Download Within- and Between-Cohort Variation in

American Journal of Epidemiology
Copyright © 2004 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved
Vol. 160, No. 8
Printed in U.S.A.
DOI: 10.1093/aje/kwh280
Within- and Between-Cohort Variation in Measured Macronutrient Intakes, Taking
Account of Measurement Errors, in the European Prospective Investigation into
Cancer and Nutrition Study
Pietro Ferrari1, Rudolf Kaaks2, Michael T. Fahey1, Nadia Slimani1, Nicholas E. Day3, Guillem
Pera4, Hendriek C. Boshuizen5, Andrew Roddam6, Heiner Boeing7, Gabriele Nagel8, Anne
Thiebaut9, Philippos Orfanos10, Vittorio Krogh11, Tonje Braaten12, and Elio Riboli1
1
Unit of Nutrition and Cancer, International Agency for
Research on Cancer, Lyon, France.
2 Hormones and Cancer Group, International Agency for
Research on Cancer, Lyon, France.
3 Institute of Public Health, University of Cambridge,
Strangeways Research Laboratory, Cambridge, United
Kingdom.
4 SERC, Institut Català d’Oncologia, Barcelona, Spain.
5 Centre for Information Technology and Methodology,
National Institute for Public Health and the Environment,
Bilthoven, the Netherlands.
6 Cancer Research UK Epidemiology Unit, Oxford University,
Oxford, United Kingdom.
7 Division of Clinical Epidemiology, German Cancer
Research Centre, Heidelberg, Germany.
8 German Institute of Human Nutrition, Potsdam-Rehbrücke,
Germany.
9 E3N-EPIC, INSERM Z521, Institut Gustave Roussy,
Villejuif, France.
10 Department of Hygiene and Epidemiology, University of
Athens Medical School, Athens, Greece.
11 Epidemiology Unit, Istituto Nazionale dei Tumori, Milan, Italy.
12 Institute of Community Medicine, University of Tromsø,
Tromsø, Norway.
Received for publication October 2, 2003; accepted for publication April 8, 2004.
Multicenter epidemiologic studies provide a unique opportunity to evaluate the association between exposure and
disease at the individual and the aggregate levels. The two components can eventually be pooled to corroborate each
other, using weights proportional to the intraclass correlation coefficient (ICC), which expresses the amount of betweencohort variability in the exposure variable compared with the total. The greater the ICC, the more the overall estimate will
reflect the between-cohort component. Dietary measurements are affected by measurement errors, particularly within a
cohort. In 1992–2000, the variability of macronutrient intake distribution before and after calibration for measurement
error in the European Prospective Investigation into Cancer and Nutrition was evaluated. A two-level, random-effects
model was used. Evaluation of macronutrient densities revealed that energy has a considerable effect on the calibration
model, leading to ICC values larger than those for the absolute intakes. Given the shrinkage of the within-center
variability, a sizable increase in the ICC was observed for protein in men and women (0.48 and 0.54, respectively) and
carbohydrates in men (0.41). Results suggest that the effect of calibration on macronutrient intake variability is greater
for the within-cohort component, thus increasing the relative importance of the between-cohort component. After
calibration, the two components had a similar weight. This observation has important implications for the analysis of
multicenter studies because the between-cohort component provides a large part of the overall heterogeneity.
calibration; diet; epidemiologic methods; multicenter studies
Abbreviations: EPIC, European Prospective Investigation into Cancer and Nutrition; ICC, intraclass correlation coefficient.
Ecologic correlation studies (i.e., those based on international comparisons) have shown strong associations between
cancer risk and average per capita consumption of specific
food groups and nutrients (1). This initial observation—
complemented by data from time-trend studies and migrant
studies, which confirm the importance of environmental and
lifestyle factors in determining cancer risk—has led to a vast
research effort to identify diet-cancer relations at the level of
single individuals (2). After hundreds of case-control
studies, this effort has gradually shifted toward those of a
Correspondence to Dr. Pietro Ferrari, IARC, 150, cours Albert-Thomas, 69372 Lyon Cedex 08, France (e-mail: [email protected]).
814
Am J Epidemiol 2004;160:814–822
Within- and Between-Cohort Macronutrient Variation 815
prospective cohort design because of concern that casecontrol studies may be biased by selection of improper
controls or by differential errors between cases and controls
in their estimates of habitual diet.
Compared with ecologic studies, investigations that use
individual-level information on dietary exposures and disease
outcome have the advantage of allowing more complex analyses, with adjustments for potential confounding factors and
evaluations of possible interactions between multiple exposures. However, a possible limitation is that random errors in
assessing individuals’ habitual dietary intake can be so large
as to virtually mask potentially existing associations with
disease risk (3, 4). The impact of random measurement error
on estimated measures of diet-disease associations depends
on the variance of errors relative to the between-person variance of true intake levels (5).
For practical and logistic reasons, the increase in statistical
power that may be obtained by increasing study size or
reducing the average magnitude of random dietary assessment errors will generally be limited. However, an additional
possibility for improving the statistical power is to increase
the true overall heterogeneity of the dietary exposures
studied, combining data from multiple studies conducted in
populations with different dietary habits. This rationale was
used to set up the European Prospective Investigation into
Cancer and Nutrition (EPIC), a multicenter cohort study on
diet and cancer conducted in 28 regional centers nested in 10
Western European countries with widely varying dietary
practices and cancer risks (6).
In a multicohort study, the overall association between diet
and cancer risk can be broken down into 1) within-cohort
relations, which reflect the associations at the individual
level in each of the cohorts; and 2) a between-cohort relation,
which captures the association between exposure and
disease risk at the aggregate level. For proper comparison, it
is crucial that corrections be made for any between-cohort
differences that may exist in systematic exposure measurement error as well as for attenuation biases in relative risk
estimates within cohorts (7, 8). For this purpose, the EPIC
project combined two types of dietary intake assessment: 1)
for all 520,000 cohort participants, a food frequency questionnaire or modified dietary history questionnaire to assess
individuals’ habitual, long-term intake levels; and 2) a
single, highly standardized, 24-hour recall of actual food
consumption during the previous day for a large representative subsample (about 37,000 subjects) of the entire EPIC
cohort (9).
To adjust for between-cohort differences in systematic
over- or underestimation in dietary intake measurements,
and to correct for attenuation bias in relative risk estimates, a
regression calibration approach can be used (7, 8, 10). With
this approach, an (approximate) estimate can be obtained of
within-cohort variation in true intake levels predicted by the
individual-level questionnaire measurements (11). Furthermore, between-cohort variation in mean intake levels can be
estimated from the population mean values of the 24-hour
recalls.
In this paper, we propose the use of a multilevel calibration
model that takes into account the two levels of evidence,
within-cohort and between-cohort, and uses information at
Am J Epidemiol 2004;160:814–822
the individual and aggregate levels. Moreover, we report
estimates of within- and between-cohort variation in intakes
of total energy and macronutrients before and after regression calibration. We show that, after considering withincohort attenuation effects, between-cohort differences
provide a considerable amount of variation in intake levels.
This observation may have important implications for the
analysis of the EPIC study as well as for the design of other
prospective studies.
MATERIALS AND METHODS
EPIC data collection
EPIC is a multicenter cohort study aimed at investigating
the association between diet and cancer across Western
Europe. Individual habitual dietary intake was assessed by
using different questionnaires developed and administered
independently in each country. To express individual dietary
intakes according to the same reference scale and to correct
the diet-disease associations for attenuation due to measurement errors, a single 24-hour dietary recall was collected by
using a software program (EPIC-SOFT) specially designed
to standardize the recall interviews (12). The 24-hour dietary
recalls are used as reference measurements and were
collected from a stratified sample of about 37,000 men and
women in the EPIC cohort (13). For the present analyses,
study subjects were recruited in 1992–2000.
Calibration model
We assumed the following measurement error model for
the baseline dietary assessment measurements (Q) and for
the 24-hour dietary recall measurement (R) in relation to the
unknown true intake (T):
Q = αQ + βQT + εQ
(1a)
R = T + εR,
(1b)
where E(εR |T) = 0, E(εQ|T) = 0, var(εR) = σεR2, and var(εQ) =
σεQ2. The association shown in model 1b—for reference
measurement R—assumes that errors are strictly random and
that variation around individuals’ unknown true intake is
totally due to within-person random variability or to random
measurement errors in reporting individuals’ diet (i.e.,
cov(εR, T) = 0). For the questionnaire measurements Q, the
coefficients αQ and βQ in model 1a express constant and
proportional scaling biases (14), whereas the residual term
εQ models the random part of measurement errors in Q and is
assumed to be uncorrelated with true intake level T (15).
To extend the calibration methodology introduced by
Rosner et al. (10), a very general form of the EPIC calibration model can be defined as
R ij = α j + λ Wj ( Q ij – Q j ) + λB Q j
+ γ W ( Z ij – Z j ) + γ B Z j + ε ij ,
where
(2)
816 Ferrari et al.
αj = α + u0j, λWj = λW + u1j
⎛ 0
⎛ u 0j ⎞
⎜
⎟ ∼N ⎜
⎝ 0
⎝ u 1j ⎠
⎞ ⎛⎜ σ 2B σ Bλ
⎟, ⎜
2
⎠ ⎝ σ
Bλ σ λ
⎞
⎟ and ε ∼ N ( 0, σ 2 ),
ij
ε
⎟
⎠
where j = 1 … G indexes the group variable, be it the recruitment region, center, or country; i = 1 … nj are individual
measurements within group j; and Q j are the group means.
The coefficients γ’s model the effect of those variables (Z)
included as confounding factors, for example, age, body
mass index, physical activity, and smoking status. Models 1a
and 1b and model 2 assume independence between random
errors in R and Q, which is unlikely to occur in self-reported
measurements as a consequence of study subjects’ tendency
to consistently under- or overestimate dietary intake (14, 16–
19). However, in the absence of a third objective measurement of intake, such as a biomarker, the assumption of independence between the errors in R and Q cannot be assessed.
One single replicate of R allows a correct estimate of the
attenuation factors (11) to be obtained in model 2. This
model has fixed-effect parameters (α, λ’s, and γ’s) and three
random effects, u0j, u1j, and εij, which give rise to four variance components reflecting, respectively, variation between
groups (σB2), variability in the calibration coefficients across
groups (σλ2), their covariance (σBλ), and within-group variation among study subjects (σε2). In model 2, λWj captures the
within-group component of the calibration model, while λB
reflects the between-group component. Multilevel models
provide a unique opportunity to simultaneously model
within- and between-group effects (20–22).
Use of a random-effects model allows the heterogeneity of
attenuation factors to be estimated parsimoniously. It also
enables evaluation of the agreement between R and Q
measurements at the aggregate level in model 2 (20, 23).
Within center, the calibration factor λWj can be used to
correct the observed (“naïve”) parameter that estimates the
association between dietary exposure and the outcome of
interest, for example, cancer incidence, by
β̂ True ( j ) = β̂ Obs ( j ) ⁄ λ̂ Wj ,
where β̂Obs ( j ) is the naïve estimate that expresses the increase
in risk for a unit increase in the exposure and is obtained by
relating cancer incidence to the Q measurements, following a
“disease model.” A mathematically equivalent way to correct
for measurement error in the diet/disease association is to estimate E(T|Q) as the predicted values of the calibration model
and use these predicted values as the imputed exposures in the
disease model.
Variability of exposure in the EPIC study
To evaluate the effect of calibration in a multicenter
setting, we focused on the distribution of predicted exposure
values, E(T|Q). As a result of the calibration model considered, the variance of predicted within-group values,
compared with Q, will shrink by a factor proportional to λWj2
(11). Using the calibration sample in the EPIC study (13), we
calculated the predicted values according to calibration
model 2. To estimate the between- and within-group variance components in the “calibrated” exposure, a mixed
model was used. Here, we refer to this model as the “evaluation” model, where the observed Q and R measurements and
the predicted intake values are each used as response variables in separate analyses, and the variance components are
estimated by modeling the group effect with a random effect
as
Y ij = α Ej + γ EW ( Z ij – Z j ) + γ EB Z j + ε Eij
αEj = αE + uE0j,
(3)
uE0j ~ N(0, σ12) and εEij ~ N(0, σ22),
with Yij being either Qij, Rij, or E(T|Q)ij estimated in model 2.
This model enables estimation of the proportion of total variation in the exposure attributable to the between-group
component, expressed by intraclass correlation coefficient
(ICC) = σ12/(σ12 + σ22). By symmetry, 1 – ICC expresses the
variability attributable to the within-group component.
Energy and macronutrient variability
We considered statistical models for intakes of protein,
lipids, carbohydrate, alcohol, and total energy. In the
absence of a standardized European nutrient database,
country-specific food composition tables were used to calculate nutrient intakes for both Q and R. Results from a recent
validation study (24) showed higher correlations between
dietary protein from R measurements and protein estimated
from urinary nitrogen, compared with Q. This finding
suggests that the dietary assessment method used probably
has a higher impact on measurement errors than use of local
nutrient databases.
To adjust for total energy, macronutrient densities (i.e.,
macronutrient intakes divided by total energy intake) were
also considered. ICC estimates were computed for nutrient
intake in Q and R measurements, as well as for predicted
values, according to model 2 for absolute intakes and density
values.
To model the group effect in the present analysis, 28
centers were considered. The macronutrient distributions
approximated a normal distribution, whereas the distribution
of alcohol intake was rather skewed. However, the present
analyses focused primarily on point estimates of fixed
effects and variance components. These quantities are fairly
robust to departures from normality in mixed models (25).
Results using log-transformed values were very similar to
those obtained by using untransformed variables.
The calibration and the evaluation models were all
adjusted for age and body mass index at the individual and at
the group levels. These variables were chosen because
previous studies (26–29) have suggested that they are significant determinants of the accuracy of Q measurements. To
allow the effect of age and body mass index to be heterogeneous across centers, random effects for these variables were
included in model 2, but the associated variance components
were not significantly different from zero, thus suggesting
Am J Epidemiol 2004;160:814–822
Within- and Between-Cohort Macronutrient Variation 817
TABLE 1. Estimated calibration coefficients and between- and within-center variance components of absolute values and densities
of macronutrient intake for men in the European Prospective Investigation into Cancer and Nutrition study, 1992–2000*
Exposure
and type†
Absolute intake value
λ B‡
λW §
σλ¶
Density value
B#
W**
ICC††
Q
248
621
0.14
R
157
833
0.03
152
278
0.23
Q
14
25
R
9
35
9
10
0.48
Q
36
78
R
30
97
30
36
0.41
Q
10
30
R
11
48
11
15
0.35
Q
8
22
R
8
32
8
17
0.18
λB‡
λW §
σλ¶
B#
W**
ICC
0.25
1.12
2.16
0.21
0.07
0.90
3.99
0.05
0.88
1.01
0.43
0.17
3.71
6.48
0.25
0.09
3.90
9.71
0.14
3.91
3.74
0.52
0.10
2.80
5.53
0.20
0.05
3.59
8.99
0.14
3.57
2.86
0.61
0.11
1.84
5.61
0.10
0.06
1.82
7.54
0.06
1.81
4.37
0.15
Energy intake
E(T |Q)
0.36
0.42
0.12
Protein
E(T |Q)
0.44
0.36
0.10
0.17
0.44
0.13
Carbohydrates
E(T |Q)
0.65
0.43
0.11
0.90
0.56
0.12
Lipids
E(T |Q)
0.03
0.46
0.13
0.85
0.48
0.17
Alcohol
E(T |Q)
1.04
0.79
0.14
0.94
0.78
0.09
* Results were adjusted for age and body mass index.
† Q, dietary questionnaire measurements; R, 24-hour dietary recall measurements; E(T|Q), predicted values.
‡ Between-center calibration factor in model 2. Refer to the Materials and Methods section of the text for an explanation of the models.
§ Within-center calibration factor in model 2.
¶ Square root of the variance component for the within-center calibration factor (λW) in model 2, consistently significantly different from zero,
according to likelihood ratio test.
# Square root of the variance component for the between-center component in evaluation model 3.
** Square root of the variance component for the within-center component in evaluation model 3.
†† ICC, intraclass correlation coefficient.
homogeneity of effects. Overall, linear fixed effects for age
and body mass index, compared with models with higherorder terms, best fit the data.
To explore the role of confounding factors in explaining
the heterogeneity of the attenuation factors, we also considered a calibration model with a cross-level interaction
between Q measurements and mean values of age and body
mass index. Doing so enables exploration of the dependency
of attenuation factors on predictors.
Gender-specific analyses were performed. Statistical analyses were carried out by using the MIXED procedure of SAS
software (30). Parameter estimates were obtained by
restricted maximum likelihood estimation. A “sandwich”
estimator was used to compute corrected standard errors
involving fixed-effects parameters, which are robust against
model misspecification (25, 31, 32) in model 2. The likelihood ratio statistic based on restricted maximum likelihood
estimation was used to test the significance of random
effects.
Am J Epidemiol 2004;160:814–822
RESULTS
In tables 1 and 2, the square root of the between- and
within-center variance components and corresponding ICC
estimates for energy intake and the macronutrient distributions of Q, R, and E(T|Q) measurements are reported for men
and women, respectively. These tables also show estimates
for the between- and within-center components of the calibration model, λB and λW, according to model 2. Results for
macronutrients are reported with and without adjustment for
energy intake.
For men and women, respectively, the calibration factors
for the within-center component ranged from 0.36 and 0.31
for protein intake to 0.79 and 0.84 for alcohol. Similar values
for the between-center λB were obtained, ranging from 0.03
for lipids in men and 0.10 for energy intake in women to 1.04
and 1.29 for alcohol in men and women. Adjustment for
energy intake had very little impact on λW, whereas betweencenter calibration factors were higher regarding energy
densities for carbohydrates in men (λB from 0.65 to 0.90) and
818 Ferrari et al.
TABLE 2. Estimated calibration coefficients and between- and within-center variance components of absolute values and densities
of macronutrient intake for women in the European Prospective Investigation into Cancer and Nutrition study, 1992–2000*
Exposure
and type†
Absolute intake value
λ B‡
λW §
σλ¶
Density value
B#
W**
ICC††
Q
194
494
0.13
R
98
617
0.02
94
191
0.20
Q
11
21
R
8
26
7
7
0.54
Q
30
65
R
16
77
16
27
0.26
Q
10
24
R
8
35
8
10
0.39
Q
3
11
R
3
17
3
8
0.12
λ B‡
λW §
σλ¶
B#
W**
ICC
0.21
1.42
2.41
0.26
0.08
1.22
4.34
0.07
1.21
1.14
0.53
0.17
3.76
6.50
0.25
0.04
3.49
9.89
0.11
3.47
3.40
0.51
0.15
3.07
5.42
0.24
0.05
3.11
8.89
0.11
3.07
2.50
0.60
0.06
0.93
3.75
0.06
0.03
1.05
5.46
0.04
1.04
3.08
0.10
Energy intake
E(T |Q)
0.10
0.37
0.11
Protein
E(T |Q)
0.44
0.31
0.09
0.43
0.45
0.10
Carbohydrates
E(T |Q)
0.41
0.41
0.11
0.76
0.49
0.12
Lipids
E(T |Q)
0.33
0.38
0.13
0.75
0.42
0.14
Alcohol
E(T |Q)
1.29
0.84
0.31
1.05
0.84
0.18
* Results were adjusted for age and body mass index.
† Q, dietary questionnaire measurements; R, 24-hour dietary recall measurements; E(T|Q), predicted values.
‡ Between-center calibration factor in model 2. Refer to the Materials and Methods section of the text for an explanation of the models.
§ Within-center calibration factor in model 2.
¶ Square root of the variance component for the within-center calibration factor (λW) in model 2, consistently significantly different from zero,
according to likelihood ratio test.
# Square root of the variance component for the between-center component in evaluation model 3.
** Square root of the variance component for the within-center component in evaluation model 3.
†† ICC, intraclass correlation coefficient.
women (λB from 0.41 to 0.76) and for lipids in men (λB from
0.03 to 0.85) and women (λB from 0.33 to 0.75). Lower
values were observed for protein in men (λB from 0.44 to
0.17).
The square root of the variance component of λWj, σλ,
which estimates the variability of attenuation factors across
EPIC centers, was consistently and significantly different
from zero for both men and women. This finding suggests
that the accuracy of Q was rather heterogeneous across the
EPIC centers.
Variance components for Q measurements showed very
similar patterns for between-center variability in men and
women. Respectively, we found low ICC values for alcohol
(0.11 and 0.06), moderately low ICCs for energy intake
(0.14 and 0.13) and lipids (0.10 and 0.15), and slightly
higher ICCs for protein (0.25 and 0.21) and carbohydrates
(0.17 for both genders).
The patterns of ICC values for R measurements were
similar, but the values were consistently lower than 0.10.
The between-center variance components were overall lower
in R than in Q, except regarding lipids in men. On the other
hand, within-center variation was considerably higher in R, a
result that very likely reflected day-to-day variability in the
single dietary measurements available for each study participant in the calibration subsample. For energy intake, withincenter variability was 1.8 and 1.6 times larger in R than in Q
measurements for men and women, respectively. For men
and women, the same ratios were 2.0 and 1.5 for protein, 1.6
and 1.4 for carbohydrates, 2.5 and 2.1 for lipids, and 2.1 and
2.4 for alcohol.
For the predicted intake E(T|Q), the between-center variability was very close, but not identical, to the variation in
the R measurements as a consequence of including a random
effect (u0j) for the center variable in model 2. The withincenter variability of E(T|Q) shrank when compared with Q
measurements.
Adjustment for energy intake had a variable impact on the
estimated variability of macronutrients. For alcohol and
Am J Epidemiol 2004;160:814–822
Within- and Between-Cohort Macronutrient Variation 819
protein in men and women, ICC values for predicted intakes
were basically unaffected. Adjustment for energy had a
moderate effect for carbohydrates in men (ICC from 0.41 to
0.52). Decidedly higher differences in ICC values were
observed for macronutrient densities regarding carbohydrates in women (ICC from 0.26 to 0.51) and lipids in men
(ICC from 0.35 to 0.61) and women (ICC from 0.39 to 0.60).
Very little change in the variance components for Q and R
measurements was observed for macronutrient density
values.
For both absolute intake and density values, ICC values
overall were between 1.5- and 3.0-fold higher for E(T|Q)
than for Q, thus indicating strong shrinkage in the withincenter variance component after calibration. For lipid
density in men, the ICC ranged from 0.20 for Q measurements to 0.61 for predicted values.
Overall, negative effects were observed in calibration
model 2 for age and body mass index. For alcohol intake in
men, the effects of body mass index were positive (data not
shown). Results from the model with a cross-level interaction to explore the heterogeneity of attenuation factors
revealed overall nonsignificant associations. However, in the
case of carbohydrate and lipid density, of alcohol intake in
men, and of protein density in women, a significant negative
interaction with body mass index was observed, thus
suggesting that centers in which mean body mass index
levels were higher tended to have lower Q accuracy (data not
shown).
DISCUSSION
In multicenter cohort studies, associations of exposure
with disease risk can be studied at two complementary
levels: within and between cohorts (countries or recruitment
centers). Proper comparability of both the within- and
between-center relations, and proper relative weighting of
these components of the overall association, requires a high
degree of standardization of exposure measurements as well
as measured disease outcomes (7). Furthermore, corrections
must be made for variable degrees of random measurement
error that, within center, will generally attenuate relative risk
estimates.
In the context of the European EPIC study on diet and
cancer, we have described the use of calibration substudies
to correct for between-cohort differences in effects of dietary
assessment errors, putting the measurements on a more standardized scale. The calibration was based on collection,
within a large, stratified subsample of the EPIC cohorts, of
standardized R measurements. Regression of these “reference” measurements on the individual-level questionnaire
data enabled estimation of predicted values E(T|Q), indicating average true intake levels for given Q measurements
that, in a second phase, can be used as predictors in a disease
model. A two-level, random-effects model was used to estimate within- and between-center calibration effects. The
calibration model included adjustments for effects of potential confounders, particularly total energy intake, body mass
index, and age.
Random effects allowed parsimonious modeling of the
heterogeneity of effects across EPIC centers without having
Am J Epidemiol 2004;160:814–822
to include a large number of interaction terms. Randomeffects models also allow the inclusion of aggregate (contextual) variables to be evaluated and the exploration of
possible sources of between-center differences in dietary
questionnaire accuracy, by making such differences dependent on aggregate-level predictors.
We decided to focus on the effects of calibration on
within- and between-center variance components of macronutrient intakes. The macronutrient composition of diet is
still a major factor of interest in the nutritional epidemiology
of cancer. In a relatively recent review (33), the macronutrients were considered comparable (e.g., in the case of
alcohol) or readily comparable after correction for different
methods of calculation and mode of expression (as for
energy, protein, and carbohydrates). In addition, a validation
study by Slimani et al. (24), recently conducted in the EPIC
cohort, supported this finding, where relatively high correlations were observed in an ecologic comparison of protein
intake estimated from urinary nitrogen and dietary protein,
estimated by R and Q measurements. A more extensive, fully
standardized European food composition database, also
comprising all detailed micronutrients, is currently being
prepared for the EPIC study (34).
The lack of common food composition tables might lead to
a larger between-center variability, as a consequence of
systematic food composition tables-specific errors,
randomly distributed across countries, but it is less likely to
affect the sizable change in ICC values observed for
predicted intakes. In the present analyses, the calibration
model took into account the geographic specificity of the
EPIC entities, by performing center-specific comparisons
between R and Q measurements, for which the same national
nutrient databases were used. After correction for measurement error, we believe that the shrinkage of within-center
macronutrient variability rather than the inflation of
between-center variability was responsible for the relative
increase in ICC values for E(T|Q) compared with Q. As
expected, between EPIC centers, mean values of E(T|Q)
were practically identical to mean levels of R. The term λB
provides an estimate of the between-center (linear) relations
between mean R and Q measurements. Rather low values of
the coefficient were observed for lipids in men and total
energy intake in women, but values approximating 1 were
observed for alcohol, particularly in men, thus suggesting
very good agreement between R and Q at the aggregate level.
Because special efforts have been made to standardize the R
measurements to estimate absolute, mean dietary intake
levels, we assume that the mean calibrated values reflect
more accurately the between-cohort variations in mean true
intake than the mean Q values do (23).
Within EPIC centers, the distributions of predicted intake
values (i.e., after calibration) generally showed strong
shrinkage toward center-specific mean values. The
shrinkage in within-center variance was proportional to the
square of within-center calibration coefficients λWj. The
shrinkage depended only very moderately on the effects of
other individual-level variables used to explain R variability,
such as age and body mass index.
The association between exposure and disease can be
decomposed as βTotal = βB ICC + βW (1 – ICC), where the β’s
820 Ferrari et al.
express the ecologic (B) and individual (W) association
between exposure and disease, provided that the withincenters relations are homogeneous. After calibration, at least
for macronutrients, a sizable weight is given to the betweencohort component of the multicenter EPIC analysis, and the
weights of the two levels of evidence of the diet/disease
association are similar.
Results from the present analyses suggest that the
between-center variability of exposure provides a sizable
amount of information and should therefore be exploited.
Analysis of the diet/disease association in the EPIC study
should provide estimates of the within- and between-center
components in a framework very similar to that of calibration model 2, where both levels of effect were estimated
simultaneously. In this type of disease model, the betweencenter component has the great advantage over a “classical”
ecologic study that it can adjust for confounding at the individual level and over an aggregate design (35) and that the
populations under study provide estimates of cancer incidence and dietary exposure.
The impact of energy adjustment on observed macronutrient variability appeared to be considerable in the calibration model. Adjustment for energy is motivated both by the
requirement to consider isocaloric models and by the partial
control of measurement error that it can bring about (36).
Within-center variances were affected more than betweencenter variance, particularly for carbohydrates and lipids,
reflecting a reduction in the error variance. Thus, the ICC
values for the macronutrient densities were considerably
larger than those for the absolute intakes.
The calibration model relies on some essential statistical
assumptions. First, the errors in R and Q measurements
must be uncorrelated. In practice, this assumption may not
be fully met, and some positive correlation between errors
may occur, for instance, because of subjects’ tendencies to
under- or overreport intake when different dietary assessment methods are used (14, 16, 37, 38). A positive correlation
between errors leads to overestimation of the within-center
calibration coefficient λW and thus to an underestimation of
shrinkage of the intake distributions within centers. The
ICC values in the present analyses may thus underestimate
the relative weight that the between-center component
should have in a full multilevel analysis. There is some
evidence that adjustment for total energy intake, either
through a linear model or by calculating nutrient densities,
can reduce the magnitude of correlation between random
measurement errors of Q measurements and R or food
consumption records (36, 39). The effective error variance
for the macronutrients is also probably reduced by energy
adjustment. This seems to happen because the error in
energy and the error in estimated macronutrient intake are
highly correlated and, after adjustment, partly “cancel”
each other (18). Thus, although the ICC values for the
absolute intakes may be considerably underestimated, the
ICC values for the energy-adjusted macronutrients can be
less affected. The latter are of greater relevance, since some
form of energy adjustment of macronutrient intakes is standard practice in epidemiology.
A second major assumption is that R provide unbiased
estimates of dietary intake at a group level; that is, αR = 0 and
βR = 1.0 in measurement-error model 1b (absence of constant
and proportional scaling biases). This assumption may be
relaxed, however, to that of homogeneous scaling biases
(i.e., homogeneous αR’s and βR’s) in all study centers, which
in practice may be reached approximately. Again, energy
adjustment may help to standardize dietary measurements
and meet at least the relaxed assumption. Under this assumption, the measurement error correction will still provide a
correct relative weighting of within- versus between-center
evidence for diet-disease associations. In addition, βR’s of
less than unity lead to underestimation of within-center calibration factors. In this context, it is possible that the correlation between errors and the reference measurement bias may
counterbalance their effects, thus leading to correct estimates
of λWj.
The calibration model in model 2 assumes fully linear
associations between response and explanatory variables.
Nonlinear calibration models are advisable solutions to
capture the lack of linearity between R and Q. The shrinkage
of the within-center variance component would characterize
the predicted macronutrient distribution, also in nonlinear
calibration models.
Measurement errors bias the association between any two
quantities, for example, the relative risk estimate in a disease
model, and cause power loss. Within single study centers,
calibration may correct for bias but does not recuperate
power losses due to random errors. Indeed, the effect of
shrinkage of exposure distributions generally outweighs the
effect of bias correction (i.e., increase) in relative risk estimates. In a multicenter study, however, calibration in principle can reduce between-center heterogeneity in dietdisease associations caused by differential impact of
measurement errors and, probably more importantly, will
lead to a less biased estimate of the overall relative risk (7,
11).
Moreover, in estimating associated variability of corrected
relative risk estimates, it is necessary to take into account the
variability in the estimate λ’s (10, 11). In the EPIC study,
this issue is minor because the large size of the calibration
study allows the calibration factor to be estimated with high
precision.
ACKNOWLEDGMENTS
The EPIC study was supported by the “Europe Against
Cancer” Programme of the European Commission
(SANCO); Ligue contre le Cancer (France); Société 3M
(France); Mutuelle Générale de l’Education Nationale;
Institut National de la Santé et de la Recherche Médicale
(INSERM); German Cancer Aid; German Cancer Research
Center; German Federal Ministry of Education and
Research; Danish Cancer Society; Health Research Fund
(FIS) of the Spanish Ministry of Health; the participating
regional governments and institutions of Spain; Cancer
Research UK; Medical Research Council, United Kingdom;
the Stroke Association, United Kingdom; British Heart
Foundation; Department of Health, United Kingdom; Food
Standards Agency, United Kingdom; the Wellcome Trust,
Am J Epidemiol 2004;160:814–822
Within- and Between-Cohort Macronutrient Variation 821
United Kingdom; Greek Ministry of Health; Greek Ministry
of Education; Italian Association for Research on Cancer;
Italian National Research Council; Dutch Ministry of Public
Health, Welfare and Sports; Dutch Ministry of Health;
Dutch Prevention Funds; LK Research Funds; Dutch ZON
(Zorg Onderzoek Nederland); World Cancer Research Fund
(WCRF); Swedish Cancer Society; Swedish Scientific
Council; Regional Government of Skane, Sweden; and the
Norwegian Cancer Society.
17.
18.
19.
REFERENCES
1. Armstrong B, Doll R. Environmental factors and cancer incidence and mortality in different countries, with special reference to dietary practices. Int J Cancer 1975;15:617–31.
2. World Cancer Research Fund/American Institute for Cancer
Research. Diet, nutrition, and the prevention of cancer: a global
perspective. Washington, DC: WCRF/AICR, 1997.
3. Walker AM, Blettner M. Comparing imperfect measures of
exposure. Am J Epidemiol 1985;121:783–90.
4. Freudenheim JL, Marshall JR. The problem of profound mismeasurement and the power of epidemiological studies. Nutr
Cancer 1988;11:243–50.
5. Armstrong B, White E, Saracci R. Principles of exposure measurements in epidemiology. Oxford, United Kingdom: Oxford
University Press, 1992.
6. Riboli E, Hunt KJ, Slimani N, et al. European Prospective
Investigation into Cancer and Nutrition (EPIC): study populations and data collection. Public Health Nutr 2002;5:1113–24.
7. Kaaks R, Plummer M, Riboli E, et al. Adjustment for bias due
to errors in exposure assessments in multicenter cohort studies
on diet and cancer: a calibration approach. Am J Clin Nutr
1994;59(suppl 1):245S–50S.
8. Plummer M, Clayton D, Kaaks R. Calibration in multi-centre
cohort studies. Int J Epidemiol 1994;23:419–26.
9. Slimani N, Ferrari P, Ocke M, et al. Standardization of the 24hour diet recall calibration method used in the European Prospective Investigation into Cancer and Nutrition (EPIC): general concepts and preliminary results. Eur J Clin Nutr 2000;54:
900–17.
10. Rosner B, Willett WC, Spiegelman D. Correction of logistic
regression relative risk estimates and confidence intervals for
systematic within-person measurement error. Stat Med 1989;8:
1051–69, 1071–3 (discussion).
11. Kaaks R, Riboli E, van Staveren W. Calibration of dietary
intake measurements in prospective cohort studies. Am J Epidemiol 1995;142:548–56.
12. Slimani N, Deharveng G, Charrondiere RU, et al. Structure of
the standardized computerized 24-hour diet recall interview
used as reference method in the 22 centers participating in the
EPIC project. Comput Methods Programs Biomed 1999;58:
251–66.
13. Slimani N, Kaaks R, Ferrari P, et al. EPIC calibration study:
rationale, design and population characteristics. Public Health
Nutr 2002;5:1125–45.
14. Kaaks R, Riboli E, Estève J. Estimating the accuracy of dietary
questionnaire assessments: validation in terms of structural
equation models. Stat Med 1994;13:127–42.
15. Kaaks R, Ferrari P, Ciampi A, et al. Uses and limitations of statistical accounting for random error correlations in the validation of dietary questionnaire assessments. Public Health Nutr
2002;5:969–76.
16. Plummer M, Clayton D. Measurement error in dietary assess-
Am J Epidemiol 2004;160:814–822
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
ment: an investigation using covariance structure models. Part
I. Stat Med 1993;12:925–35.
Day NE, Ferrari P. Some methodological issues in nutritional
epidemiology. In: Riboli E, Lambert R, eds. Nutrition and lifestyle: opportunities for cancer prevention. Lyon, France: International Agency for Research on Cancer, 2002:5–10. (IARC
scientific publication 156).
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.
Fraser GE, Stram DO. Regression calibration in studies with
correlated variables measured with error. Am J Epidemiol
2001;154:836–44.
Snjiders TAB, Bosker R. Multilevel analysis: an introduction to
basic and advanced multilevel modelling. London, United
Kingdom: Sage Publications, 1999.
Sheppard L Insights on bias and information in group-level
studies. Biostatistics 2003;4:265–78.
Begg MD, Parides MK. Separation of individual-level and cluster-level covariate effects in regression analysis of correlated
data. Stat Med 2003;22:2591–602.
Goldstein H. Multilevel statistical models. London, United
Kingdom: Arnold, 1995. (Kendall’s library of statistics 3).
Slimani N, Bingham S, Runswick S, et al. Group level validation of protein intakes estimated by 24-hour diet recall and
dietary questionnaires against 24-hour urinary nitrogen in the
European Prospective Investigation into Cancer and Nutrition
(EPIC) calibration study. Cancer Epidemiol Biomarkers Prev
2003;12:784–95.
Verbeke G, Lesaffre E. The effect of misspecifying the random-effects distribution in linear mixed models for longitudinal
data. Computat Stat Data Anal 1997;23:541–56.
Bandini LG, Schoeller DA, Cyr HN, et al. Validity of reported
energy intake in obese and nonobese adolescents. Am J Clin
Nutr 1990;52:421–5.
Heitmann BL. The influence of fatness, weight change, slimming history and other lifestyle variables on diet reporting in
Danish men and women aged 35–65 years. Int J Obes Relat
Metab Disord 1993;17:329–36.
Heitmann BL, Lissner L. Dietary underreporting by obese individuals—is it specific or non-specific? BMJ 1995;311:986–9.
Ferrari P, Slimani N, Ciampi A, et al. Evaluation of under- and
overreporting of energy intake in the 24-hour diet recalls in the
European Prospective Investigation into Cancer and Nutrition
(EPIC). Public Health Nutr 2002;5:1329–45.
SAS Institute, Inc. SAS online doc, version 8. Cary, NC: SAS
Institute, Inc, 1999.
Huber PJ. The behaviour of maximum likelihood estimators
under non-standard conditions. In: LeCam LM, Neyman J, eds.
Proceedings of the Fifth Berkeley Symposium on Mathematical
Statistics and Probability 1. Berkeley, CA: University of California Press, 1967:221–33.
White H. Maximum likelihood estimation of misspecified models. Econometrica 1982;50:1–25.
Deharveng G, Charrondière UR, Slimani N, et al. Comparison
of nutrients in the food composition tables in the nine European
countries participating in EPIC. Eur J Clin Nutr 1999;53:60–79.
Charrondiere UR, Vignat J, Møller A, et al. The European
nutrient database (ENDB) for nutritional epidemiology. J Food
Comp Anal 2002;15:435–51.
Prentice RL, Sheppard L. Aggregate data studies of disease risk
factors. Biometrika 1995;82:113–25.
Willett W. Commentary: dietary diaries versus food frequency
questionnaires—a case of undigestible data. Int J Epidemiol
2001;30:317–19.
Kipnis V, Carroll RJ, Freedman LS, et al. Implications of a new
822 Ferrari et al.
dietary measurement error model for estimation of relative risk:
application to four calibration studies. Am J Epidemiol 1999;
150:642–51.
38. Day N, McKeown N, Wong M, et al. Epidemiological assessment of diet: a comparison of a 7-day diary with a food frequency questionnaire using urinary markers of nitrogen,
potassium and sodium. Int J Epidemiol 2001;30:309–17.
39. Schatzkin A, Kipnis V, Carroll RJ, et al. A comparison of a
food frequency questionnaire with a 24-hour recall for use in an
epidemiological cohort study: results from the biomarker-based
Observing Protein and Energy Nutrition (OPEN) study. Int J
Epidemiol 2003;32:1054–62.
Am J Epidemiol 2004;160:814–822