Study sample

The Framingham Heart Study is a community-based, longitudinal, observational cohort study that was initiated in 1948 to prospectively investigate risk factors for coronary heart disease. The Framingham Offspring Study began in 1971 with recruitment of 5124 men and women who were offspring (or offspring's spouses) of the Original Cohort of the Framingham Heart Study. The design and selection criteria of the Framingham Offspring Study have been described elsewhere.[9] Participants were prospectively assessed during 7 examinations between 1971 and 2001. The consecutive examination dates were as follows: 1971-1975; 1979-1982; 1984-1987; 1987-1990; 1991-1995; 1996-1998, and 1998-2001. At each examination visit, participants underwent medical history, physical examination, anthropometry, and laboratory assessment of coronary heart disease risk factors, as previously described.[9] Framingham participants signed informed consent and the study is reviewed annually by the Boston University Medical Center Institutional Review Board.

There were 5124 participants who completed Offspring Examination 1 (in 1971-1975), of which 4989 (97%) agreed for their data to be in the open-access dataset. Of these, 1099 subjects were excluded from the present analyses because of missing education data (primarily among the participants who did not attend exams 2 or 3 when education was assessed) or being <28 years of age (n = 60) when education was assessed. Participants were restricted to those aged ≥28 years at the time education was assessed in order to allow at least 10 years from expected completion of high school (at age 18 years, on average), during which the participants could obtain higher levels of education. Consequently, 3890 subjects were included in the data analyses.

Education

The participants' own education was measured directly from Framingham Offspring Study participants at Examinations 2 (1979-1982) and 3 (1984-1987). Examination 3 education was used whenever available, otherwise the Examination 2 measure was used. In the original data, education was recorded in 6 categories of completed years of education: 0-4, 5-8, 9-11, 12, 13-16, ≥17 years. For current analyses, the participants' own education was collapsed into 3 groups: ≤12 years (reflecting high school or less), 13-16 years (indicative of some post-secondary education including technical school and college degree) and ≥17 years education (approximating those with more than an undergraduate college degree). This grouping was motivated by both (i) statistical power considerations (to ensure adequate number of participants in each category) and (ii) substantive reasons, whereby the education categories represent educational milestones recognized to influence earnings, occupation type, and socioeconomic position in society.

Blood Pressure

Each participant rested for at least five minutes before blood pressure measurement. While the participant remained seated, a physician measured SBP and DBP each twice in the left arm with a mercury-column sphygmomanometer, according to a standardized protocol [10]. The average of the two readings was used for analyses.

Covariates

Covariates were measured at each exam. A binary indicator of current cigarette smoking was determined by self-report, defined as smoking regularly in the year prior to the examination (yes/no). Alcohol consumption was evaluated by self-reported average number of alcoholic drinks (e.g. beer, wine, cocktails) per week. Body mass index was calculated as the weight in kilograms divided by the square of the height in meters (kg/m²). Current antihypertensive medication use was self-reported and modeled as a binary variable (yes/no). "Baseline age" represented age at Examination 1. "Time from baseline age" was calculated as the difference between age at a given examination and the baseline age.

Statistical Analyses

Primary analyses focused on associations of education (categorized as ≤12, 13-16, and ≥17 years, as described above) with longitudinal trajectories of SBP and DBP. Analyses relied on multivariable mixed linear models, which extend multiple linear regression to longitudinal analyses of repeated measures [11]. Accordingly, all effects reported in this study are likelihood-based estimates from mixed models, as are all 95% confidence intervals and test statistics used for inference about these estimates. The blood pressure measures from consecutive examinations represented repeated values of the continuous dependent variable. To model the dependence between repeated outcome measures, we used the autoregressive order 1 AR(1) covariance structure of the residuals, which assumed that blood pressure values measured at consecutive visits are correlated more strongly than those separated by longer time intervals [12]. All models adjusted for baseline age and time since baseline assessment (the former allowed us to adjust for potential cohort effects, such as increasing education over time in the United States). In further analyses, we additionally adjusted for several time-varying conventional risk factors for hypertension expected to be involved in potential mechanisms by which educational attainment may influence blood pressure (representing visit-specific binary indicators of current use of anti-hypertensive medication and current smoking, as well as time-varying continuous measures of alcohol consumption and body mass index). AR(1) is a standard choice for the covariance matrix in mixed model analyses of longitudinal data. Because there are no well established tests to compare fit of models' based on alternative covariance structures, the AR(1) structure is usually selected a priori. Assessment of the consistency of the AR(1) assumption is shown in the Results section. We implemented this longitudinal analysis by using PROC MIXED, with AR(1) covariance structure specified in REPEATED statement in SAS [13].

The second class of models additionally adjusted all the education effects for baseline blood pressure. Accordingly, in these models, baseline blood pressure values were not used as the outcome measure, so that the number of dependent value measures for each subject was reduced by one. Adjustment for baseline blood pressure effectively implied that we compared post-baseline trajectories of blood pressure as if participants with different education had the same baseline blood pressure. This approach allowed us to separate the antecedent effects of education, and other related factors, that might have resulted in the baseline differences in blood pressure, from the potential long-term effects of education on post-baseline changes in blood pressure.

Preliminary analyses were carried out to determine the most accurate representation of the effects of baseline age, and time since baseline. In particular, we a priori expected that both between- and within-subject effects of aging on blood pressure may be non-linear. Accordingly, we gradually expanded the basic model, with linear effects of age and time, by adding and testing first quadratic and then cubic effects of each of the two variables, while adjusting for the use of anti-hypertensive medication and for education. Furthermore, because we expected that the impact of within-subject aging (time) on blood pressure may vary depending on the baseline age, we also tested linear and quadratic interactions between age and time. All the multivariable mixed models employed in the final analyses, described below, adjusted for only those non-linear effects of age or time, and those interactions between these effects, that were statistically significant, based on Wald test with 2-tailed α = 0.05. Once the optimal representations of the effects of age and time, as well as of their interactions, were determined, these representations were used in the final analyses of the adjusted association of education with SBP and DBP. The final representation of the effects of baseline age and time from baseline for analyses on SBP was as follows: age+age²+time+time²+age*time+age*time²+age²*time. For DBP, it was as follows: age+age²+time+ time²+age*time+age*time²+age²*time+ age²*time². All analyses were sex-specific, as a formal test for sex-by-education interaction suggested that the effects of education may differ between males and females (p = 0.046 for SBP; p = 0.063 for DBP).

Prior Literature

Few studies have investigated sex-specific longitudinal trajectories of blood pressure, particularly over a substantial proportion of the life course. Diez Roux et al. in the ARIC cohort (n = 8555) aged 45 to 64 years at baseline and followed using 4 examinations over a period of 9 years, found in white participants, that education was marginally inversely associated with increases in blood pressure after adjusting for age, sex, center, medication use, and reported interactions between time and sex, and interactions between time and baseline age [4]. The 5-year change in mean SBP was 6.0 mmHg for those with <high school degree and 5.3 mmHg for those with a college degree. Further adjustment for baseline SBP somewhat reduced the association strength, to 5.9 mmHg for <high school and 5.5 mmHg for participants with college degree. Associations were weaker in black participants. Strand et al. demonstrated, in a large prospective study of 48,422 males and females aged 35-49 followed for 14 years using three examinations, that education was inversely associated with increases over time in SBP in males and females, after adjusting for year of birth [6]; socioeconomic disparities widened over time in females but not males. In a study on the Framingham Offspring cohort that included only participants aged 20-29 years at baseline (many of whom may not have completed education yet), education was not significantly associated with mean 8-year change in SBP or DBP in males or females, after adjusting for age [5]. In the CARDIA study of 2913 participants aged 18-30 years at baseline education was significantly inversely associated with mean 15-year change in both SBP and DBP [7]. Specifically for SBP, those with <high school degree had a 15-year mean increase of 8.2 mmHg versus only 0.7 mmHg for participants with >college graduate degree. However the observed associations were not adjusted for covariates [7]. Although prior cross-sectional studies suggested that associations may be stronger in females than males [3], little is known about sex-specific associations between education and blood pressure trajectories, particularly over long periods of the life course (>20 years follow-up). Finally, little is known about the effects of adjusting for use of antihypertensive medications, body mass index, alcohol consumption, smoking or other potential mechanisms that may, at least partly, mediate the impact of lower education on longitudinal trajectories of blood pressure. This study added to the literature sex-specific information demonstrating that education is inversely associated with longitudinal trajectories of mean SBP in females and males over a substantial proportion of the life course (approximately 30 years) and that association may be stronger in females than males. Furthermore, in females, lower education was associated with a higher mean post-baseline SBP even among participants with the same baseline SBP, suggesting a possible long-term impact of lower education. Adjusting for up-dated values of conventional risk factors typically reduced strengths of association, but in females the impact of lower education remained statistically significant. For DBP, association strengths were generally weaker for both females and males.

Mechanisms

The primary candidate mechanisms by which education may influence longitudinal trajectories of blood pressure involve conventional risk factors for hypertension, including smoking, obesity, blood pressure medication use, and alcohol consumption. In this study, in females, education was inversely associated with anti-hypertensive medication use, body mass index and smoking, and directly associated with alcohol consumption. In males, education was inversely associated with body mass index, alcohol consumption and smoking, and not associated with antihypertensive use. Furthermore, the estimated effects of education tended to somewhat decrease after adjusting for these potential mechanisms (particularly in males), suggesting that they may be at least partial explanatory pathways for the observed association between educational attainment and longitudinal trajectories of blood pressure. It is important to note that biases can be induced by adjusting for variables that may partly mediate the effect of exposure; therefore, these mechanistic findings should be interpreted with caution [14]. Furthermore, there remain plausible confounders unadjusted for, such as childhood socioeconomic circumstances (which are associated with adulthood education and blood pressure [15]), parental blood pressure (which may be associated with offspring education and has been related to offspring blood pressure [16]), intelligence (which is associated with educational attainment and CHD risk [17]), and early life obesity (that could affect upward social mobility via obesity discrimination particularly in women [18, 19], and is related to blood pressure in adulthood [20]. Consequently, residual confounding remains a possibility.

Low educational attainment has been demonstrated to predispose individuals to high strain jobs, characterized by high levels of demand and low levels of control, which have been associated with elevated blood pressure [21, 22]. Other related mechanisms involve stress-induced sympathetic nervous system activation due to stressful conditions outside of work, that are also associated with low educational attainment. These may be particularly important for women. It has been shown that women with low education have higher risk of co-occurring psychosocial determinants of poor health, including single-parenting, depression, income below the poverty threshold, and unemployment, compared to men with low education [23]. Consequently, low socioeconomic position may be a stronger determinant of hypertension risk in women compared with men. This may be one of the explanations for why we found a significant interaction between sex and education, and somewhat stronger associations between education and blood pressure in women than men. The extent of health care available for people of low socioeconomic position is typically less than what is available for those with high socioeconomic position, hence limiting access to treatments of hypertension [24]. Furthermore, there is evidence that people of low socioeconomic position have less healthful diets, such as lower rates of fruit and vegetable consumption, and higher salt intake, which may be additional mechanisms contributing to disparities in blood pressure [25, 26].

It has been demonstrated that although both SBP and DBP are positively associated with incidence of coronary heart disease, there are differences in the way SBP and DBP evolve over the life course. SBP tends to increase steadily with age, while DBP tends to increase until age 50 years, and to decrease steadily after that age [4, 8, 27]. The mechanisms responsible for the age-related increase in DBP among younger people likely involve an atherosclerotic increase in peripheral resistance, caused by narrowing of the smaller arteries and arterioles [8, 28]. In contrast, for older individuals, structural damage and calcification due to arteriosclerosis in the larger conduit arteries can result in loss of arterial compliance, which can cause a rise in SBP, but a reduction in DBP [8, 28]. As the burden of hypertension is greatest after the age of 50 years, and it is exceedingly uncommon to have diastolic hypertension without concurrent systolic hypertension in adults over the age of 50 years, it has been argued that SBP is by far the more important measure of the two in terms of predictive importance for population health [8].

Studies generally show consistent inverse associations between educational attainment and longitudinal changes in SBP [4, 6, 7], with the exception of young participants aged 20-29 years at baseline, followed over 8 years in the study by Hubert et al. [5] However, findings are less consistent for DBP, where studies have shown inverse [7], null [5], or even positive [4] associations between educational attainment and longitudinal changes in DBP. Our findings demonstrated fairly robust inverse associations of education with SBP, and weaker inconsistent associations with DBP. The pathophysiological mechanisms (e.g. smoking, obesity, alcohol consumption) that cause steady increases over the life course for SBP but not DBP, and also tend to be inversely associated with socioeconomic position, may explain the more consistent findings for the inverse association between education and changes in SBP rather than DBP over the life course. However, adjustment for these variables in our study appeared to account for only a small amount of the association in females, and a larger amount of the (weaker) association in males, suggesting there may be other explanatory factors, particularly in females.

Strengths and Weaknesses

Strengths of this study include having access to data on approximately 30 years of longitudinal blood pressure measurements. Furthermore, follow-up rates of the Framingham Heart Study are considered to be high for observational studies, decreasing risk of bias due to loss-to-follow-up. Finally, measurements of blood pressure were performed using methods and equipment providing good accuracy and precision, and analyses relied on statistical methods appropriate for longitudinal repeated-measures studies.

With regard to weaknesses, because the historical design of the Framingham Offspring Study reflected the population of Framingham, Massachusetts at study onset in 1948, the Original and Offspring cohorts are largely composed of white participants. Consequently, the generalizability of our findings to other races/ethnicities is uncertain. Furthermore, although we had up to 7 measurements for each covariate, we expect there to be reasonable residual confounding due to imperfect measurement of obesity (body mass index), and self-reported alcohol consumption, smoking and antihypertensive medication use.