The offspring cohort participants were aged 5–70 years at their first examination. They were next examined 8 years later and then about every 4 years through the seventh examination, followed by the eighth examination approximately 6.5 years later (2005–2008).
Details of these three studies have been reported elsewhere (10
). The studies were approved by their institutional review boards of the participating institutions. All participants provided written informed consent at each examination.
Initial eligibility criteria for these analyses included participants aged 35–54 years old and nondiabetic (defined as fasting blood glucose <126 mg/dL, and no prior history of and not on medication for diabetes) at their index or “baseline” examination (as detailed below). This yielded 12,119 participants: 8,170 from ARIC, 2,111 from CARDIA, and 1,838 from the Framingham Heart Study. Participants were further excluded if they were not African American or white (n = 30); failed to return for follow-up or did not have at least one follow-up visit to determine diabetes status (n = 709); or at the index examination were either pregnant (n = 12), did not fast >8 h (n = 330), or had missing data on systolic (SBP) or diastolic blood pressure (DBP) or any of the following cardiometabolic traits: fasting blood glucose, insulin, HDL cholesterol (HDL-C), or triglycerides, waist circumference, or BMI (n = 232). After these exclusions, 10,893 participants remained eligible. Those excluded were more likely to be African American (40.5 vs. 23.4%), not finished high school (21.1 vs. 12.4%), have higher SBP (mean 119 vs. 116 mmHg), and prevalence of hypertension (28.1 vs. 21.5%), consume more alcohol (mean 4.8 vs. 3.8 drinks/week), and be a smoker (40.8 vs. 25.2%).
For CARDIA and the Framingham Heart Study offspring cohort, their fifth examination cycle (conducted in 1995–1996 and 1991–1995, respectively) was considered the index examination, for an approximate follow-up period of 10 and 14 years, respectively. These were chosen over prior examinations to ensure a more contemporary sample while allowing for sufficient follow-up of approximately 1 decade. Because the most recent ARIC exam occurred in 1996–1999, its first examination (conducted in 1987–1989) was considered the index examination, for 9 years of follow-up.
Assessment of blood pressure, covariates, and incident diabetes
Blood pressure was measured with participants seated after a 5-minute rest using a random-zero mercury sphygmomanometer in ARIC and CARDIA, and a standard mercury-column sphygmomanometer in the Framingham Heart Study. The average of two readings was used. Three mutually exclusive blood pressure categories were established: hypertension was defined if SBP ≥140 mmHg, DBP ≥90 mmHg, or reported use of antihypertensive medication; prehypertension was defined as not having hypertension, and SBP 120–139 mmHg or DBP 80–89 mmHg; normal included SBP <120 mmHg and DBP <80 mmHg and not using antihypertensive medication.
Height, weight, and waist circumference were measured with participants in light clothing. BMI was calculated as weight in kilograms divided by the square of height in meters. Waist circumference was measured at the level of the umbilicus in ARIC and the Framingham Heart Study, and at the level of the smallest waist circumference in CARDIA. Self-reported information included race, education level, parental history of diabetes (one or both parents with diabetes), smoking status, alcohol use, and physical activity. Participants were instructed to fast overnight before providing blood specimens for measuring glucose, lipid, and insulin levels.
Participants were considered to have incident diabetes if any of the following was present at a follow-up examination: fasting blood glucose ≥126 mg/dL, casual blood glucose ≥200 mg/dL, or using insulin or oral hypoglycemic medication. Time-to-diabetes was estimated using a previously described method by Duncan et al. (13
). For cases ascertained based on blood glucose value, the incident date was estimated by linear interpolation using the glucose values at the ascertaining and previous examinations. For cases ascertained based on the use of diabetic medications, the time-to-diabetes was estimated by using their fasting glucose at the earlier visit and a slope estimated using information from all diabetic subjects who had been unaware of their status (because the fasting glucose at ascertainment for those who were on diabetic medication may have been affected by their knowledge of their diabetes status).
Baseline characteristics by study were examined using simple means and proportions. General linear models were used to compare characteristics by blood pressure categories after adjustment for age and sex (regression models for continuous traits and Poisson regression for categorical traits). In multivariable analyses, triglycerides were natural-log transformed because of their skewed distribution.
Diabetes incidence was calculated using person-years of observation. Age-specific rates were first calculated and then weighted to the standard year 2000 U.S. population to derive age-adjusted incidence rates by race, sex, and blood pressure category. Confidence intervals and trend tests across blood pressure category within race and sex groups were estimated using 2,000 bootstrap resamples. A standard normal distribution in the rates was assumed because the bias between the sample population estimates and the mean of the bootstrap estimates was less than 1%.
Proportional hazards models were fitted to assess the association of blood pressure category with incident diabetes. Tests for effect modification by sex were evaluated in race-specific models that adjusted for age and sex, whereas effect modification by race was tested in age-, race-, and sex-adjusted models. Effect modification by study was also tested. Nonproportionality of hazards over time was tested in race-specific models using time-dependent covariates for the blood pressure categories.
For multivariable analyses, Cox proportional hazards models were constructed to calculate the hazard ratio (HR) and 95% CIs to compare incident diabetes across blood pressure categories, using normal blood pressure as the reference group. The base model adjusted for age and sex. The second model then added BMI. The third, considered our primary model, further adjusted for fasting glucose, HDL-C, and triglyceride. Fasting glucose was included to account for its potential confounding effect on blood pressure and incident diabetes because its levels were positively associated with increasing blood pressure categories in both race groups. The fourth model included a longer list of cardiometabolic risk factors by also introducing fasting insulin and waist circumference. To directly compare the racial differences in the association of blood pressure and incident diabetes, we also combined both races in multivariable analyses with race-specific blood pressure categories using normotensive whites as the referent group.
Sensitivity analyses included constructing models with additional groups of covariates or varying the exclusion criteria. The additional covariates included: current smoking, alcohol use, class of antihypertensive drugs, and physical activity level, as well as education level and parental history of diabetes. Finally, because those with higher blood pressure also tend to have higher fasting glucose level, which is a strong predictor of diabetes, we performed separate multivariable analyses after lowering our exclusion threshold for baseline fasting glucose to 110 mg/dL from 126 mg/dL (i.e., excluding individuals with borderline elevated levels at the index examination).
Antihypertensive medications were classified into one of four categories: β-blockers, thiazides, ACE inhibitors or angiotensin-receptor blockers, or other single-agent medications. Combination or multiple medications were sorted into nonmutually exclusive categories (for example, someone taking a β-blocker and thiazide was included in each of those two drug classes). Questionnaires assessing physical activity were not standardized across studies. Therefore, each study’s physical activity summary score (sum of leisure, sport, and work activity scores in ARIC and CARDIA, and total physical activity in kilocalories over the past year in Framingham) was standardized (mean = 0, SD = 1) for analytic purposes. To achieve a normal distribution in the Framingham physical activity score, natural-log transformation of the original score was performed. All analyses were performed using SAS 9.1 (Cary, NC).