Kant AK, Schatzkin A, Graubard BI, Schairer C. A Prospective Study of Diet Quality and Mortality in Women. JAMA. 2000;283(16):2109-2115. doi:10.1001/jama.283.16.2109
Author Affiliations: Department of Family, Nutrition, and Exercise Sciences, Queens College of the City University of New York, Flushing, NY (Dr Kant); and the Nutritional Epidemiology Branch (Dr Schatzkin), the Biostatistics Branch (Dr Graubard), and the Environmental Epidemiology Branch (Dr Schairer), Division of Cancer Epidemiology and Genetics, National Cancer Institute, Bethesda, Md.
Context Most studies of diet and health care have focused on the role of single
nutrients, foods, or food groups in disease prevention or promotion. Few studies
have addressed the health effects of dietary patterns, which include complex
mixtures of foods containing multiple nutrients and nonnutrients.
Objective To examine the association of mortality with a multifactorial diet quality
Design and Setting Data from phase 2 (1987-1989) of a prospective cohort study of breast
cancer screening, the Breast Cancer Detection Demonstration Project, with
a median follow-up of 5.6 years.
Participants A total of 42,254 women (mean age, 61.1 years) who completed the food
frequency questionnaire portion of the survey.
Main Outcome Measure All-cause mortality by quartile of Recommended Food Score (RFS; the
sum of the number of foods recommended by current dietary guidelines [fruits,
vegetables, whole grains, low-fat dairy, and lean meats and poultry] that
were reported on the questionnaire to be consumed at least once a week, for
a maximum score of 23).
Results There were 2065 deaths due to all causes in the cohort. The RFS was
inversely associated with all-cause mortality. Compared with those in the
lowest quartile, subjects in the upper quartiles of the RFS had relative risks
for all-cause mortality of 0.82 (95% confidence interval [CI], 0.73-0.92)
for quartile 2, 0.71 (95% CI, 0.62-0.81) for quartile 3, and 0.69 (95% CI,
0.61-0.78) for quartile 4 adjusted for education, ethnicity, age, body mass
index, smoking status, alcohol use, level of physical activity, menopausal
hormone use, and history of disease (χ21 for trend=35.64, P<.001 for trend).
Conclusions These data suggest that a dietary pattern characterized by consumption
of foods recommended in current dietary guidelines is associated with decreased
risk of mortality in women.
Although many studies have examined the role of single nutrients, foods,
or food groups in the etiology of disease,1- 3
relatively little research has addressed the health effects of dietary patterns
comprising multiple interdependent dietary factors.4
Research on dietary patterns is warranted on several grounds. First, complex
diets consumed by free-living individuals do not consist of single nutrients
or foods but rather a combination of foods containing multiple nutrients and
nonnutrients. Second, intercorrelation of dietary variables makes it difficult
to isolate effects of single nutrients or foods. Third, in vivo biological
activities of nutrients are interdependent.5- 7
Finally, recommendations for disease prevention implicitly reflect the dietary-pattern
approach by emphasizing the simultaneous change of several dietary behaviors,
such as increasing fruit, vegetable, and grain intake, and decreasing fat
This study examines prospectively in a large cohort of women the relationship
of all-cause and cause-specific mortality with a measure of overall diet quality
derived from current food-based dietary guidelines.
For this study, we used data from the Breast Cancer Detection and Demonstration
Project (BCDDP), sponsored by the National Cancer Institute and the American
Cancer Society. Between 1973 and 1979, the project screened 283,222 women
aged 35 through 74 years in 29 screening centers in 27 cities throughout the
In 1979, the National Cancer Institute began a follow-up study of a
subset of the BCDDP participants (phase 1, 1979-1986).8
The subset of 64,182 women included (1) all women with pathologically confirmed
incident breast cancers identified during the screening phase (n=4275), (2)
all women with biopsy-proven benign breast disease identified during the screening
phase (n=25,114), (3) all women who had an identified abnormality on 1 or
more of the screening examinations but did not have a biopsy (n=9628), and
(4) a sample of healthy women who had no abnormality or recommendation for
a biopsy during the screening phase and matched women in groups 1 and 2 for
several criteria (n=25,165). The age distribution and the education level
of the follow-up cohort were comparable with those of all BCDDP participants.
The data were collected using a baseline telephone interview and up to 6 annual
telephone interviews until 1986.
For phase 2 follow-up (1987-1989), a questionnaire was mailed to all
surviving members of the follow-up cohort; 51,694 women responded. A 62-item
food frequency questionnaire was used to collect dietary information during
this phase. A modification of the instrument was developed by Block and coworkers.9- 11 This food frequency
questionnaire has been validated among older women and includes queries about
frequency of consumption and the size of portions consumed over the past year.
Another questionnaire was mailed to all follow-up cohort members in 1993 through
1995 (phase 3). Other information collected at phases 2 and 3 included history
of exogenous hormone use, medical history, information on end points other
than breast cancer, tobacco and alcohol use, use of vitamins, physical activity,
and updated family and reproductive history.
During each follow-up phase, women who did not respond to the mailed
questionnaire were interviewed by telephone, if possible. Extensive efforts
were made to contact women not located at phase 3, including tracing them
through the National Death Index of the National Center for Health Statistics
through December 1993.
For the purpose of analyses reported herein, phase 2 (1987-1989) was
considered the baseline. Of the 51,694 women who returned mailed food frequency
questionnaires, 9437 (18.3%) were excluded because the responses were either
grossly incomplete (missing information on >10 questions) or deemed unreliable
based on previous validation studies.9,10
Three questionnaires completed by proxies were also excluded, leaving 42,254
women in the analytic file. In this cohort of 42,254 women, 2065 deaths due
to all causes occurred between phase 2 (1987-1989) and phase 3 (1993-1995).
This included 223 deaths (10.8%) for which a death certificate was not available
but death was confirmed by other sources. The cause-of-death information was
coded as listed on the death certificate. The characteristics (age, ethnicity,
body mass index [BMI], and level of education) of subjects in the analytic
cohort were comparable with those of the women who responded to the phase
Using the 62-item questionnaire from phase 2, we developed a Recommended
Foods Score (RFS) to measure overall diet quality. The RFS is based on reported
consumption of foods recommended by current dietary guidelines.1- 3
The RFS is similar to the dietary variety score for recommended foods that
we had developed for use with the National Health Interview Survey data.12 Briefly, because current dietary guidelines emphasize
consumption of fruits, vegetables, whole grains, lean meats or meat alternates,
and low-fat dairy, we decided that all questionnaire items corresponding to
these groups would contribute to the score. Furthermore, because of measurement
error associated with amounts reportedly consumed, we designed the diet quality
measure to be independent of reported amounts.13,14
We used the following 23 food frequency questionnaire items for the RFS :
apples or pears; oranges; cantaloupe; orange or grapefruit juice; grapefruit;
other fruit juices; dried beans; tomatoes; broccoli; spinach; mustard, turnip,
or collard greens; carrots or mixed vegetables with carrots; green salad;
sweet potatoes, yams; other potatoes; baked or stewed chicken or turkey; baked
or broiled fish; dark breads like whole wheat, rye, or pumpernickel; cornbread,
tortillas, and grits; high-fiber cereals, such as bran, granola, or shredded
wheat; cooked cereals; 2% milk and beverages with 2% milk; and 1% or skim
milk. The RFS is calculated by the sum of the 23 items that subjects mentioned
they consumed at least once a week, for a maximum score of 23. The remaining
39 items on the food frequency questionnaire did not meet the criteria for
inclusion in the RFS.
The number of person-years contributed by a subject was calculated from
the date of phase 2 follow-up interview to the date of death (n=2065) or the
date last known alive (n=40,189), whichever came first. The date last known
alive was the date of the phase 3 interview for those who answered the questionnaire
(n=36,188), date of last telephone contact for nonrespondents to the phase
3 mailed questionnaire (n=1872), and date of last National Death Index search
(December 31, 1993) for nonrespondents to both mail and telephone contacts
at phase 3 (n=2129). We used Cox proportional hazards regression to examine
the independent association of the diet quality measure with mortality in
the presence of covariates with follow-up time as the underlying time metric.15 The analyses were done using the PROC PHREG procedure
in the SAS software package.16 We categorized
the RFS into approximate quartiles based on its distribution in the analytic
cohort. The risk of mortality in each of the upper 3 quartiles was compared
with the risk for the lowest RFS quartile. To evaluate the linear trend with
mortality, we entered RFS in regression models both as a continuous variable
and a scored variable ranging from 0 to 3. The trend results were essentially
unchanged when the RFS was scored as the median value in each quartile. The
proportional hazards assumption required by the Cox regression model was found
to be acceptable for the primary multivariate analysis involving quartiles
of RFS (χ22, 1.71; P=.42).
The covariates in the regression model were chosen a priori based on
potential correlates of health outcome and included the following baseline
variables: age; race; educational level; BMI; smoking status; alcohol intake;
energy intake; history of cancer, heart disease or diabetes; menopausal hormone
use status; and a physical activity measure (whether a participant engaged
in regular physical exercise long enough to work up a sweat at least once
a week). Results were similar when we examined other forms of covariates,
including a quantitative alcohol intake variable, smoking duration and number
of cigarettes smoked per day, as well as a quantitative physical activity
variable reflecting hours spent at different activity levels. Inclusion of
a weight change variable (ie, screening weight minus weight at phase 2) did
not affect the results shown.
To test for statistical interaction between the RFS and various covariates,
we entered into the regression models interaction terms reflecting the product
of RFS and each of the listed covariates.
The median follow-up time was 5.6 years. The mean (SE) of the RFS in
the analytic cohort was 11.4 (0.02). The mean age of the analytic cohort at
baseline (phase 2 interview) was 61.1 (range, 40-93) years. Table 1 presents the distribution of risk factors of mortality by
quartiles of RFS. More than 87% of the analytic cohort was white and had 12
or more years of education. Generally, subjects with higher RFS were slightly
older; more educated, physically active, likely to drink alcohol, and use
supplements regularly; less likely to be smoking currently.
Table 2 presents the mean
(SE) of intake of energy and selected nutrients by quartiles of RFS. Correlation
of RFS with intake of energy and selected micronutrients is also presented.
Generally, the RFS was positively associated with intake of energy and protein,
percentage of energy from carbohydrate, and micronutrient intake but inversely
associated with percentage of energy from fat.
Table 3 presents the age-adjusted
and multiple covariate–adjusted estimates of risk of all-cause mortality.
The multiple covariate–adjusted relative risk estimates associated with
the upper 3 quartiles of RFS in reference to the bottom quartile were 0.82
(95% confidence interval [CI], 0.73-0.92) for quartile 2, 0.71 (95% CI, 0.62-0.81)
for quartile 3, and 0.69 (95% CI, 0.61-0.78) for quartile 4, (χ21 for trend=35.64; P<.001). We
also ran these analyses with approximate decile cuts of RFS and observed a
similar trend (β, −.04 [SE, .008]; χ21
for decile trend=30.64; P<.001). The results shown
in Table 3 were unaffected by
exclusion of 223 deaths not confirmed by death certificates.
To test for a nonlinear trend, a 3-knot cubic regression spline,17 which involves 2 continuous variables, the linear
RFS variable and a nonlinear cubic expression of the RFS variable, was fit
to the data. A statistically significant nonlinear trend was noted (β
associated with the nonlinear variable was .00030 [.00012]; χ21=6.09; P=.01). Relative to analyses with only
a linear trend, the shape of the nonlinear trend showed a sharper decline
in risk of mortality for RFS values ranging from 0 to about 11 with a leveling
off for RFS values greater than 11. This observation is consistent with the
pattern of risk reduction associated with quartiles of RFS in multiple covariate–adjusted
regression models in Table 3.
The greatest decrease in risk of mortality was present going from the first
to the second quartile with leveling off between quartiles 3 and 4, for which
the RFS values are in the 12 and greater range.
To exclude the possibility that subjects with clinical disease may differ
in dietary patterns at baseline, we examined the RFS-mortality association
after exclusion of those reporting history of cancer, diabetes, or heart disease
at baseline (1193 deaths). Similarly, to exclude the possibility of those
reporting poor diets at baseline due to preclinical disease, we reexamined
the RFS-mortality association after excluding the first 2 and 3 years of follow-up.
The inverse RFS-mortality association persisted (P<.001)
after these exclusions (Table 4).
In our analyses of potential interaction between the RFS and several
covariates (education, smoking status, physical activity, alcohol intake,
BMI, menopausal hormone use status, and energy intake) in altering the RFS-mortality
association, none of the interaction terms was significant (data not shown).
Table 5 shows the relationship
between RFS and mortality from all-sites cancer, coronary heart disease, stroke,
and all other causes combined. An inverse association of RFS with mortality
from each cause was noted. For all-sites cancer, coronary heart disease, and
stroke mortality, respondents in the highest quartile of RFS had at least
30% lower risk than those in the bottom quartile.
Our study suggests that women reporting dietary patterns that included
fruits, vegetables, whole grains, low-fat dairy, and lean meats, as recommended
by current dietary guidelines, have a lower risk of mortality. Women in the
highest intake level of recommended foods had 30% lower risk of multivariate-adjusted
all-cause mortality compared with those in the lowest level. Our results provide
evidence in support of the prevailing food-based dietary guidelines and suggest
that diets complying with current dietary recommendations are indeed associated
with improved health outcome. The potential public health implications of
these findings are considerable; despite increased public awareness of the
importance of diet in decreasing the risk of chronic disease, large gaps remain
in food-based recommendations and actual dietary practices of the US population.18
Few studies have examined global measures of diet quality as it relates
to mortality. Nube et al19 reported a significant
positive association between 25-year survival and consuming a "prudent" diet,
based on consumption of 10 food items, in men but not in women. In the first
National Health and Nutrition Examination Survey (NHANES) Epidemiologic Follow-up
Study, we found diets characterized by a low diet diversity score based on
evaluation of whether each of the major food groups (fruit, vegetable, grain,
meat, and dairy) were reported to be associated with an increased risk of
all-cause mortality in both men and women.20,21
Women consuming 2 or fewer food groups daily compared with those who consume
5 had a 40% higher risk of mortality. Huijbregts et al22
have reported a 13% decrease in risk of mortality in men with healthy diet
patterns. McCullough et al23 recently noted
a weak association between health outcome and a complex index of diet quality
comprising both nutrient (mostly dietary fat related) and food group serving
recommendations in men. The measures of diet quality mentioned above,19- 23
however, are not directly comparable with the RFS, which assesses diet quality
relative to current food-based dietary recommendations. The only study reflecting
a comparable approach to diet quality is the clinical trial of the effect
of dietary patterns on blood pressure (Dietary Approaches to Stop Hypertension,
the DASH trial).24 In that study, a diet of
fruits, vegetables, low-fat dairy, whole grains, and lean meat and poultry
for 8 weeks reduced blood pressure in both hypertensive and normotensive subjects.
Conceptions of diet quality have evolved over time. Early in this century,
nutrition scientists focused on preventing nutrient deficiencies; diets that
provided the recommended intake levels of known essential nutrients and energy
were considered desirable.4 With increasing
recognition of the role of diet in prevention and promotion of chronic diseases,
dietary characteristics associated with decreased risk of chronic diseases
have been promoted.1- 3
Therefore, recent US dietary guidelines reflect current beliefs about how
nutrients, such as excess fat or foods such as fruits and vegetables, relate
to risk reduction.1- 3
The diet quality measure used in this study is based on this recent guideline.
The RFS is a relatively simple measure of the extent of healthful eating
and is portion-size independent. As is evident from Table 2, those with a high RFS had higher intake of energy and micronutrients
but a lower percentage fat energy than those with a low score. It is unlikely
that higher energy intake associated with RFS explains all the nutrient differences
noted among the RSF quartiles. For example, the mean energy intake in quartile
4 was 131% of the mean level in quartile 1; however, mean levels of dietary
fiber, vitamin C, folate, and provitamin A carotenoids in quartile 4 were
200%, 230%, 181%, and 253%, respectively, of mean levels in quartile 1. This
suggests qualitative differences in food selection in association with higher
RFS scores. Diets characterized by a low consumption of recommended foods
may have marginal intakes of several nutrients (or other biologically active
nonnutrient chemicals). Long-term marginal intakes of known essential nutrients
or poorly understood nonnutrients may not be compatible with favorable health
outcome. It is likely that the RFS-mortality association reflects a complex
interaction of multiple dietary constituents beyond the biological activity
of single nutrients.
The source of dietary information in our study was a single measure
of usual dietary intake derived from a 62-item food frequency questionnaire.
Although the food frequency questionnaire used in this study has been previously
all measurement errors inherent in this retrospective method of dietary assessment
are applicable to this instrument.13,14
The problems of dietary measurement error and underreporting of food intake
in dietary surveys have received considerable attention in recent years.13,25- 27 The
extent to which the general dietary measurement error problem affects an aggregate
dietary-pattern–based score like the RFS is unknown and merits further
research. In this study, the RFS is computed by counting selected questionnaire
items that are mentioned as having been consumed at least weekly independently
of portion size reported; therefore, the RFS is relatively unaffected by misreporting
of portion size. Use of the RFS may have allowed us to classify women, with
reasonable accuracy, into broad categories of low- or high-risk dietary behaviors.
Our analyses largely exclude the possibility that reverse causation
(due to women with preclinical disease at baseline reporting poor diets) accounts
for the results observed. Deaths occurring early in follow-up (first 2 and
3 years) were excluded without materially affecting the results observed.
Similarly, the trends observed remained significant after exclusion of women
who reported chronic conditions at baseline.
It would be premature to conclude that the observed inverse relationship
between RFS and mortality is causal. Given the observational epidemiologic
nature of our study, the possibility that RFS is a surrogate for some unknown,
poorly measured, or inadequately controlled determinant of mortality cannot
be ruled out. Smoking status, physical inactivity, alcohol use, vitamin supplement
use, and education level (a potential proxy for certain environmental exposures
or lifestyle characteristics) were all associated with RFS in this study.
Although we controlled for these and other factors as best as we could, we
cannot dismiss the possibility of residual confounding. Furthermore, given
that our cohort consists of participants in a screening study, it is possible
our results have limited generalizability. It would certainly be valuable
to see whether the RFS-mortality association holds in other large cohorts
of men and women.
Although the strategy of examining global measures of diet quality is
consistent with the complexity of diets consumed by free-living individuals,
one limitation of this approach is that it makes it difficult to elucidate
mechanisms through which the diet effect on a particular health outcome is
mediated. From a public health perspective, however, it is not essential to
wait for elucidation of every mechanism underlying health promoting activities
and interventions. The results of the cause-specific analyses confirm the
importance of dietary and nutritional factors for decreasing the risk of mortality
from leading causes of death (all sites cancer, coronary heart disease, and
stroke). The relatively weak association of RFS with all other causes of mortality
(Table 5) may reflect the nonspecific
nature of this category that includes causes unlikely to be related to diet.
The results from this large cohort of women with prospective follow-up
suggest that dietary patterns characterized by consumption of fruits, vegetables,
whole grains, low-fat dairy, and lean meat are associated with lower risk
of mortality. Given the simplicity of the diet quality score used in this
study, increasing the intake of recommended foods—without undue emphasis
on learning about hidden fat, total amount and type of fiber, or individual
vitamins and minerals—may represent a practical recommendation for improving
health. Whether the observed protection is explicitly conferred by pattern
of intake of recommended foods or reflects certain unknown factors related
to both RFS and mortality remains an open question.