Introduction

For decades, dietary recommendations for dairy foods have been based on predicted health effects of isolated individual nutrients, such as for bone health (eg, expected benefits of calcium and vitamin D), and cardiovascular disease (CVD; eg, expected harms of total fat and saturated fat). Based on these considerations, major dietary guidelines recommend low-fat dairy products and avoidance of whole-fat dairy.¹ However, neither low-fat nor whole-fat dairy foods have major effects on traditional cardiovascular risk factors² nor are appreciably linked to risk of coronary heart disease events.³ Interestingly, self-reported consumption of dairy foods has mixed associations with risk of type 2 diabetes mellitus, without consistently different results for low-fat versus whole-fat products.^4–6 For instance, several studies suggest that yogurt may be beneficial for diabetes mellitus,^4–6 whereas some large studies,^4,5 but not others,⁶ further suggest that cheese, which is highest in dairy fat, may also be protective.

Clinical Perspective on p 1654

Most previous investigations have relied on self-reported dietary information from questionnaires, which can be limited by errors in memory or subjective reporting, especially for dairy fat. In addition, dairy fat is consumed not only from whole foods such as milk, cheese, yogurt, and major dishes such as pizzas, but also throughout the food supply in smaller amounts in mixed dishes, bakery products, and prepared foods. This may increase challenges of accurately capturing the intake of dairy fat from all sources. In comparison, circulating blood biomarkers of fatty acids may provide a more complete measure of dairy fat consumption. These include specific fatty acids obtained primarily from dairy fat including the odd-chain saturated fats pentadecanoic acid (15:0) and heptadecanoic acid (17:0)^7,8 and the natural ruminant trans fat trans-palmitoleate (t-16:1n-7).⁹ These fatty acids are not endogenously synthesized and are obtained only from diet, particularly dairy, making them reasonable biomarkers of dairy fat consumption.^10,11 In addition to their role as biomarkers, it has been hypothesized that these fatty acids could have direct metabolic effects, eg, t-16:1n-7 could mimic experimentally observed effects of adipose-produced c-16:n-7 to increase muscle glucose uptake and suppress hepatic de novo lipogenesis.¹² Yet, few prospective studies have investigated these biomarkers and incident diabetes mellitus, with mixed results (see Discussion).

Thus, the effects of dairy fat, and specific dairy fatty acids, on risk of diabetes mellitus remain unclear. Understanding these relationships is crucial for informing dietary guidelines, and designing additional interventional and experimental studies, as well, to explore biological pathways and mechanisms. We tested the hypothesis that total plasma and erythrocyte biomarkers of dairy fat, 15:0, 17:0, and t-16:1n-7, are associated with a lower incidence of diabetes mellitus in 2 large prospective cohorts.

Methods

Study Design and Population

This investigation was derived from 2 large US prospective studies, the Nurses’ Health Study and Health Professionals Follow-Up Study (HPFS). Details of the cohorts and blood sample collection have been reported.¹³

We used previously measured fatty acid concentrations in stored blood used for nested case-control studies of incident CVD. Among Nurses’ Health Study and HPFS subjects who provided blood samples and were free of prevalent CVD or cancer at the time of blood sampling, we measured total plasma and erythrocyte fatty acid concentrations in 3499 men and women. Approximately 71% of participants were fasting at the time of blood collection. For present analysis focused on incident diabetes mellitus, participants with diabetes mellitus at the time of blood collection were excluded (n=166), leaving a total of 3333 participants for analysis. This included 1364 participants who developed CVD during follow-up and 1969 participants who did not develop CVD during follow-up. All analyses were first conducted in these 2 groups separately, and then combined after confirming that all associations were similar in magnitude and direction. This investigation was approved by the human subjects committees of all participating institutions, and all participants gave implied consent by providing blood samples and returning completed questionnaires.

Fatty Acids

In both cohorts, fatty acid concentrations were measured in stored total plasma and erythrocyte samples in the same laboratory using gas-liquid chromatography.^13,14 Concentrations of individual circulating fatty acids were expressed as a percentage of total fatty acids in plasma or erythrocyte membranes. Technicians and laboratory personnel were unaware of participant clinical information including disease status. Forty fatty acids in plasma and erythrocytes were quantified (Table I in the online-only Data Supplement). For this investigation, primary biomarkers of interest were pentadecanoic acid (15:0), heptadecanoic acid (17:0), and trans-palmitoleic acid (t-16:1n-7). Myristic acid, a much less specific biomarker for dairy fat that is found in dairy, beef, and some plant oils¹⁵ and also endogenously synthesized by the liver (eg, in response to excess carbohydrate intake), was secondarily analyzed. All the participants had levels of these 4 fatty acids above detectable limits.

Reproducibility of these fatty acids (intraclass correlation coefficients) in plasma and red blood cells over time has been reported.^13,16 Average interassay CVs for quality controls in plasma were 16.0% for 15:0, 7.2% for 17:0, 10.4% for t-16:1n-7, and 30.8% for 14:0. Corresponding CVs in erythrocyte measurements were 22.4%, 11.3%, 14.2%, and 39.0%.

Incident Diabetes Mellitus

In both cohorts, participants were asked to report physician-diagnosed diabetes mellitus and the calendar year of diagnosis. To validate self-reports, a supplementary questionnaire was sent obtaining further information on symptoms, diagnosis, and drug treatment. Suspected cases were labeled as confirmed if, according to the supplementary questionnaire, they met ≥1 of 3 National Diabetes Data Group criteria: (1) classic symptoms plus fasting glucose ≥140 mg/dL or random glucose ≥200 mg/dL; (2) at least 2 separate elevated plasma glucose levels (fasting ≥140, random ≥200, or 2-hour challenge ≥200 mg/dL); or (3) medical prescription of oral hypoglycemic agents or insulin. These diagnostic criteria were modified after June 1996 to incorporate lower diagnostic cut offs for fasting glucose of 126 mg/dL. Validity of supplementary questionnaire for diagnosing diabetes mellitus was confirmed in subsets of participants in each cohort by comparison with direct review of medical records: 98% were confirmed in Nurses’ Health Study, and 97% were confirmed in HPFS.^17,18

Covariates and Other Risk Factors

Data on medical history, risk factors, and lifestyle were obtained in both cohorts via validated self-administered questionnaires (online-only Data Supplement). Usual alcohol use and dietary habits were assessed through validated semiquantitative food frequency questionnaires.

Statistical Analysis

Fatty acid levels were evaluated in quartiles as indicator variables, continuously comparing the medians of the highest versus lowest quartile (87.5th versus 12.5th percentiles). Linear trend was assessed by assigning the median value of each quartile with participants and evaluating this as a continuous variable. Cox proportional hazards evaluated associations of each fatty acid with incident diabetes mellitus through 2010, with time at risk from blood collection until first event, death, or censoring at the return of the last questionnaire. Analyses were conducted separately in each cohort and then combined using fixed-effects inverse-variance–weighted meta-analysis;¹⁹ with pooled P-for-trend calculated by using generalized least-squares trend meta-analysis.²⁰ Proportional hazards assumptions were tested by an interaction term with time and not rejected for any fatty acid, with the exception of 14:0 (P-interaction=0.01 in Nurses’ Health Study, 0.01 in HPFS). Visualization of Schoenfeld residuals for each fatty acid, including 14:0, did not suggest a nonzero slope, suggesting that any potential time interaction was small.

Multivariable models were used to minimize potential confounding, with covariates included based on biological relevance, clinical interest, strength of associations with exposure or outcome, or percentage of change in the risk estimate of interest (>5%). For example, because dairy fat intake could be associated with other dietary habits that may also influence diabetes mellitus, we adjusted for fruits, vegetables, fish, meats, whole grains, sugar-sweetened beverages, polyunsaturated fat, calcium, and glycemic load. We also adjusted for biomarker levels of trans-18:1 and trans-18:2, because dairy fat can contain ruminant trans fats; and for palmitic acid (16:0) and stearic acid (18:0) synthesized by hepatic de novo lipogenesis, because these fatty acids can alter relative circulating proportions of trace fatty acids and are linked to insulin resistance and diabetes mellitus in experimental and prospective studies.^13,14,21 Dairy consumption may causally influence body mass index (BMI) – eg, yogurt and cheese^22,23 (the latter, if replacing refined carbohydrates) are linked to lower long-term weight gain and dairy intake reduces weight and fat mass in short-term trials²⁴ – suggesting that BMI could partly mediate any associations between dairy intake and diabetes mellitus. Individuals with different BMIs may also select different dairy foods, which could confound associations. We therefore separately considered BMI as a potential mediator and confounder in an additional multivariable model.

We explored effect modification by age and BMI by using the Wald test for a multiplicative interaction term, with Bonferroni-corrected significance (4 fatty acids × 2 effect modifiers: α=0.05/8=0.006). To minimize misclassification attributable to exposure changes over time, we performed sensitivity analyses restricted to the first 8 years of follow-up. To minimize reverse causation attributable to unrecognized subclinical disease or presence of clinical risk factors at the time of fatty acid measurement, we also excluded cases occurring during the first 2 years. We evaluated self-reported yogurt, cheese, and dairy fat consumption as additional covariates to assess if the associations were independent of these dietary factors outside their specific fatty acid contents. Multiple (10-fold) imputation was performed for missing continuous, and indicator variables for missing categorical covariates. Nonlinear relationships were assessed by semiparametric restricted cubic splines, excluding participants with fatty acid levels <0.5th or >99.5th percentiles to minimize the influence of outliers. Analyses were conducted using SAS 9.3 (Cary, NC), 2-tailed α=0.05.

Results

At baseline, mean±standard deviation age was 64.6±8.6 years among men (range, 48–83 years) and 60.4±6.3 years among women (range, 44–70 years; Table 1, Table II in the online-only Data Supplement). About half of participants were overweight or obese; 1 in 4 had hypertension or a family history of diabetes mellitus; 1 in 3 had hypercholesterolemia; and 1 in 12 men and 1 in 5 women were current smokers. Average total dairy consumption was 2.1 servings/d, with similar contributions from whole-fat (≈45%) and low-fat (≈55%) products.

Within each circulating lipid compartment (plasma, erythrocyte), the different fatty acid biomarkers were modestly intercorrelated: in plasma, adjusted intercorrelations (r) ranged from 0.10 to 0.59 (Table III in the online-only Data Supplement); and in erythrocytes, from 0.47 to 0.69 (Table IV in the online-only Data Supplement). As expected for these different lipid compartments, correlations between plasma and erythrocyte fatty acid biomarkers were moderate: 0.32 for 15:0, 0.44 for 17:0, 0.42 for t-16:1n-7, and 0.24 for 14:0. Partial correlations with self-reported dairy fat consumption were modest for plasma 15:0 (r=0.29), 17:0 (r=0.21), and t-16:1n-7 (r=0.22); and weaker for 14:0 (r=0.11; Table III and Figure I in the online-only Data Supplement); and similar but somewhat lower for erythrocyte fatty acids (Table IV in the online-only Data Supplement). These correlations were similar among participants who were fasting at the time of blood collection (data not shown).

Plasma Fatty Acid Biomarkers and Risk of Diabetes Mellitus

During mean±standard deviation follow-up of 15.2±5.6 years, 277 new cases of diabetes mellitus were diagnosed. After adjustment for demographics, metabolic risk factors, lifestyle, dietary habits, and other circulating fatty acids, in comparison with the lowest quartile, individuals in the highest quartile of plasma 15:0 had 44% lower risk of diabetes mellitus (hazard ratio [HR],0.56; 95% confidence interval [CI], 0.37–0.86; P-trend=0.01); of plasma 17:0, 43% lower risk (HR, 0.57; 95% CI, 0.39–0.83; P-trend<0.01); and of t-16:1n-7, 52% lower risk (HR, 0.48; 95% CI, 0.33–0.70; P-trend <0.001; Table 2). Findings were generally similar in magnitude and direction in the 2 separate cohorts, without statistically significant effect modification by sex. 14:0 was not significantly associated with diabetes mellitus (P-trend=0.36).

Evaluated continuously, 15:0 was associated with 38% lower risk (pooled HR, 0.62; 95% CI, 0.46–0.85); 17:0, with 32% lower risk (HR, 0.68; 95% CI, 0.50–0.91); and t-16:1n-7, with 46% lower risk (HR, 0.54; 95% CI, 0.40–0.73; Table 3). Similar to categorical analyses, plasma 14:0 was not associated with diabetes risk (HR, 0.82; 95% CI, 0.60–1.11).

Because BMI is a major predictor of insulin resistance and diabetes mellitus, we conducted further multivariable-adjusted analyses to evaluate the potential independent associations between BMI and dairy fatty acid biomarker levels at baseline. In both cohorts combined, higher 15:0 was not associated with BMI (Q1 versus Q4: 25.3 versus 25.3 kg/m², P-trend=0.97; Table II in the online-only Data Supplement). Higher 17:0 (25.8 versus 25.0, P-trend<0.001) and t-16:1n-7 (25.9 versus 25.2, P-trend<0.001) were associated with lower BMI, although absolute differences were modest. 14:0 was associated with higher BMI (24.9 versus 25.9, P-trend<0.001). After further adjustment for baseline BMI as a potential mediator or confounder, associations with incident diabetes mellitus were slightly attenuated for 17:0 and not appreciably altered for other fatty acids (Table 3). Results were also similar after further multivariable adjustment for conjugated linoleic acid, another ruminant-derived fatty acid (data not shown).

We further examined the potential influence of each individual covariate in the model. Adjustment for fatty acid biomarkers of de novo lipogenesis, 16:0 and 18:0,^25,26 had the greatest effect on plasma 15:0, which was no longer significantly associated with diabetes mellitus when 16:0 and 18:0 were removed as covariates (continuous HR, 1.10; 95% CI, 0.85–1.42). Conversely, removal of these fatty acid covariates had smaller influence on associations with diabetes mellitus for plasma 17:0 (HR, 0.67; 95% CI, 0.51–0.89) or t-16:1n-7 (HR, 0.68; 95% CI, 0.51–0.92).

With simultaneous mutual adjustment for all 3 dairy fat biomarkers, associations were nonsignificant for plasma 15:0 (continuous HR, 0.83; 95% CI, 0.53–1.29) and 17:0 (HR, 0.93; 95% CI, 0.64–1.36), and unchanged for t-16:1n-7 (HR, 0.59; 95% CI, 0.43–0.80).

Semiparametric Analyses

Restricted cubic splines demonstrated an inverse linear relationship for plasma 15:0, 17:0, and t-16:1n-7 and incident diabetes mellitus (Figure). Plasma 14:0 was not associated with risk (P-linearity=0.48).

• Download figure
• Download PowerPoint

Erythrocyte Fatty Acid Biomarkers

When we evaluated erythrocyte fatty acids, a lower incidence of diabetes mellitus was seen with higher levels of 17:0 (HR, 0.54; 95% CI, 0.34–0.87; P-trend <0.001), and t-16:1n-7 (HR, 0.78; 95% CI, 0.51–1.18; P-trend=0.05; Table V in the online-only Data Supplement). Erythrocyte 15:0 was not significantly associated with incident diabetes (HR, 0.83; 95% CI, 0.55–1.25; P-trend=0.63), and erythrocyte 14:0 was associated with a nonsignificant trend toward higher risk (HR, 1.56; 95% CI, 0.98–2.49; P-trend=0.13). The latter association was driven by findings in HPFS.

When these erythrocyte fatty acids were evaluated continuously, similar findings were seen for 15:0 and 17:0; t-16:1n-7 was no longer associated with lower risk (HR, 0.80; 95% CI, 0.57–1.13); and 14:0 was associated with 36% higher risk (HR, 1.36; 95% CI, 1.11–1.67; Table VI in the online-only Data Supplement). As with the categorical analyses, the latter association was driven by findings in HPFS only (data not shown). As seen with plasma 17:0, further adjustment for BMI as a mediator or confounder partly attenuated the observed inverse association for erythrocyte 17:0. Among other fatty acid covariates in the multivariable model, removal of adjustment for fatty acid biomarkers of de novo lipogenesis had little effect on these risk estimates (data not shown).

Sensitivity Analyses

When we evaluated potential effect modification, the associations of plasma and erythrocyte fatty acids with incident diabetes mellitus did not significantly vary according to differences in age or BMI (P>Bonferroni-corrected α=0.006 each). Associations were also similar after further adjustment for self-reported consumption of yogurt, cheese, or dairy fat (Table VII in the online-only Data Supplement). For example, in comparison with the association of plasma t-16:1n-7 with diabetes mellitus in the main model (continuous HR, 0.54; 95% CI, 0.40–0.73), findings were not appreciably altered by additional adjustment for reported consumption of yogurt (HR, 0.55; 95% CI, 0.41–0.73), cheese (HR, 0.53; 95% CI, 0.40–0.71), or dairy fat (HR, 0.52; 95% CI, 0.38–0.70). Results were also generally similar after excluding cases occurring within the first 2 years and restricting to the first 8 years of follow-up (Table VIII in the online-only Data Supplement).

Discussion

In 2 separate cohorts of US men and women, 3 plasma biomarkers of dairy fat, 15:0, 17:0, and trans-16:1n-7, were associated with lower risk of incident diabetes mellitus. Findings were generally similar for erythrocyte fatty acids, although with smaller and not always significant magnitudes of association. These results provide new evidence on associations of dairy-derived circulating fatty acids and risk of diabetes mellitus.

Our findings suggest that either these dairy fatty acids themselves or other correlated factors in dairy fat could reduce the risk of diabetes mellitus. For instance, short- and medium-chain saturated fats, vitamin D (a fat-soluble vitamin), omega-3 fats, or gangliosides in dairy fat could each play a role.²⁷ The magnitudes of our associations, linear dose response, independence from established diabetes risk factors and other dietary habits, and consistency of findings across sensitivity analyses argue against residual confounding as the sole explanation for our findings. These results highlight the need for intensive experimental and mechanistic evaluation on health effects of dairy fat, and dietary and potential metabolic determinants of these circulating fatty acids, as well, including their sensitivity to indicators of de novo lipogenesis.

In observational analyses, self-reported consumption of high-fat dairy foods such as cheese have beneficial or neutral associations with diabetes mellitus,^4–6 whereas dairy fat consumption and dairy fat biomarkers correlate with improved hepatic and systemic insulin sensitivity and lower hepatic steatosis.^28,29 When we included all 3 plasma fatty acids in 1 model, only t-16:1n-7 retained independent significant association with diabetes mellitus. We have hypothesized¹² that benefits could be attributable to isomeric similarities between t-16:1n-7, a natural ruminant trans fat, and cis-16:1n-7 that, when derived from adipose tissue through de novo lipogenesis, may act in feedback loops to reduce hepatic fat synthesis and increase muscle insulin sensitivity.^30,31 We wonder whether t-16:1n-7, consumed in the diet, might have similar effects.

Fatty acids 15:0, 17:0, and t-16:1n-7 are predominantly obtained from diet and not synthesized, making reverse causation because of abnormal metabolism at baseline unlikely. One small study (n=12) found that providing a mixed oil high in trans-18:1 (vaccenic acid) also increased serum t-16:1n-7, whereas human peripheral blood mononuclear cells cultured with t-18:1 incorporated small amounts of t-16:1n-7 into cellular lipids.³² These findings suggest small amounts of t-16:1n-7 could derive from partial β-oxidation (endogenous chain shortening) of t-18:1; labeled fatty acid tracer studies are needed to confirm this result. In the present study, among all dietary factors, circulating t-16:1n-7 correlated most with dairy fat rather than sources of industrial trans fat (vaccenic acid), suggesting that direct dairy consumption remains a major source. Genome-wide association studies have not identified significant genetic determinants of circulating t-16:1n-7,³³ further suggesting the absence of strong endogenous influences. Body weight and insulin resistance produce no known effects on levels of these circulating fatty acids, and findings were generally similar following adjustment for BMI. Reverse causation could play a role in behaviors, for example, if higher-risk participants with subclinical prediabetes elected to avoid whole-fat dairy products. Yet, the prospective nature of our analysis minimizes this possibility; and 15:0, 17:0, and t-16:1n-7 also remained inversely associated with diabetes mellitus after excluding cases occurring in the first 2 years.

In these cohorts, plasma fatty acid biomarkers correlated more strongly with dairy fat intake than did erythrocyte biomarkers, and were also more strongly inversely associated with diabetes mellitus. Correlations of all these fatty acids with self-reported dairy consumption were modest. This may be attributable to random or systematic errors in self-reported diet, variability of these fatty acids in different dairy foods,³⁴ laboratory error in fatty acid measures, within-person variation in diet or circulating fatty acids, or other unknown influences on bioavailability, metabolism, or incorporation into specific lipid compartments of these fatty acids. Notably, dairy fat is consumed not just as whole foods (milk, cheese, yogurt, butter) but mixed into numerous foods, dishes, and recipes as major and minor ingredients. In our analysis of National Health and Nutrition Examination Survey (NHANES) data based on detailed, product-specific dietary recalls (2005–2012), 51% of cheese and 30% of total dairy is consumed in mixed dishes, especially grain products but also mixed with meats, sweets, vegetables, and eggs (data not shown). Food frequency questionnaires that estimate dairy fat intakes from whole foods and major mixed sources (eg, pasta dishes, burritos, pizza) may not accurately capture quantities in these mixed dishes nor the multitude of smaller amounts in many other products. Thus, the observed modest correlations of 15:0, 17:0, and t-16:1n-7 with self-reported dairy fat may appropriately reflect the challenges in fully estimating dairy fat from questionnaires.

Fatty acid 14:0 had weak correlations with self-reported intakes of dairy fat, meat, and carbohydrate-rich foods (the latter being drivers of endogenous hepatic fatty acid synthesis). Plasma 14:0 was unassociated with diabetes risk; erythrocyte 14:0 was associated with increased risk in 1 cohort. Higher laboratory imprecision in 14:0 measures could limit the ability to detect associations. Trend toward higher risk could also relate to derivation of 14:0 from endogenous de novo lipogenesis, a correlate of insulin resistance and diabetes mellitus.²¹ This could also be a chance finding and should be interpreted with caution. Recent analysis in the European Prospective Investigation into Cancer and Nutrition (EPIC) study identified a positive association between plasma phospholipid 14:0 and incident diabetes mellitus.²¹ Our findings support the need for further studies on 14:0, especially on endogenous metabolic determinants of its circulating levels.

In pooled analyses of observational studies, self-reported intakes of total dairy, low-fat dairy, whole-fat dairy, milk, and cheese have each shown mixed (beneficial or neutral) associations with diabetes mellitus; yogurt consumption is more consistently associated with lower risk.^4–6 Interestingly, none of these studies identified harms of dairy, including whole-fat dairy, for diabetes mellitus. A handful of prior prospective studies have evaluated circulating biomarkers of dairy fat consumption in relation to incident diabetes mellitus (Table 4). Only 2 studies^12,35 reported on t-16:1n-7; both found substantially lower risk, consistent with our findings. Three of 5 previous studies evaluating 15:0 and 2 of 3 studies evaluating 17:0 observed significant inverse associations with diabetes mellitus; none observed higher risk. In the present analysis, t-16:1n-7 remained associated with diabetes mellitus after adjustment for the other dairy fat biomarkers, suggesting it may be particularly relevant for further investigation of its determinants and biological effects. Our findings build on and extend previous studies by evaluating these fatty acid biomarkers, including t-16:1n-7, in total plasma and erythrocytes in 2 large, well-established cohorts of US men and women and including adjustment for a range of potential confounding factors.

Our investigation has several strengths. We assessed circulating fatty acids, which may more fully capture all dietary sources of dairy fat. We evaluated 2 separate lipid compartments, in which generally similar associations support validity of the findings, especially in light of their potential differential temporal and metabolic responses to dietary fats intake. The prospective cohort design minimized reverse causation, selection bias, and recall bias. Inclusion of 2 separate cohorts including both men and women provided for replication, and pooling, as well, of cohort findings.

Potential limitations should be considered. Misclassification because of laboratory error and within-person changes over time would attenuate results, making it more difficult to detect true associations. Despite this, we were able to detect associations with diabetes mellitus, suggesting true relationships could be even stronger. Residual confounding attributable to unmeasured or mismeasured covariates cannot be excluded. However, magnitudes of associations and linear dose responses suggest that residual confounding is unlikely to fully explain the results. These cohorts comprised educated US health professionals. Yet, although absolute levels of fatty acids and rates of diabetes mellitus may vary by education, geography, or race/ethnicity, we do not suspect that biological effects of these fatty acids should differ on a relative scale across different populations.

In summary, we found that plasma 15:0, 17:0, and trans-16:1n-7 were associated with a lower incidence of diabetes mellitus in 2 separate prospective cohort studies.

Acknowledgments

Drs Yakoob, Hu, and Mozaffarian participated in project conception and development of research methods; Drs Hu and Mozaffarian obtained funding and provided oversight; Drs Yakoob, Shi, and Orav analyzed data and performed analysis; Drs Yakoob and Mozaffarian drafted the manuscript; and Drs Willett, Rexrode, Campos, Orav, Hu, and Mozaffarian provided critical feedback on revisions and other intellectual input.

Footnotes