American Journal of Epidemiology
Copyright ª 2005 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved; printed in U.S.A.
Vol. 163, No. 3
DOI: 10.1093/aje/kwj031
Advance Access publication December 7, 2005
Practice of Epidemiology
The National Cancer Institute Diet History Questionnaire: Validation of Pyramid
Food Servings
Amy E. Millen1, Douglas Midthune2, Frances E. Thompson3, Victor Kipnis2, and Amy F. Subar3
1
Department of Social and Preventive Medicine, School of Public Health and Health Professions, State University of
New York at Buffalo, Buffalo, NY.
2
Biometry Research Group, Division of Cancer Prevention, National Cancer Institute, Bethesda, MD.
3
Risk Factor Monitoring and Methods Branch, Applied Research Program, Division of Cancer Control and Population
Sciences, National Cancer Institute, Bethesda, MD.
Received for publication June 1, 2005; accepted for publication September 13, 2005.
The performance of the National Cancer Institute’s food frequency questionnaire, the Diet History Questionnaire
(DHQ), in estimating servings of 30 US Department of Agriculture Food Guide Pyramid food groups was evaluated
in the Eating at America’s Table Study (1997–1998), a nationally representative sample of men and women aged
20–79 years. Participants who completed four nonconsecutive, telephone-administered 24-hour dietary recalls
(n ¼ 1,301) were mailed a DHQ; 965 respondents completed both the 24-hour dietary recalls and the DHQ. The
US Department of Agriculture’s Pyramid Servings Database was used to estimate intakes of pyramid servings for
both diet assessment tools. The correlation (q) between DHQ-reported intake and true intake and the attenuation
factor (k) were estimated using a measurement error model with repeat 24-hour dietary recalls as the reference
instrument. Correlations for energy-adjusted pyramid servings of foods ranged from 0.43 (other starchy vegetables)
to 0.84 (milk) among women and from 0.42 (eggs) to 0.80 (total dairy food) among men. The mean q and k after
energy adjustment were 0.62 and 0.60 for women and 0.63 and 0.66 for men, respectively. This food frequency
questionnaire validation study of foods measured in pyramid servings allowed for a measure of food intake
consistent with national dietary guidance.
data collection; food; nutrition assessment; nutrition surveys; questionnaires; validation studies [publication type]
Abbreviations: CSFII, Continuing Survey of Food Intakes by Individuals; DHQ, Diet History Questionnaire; EATS, Eating at
America’s Table Study; FFQ, food frequency questionnaire; OPEN, Observing Protein and Energy Nutrition.
The food frequency questionnaire (FFQ) is a practical
approach to measuring dietary exposure in epidemiologic
studies investigating relations between diet and disease.
However, relative risk estimates derived from FFQs may
be biased or attenuated because of substantial FFQ measurement error (1, 2).
Researchers at the National Cancer Institute developed
a new cognitively based FFQ, the Diet History Questionnaire
(DHQ), which is designed to minimize measurement error
through improvements in questionnaire design, wording,
layout, and database development (3). One unique feature of
the DHQ database is the inclusion of the Pyramid Servings
Database (4). The Pyramid Servings Database allows for
the conversion of questionnaire responses into quantitative
estimates, called ‘‘pyramid servings,’’ of 30 different food
groups in the US Department of Agriculture’s original Food
Guide Pyramid, issued in 1992 (5). The latest Food Guide
Pyramid, introduced in April 2005, recommends intakes of
foods in common household measures which are directly
translated from the 1992 Pyramid Servings Database.
Correspondence to Dr. Amy E. Millen, Department of Social and Preventive Medicine, Farber Hall, Room 270, State University of New York
at Buffalo, 3435 Main Street (South Campus), Buffalo, NY 14214-8001 (e-mail: [email protected]).
279
Am J Epidemiol 2006;163:279–288
280 Millen et al.
Pyramid servings represent a standardized classification
scheme for estimating food intake consistent with dietary
guidance. Having a pyramid servings database for an FFQ is
an innovation in that component ingredients of disaggregated food mixtures are systematically assigned to the appropriate guidance-based food groups (6), thereby allowing
for assessment of the full, continuous range of reported intakes of foods. In addition, it allows for the quantitative assessment of added sugar and discretionary fat from all food
sources, previously unavailable from dietary databases. The
Pyramid Servings Database provides an additional, potentially more precise metric, different from previously used metrics of frequencies, servings (derived by different methods),
or grams, and thus warrants evaluation.
The DHQ’s ability to assess nutrient intake was previously
evaluated in the Eating at America’s Table Study (EATS) (3).
The purpose of this research was to validate the DHQ’s
ability to assess intake (in pyramid servings) of 30 US Department of Agriculture food groups and subgroups (4). We
used the EATS data to compare intakes of pyramid servings
between the DHQ and 24-hour dietary recalls (reference instrument). The standard measurement error model (7) for
nutrients is not directly applicable to foods, because of the
nonnegligible probability of zero intake on a given day for
most foods. In this paper, we present a measurement error
model for foods and show that under this model one can use
the standard estimation procedures to obtain consistent estimates of the correlation with true intake and the attenuation
factor, the multiplicative bias in the estimated log relative
risk of disease due to measurement error in the DHQ.
Evaluation of FFQ-reported food intake is important, because nutritional epidemiologists are increasingly interested
in the relations of foods and dietary patterns with disease
outcomes (8). Additionally, because dietary recommendations are phrased in terms of pyramid servings or equivalent
measures of food intake, directly estimating these from
FFQs is advantageous for epidemiologic studies, since it
will be possible to interpret findings in amounts and terms
consistent with public health food recommendations in the
new Food Guide Pyramid.
MATERIALS AND METHODS
Sample and study design
EATS, which began in August 1997, used random digit
dialing techniques to obtain a nationally representative sample of persons aged 20–79 years, balanced by gender. A detailed description of the sampling process and methods used
has been published elsewhere (3). There were 1,500 participants (n ¼ 738 women, n ¼ 762 men) who completed at least
one of four nonconsecutive, telephone-administered 24-hour
dietary recalls. Of these persons, 1,301 completed all four
24-hour dietary recalls and were eligible to receive the DHQ
by mail. Participants were also randomized into two groups
and were sent either the Block FFQ (9) or the Willett FFQ
(10) by mail, in addition to the DHQ. Participants were asked
to fill out the DHQ and the other mailed FFQ (either the
Block FFQ or the Willett FFQ) 1 month apart. Participants
were instructed to mail back the completed questionnaires.
Of the 1,301 participants who completed all four 24-hour
dietary recalls, 1,000 mailed back the DHQ.
Twenty-four-hour dietary recalls
The four 24-hour dietary recalls were scheduled to be
administered 3 months apart (one per season) and collected
between September 1997 and August 1998, using the
multiple-pass method developed for the 1994–1996 Continuing Survey of Food Intakes by Individuals (CSFII) (11).
The 24-hour dietary recall data were coded using the
Food Intake Analysis System (version 3.0), developed at
the University of Texas, which uses food codes from the
CSFII. For foods reported but not found in the CSFII database (n ¼ 14), a food which provided the best match with
regard to food description, total energy, and macronutrient
content was chosen from the CSFII data files released
through 1998 to approximate pyramid serving values.
Food frequency questionnaire
The DHQ, described previously (3), queries respondents
about their frequency of intake of 124 separate food items
and asks about portion sizes for most of these items by
providing a choice of three ranges. For 44 of the 124 foods,
1–7 additional nested questions are asked about related factors such as seasonal intake, food type (e.g., low-fat, lean,
diet, caffeine-free), fat use, or fat additions. The DHQ also
includes additional questions about the use of low-fat foods.
Of the 1,301 DHQs sent to participants, 1,000 were returned. Thirty-five were deemed incomplete and excluded
from further analysis because the respondents skipped two
or more adjacent pages of the 36-page DHQ booklet. This
left 519 women and 446 men for further analyses.
Pyramid Servings Database
We used the Pyramid Servings Database corresponding to
the 1994–1996 CSFII (11), which provides the number of
servings of each of the Food Guide Pyramid’s food groups
and the amounts of discretionary fat (in grams) and added
sugar (in teaspoons) contained in 100 g of every food included in the survey. It utilizes a recipe file to disaggregate
food mixtures into their component ingredients and assigns
them to food groups. Pyramid servings were added to the
DHQ database, in addition to the nutrient content of foods,
on the basis of methods previously described (6). Since
EATS’ 24-hour dietary recalls and the DHQ both utilized
the Pyramid Servings Database, pyramid servings of the 30
US Department of Agriculture major food groups and subgroups could be determined for both dietary measurement
tools. To our knowledge, pyramid servings have not been
added to either the Block or the Willett FFQ database,
thereby precluding similar analyses of those tools.
Statistical analyses
We evaluate the ability of the DHQ to measure food
group intakes in epidemiologic studies by estimating the
correlation coefficient (q) and the attenuation factor (k)
for each food group. q is the correlation between reported
Am J Epidemiol 2006;163:279–288
Validation of Pyramid Food Servings
intake (the DHQ) and estimated true usual intake (based on
four 24-hour dietary recalls) and is estimated in this case
from the deattenuated crude correlation between the DHQ
and the four 24-hour dietary recalls. k is the multiplicative
factor by which an estimated log relative risk of disease
would be biased because of measurement error in the DHQ.
It is equivalent to the slope of the linear regression of the 24hour dietary recall on the DHQ, and it usually falls between
0 and 1 in nutritional studies, thereby attenuating (biasing
toward the null) estimates of relative risk. Values of k close
to 0 indicate maximum attenuation, while values close to 1
indicate minimum attenuation.
We estimated correlation coefficients and attenuation factors using the following measurement error model. For individual i, let pi denote the probability of consuming an item
from a given food group on any given day. Let Ai be the
person’s usual intake on a consumption day, measured on
an appropriate scale. On the chosen scale, the person’s overall true usual intake, Ti, can be represented as the product
Ti ¼ pi Ai.
Error in the DHQ was allowed to include systematic bias
correlated with true usual intake and within-person random
variation as in the standard measurement error model (7).
The 24-hour dietary recall (our reference instrument) was
assumed to involve no misclassification of consumption
days and to report zero intake only on a nonconsumption
day. It was also assumed that, on an appropriate scale, the
24-hour dietary recall represents usual intake on a consumption day, plus additive within-person random error with
mean zero uncorrelated with true intake and error in the
DHQ. Because the repeat recalls were conducted approximately 3 months apart, it was assumed that random errors
corresponding to repeat measurements for the same person
were uncorrelated with each other.
Given the above assumptions, the measurement error
model is
Qi ¼ b0 þ b1 Ti þ ei ;
Fij ¼ 0 on a nonconsumption day j;
Fij ¼ Ai þ eij on a consumption day j;
where, for person i, Qi is the DHQ-reported intake, Ti is the
true usual intake, b0 is the intercept and b1 is the slope of the
linear regression of Qi on Ti, and ei is random within-person
error. Fij is the jth repeat 24-hour dietary recall-reported
intake, j ¼ 1, . . ., 4, and eij is the within-person random
error for repeat measurement j. Within-person errors ei
and eij are assumed to have means equal to zero and constant
variances and to be independent of each other and of Ti and
Ai. One can show that this measurement error model can be
rewritten as
Qi ¼ b0 þ b1 Ti þ ei ;
281
model is similar to the standard model (7), with the exception that neither true usual intake nor random within-person
error in the 24-hour dietary recall is normally distributed,
even if the distributions of the usual amount on consumption
days, Ai, and within-person error in the 24-hour dietary recall on consumption days, eij, are normal.
Prior to statistical analyses, we excluded observations
from the 24-hour dietary recall and the DHQ that were determined to be outliers with respect to reported total energy
intake measured on the log scale. For the remaining food
group observations, data were transformed to a scale on
which the 24-hour dietary recall had additive measurement
error on days on which the participant consumed an item
from the food group. Finally, we removed outliers among
nonzero values for each food group and instrument, measured on the transformed scale. Outliers were defined as
values outside the interval ranging from the 25th percentile
of the distribution minus two times the interquartile range to
the 75th percentile plus two times the interquartile range.
We used the functional transformation method of Eckert
et al. (12) to find the best power transformation for each
food group. So that outliers would not unduly influence
the choice of transformation, we first excluded outliers;
for this step only, we used a more conservative definition
of outliers, namely values outside the interval of the 25th
and 75th percentiles, plus or minus three times the interquartile range. Final outlier exclusions across all food
groups, instruments, and repeated applications of the 24hour dietary recall ranged from 2 to 14 for men (mean ¼ 7)
and from 9 to 24 for women (mean ¼ 13).
The measurement error model was applied to both unadjusted food intake and energy-adjusted food intake using
the density method to adjust for energy (13). Parameters
were estimated using maximum likelihood estimation, assuming that the random variables were normally distributed.
Although this normality assumption is likely to be violated,
at least with respect to such random variables as true usual
intake and within-person random error for the 24-hour dietary recall, the estimated parameters will still be consistent
(asymptotically unbiased). Since model-based estimates of
standard errors may be incorrect, we calculated bootstrap
standard errors, which do not depend on distributional assumptions (14). Since maximum likelihood estimation does
not require every person to have complete data (i.e., four 24hour dietary recalls and the DHQ), we included in the analyses all 1,500 study participants who completed at least one
24-hour dietary recall.
Because the adopted measurement error model assumes
that DHQ-reported intake follows a continuous distribution,
we excluded from the analysis three foods for which substantial percentages of respondents reported zero intake on
the DHQ: yogurt, organ meat, and soy. Alcoholic beverages,
for which rather large percentages of zero intake were also
reported on the DHQ, were further analyzed for the DHQ
consumers only.
Fij ¼ Ti þ nij ;
where nij, j ¼ 1, . . ., 4, have mean zero and constant variance
and are uncorrelated with true intake Ti, error in the DHQ,
and each other. Therefore, the adapted measurement error
Am J Epidemiol 2006;163:279–288
RESULTS
The demographic profile of the participants during the
progression of the study—from completion of at least one
282 Millen et al.
24-hour dietary recall to randomization after completion of
four 24-hour dietary recalls to completion of an FFQ—was
previously reported (3). Participants completing an FFQ, as
compared with those completing at least one 24-hour dietary
recall, were significantly more likely to be female, older, and
of White race/ethnicity. No educational differences were
seen.
Table 1 shows the mean values and standard errors for
daily intake of pyramid servings of the 30 food groups
among 497 women and 436 men who completed all four
24-hour dietary recalls and the DHQ, after exclusion of 32
participants whose values for reported total energy intake
were outliers. For women and men, reported intakes were
greater (>15 percent difference) on the DHQ than on the 24hour dietary recalls for other starchy vegetables, dark green
vegetables, total fruit, citrus fruit/melon/berries, other fruit,
and yogurt. For women, reported intakes were also greater
for total vegetables, deep yellow vegetables, legumes, other
vegetables, fish/other seafood, and milk. For men, reported
intakes were also greater for nuts and seeds and alcohol.
Reported intakes for men and women were lower (>15
percent difference) on the DHQ than on the 24-hour dietary
recalls for total grains, nonwhole grains, poultry, organ
meat, and eggs. For men, reported intakes were also lower
for red meat/poultry/fish, soy, and cheese.
Table 2 shows the estimated correlation (q) between the
DHQ and true usual intake and the attenuation factor (k) for
the transformed food group data with and without energy
adjustment, by gender. Adjusting for energy intake strengthened the correlation between truth and the DHQ for the
majority of the food groups. Among women, energyadjusted correlations ranged from 0.43 for servings of other
starchy vegetables to 0.84 for servings of milk. Among men,
energy-adjusted correlations ranged from 0.42 for servings
of tomatoes and eggs to 0.80 for servings of total dairy food.
The energy-adjusted correlation coefficients were greater
than 0.50 for all food groups except white potatoes (men
only), other starchy vegetables (women only), tomatoes,
poultry (women only), and eggs. Energy-adjusted correlation coefficients were highest (>0.65) for whole grains (men
only), total vegetables (women only), dark green vegetables,
legumes (men only), total fruit, other fruit (men only), beef/
pork/lamb, total dairy food, milk, discretionary fat (men
only), added sugar, and alcohol. The mean q across all
energy-adjusted food groups was 0.62 for women and 0.63
for men (table 3).
Energy adjustment primarily strengthened or left unchanged the attenuation coefficients. Among women, energyadjusted attenuation coefficients ranged from 0.20 for other
starchy vegetables to 0.92 for eggs (table 2). Among men,
energy-adjusted attenuation coefficients ranged from 0.41 for
other starchy vegetables to 1.32 for eggs. After energy adjustment, the attenuation coefficients were greater than 0.50
for all food groups except total grains (women only), nonwhole grains (women only), total vegetables (men only), white
potatoes (men only), other starchy vegetables, dark green
vegetables, tomatoes, legumes, poultry, fish/other seafood,
franks/luncheon meats (women only), and discretionary fat.
After energy adjustment, the mean k was 0.60 for women and
0.66 for men (table 3).
DISCUSSION
We estimated the ability of the DHQ to assess food group
intake as measured in pyramid servings for 30 US Department of Agriculture food groups. Validation of food intake
poses challenges different from those of nutrient intake
along several dimensions: 1) how to disentangle individual
foods from mixtures; 2) how to combine the nutrient contributions of solid foods and less dense liquid foods; and
3) how to relate food intake to dietary guidance. The availability of the US Department of Agriculture’s Pyramid Servings Database eloquently satisfies these challenges. To our
knowledge, this is the first FFQ validation study of food
groups measured in pyramid servings.
A further statistical challenge is the variability and nonnormal distribution of food intakes. Daily intakes of individual foods tend to be more variable than those for nutrients.
Pyramid servings may somewhat reduce this variability for
some foods, because the system disaggregates food mixtures into a smaller set of component ingredients. Therefore,
one implication of using pyramid servings in future studies
is a possible reduction in measurement error in assessing
food intake. The distributions of pyramid servings are
more similar to those of nutrients, both being represented
by continuous variables in the database. This is unlike other
methods of quantifying food intake, such as determining
frequencies or grams, the latter of which takes into account
frequency and portion size but generally does not disaggregate food mixtures.
The inclusion of a database with the ability to assess
pyramid servings is a scientific advance which may allow
researchers to more accurately assess intakes in different
food groups and may help facilitate the communication of
scientific findings to the public with reference to dietary
guidance. We have shown here that the use of pyramid
servings for FFQ databases is possible and that it has the
potential to provide a standardized measurement of food
intake from study to study.
The new Food Guide Pyramid, MyPyramid (http://www.
mypyramid.gov/), makes recommendations based on common household measurements (i.e., cups, ounces, and teaspoons). Translating between household measurements and
pyramid servings is relatively easy. This change is less a substantive one than an attempt to better communicate with
consumers. For example, in MyPyramid, 1/4 cup of raisins
is equivalent to 1/2 cup of fruit, and both are equivalent to
one pyramid serving. There are, however, a few substantive
changes, such as recommendations for the separation of oil
from discretionary fat and the inclusion of the term ‘‘discretionary calories’’ to convey the idea that after food intake
recommendations are met, people have a limited amount of
calories to consume within food groups or as sugar, alcohol,
or discretionary fat. At the time of this analysis, we used the
current Pyramid Servings Database available to us. In the
future, we will update the DHQ to the current system.
In our study, median intake estimates on the DHQ and the
24-hour dietary recalls differed for many pyramid food
groups. There are several possible reasons for this. There
may be a bias for people to report higher consumption
of more socially desirable foods (e.g., fruits, vegetables,
Am J Epidemiol 2006;163:279–288
Validation of Pyramid Food Servings
283
TABLE 1. Mean intakes and percentages of persons reporting no consumption in Food Guide Pyramid food groups among persons
who completed four 24-hour dietary recalls and the Diet History Questionnaire, Eating at America’s Table Study, 1997–1998
Women (n ¼ 497)
Food Guide Pyramid food group
24-hour dietary recalls
Mean (SE*)
% zeroy
Men (n ¼ 436)
Diet History
Questionnaire
Mean (SE)
24-hour dietary recalls
% zero
Mean (SE)
% zero
Diet History
Questionnaire
Mean (SE)
% zero
Grain group (servings/day)
Total grains
5.84 (0.10)
0
4.90 (0.12)z
0
8.54 (0.17)
0
6.76 (0.18)z
0
Whole grains
1.02 (0.04)
7
0.98 (0.03)
0
1.55 (0.07)
8
1.39 (0.05)
0
Nonwhole grains
4.82 (0.09)
0
3.92 (0.11)z
0
6.99 (0.15)
0
5.36 (0.15)z
0
Vegetable group (servings/day)
Total vegetables
3.10 (0.07)
0
3.97 (0.11)z
0
4.22 (0.10)
0
4.49 (0.14)
0
White potatoes
0.75 (0.03)
14
0.84 (0.03)
0
1.12 (0.05)
12
1.08 (0.05)
0
Other starchy vegetables
0.15 (0.01)
45
0.35 (0.01)z
0
0.23 (0.02)
46
0.38 (0.02)z
0
Dark green vegetables
0.26 (0.02)
42
0.41 (0.03)z
1
0.22 (0.02)
52
0.36 (0.03)z
1
Deep yellow vegetables
0.21 (0.01)
19
0.25 (0.01)z
0
0.27 (0.02)
23
0.25 (0.01)
0
Tomatoes
0.48 (0.02)
2
0.54 (0.02)
0
0.71 (0.03)
2
0.72 (0.03)
0
Legumes§
0.12 (0.01)
61
0.14 (0.01)z
4
0.21 (0.02)
57
0.19 (0.01)
3
Other vegetables
1.14 (0.04)
0
1.44 (0.05)z
0
1.47 (0.05)
0
1.51 (0.06)
0
Fruit group (servings/day)
Total fruit
1.48 (0.06)
3
2.40 (0.10)z
0
1.96 (0.10)
1
2.68 (0.14)z
0
Citrus fruit, melon, berries
0.75 (0.04)
6
1.16 (0.05)z
0
1.02 (0.07)
5
1.29 (0.07)z
0
Other fruit
0.73 (0.04)
11
1.25 (0.05)z
0
0.95 (0.05)
12
1.38 (0.08)z
0
Red meat, poultry, fish
3.57 (0.08)
1
3.30 (0.10)
0
5.93 (0.14)
2
5.03 (0.17)z
0
Beef, pork, lamb, etc.
1.59 (0.06)
7
1.46 (0.05)
0
2.65 (0.10)
6
2.42 (0.09)
0
Poultry
1.10 (0.05)
19
0.86 (0.04)z
0
1.60 (0.08)
18
1.06 (0.06)z
0
Fish, other seafood
0.43 (0.04)
52
0.51 (0.04)z
0
0.72 (0.06)
46
0.64 (0.04)
0
Franks, luncheon meats
0.43 (0.02)
35
0.46 (0.02)
2
0.94 (0.05)
21
0.90 (0.05)
1
Organ meat
0.03 (0.01)
96
0.01 (0.0)z
61
0.02 (0.01)
97
0.01 (0.0)z
58
0
Meat group (ounces{/day)
Lean meat
Lean meat equivalents
Eggs#
0.29 (0.02)
9
0.22 (0.01)z
0
0.43 (0.03)
5
0.26 (0.01)z
Nuts, seeds#
0.15 (0.01)
30
0.17 (0.01)
1
0.28 (0.03)
27
0.32 (0.03)z
0
Soy
0.06 (0.01)
81
0.05 (0.01)
66
0.06 (0.02)
82
0.04 (0.01)z
65
Total dairy food
1.26 (0.04)
0
1.41 (0.06)
0
1.75 (0.06)
0
1.76 (0.07)
0
Milk
0.76 (0.03)
2
0.92 (0.06)z
0
1.05 (0.05)
1
1.14 (0.06)
0
Cheese
0.44 (0.02)
7
0.39 (0.02)
0
0.66 (0.03)
6
0.54 (0.02)z
0
Yogurt
0.05 (0.01)
83
37
0.03 (0.01)
90
0.07 (0.01)z
52
Discretionary fat (g/day)
48.67 (0.93)
0
48.70 (1.24)
0
72.99 (1.45)
0
66.40 (1.98)
0
Added sugar (teaspoons**/day)
16.62 (0.50)
0
15.72 (0.66)
0
23.70 (0.76)
0
22.08 (0.88)
0
0.40 (0.04)
58
0.39 (0.05)
28
0.87 (0.07)
47
Dairy group (servings/day)
0.08 (0.01)z
Pyramid tip
Alcohol (no. of drinks/day)
1.10 (0.19)z
22
* SE, standard error.
y Percentage of persons reporting no consumption of items in this food group.
z Greater than 15% difference between the average of the four 24-hour dietary recalls and the Diet History Questionnaire sample means.
§ Legumes can be counted toward the meat group or the vegetable group, according to the Food Guide Pyramid. One serving of dry beans and
peas is equivalent to 1 ounce (28.4 g) of cooked lean meat.
{ 1 ounce ¼ 28.4 g.
# 1/3 cup of nuts, 1/4 cup of seeds, and one egg are equivalent to 1 ounce (28.4 g) of cooked lean meat.
** 1 teaspoon ¼ 5 g.
Am J Epidemiol 2006;163:279–288
284 Millen et al.
TABLE 2. Deattenuated correlation coefficients (r) and attenuation factors (l) for
correlation between the Diet History Questionnaire and true usual intake (based on four
24-hour dietary recalls) obtained using a measurement error model, unadjusted
and adjusted for total energy intake, Eating at America’s Table Study, 1997–1998
Women
Men
Food Guide Pyramid food group
q (SE*)
k (SE)
q (SE)
k (SE)
Grain group (servings/day)
Total grains
Unadjusted
0.48 (0.05)
0.32 (0.05)
0.53 (0.05)
0.41 (0.05)
Adjusted
0.59 (0.06)
0.44 (0.05)
0.63 (0.05)
0.53 (0.05)
Whole grains
Unadjusted
0.55 (0.05)
0.66 (0.07)
0.62 (0.05)
0.75 (0.08)
Adjusted
0.62 (0.05)
0.81 (0.07)
0.66 (0.04)
0.89 (0.08)
Nonwhole grains
Unadjusted
0.49 (0.06)
0.32 (0.04)
0.48 (0.05)
0.39 (0.04)
Adjusted
0.60 (0.06)
0.48 (0.05)
0.61 (0.05)
0.56 (0.05)
Vegetable group (servings/day)
Total vegetables
Unadjusted
0.49 (0.06)
0.31 (0.04)
0.46 (0.06)
0.31 (0.05)
Adjusted
0.66 (0.05)
0.53 (0.05)
0.63 (0.07)
0.45 (0.06)
White potatoes
Unadjusted
0.48 (0.06)
0.40 (0.06)
0.46 (0.08)
0.38 (0.07)
Adjusted
0.52 (0.06)
0.54 (0.08)
0.50 (0.08)
0.50 (0.09)
Other starchy vegetables
Unadjusted
0.38 (0.11)
0.16 (0.05)
0.56 (0.10)
0.36 (0.10)
Adjusted
0.43 (0.11)
0.20 (0.05)
0.53 (0.11)
0.41 (0.13)
Vegetable group (servings/day)
Dark green vegetables
Unadjusted
0.69 (0.07)
0.43 (0.05)
0.77 (0.08)
0.42 (0.06)
Adjusted
0.67 (0.06)
0.48 (0.06)
0.75 (0.10)
0.41 (0.06)
Deep yellow vegetables
Unadjusted
0.60 (0.08)
0.46 (0.07)
0.49 (0.07)
0.59 (0.16)
Adjusted
0.58 (0.09)
0.76 (0.23)
0.54 (0.08)
0.81 (0.30)
Tomatoes
Unadjusted
0.44 (0.07)
0.34 (0.06)
0.37 (0.08)
0.35 (0.07)
Adjusted
0.49 (0.08)
0.49 (0.09)
0.42 (0.08)
0.43 (0.08)
Legumesy
Unadjusted
0.59 (0.10)
0.33 (0.08)
0.53 (0.09)
0.34 (0.06)
Adjusted
0.62 (0.12)
0.33 (0.05)
0.66 (0.10)
0.44 (0.06)
Other vegetables
Unadjusted
0.52 (0.06)
0.47 (0.06)
0.49 (0.07)
0.40 (0.07)
Adjusted
0.63 (0.05)
0.67 (0.07)
0.57 (0.08)
0.52 (0.08)
Fruit group (servings/day)
Total fruit
Unadjusted
0.62 (0.04)
0.57 (0.04)
0.63 (0.05)
0.72 (0.06)
Adjusted
0.66 (0.04)
0.68 (0.05)
0.70 (0.03)
0.86 (0.05)
Citrus fruit, melon, berries
Unadjusted
0.60 (0.05)
0.61 (0.06)
0.58 (0.05)
0.83 (0.08)
Adjusted
0.63 (0.05)
0.74 (0.07)
0.61 (0.04)
0.94 (0.09)
Other fruit
Unadjusted
0.61 (0.04)
0.50 (0.05)
0.65 (0.04)
0.66 (0.07)
Adjusted
0.64 (0.04)
0.61 (0.05)
0.71 (0.04)
0.82 (0.07)
Table continues
Am J Epidemiol 2006;163:279–288
Validation of Pyramid Food Servings
TABLE 2. Continued
Women
Men
Food Guide Pyramid food group
q (SE*)
)
k (SE)
q (SE)
k (SE)
Meat group (ouncesz/day)
Lean meat
Red meat, poultry, fish
Unadjusted
0.48 (0.07)
0.39 (0.06)
0.53 (0.05)
0.51 (0.07)
Adjusted
0.58 (0.06)
0.56 (0.07)
0.61 (0.06)
0.72 (0.09)
Beef, pork, lamb, etc.
Unadjusted
0.71 (0.07)
0.58 (0.06)
0.61 (0.06)
0.55 (0.07)
Adjusted
0.79 (0.07)
0.70 (0.06)
0.72 (0.07)
0.73 (0.07)
Poultry
Unadjusted
0.45 (0.08)
0.32 (0.06)
0.60 (0.10)
0.47 (0.07)
Adjusted
0.50 (0.08)
0.41 (0.06)
0.56 (0.10)
0.46 (0.08)
Fish, other seafood
Unadjusted
0.53 (0.11)
0.37 (0.08)
0.60 (0.08)
0.46 (0.07)
Adjusted
0.53 (0.17)
0.39 (0.08)
0.60 (0.09)
0.47 (0.07)
Franks, luncheon meats
Unadjusted
0.56 (0.09)
0.33 (0.05)
0.60 (0.10)
0.43 (0.08)
Adjusted
0.54 (0.09)
0.38 (0.06)
0.64 (0.09)
0.56 (0.09)
Lean meat equivalents
Eggs§
Unadjusted
0.50 (0.08)
0.81 (0.21)
0.46 (0.07)
1.01 (0.30)
Adjusted
0.45 (0.08)
0.92 (0.25)
0.42 (0.08)
1.32 (0.52)
Nuts, seeds§
Unadjusted
0.53 (0.06)
0.51 (0.09)
0.57 (0.06)
0.62 (0.13)
Adjusted
0.51 (0.07)
0.65 (0.13)
0.56 (0.06)
0.72 (0.17)
Dairy group (servings/day)
Total dairy food
Unadjusted
0.76 (0.03)
0.70 (0.04)
0.73 (0.04)
0.63 (0.05)
Adjusted
0.78 (0.03)
0.72 (0.05)
0.80 (0.04)
0.65 (0.04)
Milk
Unadjusted
0.84 (0.03)
0.80 (0.05)
0.75 (0.03)
0.71 (0.04)
Adjusted
0.84 (0.04)
0.86 (0.05)
0.78 (0.03)
0.76 (0.05)
Cheese
Unadjusted
0.58 (0.06)
0.49 (0.05)
0.62 (0.05)
0.62 (0.06)
Adjusted
0.62 (0.08)
0.76 (0.22)
0.61 (0.06)
0.95 (0.17)
Pyramid tip
Discretionary fat (g/day)
Unadjusted
0.49 (0.06)
0.33 (0.05)
0.59 (0.05)
0.41 (0.04)
Adjusted
0.65 (0.04)
0.45 (0.03)
0.66 (0.04)
0.46 (0.04)
Added sugar (teaspoons{/day)
Unadjusted
0.72 (0.03)
0.65 (0.03)
0.66 (0.04)
0.60 (0.05)
Adjusted
0.79 (0.02)
0.74 (0.03)
0.68 (0.05)
0.67 (0.05)
Alcohol (no. of drinks/day)
Unadjusted
0.77 (0.05)
0.89 (0.12)
0.76 (0.04)
0.84 (0.07)
Adjusted
0.77 (0.04)
0.82 (0.11)
0.79 (0.03)
0.86 (0.10)
* SE, standard error.
y Legumes can be counted toward the meat group or the vegetable group, according to the Food Guide
Pyramid. One serving of dry beans and peas is equivalent to 1 ounce (28.4 g) of cooked lean meat.
z 1 ounce ¼ 28.4 g.
§ 1/3 cup of nuts, 1/4 cup of seeds, and one egg are equivalent to 1 ounce (28.4 g) of cooked lean meat.
{ 1 teaspoon ¼ 5 g.
Am J Epidemiol 2006;163:279–288
285
286 Millen et al.
TABLE 3. Mean correlation and attenuation coefficients for
correlation between the Diet History Questionnaire and four
24-hour dietary recalls, with and without adjustment for total
energy intake, Eating at America’s Table Study, 1997–1998
Mean correlation
coefficient (q)
Mean attenuation
factor (k)
Unadjusted
Adjusted
Unadjusted
Adjusted
Overall
0.58
0.62
0.52
0.63
Women
0.57
0.62
0.48
0.60
Men
0.58
0.63
0.55
0.66
fish/other seafood, yogurt, milk) and lower consumption of
less socially desirable foods (e.g., nonwhole grains, red
meat/poultry/fish, organ meat, eggs, cheese) on the DHQ
than on the 24-hour dietary recalls. In general, these findings
were consistent with findings from previous studies (15–20).
However, a few discrepancies did occur. For example, a few
studies reported added fat consumption at a higher frequency on FFQs than on 24-hour dietary recalls (16–18).
These inconsistencies may be partially explained by variation in the proportions of different types of fats consumed
by country.
Another reason that reported intakes for some food
groups differed between the DHQ and the 24-hour dietary
recalls may be that there are differences in the nature of the
two dietary assessment methods. Four random 24-hour dietary recalls are less likely to provide observations of intake
for foods that are rarely consumed or only episodically consumed. The recalls may indicate lower levels of intake in
comparison with an FFQ that queries about usual dietary
intake over the past year.
For most food groups, the deattenuated correlation coefficients and attenuation factors were greater than 0.50 for
both genders after energy adjustment; however, they were
slightly stronger, on average, among men. Unlike mean and
median values, correlation coefficients do not compare absolute intakes of foods between dietary assessment tools but
rather compare how the different tools classify individuals
with respect to food intake. Energy adjustment, on average,
improved the correlation and attenuation coefficients. This
suggests that adjusting for energy intake reduces measurement error in reporting of foods on the DHQ. Nearly all
items on the DHQ contribute calories, and therefore total
energy intake may serve as a good surrogate variable with
which to adjust for measurement error in FFQs (21).
Attenuation factors are used to estimate the degree to
which the log relative risk of disease would be attenuated
due to measurement error in the DHQ. For example, if a true
relative risk of 0.50 existed for a specific disease among
women with a high adjusted total vegetable intake as compared with a low intake, this relative risk would be attenuated by measurement error to 0.85.
The attenuation and correlation coefficients among men
were, on average, slightly higher than those among women.
This may be explained by differential measurement error by
gender. In the Observing Protein and Energy Nutrition
(OPEN) Study, which also used the National Cancer Insti-
tute DHQ, Subar et al. (21) found that women underreported
their total energy intake more than men did.
A number of FFQ-type food validation studies have been
conducted among adults in Europe (15–18, 22–28), Asia
(29–36), the United States (19, 20, 37, 38), Africa (39,
40), and South America (41). While it is recognized that
differences in instruments and study samples used among
these various studies may have caused differences in results,
another factor deserving of attention is differences in analytical methods.
To our knowledge, our study is the first food validation
study to use pyramid servings; no other studies reviewed in
this manuscript measured food intake with this unit. Thus, it
is possible that differences in study results could be explained by differences in the units used to express food intake: frequencies (the number of times a food is eaten within
a specified time period (day, week, etc.), without reference
to portion size) or servings (a count of the number of times
a specific portion size of a food is eaten within a specified
time period) (15, 19, 20, 29, 30, 37); grams (16–18, 22–28,
31–36, 38–40); percentage of total calories (41); or, as in
this case, pyramid servings. Different studies may accentuate different metrics when assigning foods to specific food
groups. For example, Flagg et al. (38) analyzed the sum of
dairy food intake (milk þ yogurt þ cheese), combining
liquid and solid dairy products, whereas Feskanich et al.
(19) differentiated between the different types of dairy foods
(i.e., milk, yogurt, cheese, cream cheese, etc.).
It appears that investigators in past FFQ validation studies
considered added sugar to be the addition of raw sugar to
prepared foods. In this study, pyramid servings of added
sugar were based on all sources, including sugar added by
food manufacturers to cereals, baked goods, etc. Similarly,
pyramid servings of discretionary fats included fats added to
foods during cooking and at the table, as well as fats exceeding those present in lean cuts of meat and poultry. The
food groups ‘‘discretionary fats’’ and ‘‘added sugars’’ are
unique additions to a foods database, because they express
fat and sugar intake behavior as measures, within the dietary
guidance concept. Individuals have choices in how to
‘‘spend’’ their fat and sugar allowances.
Finally, the search for consistency between studies is
hampered by the presentation of Spearman correlation coefficients or unadjusted Pearson correlation coefficients
(16–18, 22, 23, 26, 27, 29, 31–34, 39, 40), which are not
directly comparable to deattenuated Pearson correlation coefficients, which correct for within-person variation in the
reference instrument.
The results of this validation study may not be completely
generalizable to other populations. Our sample consisted
primarily of White, well-educated persons who were willing
to complete four 24-hour dietary recalls followed by two
FFQs. This is a limitation of many validation studies.
In this study, we assumed that error in the 24-hour dietary
recalls included no misclassification of nonconsumption
days and, for nonzero reported amounts, was unbiased and
contained only within-person random error uncorrelated
with errors in the DHQ. These assumptions may be unwarranted for self-reported reference instruments such as
24-hour dietary recalls and dietary records. The OPEN
Am J Epidemiol 2006;163:279–288
Validation of Pyramid Food Servings
Study showed that there is reporting bias for energy and
protein intake in both the DHQ and 24-hour dietary recalls,
and that people systematically differ in their reporting accuracy (21). Therefore, all dietary reference instruments
could involve systematic error at the individual level, and
this error could be correlated with its counterpart in the
DHQ. Thus, our results may reflect overestimation of the
correlations with true intake and underestimation of true
attenuation (42).
With respect to absolute energy and protein intakes, as the
OPEN biomarker study showed, the DHQ has significant
measurement error, especially in the direction of underreporting (21). This is probably true for many FFQs. However,
after adjustment for total energy intake, protein intake
showed far less measurement error in comparison with unadjusted values. This observation suggests that energy adjustment may minimize the effects of measurement error in
FFQs (42), although the extent to which the results for protein can be extended to individual foods or food groups is
unknown.
In summary, the results of this study demonstrate for the
first time the use of a pyramid servings database for FFQs.
The validity of the US Department of Agriculture Pyramid
Servings Database was comparable or superior to that of
previous FFQ food group validation studies using less precise food grouping methods. The use of a pyramid servings
database in nutritional epidemiologic research provides substantial advantages in measuring consumption in individual
food groups, as well as consumption of sugar and discretional fat. The inclusion of a pyramid servings database in
the DHQ provides an innovative and advanced method of
assessing food intake behavior for use in nutrition surveillance, assessment, or analysis of diet-disease associations.
The DHQ is publicly available online at a National Cancer
Institute website (http://riskfactor.cancer.gov/DHQ/).
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
ACKNOWLEDGMENTS
Conflict of interest: none declared.
18.
19.
REFERENCES
1. Freedman LS, Schatzkin A, Wax Y. The impact of dietary measurement error on planning sample size required in a cohort
study. Am J Epidemiol 1990;132:1185–95.
2. Freudenheim JL, Marshall JR. The problem of profound mismeasurement and the power of epidemiological studies of diet
and cancer. Nutr Cancer 1988;11:243–50.
3. Subar AF, Thompson FE, Kipnis V, et al. Comparative validation of the Block, Willett, and National Cancer Institute food
frequency questionnaires: The Eating at America’s Table
Study. Am J Epidemiol 2001;154:1089–99.
4. Cook A, Friday JE. Pyramid Servings Database for USDA
survey food codes, version 2.0. Washington, DC: Agricultural
Research Service, US Department of Agriculture, 2004.
(http://www.barc.usda.gov/bhnrc/cnrg/section1.htm).
5. Shaw A, Fulton L, Davis C, et al. Using the Food Guide Pyramid: a resource for nutrition educators. Washington, DC:
Am J Epidemiol 2006;163:279–288
20.
21.
22.
23.
287
Center for Nutrition Policy and Promotion, US Department
of Agriculture, 2005. (http://www.nal.usda.gov/fnic/Fpyr/
guide.pdf).
Cleveland LE, Cook DA, Krebs-Smith SM, et al. Method for
assessing food intakes in terms of servings based on food
guidance. Am J Clin Nutr 1997;65(suppl):1254S–63S.
Freedman LS, Carroll RJ, Wax Y. Estimating the relation between dietary intake obtained from a food frequency questionnaire and true average intake. Am J Epidemiol 1991;134:
310–20.
Hu FB. Dietary pattern analysis: a new direction in nutritional
epidemiology. Curr Opin Lipidol 2002;13:3–9.
Block G, Woods M, Potosky A, et al. Validation of a selfadministered diet history questionnaire using multiple diet
records. J Clin Epidemiol 1990;43:1327–35.
Willett WC, Sampson L, Stampfer MJ, et al. Reproducibility
and validity of a semiquantitative food frequency questionnaire. Am J Epidemiol 1985;122:51–65.
Tippett KS, Cypel YS. Design and operation: the Continuing
Survey of Food Intakes by Individuals and the Diet and Health
Knowledge Survey, 1994–96. Washington, DC: Agricultural
Research Service, US Department of Agriculture, 1998.
(Document no. SuDoc A 77.516:96-1).
Eckert RS, Carroll RJ, Wang N. Transformations to additivity
in measurement error models. Biometrics 1997;53:262–72.
Willett W, Stampfer M. Implications of total energy intake for
epidemiologic analyses. In: Kelsey JL, Marmot MG, Stolley
PD, et al, eds. Nutritional epidemiology. New York, NY:
Oxford University Press, 1998:273–301.
Efron B, Tibshirani RJ. An introduction to the bootstrap.
London, United Kingdom: Chapman and Hall Ltd, 1993.
Johansson I, Hallmans G, Wikman A, et al. Validation and
calibration of food-frequency questionnaire measurements in
the Northern Sweden Health and Disease cohort. Public Health
Nutr 2002;5:487–96.
Pisani P, Faggiano F, Krogh V, et al. Relative validity and reproducibility of a food frequency dietary questionnaire for use
in the Italian EPIC centres. Int J Epidemiol 1997;26(suppl 1):
S152–60.
Ocke MC, Bueno-de-Mesquita HB, Goddijn HE, et al. The
Dutch EPIC food frequency questionnaire. I. Description of
the questionnaire, and relative validity and reproducibility for
food groups. Int J Epidemiol 1997;26(suppl 1):S37–48.
EPIC Group of Spain. Relative validity and reproducibility of
a diet history questionnaire in Spain. I. Foods. Int J Epidemiol
1997;26(suppl 1):S91–9.
Feskanich D, Rimm EB, Giovannucci EL, et al. Reproducibility and validity of food intake measurements from a semiquantitative food frequency questionnaire. J Am Diet Assoc
1993;93:790–6.
Salvini S, Hunter DJ, Sampson L, et al. Food-based validation
of a dietary questionnaire: the effects of week-to-week variation in food consumption. Int J Epidemiol 1989;18:858–67.
Subar AF, Kipnis V, Troiano RP, et al. Using intake biomarkers to evaluate the extent of dietary misreporting in a large
sample of adults: The OPEN Study. Am J Epidemiol 2003;
158:1–13.
Bohlscheid-Thomas S, Hoting I, Boeing H, et al. Reproducibility and relative validity of food group intake in a food frequency questionnaire developed for the German part of the
EPIC project. European Prospective Investigation into Cancer
and Nutrition. Int J Epidemiol 1997;26(suppl 1):S59–70.
van Liere MJ, Lucas F, Clavel F, et al. Relative validity and
reproducibility of a French dietary history questionnaire. Int J
Epidemiol 1997;26(suppl 1):S128–36.
288 Millen et al.
24. Bonifacj C, Gerber M, Scali J, et al. Comparison of dietary
assessment methods in a southern French population: use of
weighed records, estimated-diet records and a food-frequency
questionnaire. Eur J Clin Nutr 1997;51:217–31.
25. Mannisto S, Virtanen M, Mikkonen T, et al. Reproducibility
and validity of a food frequency questionnaire in a casecontrol study on breast cancer. J Clin Epidemiol 1996;49:
401–9.
26. Elmstahl S, Riboli E, Lindgarde F, et al. The Malmö Food
Study: the relative validity of a modified diet history method
and an extensive food frequency questionnaire for measuring
food intake. Eur J Clin Nutr 1996;50:143–51.
27. Goldbohm RA, van den Brandt PA, Brants HA, et al. Validation of a dietary questionnaire used in a large-scale prospective
cohort study on diet and cancer. Eur J Clin Nutr 1994;48:
253–65.
28. Pietinen P, Hartman AM, Haapa E, et al. Reproducibility and
validity of dietary assessment instruments. II. A qualitative
food frequency questionnaire. Am J Epidemiol 1988;128:
667–76.
29. Sauvaget C, Allen N, Hayashi M, et al. Validation of a food
frequency questionnaire in the Hiroshima/Nagasaki Life Span
Study. J Epidemiol 2002;12:394–401.
30. Forman MR, Zhang J, Nebeling L, et al. Relative validity of
a food frequency questionnaire among tin miners in China:
1992/93 and 1995/96 diet validation studies. Public Health
Nutr 1999;2:301–15.
31. Ogawa K, Tsubono Y, Nishino Y, et al. Validation of a foodfrequency questionnaire for cohort studies in rural Japan.
Public Health Nutr 2003;6:147–57.
32. Ishihara J, Sobue T, Yamamoto S, et al. Validity and reproducibility of a self-administered food frequency questionnaire
in the JPHC Study Cohort II: study design, participant profile
and results in comparison with Cohort I. J Epidemiol 2003;
13(suppl):S134–47.
33. Sasaki S, Ishihara J, Tsugane S. Reproducibility of a selfadministered food frequency questionnaire used in the 5-year
follow-up survey of the JPHC Study Cohort I to assess food
and nutrient intake. J Epidemiol 2003;13(suppl):S115–24.
34. Lee KY, Uchida K, Shirota T, et al. Validity of a self-administered food frequency questionnaire against 7-day dietary
records in four seasons. J Nutr Sci Vitaminol (Tokyo) 2002;
48:467–76.
35. Tokudome S, Imaeda N, Tokudome Y, et al. Relative validity
of a semi-quantitative food frequency questionnaire versus 28
day weighed diet records in Japanese female dietitians. Eur
J Clin Nutr 2001;55:735–42.
36. Wakai K, Egami I, Kato K, et al. A simple food frequency
questionnaire for Japanese diet—Part I. Development of
the questionnaire, and reproducibility and validity for food
groups. J Epidemiol 1999;9:216–26.
37. Smith-Warner SA, Elmer PJ, Fosdick L, et al. Reliability and
comparability of three dietary assessment methods for estimating fruit and vegetable intakes. Epidemiology 1997;8:196–201.
38. Flagg EW, Coates RJ, Calle EE, et al. Validation of the
American Cancer Society Cancer Prevention Study II Nutrition Survey cohort food frequency questionnaire. Epidemiology 2000;11:462–8.
39. Parr CL, Barikmo I, Torheim LE, et al. Validation of the second version of a quantitative food-frequency questionnaire for
use in western Mali. Public Health Nutr 2002;5:769–81.
40. Torheim LE, Barikmo I, Hatloy A, et al. Validation of a quantitative food-frequency questionnaire for use in western Mali.
Public Health Nutr 2001;4:1267–77.
41. Rodriguez MM, Mendez H, Torun B, et al. Validation of a
semi-quantitative food-frequency questionnaire for use among
adults in Guatemala. Public Health Nutr 2002;5:691–9.
42. Kipnis V, Subar AF, Midthune D, et al. Structure of dietary
measurement error: results of the OPEN biomarker study. Am
J Epidemiol 2003;158:14–21.
Am J Epidemiol 2006;163:279–288