Uncertainty about the validity of dietary self-report instruments due to random and systematic measurement errors is one of the major concerns in nutritional epidemiology because these dietary assessment errors may attenuate risk estimates and obscure diet–disease associations(Reference Kaaks and Riboli1–Reference Schatzkin, Kipnis and Carroll3). In multi-centre studies, this methodological issue is amplified due to possible between-centre differences in systematic over- or underestimation in dietary intake measurements(Reference Kaaks, Plummer and Riboli4–Reference Slimani, Kaaks and Ferrari6).
Self-reported diet tends to be biased towards underestimation of energy intake (EI)(Reference Hill and Davies7, Reference Livingstone and Black8) and specific macronutrients(Reference Livingstone and Black8–Reference Subar, Kipnis and Troiano10), although a relative overestimation of foods that are perceived as socially desirable (and related nutrients) may also occur(Reference Livingstone and Black8, Reference Michels, Welch and Luben11). Dietary assessment errors would be less problematic and could be easily corrected for if they were randomly distributed across individuals and food groups but, unfortunately, misreporting appears to occur at higher rates in certain subgroups of the population and with certain foods(Reference Heitmann and Lissner12, Reference Poppitt, Swann and Black13). Having a high BMI has been identified as the most consistent factor that is associated with under-reporting of EI(Reference Hill and Davies7, Reference Livingstone and Black8, Reference Ferrari, Slimani and Ciampi14). However, not all overweight or obese persons under-report their diet(Reference Livingstone and Black8) and associations between under-reporting and BMI may also depend on the method of dietary assessment used(Reference Subar, Kipnis and Troiano10, Reference Lissner, Troiano and Midthune15, Reference Scagliusi, Ferriolli and Pfrimer16).
Considering the ongoing epidemic of obesity and of diseases associated with obesity worldwide, it is important that BMI-related reporting errors in dietary assessment are addressed. For example, evaluating the patterns of reporting errors among population subgroups as defined by their BMI and whether such errors vary across populations with diverse dietary habits may help to refine measurement error correction methods and aid in the interpretation of nutritional epidemiological findings. Furthermore, detailed analysis of the shortcomings of dietary self-report may eventually help to improve dietary assessment instruments.
Dietary biomarkers do not rely on self-report and thus fulfil the criterion of an independent and objective measure of dietary intake against which the validity of dietary self-report instruments can be tested(Reference Kaaks and Riboli1). Of particular interest are biomarkers that provide an estimate of absolute dietary intake, based on the urinary recovery of nutrients or metabolites from diet. Those so-called ‘recovery’ biomarkers(Reference Kaaks, Ferrari and Ciampi17) include N and K from 24 h urinary measurements that allow calculation of daily intake of protein and K, respectively(Reference Bingham and Cummings18, Reference Tasevska, Runswick and Bingham19).
Relatively few methodological studies have used these biomarkers to evaluate the validity of dietary self-report instruments in relation to obesity simultaneously for 24 h dietary recalls (24-HDR) and dietary questionnaires (DQ)(Reference Lissner, Troiano and Midthune15). Moreover, previous studies were not able to examine whether possible differences in misreporting between countries depend on BMI.
In an earlier cross-sectional study, protein biomarker data from the same twelve centres as those included in the present analysis were used to validate 24-HDR measurements as a reference method for between-cohort calibration(Reference Slimani, Bingham and Runswick9). Despite the presence of measurement error in this reference method, it was concluded that standardised 24-HDR measurements are useful to discriminate population mean estimates and to partially correct the diet–disease associations for attenuation due to measurement errors in baseline DQ measurements(Reference Slimani, Bingham and Runswick9). In the present study, we propose to extend these analyses with the aims (1) to evaluate BMI-related reporting errors on 24-HDR and DQ by using recovery biomarkers for protein and K intake and (2) to examine whether the effect of BMI differs between six European countries.
Subjects and methods
Subjects
The study population was a convenience subsample of the European Prospective Investigation into Cancer and Nutrition (EPIC) study(Reference Riboli, Hunt and Slimani20) and has been described elsewhere(Reference Slimani, Bingham and Runswick9). In brief, subjects from the following twelve geographical areas (EPIC centres) were re-contacted between 1995 and 1999 and asked to collect a 24 h urine specimen: Paris (France); Varese, Turin, Florence, Naples and Ragusa (Italy); Greece; Cambridge and Oxford (UK); Heidelberg and Potsdam (Germany); Bilthoven (The Netherlands). Within countries, the data of different centres were combined for the present analysis except for the two UK centres, where the Oxford cohort includes vegetarians, vegans and other health-conscious individuals.
A total of 1386 volunteers were recruited in the study, but a number of subjects were excluded: 186 having invalid collections of the 24 h urine sample as determined by para-aminobenzoic acid (PABA); forty-five who did not complete both self-report instruments (i.e. 24-HDR and DQ); eight having missing values on height, weight and lifestyle baseline data or extreme DQ EI:energy requirement ratios (1 % of the extremes, centre and sex specific, a routine exclusion made on the EPIC baseline questionnaire data); fifty-two subjects reported being on a diet during the 24-HDR. Furthermore, nine subjects with implausible values of height ( < 130 cm), BMI ( < 16 kg/m2), waist circumference ( < 40 cm or >160 cm), the combination of a waist circumference ( < 60 cm) and BMI (>25 kg/m2) and with missing BMI were excluded. Overall, a total of 300 subjects (21·6 %) were excluded, giving a final sample size of 1086 subjects for the present study.
The present study was conducted according to the guidelines laid down in the Declaration of Helsinki and all procedures involving human subjects were approved by the local ethics committees in the participating countries and the International Agency for Research on Cancer Institutional Review Board. Written informed consent was obtained from all subjects.
Self-reported diet and anthropometric measurements
Information on ‘habitual’ dietary intake was assessed at baseline for each individual at recruitment to the EPIC cohorts using semi-quantitative DQ (France, The Netherlands, Germany, Greece and Italy, except Naples) or a semi-quantitative FFQ (UK, Naples) developed and validated in each participating country(Reference Riboli, Hunt and Slimani20, Reference Margetts and Pietinen21).
A single 24-HDR interview was collected from a subsample of each cohort as the dietary reference method for calibration purposes. Trained dietitians conducted all (unannounced) interviews face to face. Food portion sizes were estimated using a common picture book and other available methods such as standard units and household measures(Reference Slimani, Kaaks and Ferrari6). In contrast to the baseline DQ, the 24-HDR interviews were standardised across countries, using the computerised program EPIC-Soft® (International Agency for Research on Cancer, Lyon, Rhone-Alpes, France) with the same structure and interview procedure across countries. More details on the concept of standardisation are given elsewhere(Reference Slimani, Ferrari and Ocké22).
Country-specific nutrient databases were used to estimate protein, K and EI from baseline DQ. Respective intakes from 24-HDR were estimated using the standardised EPIC Nutrient Database. The EPIC Nutrient Database project outlines in detail the methods used to standardise the national nutrient databases across countries(Reference Slimani, Deharveng and Unwin23).
BMI was calculated from measured weight and height and categorised into normal weight (BMI < 25 kg/m2), overweight (BMI 25 to < 30 kg/m2) and obese (BMI ≥ 30 kg/m2). The category < 25 kg/m2 included eleven underweight subjects with a BMI below 18·5 kg/m2 (1 % of the sample). For 105 subjects for whom more than 1 year elapsed between completing the questionnaire and anthropometric measurements and for one subject for whom no measured information was available, self-reported data from their questionnaires were used.
24 h urine samples and recovery biomarkers for protein and potassium intake
As the primary purpose of the 24 h urine collections was to validate 24-HDR, the urine samples were collected on the day of the 24-HDR interview or due to organisational reasons within a maximum of 6 d afterwards. The DQ were already collected before the 24-HDR, thus the time interval between collecting the DQ and urine samples was longer and varied from a few days to < 2 months in Bilthoven and the German centres to 2–5 years later in Turin (men), France and Naples(Reference Slimani, Bingham and Runswick9).
The subjects provided one 24 h urine collection following a standardised protocol(Reference Bingham, Cassidy and Cole24). In order to verify completeness, each participant was given three 80 mg tablets of PABA (PABAcheck; Laboratories for Applied Biology, London, UK) to be ingested on the day of urine collection(Reference Bingham and Cummings18). The urine specimens were analysed centrally in Cambridge according to the standard procedures: total N was measured by the Kjeldahl technique (Tecator 1002; Perstorp Analytical Limited, Bristol, UK)(Reference Bingham, Cassidy and Cole24); K by using an IL 943 flame photometer (Instrumentation Laboratory, Warrington, UK)(Reference Tasevska, Runswick and Bingham19); PABA by using colorimetry(Reference Bingham, Williams and Cole25).
The 24 h urine collections containing 85–110 % of the PABA marker were considered complete. Urine specimens containing < 70 % PABA recovery (when fewer than three tablets have been taken and/or urine collection incomplete) and above 110 % (might happen when subjects take drugs that interfere with the colorimetric analysis) were excluded from the analysis. Urine samples with 70–84 % PABA recovery were used after correction for urinary electrolytes up to the expected values at 93 % of PABA recovery(Reference Johansson, Bingham and Vahter26).
To calculate individual protein intake (g/d) from N excretion, 24 h urinary N was divided by 0·81 and then multiplied by 6·25(Reference Bingham and Cummings18). Based on published studies that compared K intake with urinary excretion, the values of 24 h urinary K were divided by 0·81 (see the Discussion section) to convert urinary K to dietary K (g/d).
Statistical analysis
Dietary misreporting was defined as the individual log-ratio (log of the ratios of untransformed values) of reported intake to biomarker measurements of protein or K.
At the group level, log-ratios were back-transformed and reported as geometric means across groups. The magnitude of a deviation below or above the ratio of 1 defined the degree of under- or over-reporting. The P value for trend across BMI groups was calculated from individual-level regression. A paired t test was used to test differences between individual biomarker and dietary measurements within each pair. The means of the ratios were adjusted for age and sex and the 24-HDR data were additionally weighted for day of the week and season of recall using a generalised linear model.
Multiple linear regression analyses were used to predict mean log-ratios of reported intake to biomarkers of protein or K by BMI and to control for a priori defined covariates including age, sex, education and country (centre). We also performed the regression of each of the urinary measurements (biomarkers) v. the respective dietary measurements and BMI to estimate the proportion of the total variability (adjusted R 2) of the biomarkers that can be accounted for by each of the dietary instruments after adjusting for age, sex and country. Urinary and dietary measurements were log-transformed to improve fit of the data (BMI values were untransformed). To show how much of the variability is accounted for by BMI, R 2 values with and without adjustment for BMI were also calculated.
Sensitivity analyses were performed by excluding subjects with reported EI that were unlikely to represent either habitual intake or a randomly low intake according to Goldberg's tables for lower and upper limits of EI over BMR (EI:BMR). Energy under-reporters, i.e. ‘low-energy reporters (LER)’, have EI:BMR ratios that differ from an assumed moderate physical activity level of 1·55 by more than − 2 sd and energy over-reporters differ by more than +2 sd. Variation in BMR and physical activity level, daily variation in EI and the number of days of diet assessment were taken into account(Reference Goldberg, Black and Jebb27). Thus, different cut-offs were used for 24-HDR (one single observation per individual) and DQ (long-term habitual diet). Participants with an EI:BMR ratio lower than 0·88 or higher than 2·72 were excluded from the analysis of the 24-HDR: 110 under-reporters (10·1 %) and thirteen over-reporters (1·2 %). On DQ, corresponding EI:BMR values were 1·14 for the lower limit and 2·10 for the upper limit, leading to exclusions of 271 under-reporters (25·0 %) and 73 over-reporters (6·7 %).
All reported P values are two-sided and statistical analyses were performed with SAS version 9.1 (SAS Institute, Cary, NC, USA) or STATA 11.1 (StataCorp, College Station, TX, USA).
Results
The study population consisted mainly of middle-aged people (range of mean age across centres: 51–60 years), except for men and women in The Netherlands and in Ragusa, who were younger (mean age in both centres: 48 years). The sample was 58 % female, reflecting the different sex distributions in the EPIC study (approximately 70 % women). When all countries were combined, 13 % of the subjects were obese, 37 % were overweight and 50 % had a normal weight (Table 1). There was a large variation in the distribution of obesity across countries, ranging from 1 % in France to 43 % in Greece (data not shown). The personal characteristics and the intake of energy and macronutrients of the study population are summarised in Table 1. Table S1 of the supplementary material (available online at http://www.journals.Cambridge.org/bjn) informs about the same variables, but with under- and over-reporters of EI excluded.
(Mean values with their standard errors)
* Calculated from 24 h dietary recalls.
† Normal weight (BMI < 25 kg/m2), n 541; overweight (BMI 25 to < 30 kg/m2), n 406; obese (BMI ≥ 30 kg/m2), n 139.
Protein under-reporting by BMI group
Tables 2 and 3 show the geometric means of protein intake based on dietary self-report instruments (i.e. 24-HDR and DQ) and biomarker measurements as well as the respective ratios. The ratios of reported intake to biomarker show a strongly significant (P = 0·004) trend to a greater degree of under-reporting of protein intake with increasing BMI using 24-HDR, for men and women of all countries combined. For DQ-derived protein, the estimated trend was slightly less in magnitude but also strongly significant (P = 0·002). However, once under- and over-reporters of EI were excluded according to the Goldberg cut-off, there was no evidence of a systematic trend for under-reporting by BMI for either dietary method.
(Mean values, 95 % confidence intervals, number of subjects and percentages)†
PBM, protein biomarker; 24-HDR, 24 h dietary recall; DQ, dietary questionnaire.
* Normal weight (BMI < 25 kg/m2); overweight (BMI 25 to < 30 kg/m2); obese (BMI ≥ 30 kg/m2).
† All values are adjusted for age and sex; 24-HDR are additionally weighted for day of the week of recall and season.
‡ Calculated from individual-level regression using log-ratios.
§ Protein biomarker = urinary N/0·81 × 6·25.
∥ Paired t test of mean of differences between individual biomarker and dietary measurements (24-HDR and DQ) calculated from log-transformed values.
¶ Exclusion of under- and over-reporters of energy intake according to Goldberg et al. (Reference Goldberg, Black and Jebb27) assuming a physical activity level of 1·55.
(Mean values, 95 % confidence intervals, number of subjects and percentages)†
PBM, protein biomarker; 24-HDR, 24 h dietary recall; DQ, dietary questionnaire.
* Normal weight (BMI < 25 kg/m2); overweight (BMI 25 to < 30 kg/m2); obese (BMI ≥ 30 kg/m2).
† All values are adjusted for age; 24-HDR are additionally weighted for day of the week of recall and season.
‡ Calculated from individual-level regression using log-ratios.
§ Protein biomarker = urinary N/0·81 × 6·25.
∥ Paired t test of mean of differences between individual biomarker and dietary measurements (24-HDR and DQ) calculated from log-transformed values.
¶ Exclusion of under- and over-reporters of energy intake according to Goldberg et al. (Reference Goldberg, Black and Jebb27), assuming a physical activity level of 1·55.
Considering men and women separately, the situation regarding a greater degree of under-reporting with increasing BMI was approximately the same for both sexes, although the trend in men was non-significant (P = 0·18), presumably due to stratification leading to reduced power. For women, but not for men, the DQ yielded mean protein estimates closer to the biomarker measurements across BMI groups than did the 24-HDR (Table 3).
The detailed results for protein and K stratified by country are given in Tables S2 and S3 and presented graphically in Fig. S1 (supplementary material for this article can be found at http://www.journals.Cambridge.org/bjn).
Individual-level associations between protein under-reporting and BMI
In the regression analysis, BMI was a significant predictor of the log-ratio of reported intake to biomarker of protein for both dietary methods, after controlling for age, sex, education and country. With each unit (1 kg/m2) increase in BMI, the log-ratio decreased by − 0·017 (95 % CI − 0·018, − 0·016) and − 0·013 (95 % CI − 0·015, − 0·012) on 24-HDR and DQ, respectively (both P < 0·0001). After excluding under- and over-reporters of EI, BMI-related bias was attenuated to − 0·007 (95 % CI − 0·008, − 0·006) and − 0·008 (95 % CI − 0·009, − 0·006) for the 24-HDR and DQ, respectively (both P < 0·0001). For example, for the 24-HDR, an approximately 8 % increase in protein under-reporting for a five-unit increase in BMI would be attenuated to approximately 3 %.
The BMI effect on protein under-reporting did not differ for men and women as tested by interaction between BMI and sex (P = 0·74 on 24-HDR; P = 0·20 on DQ). Similarly, there was no evidence for an interaction between BMI and country on protein under-reporting (P = 0·15 on 24-HDR; P = 0·29 on DQ). Additional tests to compare the significance of the main effect for country with and without adjustment for BMI showed that the country effect on the validity for the two instruments was independent of the BMI effect for both protein and K (data not shown).
The coefficient of determination was marginally higher on 24-HDR when compared with DQ when performing the regression of protein biomarker v. reported intake and BMI (adjusted R 2 0·33 v. 0·31). The partial R 2 of BMI was 0·06 and 0·04 on 24-HDR and DQ, respectively; thus, BMI explained 6 and 4 % of the variability in protein biomarker.
Potassium under-reporting by BMI group
Table 4 shows a borderline significant (P = 0·05) trend of K under-reporting with increasing BMI using 24-HDR, for men and women of all countries combined. After stratification by sex, a non-significant tendency towards BMI-related under-reporting remained for both men and women (Table 5).
(Mean values, 95 % confidence intervals, number of subjects and percentages)†
KBM, potassium biomarker; 24-HDR, 24 h dietary recall; DQ, dietary questionnaire.
* Normal weight (BMI < 25 kg/m2); overweight (BMI 25 to < 30 kg/m2); obese (BMI ≥ 30 kg/m2).
† All values are adjusted for age and sex; 24-HDR are additionally weighted for day of the week of recall and season.
‡ Calculated from individual-level regression using log-ratios.
§ Potassium biomarker = urinary K/0·81.
∥ Paired t test of mean of differences between individual biomarker and dietary measurements (24-HDR and DQ) calculated from log-transformed values.
¶ Exclusion of under- and over-reporters of energy intake according to Goldberg et al. (Reference Goldberg, Black and Jebb27) assuming a physical activity level of 1·55.
(Mean values, 95 % confidence intervals, number of subjects and percentages)†
KBM, potassium biomarker; 24-HDR, 24 h dietary recall; DQ, dietary questionnaire.
* Normal weight (BMI < 25 kg/m2); overweight (BMI 25 to < 30 kg/m2); obese (BMI ≥ 30 kg/m2).
† All values are adjusted for age; 24-HDR are additionally weighted for day of the week of recall and season.
‡ Calculated from individual-level regression using log-ratios.
§ Potassium biomarker = urinary K/0·81.
∥ Paired t test of mean of differences between individual biomarker and dietary measurements (24-HDR and DQ) calculated from log-transformed values.
¶ Exclusion of under- and over-reporters of energy intake according to Goldberg et al. (Reference Goldberg, Black and Jebb27), assuming a physical activity level of 1·55.
On DQ, there was no significant trend across BMI groups either overall or considering men and women separately. Exclusion of under- and over-reporters of EI removed any evidence of a systematic trend also on 24-HDR.
Individual-level associations between potassium under-reporting and BMI
From regression analysis of the log-ratio of reported K intake to K biomarker on BMI, both dietary self-report instruments showed a similar inverse relationship with BMI, controlling for age, sex, education and country. With each unit increase (1 kg/m2) in BMI, the log-ratio decreased by − 0·007 (95 % CI − 0·009, − 0·006; P < 0·0001) using the 24-HDR and by − 0·002 (95 % CI − 0·004, − 0·001; P < 0·01) for the DQ. Exclusion of under- and over-reporters of EI attenuated the BMI-related bias on 24-HDR to − 0·002 (95 % CI − 0·004, − 0·001; P < 0·01), and it was no longer statistically significant for the DQ (β = 0·001; 95 % CI − 0·001, 0·003; P = 0·30).
As for protein, the BMI effect on K under-reporting did not differ for men and women; tested by interaction between BMI and sex (P = 0·22 on 24-HDR; P = 0·29 on DQ). Similarly, there was no evidence of heterogeneity in the BMI effect between countries as measured by interaction between BMI and country on K under-reporting (P = 0·75 on 24-HDR; P = 0·29 on DQ).
The coefficient of determination was substantially higher for 24-HDR when compared with DQ when performing the regression of K biomarker v. reported K intake and BMI (adjusted R 2 0·28 v. 0·15). BMI explained only about 1 % of the variability in K biomarker on both dietary methods.
Low-energy reporters and BMI
The probability that a subject was identified as a LER according to the Goldberg cut-off increased by 66 % for each five-unit increase in BMI adjusted for age, sex, education and country (OR from logistic regression: 1·66; 95 % CI 1·31, 2·10; P < 0·0005).
Discussion
The focus of the present study was mainly on possible between-country differences in BMI-related dietary reporting errors because the same dietary self-report instrument, even if standardised, administered in diverse populations could lead to errors varying both in size and direction. Differential misreporting related to BMI and other person-specific biases have been shown to seriously distort measurement error correction procedures in nutritional epidemiological studies(Reference Kipnis, Subar and Midthune2, Reference Subar, Kipnis and Troiano10). This problem would be amplified in a multi-centre study setting if the effect of BMI differed between centres (countries).
From the ratios of self-reported intake to biomarker measurements, a trend towards increasing under-reporting of both protein and K intake with increasing BMI was apparent on both dietary self-report instruments (i.e. 24-HDR and DQ). These observed differences in the validity of reporting protein and K intake with changing BMI are similar to those described in previous studies, which generally were conducted among less heterogeneous population groups(Reference Heitmann and Lissner12, Reference Lissner, Troiano and Midthune15, Reference Bingham, Cassidy and Cole24, Reference Zhang, Temme and Sasaki28, Reference Neuhouser, Tinker and Shaw29). The present study differed from the previous ones in that it recruited participants from twelve European regions (six European countries) with a large geographical variation and with a greater diversity in their dietary patterns(Reference Slimani, Fahey and Welch30, Reference Freisling, Fahey and Moskal31). The results indicate that the effect of BMI on protein and K under-reporting did not differ between countries, neither using the 24-HDR nor the DQ. This is encouraging and suggests that measurement error correction procedures that adjust for BMI-specific biases should perform equally well across diverse populations in multi-centre studies, at least in the EPIC study. The differences in under-reporting of protein intake between the countries included in the present analysis have been described in Slimani et al. (Reference Slimani, Bingham and Runswick9).
The present study also shows that exclusion of individuals who probably misreported EI according to Goldberg et al. (Reference Goldberg, Black and Jebb27) removed any evidence of BMI-related bias at the group level and at least attenuated its degree at the individual level for both protein and K on both dietary assessment methods. The main explanation for this finding is that among overweight and obese individuals, there was a higher proportion of ‘LER’ who also under-reported protein and K. This may in turn either be explained by a higher sensitivity of the Goldberg cut-off in individuals with a higher BMI as they possibly under-report from lower energy requirements due to lower physical activity such that their EI:BMR ratios are below the cut-off for an assumed physical activity level of 1·55 or by less differential (mis-) reporting of foods. However, the latter seems rather unlikely as dietary under-reporting by obese individuals has already been described as being selective(Reference Heitmann and Lissner12). Although the Goldberg cut-off has limitations when applied at the individual level because of its low sensitivity(Reference Black32), the present results suggest that its application may be an option to ameliorate the problem of differential misreporting with increasing BMI. It may therefore be conceivable to apply exclusions to overweight and obese individuals only as a straightforward procedure to minimise bias in a dataset. As recommended by Black(Reference Black32), individuals should also be assigned into appropriate physical activity level categories for calculating the Goldberg cut-off in order to increase its sensitivity, which, however, was out of the scope of the present study. On the other hand, it might be impossible to exclude about one-third of a sample from the analysis, which would be the case according to the present DQ data. For 24-HDR, the problem seems not as extreme, but exclusion of about 10–20 % of a sample would still considerably reduce the power of a study and may specifically exclude (obese) subjects whose diets are of particular interest in relation to diseases.
Using energy-adjusted nutrient intake is an alternative approach for bias correction. Stallone et al. (Reference Stallone, Brunner and Bingham33) compared the performance of both approaches in assessing sociodemographic differences in nutrient intake. They argued that energy adjustment was preferable to excluding LER, mainly because exclusions were highest among lower sociodemographic levels leading to selection bias. Restricting the analysis to plausible energy reporters only may, of course, reduce external validity or generalisability of study findings. However, a primary requirement of a generalisable study is internal validity(Reference Franceschi, Plummer, Ahrens and Pigeot34), which seems to be increased by the exclusion of LER. Whether more complex measurement error models, which adjust for BMI-specific bias(Reference Ferrari, Slimani and Ciampi14), would be a way out of that dilemma needs further evaluation.
Under-reporting of protein intake is also indicative of a degree of under-reporting of EI, while the use of LER does not necessarily identify under-reporting of protein intake to the same extent due to differential reporting of foods(Reference Livingstone and Black8). This implies that energy-adjusted protein intake may still differ between LER and non-LER. In contrast to previous studies(Reference Bingham, Cassidy and Cole24, Reference Voss, Kroke and Klipstein-Grobusch35), we found similar energy-adjusted protein intake (% kJ) across BMI groups, confirming that energy adjustment may indeed remove some of the measurement error(Reference Kipnis, Subar and Midthune2). This is in line with findings of Lissner et al. (Reference Lissner, Troiano and Midthune15), who showed that correlations between reported protein intake and protein biomarker in obese groups were improved relative to non-obese after energy adjustment. However, this does not exclude non-measured relative changes among the other macronutrients – i.e. fat, carbohydrate and alcohol – due to possible differential reporting of foods.
Compared with protein intake, BMI-related under-reporting of K intake was of a smaller magnitude at the individual level on both instruments, suggesting some food-specific misreporting. The weaker or absent group-level evidence of BMI-related K under-reporting is presumably attributable to loss of power due to categorisation. We explored K misreporting also by evaluating under-reporting of the ratio of K:protein as a proxy for energy-adjusted K intake, which confirmed that relative to protein, K under-reporting was less in magnitude, but still significantly predicted by BMI at the individual level on both methods and in both sexes (data not shown). Differential reporting of nutrients is possibly related to differential reporting of its food sources. In the EPIC study populations, the major food sources of protein and K differ to some extent, thus differential under-reporting of these nutrients is conceivable. Food sources for K from diet are fruits and vegetables (approximately 26 %), non-alcoholic beverages (approximately 14 %), cereals and products (approximately 13 %), meats and products (approximately 13 %) and dairy products (approximately 12 %)(Reference Welch, Fransen and Jenab36). In contrast, protein sources are between 55 and 73 % of animal origin (not counting the UK health-conscious group) and between 24 and 39 % of plant origin, mostly from cereals (approximately 20 %) and less from fruits and vegetables (approximately 6 %)(Reference Halkjaer, Olsen and Bjerregaard37). Although it was not possible to determine whether specific foods were systematically under- or over-reported, we know from the literature that DQ tend to overestimate specific foods, particularly vegetables(Reference Michels, Welch and Luben11, Reference Goldbohm, van den Brandt and Brants38) and that LER tend to report more socially desirable foods, including fruits and vegetables(Reference Livingstone and Black8). We may speculate that these ‘good’ foods, which tend to have a higher K content, are more often over-reported. However, as with almost all other nutrients, K intake is also related to EI (Pearson's r between intake of K and energy on 24-HDR and DQ were both 0·64, both P < 0·001). Thus, the proportion of LER who simultaneously under-report K intake is still higher when compared with LER who differentially report K intake and those individuals are more likely to be overweight or obese. The latter is supported by the results of Zhang et al. (Reference Zhang, Temme and Sasaki28), who showed that under-reporting of K intake was more prevalent in subjects with a higher BMI while subjects with a lower BMI more frequently over-reported.
The differential reporting of protein and K intake might also be attributable to differences in the validity of the two respective biomarkers. There is more uncertainty about the proportion of K intake excreted into urine when compared with N. We found a considerable between-study variation of urinary K excretion (0·76–0·89)(Reference Tasevska, Runswick and Bingham19, Reference Mickelsen, Makdani and Gill39, Reference McDonald and Margen40–Reference Dériaz, Thériault and Lavallée44), and we used the crude mean (0·81) of those studies to correct for extra-renal losses of K. If this factor was, for example, too large, then K excretion would be underestimated, leading to a false conclusion that K intake was not or to a lesser extent under-reported. However, the use of different correction factors (data not shown) did not change the strength of the association between BMI and K under-reporting. This has similarly been shown by Zhang et al. (Reference Zhang, Temme and Sasaki28).
Similar to the results of the Observing Protein and Energy Nutrition study(Reference Schatzkin, Kipnis and Carroll3, Reference Subar, Kipnis and Troiano10), we found that a single 24-HDR predicted protein intakes at the individual level almost as well as DQ across BMI categories including obese subgroups and outperformed DQ in predicting K intakes (results of stratification by BMI group are not shown, but they are similar to the combined analysis). Differences in the performance of the two dietary methods in the present study may be explained by the high degree of standardisation of the 24-HDR across countries within the EPIC calibration study when compared with country-specific DQ. For other nutrients, with higher daily within-person variation in intakes, DQ might perform better relative to 24-HDR.
The two dietary self-report instruments differed slightly between men and women in estimating mean intakes, although the effect of BMI on misreporting was the same for both sexes. In women, the DQ seems to yield higher mean intakes of protein and K that were closer to the biomarker-based measurements across BMI categories when compared with 24-HDR.
Among the limitations of the study are that the convenience sample was not recruited from all centres of the EPIC study. Thus, caution should be taken in the extrapolation of the data to the whole EPIC cohort or general population. Furthermore, due to the small number of participants at the country (centre) level, results may be less consistent in the stratified analysis. Another limitation concerns the collection of the 24 h urine specimens. At least 189 (13·6 %) subjects did not succeed in obtaining a 24 h urine collection. In addition, a large proportion of these subjects were obese. Although the time interval between collection of urinary and dietary measurements varied considerably across centres, particularly on DQ, respective effects were estimated to be probably very modest(Reference Slimani, Bingham and Runswick9).
In conclusion, under-reporting of protein and K intake appears to be predicted by BMI, but this effect may be driven mostly by individuals who probably misreported EI. The BMI effect on protein and K under-reporting seems to be the same across countries despite their diverse dietary patterns and other cultural differences. These results may aid in the interpretation of nutritional epidemiological findings and suggest that future analyses of EPIC data should be accompanied by a sensitivity analysis to explore the robustness of results obtained with and without exclusion of energy misreporters on DQ and, in case regression calibration is used, on 24-HDR.
Acknowledgements
PANACEA (Physical Activity, Nutrition, Alcohol, Cessation of smoking, Eating out of home And obesity) uses data collected as part of the EPIC, co-financed by the European Commission (PANACEA no. 2005328). The present study was financially supported by the European Commission: Public Health and Consumer Protection Directorate 1993–1994; Research Directorate General 2005–; Ligue contre le Cancer, Societé 3M, Mutuelle Générale de l'Education Nationale, Institut National de la Santé et de la Recherche Médicale (INSERM) (France); German Cancer Aid, German Cancer Research Center, Federal Ministry of Education and Research (Germany); Danish Cancer Society (Denmark); Health Research Fund (FIS) of the Spanish Ministry of Health, the participating regional governments and institutions and the ISCIII (Instituto de Salud Carlos III) of the Spanish Ministry of Health (RETICC DR06/0020) (Spain); Cancer Research UK, Medical Research Council, Stroke Association, British Heart Foundation, Department of Health, Food Standards Agency, the Wellcome Trust (UK); Greek Ministry of Health and Social Solidarity, Hellenic Health Foundation and Stavros Niarchos Foundation (Greece); Italian Association for Research on Cancer, National Research Council (Italy); Dutch Ministry of Public Health, Welfare and Sports (VWS), Netherlands Cancer Registry (NKR), LK Research Funds, Dutch Prevention Funds, Dutch ZON (Zorg Onderzoek Nederland), World Cancer Research Fund (WCRF) (The Netherlands); Swedish Cancer Society, Swedish Scientific Council, Regional Government of Skane (Sweden); Nordforsk (Norway). H. F. contributed to the analysis and interpretation of the data and wrote the manuscript, taking into account comments from all co-authors. M. M. E. v. B., C. B., G. B., A. M. M., P. H. P., T. N., S. R. and N. S. contributed to the conception, analysis and interpretation of the data and drafting of the manuscript. E. R. is the overall coordinator of the EPIC study and all other co-authors were local EPIC collaborators involved in the collection of dietary and other data. All of them gave input on drafting of the manuscript and interpretation of the results. All authors read and approved the final version. The authors declare no conflict of interest.
