Nutrition Elizabeth
epidemiology:
how do we know what they ate?3
Barreti-Connor
ABSTRACT It is generally believed but difficult to prove that diet plays a role in the risk of various diseases. This paper reviews strengths and deficiencies of select diet-assessment methods
used
in epidemiologic
studies
with
particular
information
on recent
intake
but does
reference
not represent
usual intake. Food frequency questionnaires provide better estimates of usual diet but are less quantitative and subject to problems ofrecall and seasonality. No method is universally the best. Lack of an expected diet-disease association may reflect exposure misclassification, inadequate statistical power, or limited
range
ofthe
nutrients
assessment methods, studies fracture have had surprisingly
studied.
Given
the differences
of dietary calcium similar results.
in diet-
and osteoporotic Am J Clin Nutr
199 1;54:182S-7S.
KEY WORDS osteoporosis
Dietary
assessment,
reproducibility,
validity,
Introduction The association creasing concern
of behavior with subsequent health is of into both the scientific and the lay communities.
No health-related
behavior
has generated
more
universal
atten-
tion than diet because everyone eats and has, therefore, a vested interest. Analytic studies ofdiet and current or future disease in human beings are the core of nutritional epidemiology. In case-control studies, people with and without disease are compared with regard to what they now say was their previous or usual diet. In cohort
studies,
at baseline,
diet and
is determined
is compared
in a population with
diets
of people
free of disease who
later
do
or do not develop the disease. Minor variations on these two themes exist but all are dependent on the assessment of diet in individuals.
In the l980s a great deal was written about the merits and demerits of specific methods for ascertaining diet for epidemiologic studies. Didactic and often unsubstantiated claims were made about the relative value of different methods. Grants were awarded or denied on the basis of the prejudices of reviewers with regard to the proposed diet-assessment tool. This paper reviews selected diet-assessment methods and their strengths and deficiencies, with particular reference to their use in the study of diet and osteoporosis. 1 82S
Am iC/in
Nuir
diet-assessment
methods
Clues to diet-disease associations have often been derived and geographic comparisons of food-disappearance
from
cross-cultural
to their use in the study of osteoporosis. Direct observation or weighed food records are useful primarily as validation for less intrusive methods. Complete food history by interview or food diary (by self report) is expensive and time consuming. A 24-h diet recall obtained by a trained dietitian can provide accurate, quantitative
Individual
199 l;54:182S-7S.
Downloaded from https://academic.oup.com/ajcn/article-abstract/54/1/182S/4690957 by guest on 06 March 2018
rates
vs
mortality.
Associations
observed
in
such
ecological
studies dividual
must be confirmed by studies of individual diet vs indisease. The five main methods of diet assessment in individuals used for epidemiologic research are summarized in Table 1. All have serious flaws with regard to cost, representativeness,
quality
behavior
change.
Direct
of quantitive
observation
estimates,
and weighed-food
and/or
study-induced
records
Direct observation or in-home weighed-food records are the only methods of diet assessment applicable to free-living populations that assure the quantitative and qualitative validity of all nutrients
consumed.
Both
are
usually
too
expensive
(in the
context of the sample size required) for epidemiologic studies ofdiet and disease. If more than 1 d is necessary to assess usual diet, direct observation ofdiet is less likely to be an option than are
weighed-food
records.
In-home
weighed-food
records
may
work particularly well in countries where participants are accustomed to recipe units given by weight rather than by measure. When direct observation is used as a “gold standard” to validate reported recent intake, neither the amount nor the direction ofthe error is predictable. Both over- and underestimation have been reported. Two studies comparing direct observation with weighed-food analysis are shown in Table 2 (1, 2). It can be seen that, contrary to popular expectation, overweight women may overestimate their intake (2). Although
direct
observation
or weighed-food
records
accu-
rately represent current intake, they may not reflect usual intake. The need to weigh and record intake may lead to a reducedcalorie or more monotonous diet. Further, when studied in the home, subjects know their diet is being observed, directly or indirectly,
by record
review.
Being
observed
is liable
to induce
behavior change. Nearly everyone in Westernized cultures has some knowledge of how they should eat and could be tempted to have a better diet to impress the observer. Few of us would allow
anyone
ence
to monitor
cloaked
in the
our feeding
aura
and
of hotdogs
authority
of medical
and
to our children.
chips
sci-
From the Department of Community and Family Medicine, University of California, San Diego, La Jolla, CA. 2SupportedbyNlH/NIA 1 R37AG07181. 3Address reprint requests to E Barrett-Connor, Department of Community and Family Medicine, University of California, San Diego, M-007, La Jolla, CA 92093-0607. Printed
in USA.
© 1991 American
Society
for Clinical
Nutrition
NUTRITIONAL TABLE 1 Individual diet-assessment
methods
for epidemiologic
Method
*
research
Expensive
Observation Diet history Diet diary or record 24-h diet recall Food frequency questionnaire
Behavior
change
Quantitative
Representative
Yes, very Yes, very Yes Yes
Yes No Yes No
Yes Yes Semiquantitative Yes
No
No
Semiquantitative
*
? Yes ? No Yes
Usual.
Food
history
The
“next-best-thing” in epidemiologic diet assessment is the food history, usually practiced as a refinement of the method described by Burke (3) in the 1940s. The quality of food-history data was considered to be one of the reasons why an association between dietary fat and cholesterol and coronary heart disease could be shown within a population in the Chicago Western Electric study, where most other within-population studies using less extensive dietary data have failed to show an probably
association
This usual
(4).
approach, and
foods,
interview
which includes data on food
by a specially
trained
on the quality of the interviewer plication. Concordance of other
TABLE 2 Mean percentage intake
error in reported
a 24-h
diet recall,
preparation,
a history
requires
of
a 1-2
estimate
of quantity
vs observed
Percentage
error
group*
Combined
main dishes
-29.2
Dairy products Vegetables Fruits Salads Cereals Breads Starches Soups Desserts
-5.7 -22.5 -2.4 -53.0 +11.9
Foodt Cottage cheese Roast turkey Green beans Boiled ham Cooked spaghetti
different
by 86 healthy
As shown
from
estimates
derived
from
of which resembled is used as the gold
methods
validated
records Food
not
by observation
record
women
weighed
diet
d, can
theo-
in that
when
the
food
record
is most
accurate
for 3-7
participants
are
trained
by dietitians in how to estimate quantity and record intakes. As shown in Table 4, intake from a 7-d food record does not always parallel the intake based on a food history (6). One problem with the food record is compliance. Writing down everything soon gets tedious and the characteristics (dietary and otherwise) of people who will do so are apt to differ from those ofpeople who will not. One could argue that only the most compulsive would actually complete a 7-d diary and that conwith
another
diet
method
might
be much
greater
in
such individuals than that for a total study population. Another problem is a training effect, a change in food intake due to participation. Recording all food consumed for > 1 or 2 consecutive days is a well-known behavior-modification method to reduce intake. Snacks and condiments, high in calories, fat, and sugar, mayjust not be worth the trouble when keeping a 7-cl diet record. Nevertheless, the food record is often used as the gold standard for validating other methods when neither a food history nor observation is possible. diet recall
quantitatively
TABLE Average method
3 daily joule intake
The
24-h
diet
recall
was designed (5-8).
When
of 400 women
Method
to assess correctly
recent
based on diet-assessment
Energy
24-h recall
6760
Current
9084
diet history
Past diet history 4-drecord *
Reference
5.
nutrient
performed
kJ/d
(1).
Downloaded from https://academic.oup.com/ajcn/article-abstract/54/1/182S/4690957 by guest on 06 March 2018
or multiple
the other
This is a false savings,
intake
(2).
or a 4 d food
or diary
+6.0
by 30 overweight
recall
each other. Nevertheless, standard against which
A food record or diary, usually obtained retically avoid the costs ofan interviewer.
+260.0 + 10.0 + 120.0 +7.5 postpartum
a 24-h
(5), was
are measured.
+85.0 +70.0
women
from
record, both food history
-17.8 -51.8 -30.8
+95.0 +25.0
3 in a study by Morgan past and current food history
in Table
estimated
Twenty-four-hour
+23.3
Potato chips Blueberries Slice of bread Orange juice
t As reported
variable.
intake
-19.9
Cola drink
As reported
is quite
caloric
cordance
%
Food
h
nutritionist. It is very dependent and is too costly for wide apmethods with the diet history
Food group*
*
l83S
EPIDEMIOLOGY
9561 7451
by
a
184S
BARRETT-CONNOR
TABLE 4 Comparison of percentage 24-hour recall and between
The degree differences between Burke history 24-hour recall and 74 record*
and
the
study
used in two
Massachusetts (n = 28)
history and 24-h recall (%) Energy Protein Calcium Phosphorus Iron 7-Day record and 24-h recall (%) Energy Protein Calcium Phosphorus Iron
New York (n = 51)
Rhode
(n
Island 87)
=
Burke
Also,
+23.3 +20.1 +21.5 +20.9 +17.3
+9.7 -7.2 +0.1 -3.7 -10.9
+2.4 +1.1 +13.1 +3.1 +7.7
+6.5 +1.9 +2.5 + 1.3 +1.7
Reference
to assess this method interviewerare
term memory required and the quantitative intake. As noted by Balogh et al (7) it should
estimates be almost
that
most
ate
than
they
have
people
The
diet recall participant, recall
that
memory
that
major
details
without
eat.
shows
may be more
people
with
method
remarkable
and
diet
concordance
history
in a third.
reproducible
than
Bernardo study, adults continued
those who to drink
who
tend
drink
unpublished
milk
observations,
to do
1988).
questionnaires
representative food
picture
ofchronic
frequency
become
of diet disease
would
than
questionnaire
to obtain of usual
questionnaires items that were
be expected
would
to
a single
day’s
many
years
was devised
a self-administered, inexpensive, intake (13-16). Initially, food fre-
were very short, with a limited selected to test a single hypothesis.
considerably
longer
for use in cohort
number They
studies,
where
disadvantage
change
diet could
of the 24-h
recall
of a psybe asked.
is the inability
of
lack of reprethat a single
is worthless for epidemiologic research recalls are certainly not recommended
when
forms
entry,
eliminating
are
As noted
above,
typically
asked
years,
is more
(8-12). to de-
Costs
designed
(such
greatly
reduced
to be scanned
the need
the other for
are
because
the
directly
to computer
for manual coding and keypunching. major advantage is that food frequency,
past
year,
representative
sometimes
of usual
be expected
by increasing as placing
the
for
intake
more
than
to be. This tends chances
in quintiles
address
remote
a short
1- or
to reduce
of correctly
by usual
ranking
intake),
more likely to reflect a diet-disease association (17, The food frequency method is not without problems, The order ofthe listing is arbitrary but may influence Completion questionnaire
the
by a self-administered questionnaire, interviewers. Costs are further reduced
or recall could
tionnaire
be excluded.)
diet. This to conclude
representativeness.
data are usually obtained without need for trained
subjects
by notifying
would not
associations
misclassification
and qualor remote
is more
diet-disease
3-d record
It is assumed,
be improved
of yesterday’s
behavior
better
only once from effect. (Some
warning
would
a single day’s intake to describe usual sentativeness has led many investigators 24-h diet recall Twenty-four-hour
usually
short-
of food an axiom
yesterday
is usually obtained there is no training
a diet
protocol,
they
they
the
is both quantitatively is the remembered usual
that
this
what
what
test and
participant
Under
remember
remember
the
argued
the
can
that the information more accurate than
chological
and
the
may be sought. Currently popular food frequency questionnaires include well over 100 food items and may be self- or interviewer-administered. The major advantages of the food frequency method are cost
than is the traditional diet history. biggest advantages of the 24-h recall
diet. Because an unprepared
older
predictor
The
have
foods
with
recall
remarkable
(E Barrett-Connor,
quency offood
and
can
and
varies
and
by 24-h
in the Rancho intake as young
ago in an attempt and rapid estimate
+0.9 +4.3 +11.9 +8.2 +3.4
trained dietitian using food models and containers quantity, the interview takes 30-60 mm. Therefore, is relatively expensive, although less so, and less
therefore, itatively
of some
A more
6.
dependent, The two
the intake
be a better
many
*
areas
Foodfrequency
diet.
intake
geographic
as they grew
of diet
4, for example,
dietary
others. For example, had moderate milk
milk +21.1 +23.8 +20.6 +23.8 +32.2
Table
in the
so daily
or representativeness
monotony
for validation.
differences
Difference
Nutrient
of reproducibility
populations’
which
is
18). however.
responses.
of even a short nonquantitative self-administered requires a certain level of literacy. If the quesis very
short,
the
limited
only one or two specific
number
hypotheses,
of food which
items
can
is not efficient
study. The need to list specific foods also tends to make the questionnaire fairly culture specific. Food frequency questionnaires for Japanese-American men include mochi-gashi, duri-manju, and monaka whereas the Oxford (England) quesfor a cohort
tionnaire
asks
about
spotted
dog.
Because
there
is a limit
to the
tect actual deficiency states in individuals, because most vitamins and trace minerals can vary from day to day and still be adequate overall. Similarly, a 24-h recall is quite misleading ifone wishes to examine a particular food, such as fish, that is not eaten daily. Estimates aging. 24-h
of reliability
Both recalls
Beaton were
analyzed to evaluate
habitual
major
nutrients
are
Liu et al (9) found
to reliably
place
subjects
also
discour-
that
multiple
in the
for some nutrients; for calcium the d. VanStaveren et al (12) used tissue
for the ratio the extent
of the number between 3 and the
et al (8) and required
quintile of intake was 1 7-19 recall
for
ofpolyunsaturated ofdietary-fat
of 24-h recalls 7 recalls were
fat intake
to saturated misclassification
same
number biopsies fatty
TABLE
5 Estimates of probabilities observations used
of misclassification
Number of dietary measurements averaged (n = 57)
Adjacent
acids
1 3 7
as a function
per subject. They concluded that necessary to adequately estimate
of an individual
(Table
5).
Downloaded from https://academic.oup.com/ajcn/article-abstract/54/1/182S/4690957 by guest on 06 March 2018
*
Reference
12.
category
for specified
number
Opposite
of
category
(p)
(q)
0.382 0.316 0.237
0.184 0.132 0.132
NUTRITIONAL
EPIDEMIOLOGY
30% clearly
TABLE 6 Intraclass correlation coefficients, measuring within individual agreement of daily nutrient estimates by two different methods in 40 young women * Comparison nutrient
14 vs 74 record
34
onstrate
*
Reference
0.45 0.42f 0.46 0.58 0.63
0.79 0.76t 0.74 0.90f 0.89f
0.09 0.02 0.04 0.19 0.24
t FFQ, food frequency
questionnaire.
number
offoods
that
may be missed. space to include completeness
can
be listed,
some
Although some diet items not
of such
surprising
dietary
of the questionnaires included in the food
responses
habits
provide list, the
is unknown.
Self-administered food frequency methods are at best semiquantitative because only fixed or subjective definitions of small, medium, or large portions are possible. The combination of missing
foods
and
semiquantitative
methods
limits
for
selected
vitamins
and
antioxidants,
there
is no
easy
way to confirm the usual food intake of most nutrients. Concordance of results based on small groups of more extensively studied subjects raises questions about the representativeness of such compliant individuals(15). Comparison with other methods provides divergent results, and does not indicate which of these results
are correct.
As shown
in Table
6, from
a study
by Stuff
et al (19), correlation coefficients with a 7-d diet record were better for 1- or 3-d records (obtained from the same 7-d diet record) than for a food frequency questionnaire. It is important to note that reproducibility, also called reliability,
is not
the
same
as validity.
There
is no question
that
the
frequency questionnaires is higher than for the 24-h recall, but a part ofthe improved reliability is an artifact. Because reproducibility is in part a funcrepresentativeness
of intake
tion of the precision recalls increase with
example, reported to vary less from Similarly, tion size
a more
by using
food
of the data, differences decreasing simplicity
between repeated of the question. For
consumption of green vegetables day to day than is consumption
any instrument has less variability
quantitative
calcium
and
sorption
and excretion
and
the
protein,
Other
kinds
that affords and more
is expected of broccoli.
tabank
assessment
and the coders.
Downloaded from https://academic.oup.com/ajcn/article-abstract/54/1/182S/4690957 by guest on 06 March 2018
foods,
and
metabolism.
supports
the study there
are
for major
drink drinks,
resulting
States
vitamin
when
single
etc.
The
the
Even
it is difficult
D, because
is fortified
the
with
come
have
striking
a good
as for
differences
in ab-
in the
vehicle
diet.
and discordant, nutrients
are
is high
7) (20). Suppose that bones. Is it the calcium that
prevent
to study major
Vitamin
bio-
nutrient,
on the calcium
concordant (Table
Nutrients
putative
oxalate,
correlation
raise the question may
that are dependent
phosphate,
In the United this
of itemized
which ofa
of associations,
dietary
source
in all
people or the
osteoporosis?
calcium
separate
of calcium
D. Conversely,
is milk
people
very little milk may drink considerably more or alcohol. This could lead to the mistaken
who
coffee, soft impression
that one of these beverages increases the risk of osteoporosis, when, in fact, the critical variable is the low milk intake. these
and
covariances.
terpretation
confounding
are
Sophisticated
are lacking.
sophisticated
solutions
for
For example,
tribution, adjusting
data. than
or whether one makes any for the other. This technique
When one item is more another, it may assume
Because
so many
nutrients
for
and
in-
the relative
by putting makes the
contribution is only
accurately an artificial
terms
analysis
sometimes
contribution of two nutrients is assessed in a multivariate model to see which one
them larger
both con-
to risk after as good as the
recalled or quantitated priority.
can be derived
(the simplest
outputs
usually give at least protein, simple and complex carbohydrate, saturated and unsaturated fatty acids, and several vitamins and minerals)
and
because
the complexity
diverse associations biologically with multiple testing. Unless
gested based
by other
data
in animals
on an a priori
association
will
out
ofbiologic
plausible,
there
by chance.
real association should be sought, relationship between the amount
processes
there
the diet-disease or humans,
hypothesis,
fall
is also
association or in other
is always
the
Additional
eg, showing of a nutrient
makes
a problem
risk
evidence
is sugwords, that
an
for
a
a dose-response and the risk of
disease.
The converse be missed,
risk, that a clinically
probably
TABLE 7 Intercorrelations Lipid Research
is even
Nutrient
*
greater.
important This
association
is because
of major nutrients for men aged 20-59 Clinics’ 24-h diet recall* pairs
the qualitative
y,
r
Protein X carbohydrate Protein X fat Protein X alcohol Carbohydrate X fat Carbohydrate X alcohol
0.48 0.72 0.05 0.58 0.05
Fat
0.02
X
alcohol
as good
Reports
other
adiposity)
few or no options for porreproducibility than does
as the food-composition daof intracoder variation of up to
should
calories
and
to dem-
Computerized
is an advantage
be studied.
with
osteoporosis,
ignored.
from
disease.
should
better
recall.
is only
and
but extreme diet patterns who love ice cream have
Discussion Diet
mechanism
Multicollinearity
the accuracy
of the estimated caloric intake. Because calories may be an independent risk factor and are often used in the analysis to correct for individual variation and for exercise, the lack of accurately assessed calories is not inconsequential. Perhaps the most significant problem with a food frequency questionnaire is uncertain validity. For many nutrients of interest, such as calcium, which is under homeostatic control, no biochemical assay of serum is useful. Aside from biochemical assays
are eaten
(or
for the ability
and
databanks
or foods
on absorption
often
P 0.05) association ofdietary calcium with hip fracture in a recent population-based prospective study in the United Kingdom (P
(27) and
0.008)
=
(28)
the significant
Kingdom
study
independent
in a US study,
risk (for the lowest
tertile
had
ofcalcium
confidence
inverse
association
in Table
shown
intake)
limits
10. The relative ofO.7 in the United
ofO.l-3.9,
which
include
the statistically significant lower risk ofO.4 for the highest tertile ofcalcium intake in Rancho Bernardo. Exactly comparable relative risks cannot be calculated with the data provided in the United
Kingdom
study,
but
these
data
could
be interpreted
as
showing rather similar results, given the study differences. In conclusion, the answer to the question, “How do we know what they ate?” is that we don’t, exactly, and that no widely applicable method is a priori better at making an estimate. Recent excellent, concise reviews and entire books dealing with the methods and limitations of nutritional assessment (see refs 2932) demonstrate
and Trulson
TABLE
that
little
(33) wrote,
has changed
“In general
since
one must
when
conclude
intake and risk of hip fracture
Britain* Calcium
quintile
1 (low) 2 3 4 5 (high) Mean calcium intake of control subjects (mg) Interquartilerange *
1960,
Young
that,
on
9
Case-control studies of calcium Britain and Hong Kong
of the
nutrients of interest within the study population is not known until the diet data are collected. For example, a paper reporting no association of dietary fat with breast cancer in > 85 000 women (23) was criticized because nearly all of the women in this very large cohort were in a narrow range of high fat intake. On the basis of ecologic studies, within this range only a small increment in relative risk would have been expected (24). Consistency of results is usually an important criterion for causality in epidemiologic studies ofassociations. In nutritional epidemiology, sociation
one
in Table 9. Study subjects
not consistent and stepwise, it should be noted that the dietary calcium in Hong Kong was ascertained entirely from a nineitem food frequency questionnaire whereas that from Britain used a six-item questionnaire. (A cross-cultural comparison to determine if the Chinese had higher fracture rates in the face of such limited dietary calcium cannot be obtained with the casecontrol design, which ascertains neither incidence nor prevalence of disease.)
be obtained from study to study, particularly in case-control studies where exposure misclassification may not be randomly distributed and where other biases may exist. Persons with diseases prevented or modified by diet often assume a better diet after diagnosis, but recall of the remote (prediagnosis) diet is colored by current diet (22). Hospital control subjects may have changed their diets as a result of the condition for which they were hospitalized. There is frequent failure to obtain diet data for the age or interval of interest, when this is known. In the cohort model, the diet is ascertained before disease onset but the method rarely allows assessment of interim diet or other behavior changes that may be relevant. For studies of osteoporosis, diet in young adult life may be the critical period but collecting
(known
average
the other
8) (21). means
studies,
are shown
ethnicity
but
lead to misclassifimisclassification
tends to bias all associations towards the null. any observed association usually underestimates
case-control
Britain(26),
more protective than was diet. In both studies, quartile of calcium intake were least likely
21.
and quantitative errors cation of the exposure.
true
recent and by number off ood records vs true odds ratio*
Reference
26. n
=
in
Hong Kongt
Men
Women
Men
Women
6.2f 5.8 3.3 6.2 1.0
1.2 1.4 1.1 1.2 1.0
2.1 1.4 1.7 1.5 1.0
1.9 1.9 1.1 1.2 1.0
651 467-799
177 75-226
168 75-176
843 560-1042
300 cases and 600 control
subjects.
t Reference 25. n 400 cases and 800 control subjects. f Odds ratio based on setting the highest calcium quintile =
erence
risk at 1.0.
as the ref-
NUTRITIONAL TABLE
EPIDEMIOLOGY
10
Two prospective
studies of dietary calcium
and the risk of hip
fracture* Rancho
Bernardo, Diet method Design Number in study
Number
*
References
t Relative Relative
United
24-h recall Cohort
33
42
50-79