Nutritional epidemiology and the Women's Health Initiative: a review

ABSTRACT

The dietary modification (DM) clinical trial, within the Women's Health Initiative (WHI), studied a low-fat dietary pattern intervention that included guidance to increase vegetables, fruit, and grains. This study was motivated in part from uncertainty about the reliability of observational studies examining the association between dietary fat and chronic disease risk by using self-reported dietary data. In addition to this large trial, which had breast and colorectal cancer as its primary outcomes, a substantial biomarker research effort was initiated midway in the WHI program to contribute to nutritional epidemiology research more broadly. Here we review and update findings from the DM trial and from the WHI nutritional biomarker studies and examine implications for future nutritional epidemiology research. The WHI included the randomized controlled DM trial (n = 48,835) and a prospective cohort observational (OS) study (n = 93,676), both among postmenopausal US women, aged 50–79 y when enrolled during 1993–1998. Also reviewed is a nutrition and physical activity assessment study in a subset of 450 OS participants (2007–2009) and a related controlled feeding study among 153 WHI participants (2010–2014). Long-term follow-up in the DM trial provides evidence for intervention-related reductions in breast cancer mortality, diabetes requiring insulin, and coronary artery disease in the subset of normotensive healthy women, without observed adverse effects or changes in all-cause mortality. Studies of intake biomarkers, and of biomarker-calibrated intake, suggest important associations of total energy intake and macronutrient dietary composition with the risk for major chronic diseases among postmenopausal women. Collectively these studies argue for a nutrition epidemiology research agenda that includes major efforts in nutritional biomarker development, and in the application of biomarkers combined with self-reported dietary data in disease association analyses. We expect such efforts to yield novel disease association findings and to inform disease prevention approaches for potential testing in dietary intervention trials. This trial was registered at clinicaltrials.gov as NCT00000611.

Introduction

In 1983 the National Cancer Institute (NCI) initiated a feasibility phase of a low fat dietary intervention trial for breast cancer prevention (1). This Women's Health Trial was motivated by international comparisons and time trends in breast cancer rates and by experiments in rodents that reported adverse effects from high-fat diets. The few existing case–control or cohort studies on dietary fat and breast cancer at that time tended to be mixed in their findings, but a later summary analysis of cohort studies (2) offered little support for the dietary fat and breast cancer hypothesis. This trial did not proceed to full scale at that time, largely because of concerns about adherence to dietary intervention goals. The investigators subsequently submitted an R01 proposal to fund a full-scale trial. This proposal was favorably reviewed, though the NCI committed funds only for an additional feasibility study of a low-fat dietary intervention among racial/ethnic minority women. While this feasibility study (3) was ongoing, Bernadine Healy came into her role as NIH Director in 1991 and obtained funding for the Women's Health Initiative (WHI), which included a low-fat dietary pattern intervention trial and trials of the benefits and risks of menopausal hormone therapy, among other components.

WHI program development by an NIH working group led to the initiation of WHI in 1992 (4). During that process, the low-fat trial design evolved to include colorectal cancer in addition to breast cancer as a primary outcome, with coronary artery disease as a secondary outcome. The dietary intervention continued to focus on fat reduction, but with related encouragement to increase vegetables, fruits, and grains. Recruitment into this Dietary Modification (DM) trial took place from 1993–1998 and included 48,485 participants, 40% of whom were randomly assigned to the low fat dietary pattern intervention group, whereas 60% were assigned to a usual diet comparison group (DM-C). A further 93,676 participants were enrolled in a companion prospective cohort Observational Study (OS). All participants were postmenopausal and in the age range 50–79 y when enrolled at 40 US Clinical Centers.

In the context of dietary approaches to prevent chronic diseases among postmenopausal women, no single clinical trial could hope to test the many dietary changes that may merit consideration. Furthermore, cost and logistical challenges imply that only a few full-scale dietary intervention trials with clinical disease outcomes are practical at any point in time. Hence, there is strong incentive to make observational studies as reliable as possible for the development of diet and health information. Consequently, over the past 15 y, the WHI research group has been carrying out a substantial effort to apply, and more recently to develop, nutritional intake biomarkers. We summarize results of these efforts, along with DM trial results, here.

Results from the WHI Dietary Intervention Trial

Dietary change

The intervention in the DM trial emphasized total fat reduction. Eligibility required all enrollees to have a dietary percentage of energy from fat of 32% or larger on FFQ screening. In part, the fat reduction goals were to be achieved by increasing servings per day of vegetables and fruit to ≥5 and of increasing servings per day of grains to ≥6. The DM trial was a component of a larger clinical trial program that also included menopausal hormone therapy trials of estrogens alone among women who were posthysterectomy and of combined estrogens plus progestin among women with uteruses, as well as a trial of calcium and vitamin supplementation, in a partial factorial design.

Over a median 8.5-y intervention period, fat reduction in the DM intervention group occurred largely through replacement of fat by carbohydrate. Changes in total energy intake or body weight were not intervention goals. FFQ estimates of average daily intakes in the intervention and comparison groups at 1-y postintervention (5–8) were, respectively, 24.1% and 35.1% for percentage of energy from total fat, 58.4% and 47.9% for percentage of energy from carbohydrate, 17.7% and 16.8% for percentage of energy from protein, 5.1% and 3.9% for servings/d of vegetables and fruit, and 5.4% and 4.7% for servings/d of grains. All differences between randomly assigned groups were highly significant (P < 0.001). Fractional reductions similar to those for total fats were reported for saturated, trans, polyunsaturated, and monounsaturated fats. Fractional increases similar to those for total carbohydrates were observed for sugars and other carbohydrates. Fractional increases similar to those for combined vegetables and fruit were observed for vegetables and fruit individually; and similarly for total grains were observed for whole and other grains. Fiber and carotenoid consumptions were comparatively higher in the intervention group, as was also the case for consumption of calcium, dairy products, and fish, whereas red meat and nuts consumptions were somewhat lower among intervention group participants (all P < 0.001). Total FFQ daily energy intake averaged 1520 kcal and 1612 kcal in the intervention and comparison group (8). Total energy assessment is a recognized weak aspect of FFQ as well as other dietary self-report approaches, and a sustained energy difference of this magnitude does not seem to be consistent with participant body weights that were lower in the intervention group by only ∼2.2 kg at 1 y after being randomly assigned, and much of this difference had dissipated by the end (31 March 2005) of the trial intervention period (9). The dietary difference between intervention and comparison groups tended to be modestly reduced later in the intervention period. For example, at 6 y following random assignment the FFQ intervention and comparison groups daily averages were 28.6% and 37.0% for percentage of energy from fat, 54.1% and 45.8% for percentage of energy from carbohydrate, 17.7% and 17.2% for percentage of energy from protein, 5.1% and 3.8% for servings of vegetables and fruit, and 4.5% and 4.2% for servings of grains. These and the other dietary intake features described above were all evident (P < 0.001) at 6 y following random assignment, with the exceptions of dairy products and fish for which differences between randomly assigned groups at this later time were not significant (P > 0.1). FFQ total energy intake averaged 1460.5 and 1571.0 kcal/d in intervention and comparison groups at 6 y following random assignment (P < 0.001).

The DM trial was statistically powered (5, 10) based on the difference in percentage of energy from fat between randomly assigned groups over a planned 9-y trial intervention period. About 70% of the planned difference in percentage of energy from fat was achieved, according to FFQ data, leading to nontrivial study power reductions under the remaining design assumptions (e.g., power for breast cancer reduced from 80% to about 63%).

Clinical outcome results

Distributions of the characteristics of DM trial participants at recruitment during 1993–1998 have been published (5–8), with no major differences between randomly assigned groups. About 37%, 47%, and 16% of participants were aged 50–59, 60–69, and 70–79 y, respectively. About 11% were self-reported African American, 4% Hispanic, 0.5% American Indian, 2% Asian/Pacific Islander, and 81% white race/ethnicity; 78% had a high school diploma or higher education; and 44% were current and 14% were former menopausal hormone users. Average BMI (kg/m²) was 28.2; average blood pressure was 128/76 mm Hg; 47% of participants were hypertensive, 7% were current smokers, 43% were posthysterectomy, 21% had bilateral oophorectomy, 4.5% had diabetes, and 11.6% had elevated cholesterol requiring medication, with ∼6% taking statins. All participants were free of prior breast cancer and had recent negative mammograms, and ∼18% had a family history of breast cancer. All participants were also free from a personal history of colorectal cancer. Most DM trial participants (about 83.5%) were not enrolled in the WHI hormone therapy trials. About half subsequently enrolled in the WHI calcium plus vitamin D trial, mostly at 1 y following the DM trial random assignment. Reference (8) contains a participant flow diagram for the DM trial.

As with previous analyses (8), HRs updated here contrast the intervention and comparison groups using Cox regression with baseline rates stratified according to age group, race/ethnicity, hysterectomy status, prior history of the outcome under analysis (if applicable), randomly assigned status in the WHI hormone therapy trials (estrogen alone, estrogen-alone placebo, estrogen plus progestin, estrogen plus progestin placebo, not randomly assigned), and study phase (intervention and postintervention, time dependent). For a specified clinical outcome the time to response is days from random assignment to first relevant clinical event, whereas times for noncases were censored at the earliest of end of the study phase under analysis, loss to follow-up, or death.

Figure 1 presents results for the major clinical outcomes in the DM trial. The outcome categories are the same as in the 2019 DM trial summary report (8) except for death from breast cancer, which was subsequently reported (11). The analyses in Figure 1 include outcomes through 30 September 2018, the most recent date for which outcome data, including that from National Death Index matching, are complete enough for reliable intervention group comparisons. This represents an additional 24 mo of follow-up, and an increase of 24% (from 13,498 to 16,698) in deaths, compared with our previous reports (8, 11).

FIGURE 1.

Monitored and other important outcomes in the WHI Dietary Modification Trial (n = 48,835) during its 8.5-y (median) intervention period (A), and over cumulative follow-up (B) of 13.4 y (median) for CVD outcomes and ∼20 y (median) for other outcomes. Summary statistics are shown for randomly assigned groups. HRs, 95% CIs, and significance levels (P values) derive from Cox regression models with baseline hazard stratified on age at randomization (50–54, 55–59, 60–69, 70–79), self-reported ethnicity (white, black, other), hysterectomy status (yes, no), prior disease (if applicable), random assignment status in the WHI hormone therapy trials, and study phase (intervention, extension phase 1, extension phase 2; time-dependent). Time to event is time from random assignment. P values are from a score (log-rank) test. Analyses for diabetes outcomes were restricted to participants without prevalent diabetes at baseline (n = 45,595). Note that results for CHD have been shown (14) to be confounded by post–random assignment use of statins. As elaborated in the narrative, CHD incidence is lower in the intervention than the comparison group in the subset of baseline healthy, normotensive women where evidence of confounding was lacking. CHD, coronary heart disease; CVD, cardiovascular disease; WHI, Women's Health Initiative.

The primary (invasive) breast cancer outcome at the end of the trial intervention period (Figure 1) has an HR of 0.92 (95% CI: 0.84, 1.01), with a corresponding P value of 0.09, based on follow-up that included 671 and 1093 incident cases in the intervention and comparison groups. This HR reduction was about 70% of that projected in the trial design, consistent with the lesser differential between randomly assigned groups in percentage of energy from fat compared with design assumptions. Several further analyses are relevant to the interpretation of the breast cancer data: intervention participants having a higher compared with a lower dietary percentage of energy from fat at baseline, based on 4-d food records, made larger reductions in fat intake and provided nominally significant evidence of breast cancer risk reduction, and the breast cancer HR varied significantly with baseline percentage of energy from fat (5). Also, there was significant HR variation with breast tumor estrogen receptor (ER)/progesterone receptor (PR) status, with evidence of risk reduction for ER+/PR− tumors in the intervention group (5, 8). Serum estradiol concentrations decreased among intervention but not among comparison group participants (5). Finally, there was a nominally significant reduction in the composite outcome of breast cancer followed by all-cause mortality (12). In comparison, there was little evidence for any intervention influence on the coprimary colorectal cancer incidence (P = 0.45), and little evidence overall of influence on the secondary coronary heart disease outcome (P = 0.61), defined as myocardial infarction or coronary death. A global index—defined as time to the earlier of breast cancer, colorectal cancer, CHD, or death from any cause—used for trial monitoring was also not significantly different between randomly assigned groups, and all-cause mortality did not differ significantly between the groups. Figure 1 shows randomly assigned group comparisons for several other important outcomes, with differences in favor of the intervention for breast cancer followed by death as previously noted, and for diabetes requiring insulin (13). Much of the evidence in favor of intervention benefit for these outcomes arose from the subset (n = 18,567) of participants who were obese at enrollment, though HR interactions with obesity were not significant (8).

It is noteworthy that the trial took place during a time of rapid increase in the use of statins. Periodic medication inventories showed evidence of post random assignment confounding of CHD results by statin use, both among the small fraction of participants having prior cardiovascular disease (CVD) and among participants without prior CVD who were hypertensive at baseline. In comparison, among normotensive participants without prior CVD, where there was no evidence of post random assignment confounding, the CHD HR (95% CI) was 0.70 (0.56, 0.87) during the intervention period (14). Also, LDL cholesterol was reduced at 1-y post random assignment among baseline healthy participants as expected, but actually increased among participants with prior CVD (interaction P = 0.01). An analysis in the subset of baseline healthy, normotensive participants (15) suggested that CHD risk was especially reduced among intervention group women who made relatively large increases in percentage of energy from protein.

Importantly, Figure 1 also shows clinical outcome results over a cumulative follow-up period as long as 21 y (median) for many outcomes. Even though the median intervention period was only 8.5 y, nominally significant reductions in death attributed to breast cancer, in breast cancer followed by death from any cause, and in diabetes requiring insulin (but not total diabetes) are evident over this much longer follow-up period. Furthermore, though not shown in Figure 1, the reduction in CHD incidence among normotensive, healthy participants remains significant (8) over longer-term cumulative follow up.

Interactions of intervention compared with comparison group HRs with other WHI interventions or with participant characteristics were generally not evident, with the previously mentioned breast cancer HR dependence on baseline diet percentage of energy from fat, and the CHD dependence on prior CVD and baseline hypertension, as the major exceptions.

DM trial summary

The totality of DM trial findings, especially over cumulative follow up, is favorable in terms of chronic disease risk reduction among postmenopausal women for the low fat dietary pattern intervention. Favorable influences on breast cancer—related outcomes, on diabetes-related outcomes in spite of a substantial increase in carbohydrate intake, and on CHD in spite of an intervention that focused on total fat reduction rather than on fat components (e.g., saturated fat, trans-fatty acid)—that may be particularly relevant to CVD risk. Furthermore, no adverse intervention effects on disease outcomes were identified. Note also that dietary changes made by DM trial participants led to significantly higher (i.e., better) index scores in the intervention compared with the comparison group for each of the Healthy Eating Index-2005, the Dietary Approaches to Stop Hypertension (DASH), the Alternative Healthy Eating Index-2010, and Alternative Mediterranean Diet (15). The difference was particularly striking for DASH index scores (15).

In summary, the low fat dietary pattern intervention studied in WHI, though not designed as an optimal diet for improving health in general, evidently had overall health benefits that exceeded risks in this study population. Moreover, this intervention involved moderate dietary changes that are likely to be achievable by many.

Results from WHI Studies of Dietary Biomarkers and Chronic Disease Risk

WHI biomarker studies

As already noted, the cost and logistics of full scale dietary intervention trials with chronic disease outcomes are such that few can be afforded—implying a substantial reliance on observational nutritional epidemiologic studies or on smaller scale intervention trials with intermediate outcomes—for human data on diet and chronic disease risk. However, observational studies of diet and chronic disease are themselves challenging because of the potential for confounding inherent to observational studies generally and, importantly, because of the difficulty in reliably assessing the diet of individuals, either over a short term or over decades that may be pertinent to chronic disease risk determination. Objective biomarkers of dietary intake could allow a substantial strengthening of evidence compared with observational studies that rely only on self-reported diet, but there are established intake biomarkers for only a few components of diet, including total energy, protein, sodium, and potassium. In part for this reason the WHI conducted a Nutrition Biomarker Study (2004–2006) among 544 DM trial participants (50% in intervention group) at 12 WHI Clinical Centers (16), that included a doubly labeled water (DLW) assessment of total energy consumption (17) and a urinary nitrogen (UN) assessment of total protein consumption (18). Intervention group participants underestimated total energy consumption using FFQs by about 100 kcal/d more than did the comparison group, giving an explanation for the apparent discrepancy between FFQ energy intake and weight change during trial cohort follow up. To date we have been unable to identify a biomarker meeting our criteria (described below) for total fat consumption, or for percentage of energy from fat, precluding an objective assessment of adherence to the hypothesized percentage of energy from fat difference between intervention and comparison groups in the DM trial.

The Nutrition Biomarker Study led to observational disease association studies using “biomarker-calibrated” intake using a regression calibration approach (19–21), as did a separate Nutrition and Physical Activity Assessment Study (22) among 450 OS participants at 9 WHI Clinical Centers (2007–2009). This study also included DLW and UN assessments, as well as indirect calorimetry for resting energy expenditure assessment, in conjunction with FFQs, 4-d food records, and 3 24-h dietary recalls. As elaborated below, this latter research project was continued (2010–2014) through the conduct of a feeding study among 153 WHI participants in the Seattle area for the express purpose of biomarker development (23). The source of data for novel biomarkers for nutrients, foods, or dietary patterns was expanded (2015–present) to include serum and urine metabolomics profiles. The use of both established and novel biomarkers for disease association analyses in WHI cohorts continues today either through the direct assessment of biomarkers in stored specimens or through biomarker calibration of dietary self-reports.

Biomarker calibrated disease association analyses

The DLW energy assessment is not obtainable from stored specimens. It requires an active and somewhat expensive protocol over a 2-wk period. It is therefore not practical to make this crucial assessment for all members of large epidemiologic cohorts. However, it may be possible to explain much of the variation in energy intake in a study population by linear regression of DLW biomarker values from a subset of a study cohort on self-reported dietary assessment, along with other participant characteristics that are correlated with energy intake or that are related to biases in self-reported energy. This turned out to be the case in both WHI biomarker studies, whether using FFQ, 4-d food records, or 24-h dietary recalls. After allowing for temporal biomarker variation using reliability subsamples, the majority of variation in (log-transformed) DLW-derived energy intake could be explained by (log-transformed) self-reported intake along with BMI, age, and race/ethnicity, with BMI contributing much to the regression R² (22). Furthermore, overweight and obese participants were found to underestimate energy intake by 30–40% (16, 22), whereas normal weight participants did not. Similar systematic biases have been reported for multiple other populations [e.g., in the Observing Protein and Energy Nutrition (OPEN) Study (24)]. Measurement error in self-reported protein intake based on comparison with UN measured protein intake also related positively to BMI, but less so than for energy. Percentage of energy from protein was somewhat overestimated using these self-report tools (16, 22).

The biomarker-based equations just mentioned generate estimates of energy and protein intake from self-reported dietary data and personal characteristics throughout the larger WHI cohorts. The linear regression development of these “calibration” equations make it plausible to regard the biomarker log intake as calibrated log intake plus measurement error that is independent of the self-reported dietary intake and of the set of participant characteristics considered. Then, regarding the actual log intake as biomarker log intake plus random error that is independent of self-reported intake measurement error and these same variables, and making certain normality assumptions, allows one to insert the calibrated log intake values into Cox regression (25) to estimate the relation between actual intake and disease HR. Note that variables included in disease risk analyses to control confounding need also to be considered for inclusion in calibration equation development to avoid HR biases, and that specialized procedures for variance estimation in the Cox regression procedure are needed to acknowledge statistical variation in the calibration equation coefficients (20,21).

This regression calibration approach was applied to several clinical outcomes in the DM-C and the OS with log-HR assumed to be linearly related to calibrated log intake values. Strong positive associations arose for total energy in relation to CVD, cancer, and diabetes using DLW-calibrated intake (26–28). Furthermore, an objective measure of activity-related energy expenditure (AREE) was defined as the difference between DLW energy and resting energy expenditure from indirect calorimetry, and these values were used in the NPAAS cohort to develop a calibration equation for log-AREE in terms of self-reported leisure physical activity and personal characteristics (29). As shown in Table 1, configured from (30), energy consumption is strongly positively related to the risk of the outcomes just mentioned, whereas AREE is inversely related to these outcomes in joint analyses in the OS cohort. Note from the right side of Table 1 that these associations are either severely attenuated or are completely absent in corresponding analyses using self-reported (FFQ) energy and leisure physical activity assessments without biomarker calibration.

TABLE 1.

Estimated HR for 20% increments in total energy and in AREE with and without biomarker calibration to correct for measurement error in self-report assessments, in the WHI Observational Study from baseline (1994–1998) to 30 September 2010, adapted from Zheng et al. (30)¹

	With biomarker calibration				No biomarker calibration
	Energy		AREE		Energy		AREE
Outcome category (cases, n)	HR2	95% CI2	HR	95% CI	HR	95% CI	HR	95% CI
CHD (1660)	1.57	1.19, 2.06	0.78	0.65, 0.95	1.00	0.98, 1.02	0.99	0.97, 1.01
Stroke (1462)	1.36	1.05, 1.76	0.83	0.69, 0.99	0.97	0.95, 1.00	0.99	0.98, 1.01
Heart Failure (780)	3.51	2.12, 5.82	0.57	0.41, 0.79	1.04	1.01, 1.08	0.97	0.95, 1.00
Total CVD3 (4212)	1.49	1.23, 1.81	0.83	0.73, 0.93	1.00	0.99, 1.01	1.00	0.99, 1.01
Breast cancer (3798)	1.47	1.18, 1.84	0.82	0.71, 0.96	1.01	0.99, 1.02	1.00	0.99, 1.01
Colon cancer (677)	1.86	1.18, 2.93	0.83	0.66, 1.04	1.00	0.96, 1.03	1.00	0.97, 1.03
Rectum cancer (103)	2.75	1.10, 6.83	0.51	0.27, 0.99	1.01	0.92, 1.10	0.99	0.93, 1.05
Total invasive cancer (9227)	1.43	1.17, 1.73	0.84	0.73, 0.96	1.01	1.00, 1.02	0.99	0.99, 1.00
Diabetes (6494)	4.17	2.68, 6.49	0.60	0.44, 0.83	1.06	1.04, 1.07	1.01	1.00, 1.02

¹Cohort sizes varied from 56,390 for cardiovascular diseases, to 64,055–68,712 for cancers, to 72,724 for diabetes, depending on extent of missing data for on variables used for calibration equations and for disease risk modeling. AREE, activity-related energy expenditure; CHD, coronary heart disease; CVD, cardiovascular disease; WHI, Women's Health Initiative.

²HR estimates and 95% CIs are based on Cox models with baseline hazard rates stratified on age at enrollment in 5-y categories, and with adjustment for a disease-specific set of potential confounding factors.

³Total CVD comprises CHD + coronary artery bypass graft/percutaneous coronary intervention + total stroke.

From the striking Table 1 analyses one can estimate, for example, that a 10% reduction in energy intake in conjunction with a 10% increase in AREE is associated with HRs of 0.75, 0.77, and 0.37 for total CVD (defined as CHD plus stroke plus coronary bypass graft/percutaneous coronary intervention), total invasive cancer (exclusive of nonmelanoma skin cancer), and type 2 diabetes, respectively. A caveat with these analyses is that these HRs become much closer to the null when BMI is included as a confounder in the disease risk model. We assume that this happens because BMI is a strong mediator of associations of energy intake and AREE with disease risk. However, some residual confounding by BMI is also a possibility and available data do not allow one to distinguish between these 2 possible roles for BMI. Regardless, the magnitudes of the Table 1 associations are such that further information on the dependence of chronic disease risk on energy consumption and activity patterns deserves a very high priority in future nutrition and activity epidemiology research.

Similarly, studies of UN-calibrated protein intake in relation to various clinical outcomes in WHI cohorts show substantial favorable associations—for example with frailty (31) and physical function (32)—that are much attenuated if analyses are instead based on FFQ protein intake without biomarker calibration.

Results from WHI Nutrition Biomarker Development Studies

The nutrition and physical activity assessment feeding study

As already noted, we are hindered in the application of a biomarker approach to objective intake assessment in nutritional epidemiologic contexts by the few nutrients, foods, or dietary patterns for which there is an established intake biomarker. The established biomarkers mentioned above were developed in human feeding studies, either in metabolic wards or among free-living persons. Development typically involves feeding a small number of specialized diets and by recovering provided nutrients or their metabolites in urine. Rather than feeding certain prescribed diets, the WHI feeding study performed in 153 WHI participants in the Seattle area (23) involved providing foods and beverages that were intended to approximate the participant's usual diet, over a 2-wk feeding period. This design was chosen so that intake variations in the study population would be retained, body weight would be maintained, and disruptions to relevant blood and urine measures would be minimized over the 2-wk feeding period. The individualized diet specifications started with 4-d food records, then adjusted for certain recognized biases (e.g., in total energy) as well as for other pertinent factors based on nutritionist interviews with participants. Foods having well-characterized nutrient composition were used to the extent practical. We developed potential biomarkers by linearly regressing (log-transformed) provided intake on corresponding (log-transformed) serum and 24-h urine measures based on specimens collected near the end of the feeding period.

In an initial report from this feeding study (23), biomarkers were presented for several micronutrients and for certain other nutritional variables based on serum concentrations and personal characteristics. Log-transformed DLW energy and UN protein “explained” about 50% and 40% of the variation in log-daily averaged intake of energy and protein, respectively, over the 2-wk period. These established biomarkers provide benchmarks, leading us to choose a regression R² of 36% or larger—along with more general sensitivity and specificity considerations—as criteria for an acceptable intake biomarker in this feeding study context. Though admittedly somewhat arbitrary, this R²>36% (R >0.6) criterion is intended to convey a quantitative notion of “nearly as strongly correlated” with feeding study intake as are the established intake biomarkers.

To enhance the potential for development of a broad range of nutritional biomarkers we added serum and urine metabolomic profiles to the feeding study database in recent years. These relatively high-dimensional data show value for the development of additional macronutrient biomarkers. For example, we have recently proposed biomarkers for protein density (percentage of energy from protein) and for carbohydrate density, along with a biomarker for (absolute) carbohydrate, and a biomarker for (absolute) protein that enhances the UN biomarker. Each of these biomarkers is strongly dependent on the metabolomic measures.

Novel biomarker application

Some of the micronutrients for which biomarkers were proposed (23) involved serum micronutrient concentrations that were measured routinely at baseline in a WHI Measurement Precision Study of 5488 participants. Specifically, intake biomarkers for α and β carotene, lutein plus zeaxanthin, and α tocopherol were determined from baseline serum concentrations and corresponding personal characteristics. Though associations of these biomarkers with chronic disease in this subcohort were somewhat modest, lower risks for CVD outcomes were detected with higher intake of each of the carotenoids, in contrast to a somewhat higher risk with α tocopherol. Certain cancer outcomes and diabetes were also lower risk at higher carotenoid intake, but not at higher α tocopherol intake (33).

Table 2 shows some results, abstracted from (34), from applying our novel biomarkers for protein density and carbohydrate density. These biomarkers, which depended primarily on blood and urine metabolomic profiles, were estimated as meeting a cross-validated 36% R² criterion for explaining feeding study provided intake variations, after allowing for temporal biomarker variation. Log-transformed biomarkers in the Nutrition and Physical activity Assessment Study were regressed linearly on corresponding log-transformed FFQ intake values and personal characteristics to produce calibration equations for these macronutrient variables. We used these equations to calculate calibrated intake values in the OS and the DM-C trial cohorts and associated these with CVD, cancer, and diabetes in the combined cohorts. Table 2 shows HR estimates and 95% CIs for a 20% increment in protein density and carbohydrate density for some major outcomes, in analyses that also included DLW-calibrated total energy intake. These analyses suggest lower risks of cancers, CVDs, and diabetes among participants having diets relatively high in carbohydrate density, whereas corresponding associations using only self-reported dietary information are much attenuated toward the null and do not suggest comparable public health importance. These analyses suggest an important role for dietary composition in determining chronic disease risk in addition to the previously mentioned likely important role for total energy consumption. As described above, the DM intervention increased carbohydrate density by ∼20%, along with many other dietary changes, though changes in protein density and total energy were evidently small. One can see that intervention influences on clinical outcomes in Figure 1 include similarities, as well as some differences, with calibrated carbohydrate density HRs in Table 2. It is also worth commenting that the Table 2 HRs were essentially unchanged when BMI was added as a potential confounding factor to the disease risk model.

TABLE 2.

Disease HRs and 95% CIs for a 20% increment in macronutrient variables, with and without biomarker calibration of food frequency questionnaire assessments, in analyses that also included total energy intake, in WHI cohorts (n = 81,894) of postmenopausal US women (1993–2020)¹

	With biomarker calibration				Without biomarker calibration
	Protein density		Carbohydrate density		Protein density		Carbohydrate density
Disease category2 (cases, n)	HR3	95% CI3	HR	95% CI	HR	95% CI	HR	95% CI
Breast cancer (5139)	0.96	0.88, 1.05	0.83	0.74, 0.93	0.99	0.97, 1.02	0.95	0.92, 0.97
Colon cancer (1060)	0.84	0.70, 1.00	0.84	0.69, 1.02	0.95	0.90, 1.01	0.95	0.91, 1.00
Rectum cancer (158)	0.60	0.32, 0.97	1.35	0.67, 2.71	0.85	0.73, 0.99	1.08	0.91, 1.28
Total invasive cancer (2804)	0.94	0.89,0.99	0.87	0.81, 0.93	0.99	0.97, 1.00	0.96	0.94, 0.98
CHD (2869)	1.01	0.90, 1.14	0.81	0.69, 0.95	1.01	0.97, 1.04	0.95	0.92, 0.98
Stroke (2425)	0.96	0.84, 1.08	0.83	0.71, 0.97	0.99	0.95, 1.02	0.96	0.93, 0.99
Total CVD4 (6964)	1.00	0.93, 1.08	0.87	0.78, 0.97	1.00	0.98, 1.02	0.96	0.92, 0.98
Diabetes (12,145)	1.17	1.09, 1.25	0.73	0.66, 0.80	1.05	1.03, 1.07	0.90	0.89, 0.92

¹CHD, coronary heart disease; CVD, cardiovascular disease; DM-C, dietary modification comparison group; OS, observational study; WHI, Women's Health Initiative.

²Cohort sizes for specific analyses reduced from n = 81,894 vary according to the number of missing values in variables included in calibration equations or in disease risk models.

³HR estimates and 95% CIs are based on Cox models with baseline hazard rates stratified on study component (DM-C or OS), hormone therapy trial status (estrogen plus progestin, estrogen plus progestin placebo, estrogen alone, estrogen alone placebo, not randomized), age at enrollment (50–54, 55–59, 60–64, 65–69, 70–74, ≥75 y) and self-reported race/ethnicity, and with adjustment for a disease-specific set of potential confounding factors.

⁴Total CVD comprised CHD + coronary artery bypass graft/percutaneous coronary intervention + total stroke.

Discussion

The WHI, since its inception nearly 30 y ago, has included a major emphasis on diet and nutrition in the largest cohort of postmenopausal US women ever assembled to evaluate related chronic disease health benefits and risks. We have presented reports from this research in a wide variety of venues, including medical, nutritional, and epidemiologic journals, and conferences and symposia worldwide. Here we briefly reviewed results from these studies—including recent results from long-term follow up in the massive DM trial of a low-fat dietary pattern—and from an active program in the use and development of dietary intake biomarkers for strengthening observational diet and disease association studies. Our review indicates that certain breast cancer, diabetes, and CHD benefits result from the low-fat dietary pattern intervention implemented in the DM trial. The origin of the DM dietary intervention, with its emphasis on total fat reduction, focused primarily on cancer prevention. Interventions tailored to contemporary hypotheses concerning diet quality and nutrient density may be able to achieve these and additional health benefits. Importantly, the DM trial provides randomly assigned trial support for the potential of dietary composition changes to reduce chronic disease risk among postmenopausal US women. Collectively, WHI nutrition research efforts reinforce the importance of diet in determining chronic disease risk, with findings that can inform future research methods and objectives for diet chronic disease prevention.

In terms of methods, the DM trial shows that a complex dietary and behavioral intervention can be implemented on the scale needed for useful chronic disease outcome comparisons. Intervention activities, and trial participation in general, were favorably regarded by participants, and recruitment into this clinical trial component was relatively easy compared with some other WHI components. Adherence to intervention goals was a challenge, with only ∼70% of the hypothesized difference between intervention and comparison groups in percentage of energy from fat achieved, according to FFQ data. The dietary difference achieved appeared to be substantially maintained over the 8.5-y intervention period. The study design called for intervention compared with comparison group HR reductions to be linearly achieved over a 3-y period for cardiovascular incidence outcomes, and over a 10-y period for cancer and mortality outcomes. In comparison, some observed HR reductions did not reach nominal significance until >10 y of intervention and follow up.

We think there is a potentially important role for randomized dietary pattern intervention trials with chronic disease outcomes in the future nutritional epidemiology research agenda. Our understanding of the health benefits and risks of a low-fat dietary pattern intervention would be quite different without this trial. For example, the purported demonstration of a lack of influence of a low-fat diet on breast cancer is sometimes described as one of the singular achievements of observational nutritional epidemiology over recent decades. However, this finding is called into question by the nominally significant reductions in death following breast cancer—and even in death attributed to breast cancer—in the DM trial intervention group. Similarly, there has been considerable encouragement toward low-carbohydrate, high-fat diets in the last couple of decades for weight loss and health improvement. However, the relatively low-fat, high-carbohydrate diet intervention implemented in the DM trial led to intermediate-term weight loss rather than gain, and produced evidence of benefit for the more severe form of diabetes requiring insulin. These issues will each require further research, perhaps using observational studies having pertinent intake biomarkers. In the meanwhile, as with the use of menopausal hormone therapy, we think current understanding, and the identification of priority outstanding research questions are much enhanced by the WHI studies. Of course, future randomized dietary intervention trials need to be reserved for interventions having substantial public health potential. Accordingly, there is need for a close integration of intervention trial planning with basic nutritional biology, high-quality observational studies, and small-scale intervention trials with intermediate outcomes, leading to occasional full-scale intervention trials when preparatory work is strong and public health implications are substantial.

Relative to the substance of the future nutritional epidemiology research agenda, one can distinguish topics related to total consumption from topics related to dietary composition. For the former, WHI biomarker studies indicate that total energy consumption could be strongly related to the risk for many important chronic diseases if body fat accumulation over the lifespan is primarily driven by overconsumption in conjunction with a sedentary lifestyle. On the other hand, if a high-energy diet is primarily the response to body fat accumulation, then total energy consumption would be less pertinent to chronic disease risk. Additional research on this crucial topic is needed because available self-reporting tools are quite inadequate for energy intake estimation, leaving a major hole in mainstream nutritional epidemiology research, and to a very limited emphasis on overall dietary consumption in international reviews and dietary recommendations, which tend to focus almost exclusively on dietary composition. Further insight into whether total energy intake is a driver or passenger on the road to chronic disease occurrence may be obtainable through longitudinal studies of DLW-calibrated intake and body mass over portions of the lifespan and—in conjunction with a practical intervention for making and maintaining a meaningful energy intake reduction—may ultimately merit an energy reduction intervention trial.

Concerning dietary composition, the further development and use of pertinent intake biomarkers could accelerate interesting recent work on macronutrient composition, and macronutrient substitution, in relation to chronic disease incidence and mortality (e.g., 35–38). For example, biomarkers for animal and plant protein, and for carbohydrate according to quality could allow our calibrated intake Table 2 analyses to be extended and more fully understood. The development and application of dietary biomarkers for nutrients, foods, and eating patterns is an important future research enterprise that will benefit from additional human feeding studies of varying designs. Substantial NIH support for such studies is anticipated following the recent strategic planning for nutrition research during 2020–2030 (39). The WHI is contributing to this enterprise through its novel feeding study design and through the application of established and novel intake biomarkers to chronic disease risk association analyses in WHI cohorts.

The future nutritional epidemiology research agenda can continue to benefit from the availability of both self-report and biomarker measures of intake in disease association studies. This agenda could also include studies that combine small-scale intervention trials having high-dimensional intermediate outcomes (e.g., metabolomic profiles in blood and urine), with large cohort studies having the same set of intermediate measures in conjunction with clinical disease outcome data. Collectively, these opportunities bode well for an exciting research enterprise and for important new insights concerning dietary modification and health improvement during the upcoming few years.

Notes

Supported by the National Heart, Lung, and Blood Institute, National Institutes of Health, US Department of Health and Human Services (contracts HHSN268201100046C, HHSN268201600001C, HHSN268201600002C, HHSN268201600003C, HHSN268201600004C, and HHSN271201600004C); and National Cancer Institute grants R01 CA119171 and P30 CA15704, and NIH instrumentation grant S10 OD021562.

Abbreviations used: AREE, activity-related energy expenditure; CAD, coronary artery disease; CVD, cardiovascular disease; DLW, doubly labeled water; DM, dietary modification; DM-C, dietary modification comparison group; ER, estrogen receptor; NCI, National Cancer Institute; OS, observational study; PR, progesterone receptor; UN, urinary nitrogen; WHI, Women's Health Initiative.

Data Availability

Data, codebook, and analytic code used in this report may be accessed in a collaborative mode as described on the Women's Health Initiative website (www.whi.org).