S71 STEFANIE BOHLSCHEID-THOMAS,* INA HOTING,** HEINER BOEING AND JÜRGEN WAHRENDORF*

Transcription

1 International Journal of Epidemiology International Epidemiological Association 1997 Vol. 26, No. 1(Suppl.1) Printed in Great Britain Reproducibility and Relative Validity of Energy and Macronutrient Intake of a Food Frequency Questionnaire Developed for the German Part of the EPIC Project STEFANIE BOHLSCHEID-THOMAS,* INA HOTING,** HEINER BOEING AND JÜRGEN WAHRENDORF* Bohlscheid-Thomas S (Division of Epidemiology 0345, German Cancer Research Centre, Im Neuenheimer Feld 280, Heidelberg, Germany), Hoting I, Boeing H, Wahrendorf J. Reproducibility and relative validity of energy and macronutrient intake of a food frequency questionnaire developed for the German part of the EPIC project. International Journal of Epidemiology 1997; 26 (Suppl. 1): S71 S81. Background. For use in the German part of the European Prospective Investigation into Cancer and Nutrition (EPIC) a self-administered, optically readable food frequency questionnaire (FFQ) was developed to assess individuals usual food and nutrient intake over the past year. The food list comprised 158 food items for which the typical portion size, the consumption frequency (1 6 times) and the time period (day, week, month, or year) were requested. This paper reports the results of the reproducibility and relative validity of nutrient intakes obtained by this instrument. Methods. The study was carried out from October 1991 to October During this period, 104 men and women, aged years, completed a 24-hour recall once a month, and the FFQ in the middle and at the end of this study. A short questionnaire on global consumption patterns was used to adjust the food consumption frequency given in the FFQ. Results. Reproducibility correlations varied from 0.59 for saturated fat to 0.88 for alcohol, with most values falling between 0.60 and Energy-adjustment decreased the observed correlations for all nutrients. Correlations between nutrient intake values from the 12-day average of dietary recalls and the FFQ corrected for food consumption frequency ranged from 0.42 for polyunsaturated fat to 0.88 for alcohol, followed by 0.65 for dietary fibres. After de-attenuation and energy-adjustment the range of correlations was somewhat wider ( , without alcohol). With regard to classification, on average, 36% of subjects fell into the same quintile, and 75% into within-one quintile when classified by the frequency-corrected FFQ compared to the quintiles based on recalled data. Conclusions. In summary, the newly developed FFQ showed fairly good reproducibility and relative validity for most evaluated nutrients. The between-subject variation of nutrients in the German cohort (including the participants from East Germany) is likely to be greater than that among the subjects of the pilot phase. A revised version of the FFQ corrected for food consumption frequency is used in the EPIC study. Keywords: food frequency questionnaire, diet recalls, validity, reproducibility, nutrient intake measurements, Germany, EPIC A self-administered, optically readable food frequency questionnaire (FFQ) was considered the most suitable method to assess individual food and nutrient intake in the German part of the European Prospective Investigation into Cancer and Nutrition (EPIC). 1 An FFQ can be administered easily and optical reading speeds up data processing. The ability to measure individual diet * Division of Epidemiology 0345, German Cancer Research Centre, Im Neuenheimer Feld 280, Heidelberg, Germany. ** Institute for Medical Information Processing, Biometry and Epidemiology, University of Munich, Germany. Unit of Medical Epidemiology, German Institute of Human Nutrition Potsdam-Rehbrücke, Germany. S71 and nutrient intake needs to be evaluated before an instrument can be used in a study. During the pilot phase of EPIC the FFQ was tested for its reproducibility and validity. Reproducibility was assessed by administering the same dietary questionnaire 6 months later. The relative validity was tested by comparing the FFQ data with the reference measurements of hour dietary recalls. This paper reports the results of the reproducibility and relative validity of energy and macronutrient intake assessed by our FFQ. Results of the reproducibility and validity of food group intakes, and results of the absolute validity of vitamin intakes are given in two accompanying papers. 2,3

2 S72 INTERNATIONAL JOURNAL OF EPIDEMIOLOGY SUBJECTS AND METHODS Subjects According to the eligibility criteria for the EPIC study, men and women aged years were selected for the pilot study. The study participants were recruited in Germany from the local health insurance institution, AOK Heidelberg (Allgemeine Ortskrankenkasse). Of 528 subjects who received a letter of invitation, 115 people agreed to participate in the reproducibility and validity study. These participants motivated an additional eight people to take part in this study. The eight were included in this study because they met all criteria for recruitment. From the 123 participants, 49 men and 55 women completed the study. Study Design The study was conducted in Heidelberg over a one-year period (October 1991 to October 1992). During the data collection period a 24-hour dietary recall was taken every month. The 12 days of recall were balanced across the days of the week for each participant, with the exception of Friday and Saturday which could not be recalled for logistical reasons. The FFQ was completed twice by the participants with a 6-month interval (February and August 1992). In January 1993 all participants were asked to complete a short questionnaire on general consumption patterns in which the consumption frequency per week of 14 selected food groups was requested. A more complete description of the study design can be found elsewhere. 2 Food Frequency Questionnaire (FFQ) The FFQ which was developed for self-administration and optical reading, asked for usual food intake over the past year. The food list comprised 158 food items which were originally selected according to their contribution to the inter-individual variability of food intake. For each food item, the questionnaire asked for the typical portion size, the consumption frequency (1 6 times) and the time period (day, week, month, or year). For food items which were not commonly consumed in household measures, the FFQ included coloured photographs to help participants in describing their usual portion sizes. Questions about the type of fat for cooking, the fat content of milk products, and added sugar or milk to hot beverages were also integrated into the FFQ. Food consumption in season was requested for fresh fruit and tomatoes. A more detailed description of the development of this FFQ is given in the companion paper. 2 The FFQ was completed at home without further instructions, and then reviewed by an interviewer together with the participant. Food coding and the calculation of nutrient intake was done using the Federal Food Code. 4 The nutrient content of each food item in the FFQ was estimated as a weighted average. The weights were taken from the 24-hour dietary recall data. The information given in the questionnaire on general consumption patterns was used to adjust for possible over- or underestimation of the consumption frequencies reported in the FFQ. The global consumption frequency given in the short questionnaire was compared to the sum of the consumption frequencies for single food items belonging to a food group calculated from the FFQ. If the sum of the frequency of the individual food items differed from the global consumption frequency, an adjustment was made for the frequencies in a linear manner. Weight factors were calculated as the general consumption frequency divided by the sum of the single frequencies. The adjustment was done by multiplying the single frequencies by their weights. Since the short questionnaire was administered only once, the correction was done only for the second FFQ which covered the same time period. 24-hour Dietary Recalls The 24-hour dietary recalls were carried out by two trained interviewers at the German Cancer Research Centre in Heidelberg. The participants were asked to recall every food item consumed the previous day. Serving sizes were estimated by means of coloured photographs (as described above), and by household measures. The intake of vitamin and mineral supplements, drugs, and other medications was also asked. All foods consumed were coded and transformed to nutrient intakes on the basis of the Federal Food Code. 4 Statistical Analysis Mean daily nutrient intakes, mean differences, and the standard deviation of the differences (SD diff ) were computed for the FFQ and for the 24-hour diet recall. 5 Nutrient intakes were logarithmically transformed to improve their distributions towards normality. Pearson correlation coefficients were calculated to assess the reproducibility, and to evaluate the relative validity of the questionnaires (FFQ2, FFQ2corr) compared with the 24-hour diet recall. Correlation coefficients were computed as unadjusted values, and then adjusted values for attentuation due to within-person variability in the 24-hour diet recall, 6 and de-attenuated values adjusted for total energy intake. Energy adjustment was done using the regression approach. 7 The distribution of nutrient intakes derived from the 24-hour diet recall and

3 RELATIVE VALIDITY OF MACRONUTRIENT INTAKE, GERMANY S73 TABLE 1 Comparison of measurements of mean daily energy and nutrient intake estimated by 12 dietary recalls and by dietary questionnaires. Data were provided by 104 individuals Nutrients FFQ1 a FFQ2 b FFQ2corr c 24HDR d mean SD mean SD mean SD mean SD Energy (kj) Energy (kcal) Total protein (g) Total fat (g) Saturated fat (g) Monounsaturated fat (g) Polyunsaturated fat (g) Cholesterol (mg) Total carbohydrates (g) Monosaccharides (g) Disaccharides (g) Polysaccharides (g) Dietary fibre (g) Alcohol (g) % energy from protein % energy from fat % energy from carbohydrates % energy from alcohol a First food frequency questionnaire. b Second food frequency questionnaire. c Second food frequency questionnaire corrected for food consumption frequency. d Dietary recalls. from the FFQ were divided into quintiles to evaluate the effect of measurement error on the ranking of individuals. Cutpoints were defined separately for the FFQ and the 24-hour diet recall. The mean intakes from the 24-hour diet recall data were estimated for the subjects in each quintile defined by the FFQ intakes. The degree of misclassification was presented as the proportion of subjects classified by the reference method into the same, the adjacent, or even the extreme quintile. Misclassification into the extreme quintile comprises both misclassification from the first to the fifth quintile, and vice versa; from the fifth to the first quintile. RESULTS Reproducibility of Questionnaire Measurements Comparison of measurements of mean daily energy and nutrient intakes estimated by FFQ1 and FFQ2 are presented in Table 1. For 10 out of 13 nutrients, we found differences in mean daily intakes within a range of ± 10%. For the remaining three nutrients, the FFQ2 gave 11 14% lower values than the FFQ1. Mono- and polyunsaturated fat, as well as mono- and disaccharides showed lower values for FFQ2 than for FFQ1. There were no gender differences in per cent overestimation in nutrient intake found (results not presented). Table 2 shows the mean difference and the SD diff calculated from the two FFQ. Nine nutrients showed a mean difference significantly different from zero. Only cholesterol, polysaccharides, dietary fibre and alcohol were not significantly different from zero. For women nearly all nutrients showed mean differences significantly different from zero, while for men no significant mean differences were found, except for polyunsaturated fat. The SD diff is a measure of random withinsubject error. In general, the SD diff was large for all nutrients. The largest variability with respect to the average mean intake from both questionnaires was observed for monosaccharides and alcohol. No gender differences in SD diff were found (data not presented). Pearson correlation coefficients between the FFQ are shown in Table 3. The correlation coefficients ranged from 0.59 for saturated fat to 0.88 for alcohol, with most values between 0.6 and 0.7. Energy-adjustment slightly reduced the correlation coefficients for all nutrients. After energy-adjustment the correlations varied from 0.37 for polysaccharides to 0.68 for disaccharides.

4 S74 INTERNATIONAL JOURNAL OF EPIDEMIOLOGY TABLE 2 Mean difference and standard deviation of the differences estimated by 12 dietary recalls and by dietary questionnaires. Data were provided by 104 individuals Nutrients Mean difference SD diff e FFQ1 a versus 24HDR c versus 24HDR versus FFQ1 versus 24HDR versus 24HDR versus FFQ2 b FFQ2 FFQ2corr d FFQ2 FFQ2 FFQ2corr Energy (kj) f f Total protein (g) 6.8 f 11.8 f Total fat (g) 10.5 f 8.4 f Saturated fat (g) 3.8 f Monounsaturated fat (g) 3.6 f f Polyunsaturated fat (g) 1.8 f 1.5 f 0.9 f Cholesterol (g) Total carbohydrates (g) 25.4 f 60.5 f 37.7 f Monosaccharides (g) 7.7 f 14.3 f Disaccharides (g) 11.7 f 25.1 f 18.6 f Polysaccharides (g) f 16.7 f Dietary fibre (g) f 3.1 f Alcohol (g) a b c d e First food frequency questionnaire. Second food frequency questionnaire. Dietary recalls. Second food frequency questionnaire corrected for food consumption frequency. Standard deviation of the differences. f Mean difference significantly different from zero (P 0.05). The analysis by sex (data not shown) revealed higher reproducibility correlations for all nutrients for men than for women, when no adjustment was done. Energy-adjustment reduced all correlations in men, while in women for total protein, polyunsaturated fat, cholesterol, dietary fibre, and alcohol higher correlations were observed. Gender differences in energy-adjusted nutrient intake estimates were found for total protein, monounsaturated fat, cholesterol, and disaccharides. For those nutrients we observed correlations for men 0.52, 0.36, 0.41, 0.76, and for women we found values of 0.71, 0.53, 0.77, 0.53 respectively. Relative Validity of Questionnaire Measurements A comparison between the mean daily nutrient intakes estimated by each of the FFQ and by the average of the 12 dietary recalls is presented in Table 1. Both FFQ gave higher nutrient intake estimates than the 24-hour diet recall. This overestimation was larger for the FFQ1 than for the FFQ2. Looking at FFQ1, only alcohol showed a difference smaller than 10%, while the other nutrients were overestimated by 17 63%. However, the FFQ2 measurements were within 10% for five nutrients compared to the 24-hour diet recall. The remaining eight nutrients gave values 13 41% higher than the 24-hour diet recall estimates. Energy (17%), total protein (16%), total carbohydrates (27%), monosaccharides (41%), disaccharides (39%), polysaccharides (15%), and dietary fibre (37%) were overestimated by the FFQ2. No gender differences were found. The mean difference and the SD diff were calculated only for the FFQ2, since it covered the same time period as the 24-hour diet recall (Table 2). Both statistical criteria gave information about the presence of within-person reporting bias. A mean significantly different from zero was observed for energy, total protein, total fat, polyunsaturated fat, total carbohydrates, mono-, di-, and polysaccharides, as well as for dietary fibre. Analysis by sex showed significant mean differences for most nutrients for men, but for women only seven nutrients showed significant mean differences (data not presented). The SD diff were large for all nutrients. The largest variabilities in differences between methods with regard to the average mean intake from both methods were seen for mono- and disaccharides, dietary fibre, and alcohol. There were no gender differences in SD diff (data not shown). Pearson correlation coefficients between the FFQ2 estimates and the 12-day average from the 24-hour diet recall are presented in Table 3. Correlations were only moderate ( ) for unadjusted variables, with the exception of alcohol (r = 0.88). Table 4 presents the

5 RELATIVE VALIDITY OF MACRONUTRIENT INTAKE, GERMANY S75 TABLE 3 Pearson correlation coefficients a of energy and nutrient intake between the first and the second dietary questionnaire and between questionnaire and recall measurements for 104 individuals Nutrients Correlation coefficients b FFQ1 c versus FFQ1 versus 24HDR e versus 24HDR versus 24HDR versus 24HDR versus 24HDR versus 24HDR versus FFQ2 d FFQ2 FFQ2 FFQ2 FFQ2 FFQ2corr f FFQ2corr FFQ2corr unadjusted energy adj g unadjusted deatt h deatt, unadjusted deatt deatt. energy adj i energy adj Energy (kj) Total protein (g) Total fat (g) Saturated fat (g) Monounsaturated fat (g) Polyunsaturated fat (g) Cholesterol (mg) Total carbohydrates (g) Monosaccharides (g) Disaccharides (g) Polysaccharides (g) Dietary fibre (g) Alcohol (g) a b c d e 95% confidence intervals for N = 104 are 0.10, 0.23 for r = 0.1; 0.01, 0.38 for r = 0.2; 0.12, 0.47 for r = 0.3; 0.25, 0.55 for r = 0.4; 0.34, 0.63 for r = 0.5; 0.46, 0.71 for r = 0.6; 0.59, 0.79 for r = 0.7; 0.72, 0.86 for r = 0.8; 0.86, 0.93 for r = 0.9. g h Based on log e -transformed values. First food frequency questionnaire. Second food frequency questionnaire. Dietary recalls. Second food frequency questionnaire corrected for food consumption frequency. Correlation coefficients energy-adjusted. Correlation coefficients de-attenuated (corrected for within-person variation in dietary recall data). Correlation coefficients energy-adjusted and de-attenuated. f i

6 S76 INTERNATIONAL JOURNAL OF EPIDEMIOLOGY TABLE 4 Variance components of the logarithms of nutrient intakes estimated by 12 dietary recalls. Data were provided by 104 individuals Nutrients 24HDR a Within- Between- Variance person person ratio b variation variation Energy Total protein Total fat Saturated fat Monounsaturated fat Polyunsaturated fat Cholesterol Total carbohydrates Monosaccharides Disaccharides Polysaccharides Dietary fibre Alcohol a b Dietary recalls. Variance components were calculated as within-person variance divided by between-person variance. within- and between-person variabilities, as well as the variance ratios. We used the log e -transformation to obtain approximately normal distributed data in order to use standard methods of variance component analysis. 8 High variance ratios were found for total fat, saturated fat, mono- and polyunsaturated fat, cholesterol, and monosaccharides; due to the high withinperson variabilities and low between-person variabilities. The adjustment of correlation coefficients for withinperson variability in the 24-hour diet recall improved all correlations, but they were still modest ( , without alcohol). The greatest effect of correction for within-person variability was seen for the nutrients with the largest variance ratios. The de-attenuated and energy-adjusted correlations ranged from 0.32 for polyunsaturated fat to 0.95 for alcohol, with most values between 0.50 and The FFQ2 provided reasonably good correlations for total fat (0.67), saturated fat (0.74), dietary fibre (0.66), and alcohol (0.95). Separate analyses for men and women (data not presented) showed higher correlations in men than in women, when unadjusted values were used. Adjustment for withinperson error improved all correlations both for men and for women, but the correlations in men were still higher than in women. De-attenuated and energy-adjusted correlations were higher than de-attenuated correlations for both sexes. Gender differences were still obvious for some nutrients, namely for polyunsaturated fats, disaccharides, and dietary fibre. Table 5 describes the actual level, and the range, of exposure in nutrient intakes derived from 24-hour diet recall. The range of nutrient intakes was modest (2.0 to 2.5-fold) for most nutrients. Monosaccharides (3.0- fold), disaccharides (4.2-fold), and alcohol (103.0-fold) showed the largest exposure differences. To assess the nutrient intake ranges that could be discriminated by FFQ2, we calculated the nutrient intake from the 24-hour diet recall using the quintiles defined by the FFQ2. For all nutrients, we observed nutrient intake ranges less than for quintiles defined by the 24-hour diet recall themselves (1.3 to 2.0-fold, without alcohol), and for some nutrients we found a non-linear relationship. For example, the FFQ2 was not able to discriminate between the third (38.3 g) and the fourth quintiles (36.6 g) of monosaccharide intake, but could nevertheless discriminate a 1.5-fold range. Nutrients with a non-linear relationship were energy, saturated and polyunsaturated fat, mono- and disaccharides, as well as dietary fibre. The proportion of participants classified by the FFQ2 into the same 24-hour diet recall quintile ranged from 25% for total protein to 58% for alcohol, followed by 34% for polysaccharides (Table 6). The classification into the adjacent category was on average 37%. Gross misclassification was rare, with a maximum of 6% for monosaccharides. Classification into the same category was slightly better in males than in females. Classification within-one category was comparable between men and women. Adjustment for Consumption Frequency The effects of adjustment for food consumption frequency on the mean nutrient intake, the mean difference, and the SD diff are presented in Tables 1 and 2. After correction most nutrients showed differences of less than 10% compared to the 24-hour diet recall data. Overestimation was obvious for total carbohydrates (17%), disaccharides (29%) and polysaccharides (13%), and dietary fibre (13%). No gender difference was observed. A mean difference significantly different from zero was seen for mono- and polyunsaturated fats, total carbohydrates, di- and polysaccharides, and dietary fibre. The effect of frequency correction on the mean difference was better in men than in women. After frequency correction, we found only four nutrients with a mean difference significantly different from zero in men, but in women there were seven nutrients significantly different from zero. For all nutrients the SD diff were reduced after correction, indicating a decrease in within-subject measurement error.

7 RELATIVE VALIDITY OF MACRONUTRIENT INTAKE, GERMANY S77 TABLE 5 Mean daily energy and nutrient intake derived from hour dietary recalls for quintiles of intake defined by the recalls and by quintiles defined by the second food frequency questionnaire (N = 104 individuals) Nutrients Mean 24HDR, a 24HDR-cutpoints Mean 24HDR, FFQ2 b -cutpoints Mean 24HDR, FFQ2corr c -cutpoints Quintile Energy (kj) Total protein (g) Total fat (g) Saturated fats (g) Monounsaturated fat (g) Polyunsaturated fat (g) Cholesterol (mg) Total carbohydrates (g) Monosaccharides (g) Disaccharides (g) Polysaccharides (g) Dietary fibre (g) Alcohol (g) a b c Dietary recalls. Second food frequency questionnaire. Second food frequency questionnaire corrected for food consumption frequency.

8 S78 INTERNATIONAL JOURNAL OF EPIDEMIOLOGY TABLE 6 Comprison of food frequency questionnaire with daily energy and nutrient intakes derived from hour dietary recalls, based on joint classification by quintiles (n = 104 individuals) Nutrients FFQ2 a versus 24HDR b FFQ2corr c versus 24HDR same category adjacent category extreme category same category adjacent category extreme category (%) (%) (%) (%) (%) (%) Energy (kj) Total protein (g) Total fat (g) Saturated fat (g) Monounsaturated fat (g) Polyunsaturated fat (g) Cholesterol (g) Total carbohydrates (g) Monosaccharides (g) Disaccharides (g) Polysaccharides (g) Dietary fibre (g) Alcohol (g) a b c Second food frequency questionnaire. Dietary recalls. Second food frequency questionnaire corrected for food consumption frequency. The Pearson correlations between the FFQ2corr and the 24-hour diet recall are shown in Table 3. The correlation coefficients increased substantially after frequency correction for all nutrients, except for alcohol intake which was not corrected. The correlations ranged from 0.42 for polyunsaturated fats to 0.65 for disaccharides. If the frequency corrected correlations were additionally adjusted for within-person variability all values improved again ( ). Energy-adjusted and de-attenuated correlations were somewhat lower, except for total fat and saturated fat, polysaccharides, dietary fibre, and alcohol. Gender differences (data not shown) were found for mono- and polyunsaturated fat, as well as for disaccharides. For those nutrients we observed higher correlations for males than for females. The range in nutrient intake that could be discriminated by the FFQ increased slightly after frequency correction, indicating a reduction in ranking errors (Table 5). For example, the FFQ2 discriminated a 2.0-fold difference of disaccharide intake, while after correction the measurable contrast increased to a 2.6- fold difference. The comparison based on classification into quintiles is summarized in Table 6. For energy, total protein, cholesterol, total carbohydrates, di- and polysaccharides, as well as dietary fibre, the proportion of subjects equally classified was improved. The remaining nutrients showed a slight reduction in the proportion of participants classified into the same category. Overall, the classification into the within-one category was improved for all corrected nutrients (on average 74.5%). DISCUSSION Study Subjects We have evaluated the reproducibility and relative validity of nutrient intake assessed by a newly developed FFQ among 104 men and women from the Heidelberg area. A matter of potential concern in the interpretation of the results is the representativeness of the participants of the pilot study relative to the entire cohort. After German unification, it was decided to establish a second study centre in the eastern part of Germany in order to increase the heterogeneity of food and nutrient intake of the German cohort. In this East German cohort a separate validity study has started. The low participation rate (22%) in the present validation study indicates the selection of highly motivated subjects. Thus, the population for this study may have enhanced the reproducibility and validity found for our FFQ. Reproducibility Overall, the reproducibility of our FFQ completed 6 months apart was good or moderately good, with Pearson correlations for most nutrients varying between

9 RELATIVE VALIDITY OF MACRONUTRIENT INTAKE, GERMANY S and The most repeatable estimate was obtained for alcohol intake (0.88) followed by the intake of disaccharides (0.74), while the lowest correlation was observed for saturated fat (0.59). Energy-adjustment reduced the reproducibility correlations for all evaluated nutrients. A substantial reduction was found for total fat, monounsaturated fat, cholesterol, total carbohydrates, mono- and polysaccharides, as well as for alcohol. Pearson correlation coefficients depend on both the within-subject random error of the FFQ and the between-subject variability in nutrient intakes. Energyadjustment decreased the between-subject variability in nutrient intakes due to controlling for total energy intake, and therefore resulted in lower correlation coefficients. The reproducibility of FFQ has been examined under a wide variety of conditions which complicates the comparison of results between studies. Previous studies reviewed by Willett 9 have shown that the FFQ method gives reproducible nutrient intake estimates for middleaged subjects, with correlations usually ranging from 0.50 to More recent studies confirm this finding Bueno-de-Mesquita et al. 10 conducted two reproducibility studies with different designs. In the first study, repeated questionnaires covering the same time period were obtained using 63 males and females aged years. The original FFQ referred to a time period approximately one year previously and the second one covered the food intakes of the preceding 2 years. The correlations ranged from 0.62 for β-carotene to 0.85 for energy, and 0.91 for ethanol. In the second study, they examined the agreement between original and repeated interviews when the time interval and the period of interest was one year. With this study design, they found correlations between 0.48 for β-carotene and 0.76 for energy. Rimm et al. 11 assessed the reproducibility of an 131-food item semiquantitative FFQ used in a cohort study among male health professionals aged years. The FFQ was administered by mail twice to a sample of 127 participants at a oneyear interval. Intra-class correlations for nutrient intake ranged from 0.47 for vitamin E (without supplements) to 0.80 for vitamin C (including supplements). Martin- Moreno et al. 12 tested a self-completed FFQ with 118 food items in a population of 147 Spanish women aged years. The FFQ with no portion size information were filled in twice with a one-year period between estimates. Pearson correlations ranged from 0.51 for saturated fat to 0.88 for alcohol, followed by 0.73 for energy. Longnecker et al. 13 evaluated the reproducibility of the Harvard-Willett questionnaire with 116 food items. The study population (middle-aged males and females), and also the time interval between repeated administration (6 12 months) were quite similar to our study, but their FFQ lacked variable portion sizes and responses were limited to nine response categories. The Pearson correlation coefficients varied between 0.39 for energy intake to 0.81 for alcohol, with a median correlation of In the pilot phase of the Greek component of the EPIC study, an FFQ with 190 food items was evaluated for its reproducibility and validity with a nearly identical design as our study. 14 Their FFQ looked somewhat different, because portion sizes were estimated by natural units or standard quantities, and frequency was asked as number of times per month, week, or day. Intra-class correlations varied from 0.38 for mono- and disaccharides to 0.78 for alcohol. Bittoni et al. 15 assessed the reproducibility of a self-administered FFQ developed at the National Cancer Institute (Bethesda, MD) in a population of breast cancer cases and controls. The FFQ included 98 food items, and asked either for foods consumed during the past year (for controls) or for foods consumed during the year prior the illness (for cases). The women with a shorter time interval had higher coefficients ( ). Ajani et al. 16 examined the reproducibility of a semiquantitative FFQ designed for the use in an eye disease case-control study by re-administration of the same FFQ approximately months later. The FFQ including 60 food items was self-administered by 325 men and women aged years. Pearson correlations for the log of calorie-adjusted nutrient intake ranged from 0.38 to 0.78, with nearly all correlations between 0.5 and 0.8. In the Netherlands, Goldbohm et al. 17 conducted a reproducibility study in which the 5-year reproducibility of a self-administered 150-food item FFQ was determined. Energy-adjusted test-retest correlations ranged from 0.42 for selenium intake to 0.89 for alcohol intake. Results observed in our study confirmed the findings of the other studies, with Pearson correlations ranging from 0.59 for saturated fat to 0.88 for alcohol, followed by disaccharides (0.74). The median of reproducibility correlations was Relative Validity This study also examined the validity of the FFQ relative to hour diet recalls carried out over a one-year period to account for seasonal as well as shortterm variability. Twenty four hour dietary recalls are sometimes criticized 18 because they rely on the subject s ability to remember all foods consumed during the previous day. Subjects may remember foods which were not really eaten or may forget to recall some foods actually consumed. Comparison of the 24-hour diet recall method with more objective measurements such as the doubly-labelled water method demonstrated

10 S80 INTERNATIONAL JOURNAL OF EPIDEMIOLOGY relative underreporting of habitual energy intake by 24-hour diet recall. 19 The 24-hour diet recalls collected in this study were checked for underreporting by calculating the ratio of energy intake to basal metabolic rate. 20 From the 104 participants, 12 subjects were identified as underreporters of habitual energy intake. Since, there were no big differences between the results with and without these underreporters, the analyses presented here are based on all participants. Another point which must be considered in the interpretation of our results is the possible bias induced by not collecting data from Friday and Saturday. For logistic reasons it was not possible to collect 24-hour diet recalls for these two days. Friday and Saturday are typical consumption days for salty snacks, fish, soups, and alcoholic beverages. It is possible that the 24-hour diet recalls therefore underestimate nutrient intakes, and especially energy intake. The short questionnaire on general consumption patterns was used to correct the frequencies given in the FFQ2. This correction improved the estimates of individual energy and nutrient intake compared to the 24-hour diet recall. In particular, we found smaller differences in estimated mean intakes, lower SD diff, higher correlation coefficients, and a reduction in ranking error as well as misclassification. The following discussion, therefore, refers to the results obtained by the FFQ2corr. Mean nutrient intakes were overestimated by the FFQ2corr relative to the 24-hour diet recall; by 17% for total carbohydrates, 29% for disaccharides, 13% for polysaccharides, and 13% for dietary fibre. No meaningful differences (± 10%) in population estimates were obtained. De-attenuated correlations between methods were good, and varied from 0.49 for polyunsaturated fat to 0.69 for dietary fibre (excluding alcohol r = 0.94). Values were somewhat lower after energy-adjustment, except for total fat, saturated fat, polysaccharides, dietary fibre, and alcohol. The correlations observed in this study were in the same range as those reviewed by Willett 9 as well as those reported by more recent studies ,21,22 Rimm et al. 11 validated their FFQ against two one-week diet records spaced approximately 6 months apart. Correlation coefficients between energy-adjusted nutrient intakes measured by the diet records and by the second FFQ ranged from 0.28 for iron (without supplements) to 0.86 for vitamin C (with supplements). The correlations were higher after adjusting for week-to-week variation in dietary records. The FFQ in the study from Martin-Moreno et al. 12 was validated against 4-day records which were obtained at 3-month intervals. Pearson correlations between the diet record and the second FFQ ranged between 0.35 for monounsaturated fat and vitamin A to 56 for vitamin C, except for alcohol (0.89). Adjustment for week-to-week variation in diet record intake increased all correlations. In the Malmö food study, 21 a self-administered food use questionnaire with 250 food items was tested. They used as a reference method 18 days of weighed food records. Their results showed Pearson correlations with the reference method in the order of 0.40 to 0.60 for most nutrients after adjustment for energy intakes. The Harvard-Willett instrument 13 was compared to multiple one-day diet records (mean = 5 days during a 6-month to one-year period) in subjects from South Dakota and Wyoming. Energy-adjusted Pearson correlations between the intakes estimated by diet records and the second FFQ were between 0.23 for vitamin C and 0.57 for calcium (median = 0.42). Adjustment for measurement error increased all correlations (median = 0.52). In the Netherlands, Goldbohm et al. 22 validated a selfadministered FFQ used in a large cohort study on diet and cancer. A dietary record kept over three 3-day periods, 4 to 5 months apart, served as reference method. Pearson correlations between nutrient intakes assessed by the record and the FFQ that was completed afterwards ranged from 0.40 for vitamin B 1 to 0.86 for alcohol intake, with correlations for most nutrients between 0.60 and Adjustment for energy intake and sex did not materially affect these correlations. Correction of correlation coefficients for attenuation by day-to-day variance in the record data improved them by 0.07 points on average. In the Greek pilot phase of the EPIC study, 15 the FFQ was compared with the mean of hour diet recalls as in our study. The Greek researchers found, for males, energy-adjusted correlations ranging from 0.23 for polyunsaturated fat to 0.52 for cholesterol and 0.69 for alcohol. For women they showed energy-adjusted correlations between 0.04 for β-carotene and 0.63 for retinol. De-attenuated correlations were somewhat higher. The approximate twofold variability in nutrient intake observed in our study population was modest. The high degree of measurement error relative to the low between-subject variation in nutrient intakes resulted in ranking errors, misclassification, and decreased the measurable contrasts. For some nutrients the FFQ was not able to discriminate satisfactorily between each quintile of intake. However, it can discriminate between biologically meaningful ranges of intake. In summary, the newly developed FFQ showed fairly good reproducibility and relative validity for most evaluated nutrients. The between-subject variation of nutrients in the German cohort (including the participants from East Germany) is likely to be greater than

11 RELATIVE VALIDITY OF MACRONUTRIENT INTAKE, GERMANY S81 that among the subjects of the pilot phase. A revised version of the FFQ corrected for food consumption frequency is used in the EPIC study. ACKNOWLEDGEMENTS The authors would like to thank Bruno Krüger from the local health institution AOK Heidelberg for allowing us to recruit AOK members, Renate Birr for doing the interviews, all nutritionists who did the food coding, and especially Dr Gabriele Perez-Nelius who revised the food coding. We are grateful to Günther Laib and Elke Bauer for their help in data management and writing the computer programs. This work was part of the EPIC-project (European Prospective Investigation into Cancer and Nutrition), supported by the Europe Against Cancer Program of the CEC (grant No. 91 CVV ). REFERENCES 1 Riboli E. Nutrition and cancer: Background and rational of the European Prospective Investigation into Cancer and Nutrition (EPIC). Ann Onc 1992; 3: Bohlscheid-Thomas S, Hoting I, Boeing H, Wahrendorf J. Reproducibility and relative validity of food group intake in a food frequency questionnaire developed for the German part of the EPIC-project. Int J Epidemiol 1997; 26 (Suppl. 1): S59 S70. 3 Boeing H, Bohlscheid-Thomas S, Voss S, Schneeweiss S, Wahrendorf J. The relative validity of vitamin intakes derived from a food frequency questionnaire compared to 24-hour recalls and biological measurements: Results from the EPIC pilot study in Germany. Int J Epidemiol 1997; 26 (Suppl. 1): S82 S90. 4 Polensky W. In: Zusammenarbeit mit dem Bundesgesundheitsamt. Institut für Sozialmedizin und Epidemiologie. Die Bundeslebensmittelschlüsseldokumentation. 1989, Version II.1. 5 Altman D G. Practical Statistics for Medical Research. Chapman & Hall, London: Liu K, Stamler J, Dyer A, McKeever J, McKeever P. Statistical methods to assess and minimize the role of intra-individual variability in obscuring the relationship between dietary lipids and serum cholesterol. J. Chron Dis 1978; 31: Willett W, Stampfer M J. Total energy intake: implications for epidemiologic analyses. Am J Epidemiol 1986; 124: Hartung J. In: Hartung J, Elpelt B, Klösener K H (eds). Statistik: Lehr- u. Handbuch d. Angewandten Statistik. München, Wien: Oldenburg, Willett W C. Nutritional Epidemiology. New York: Oxford University Press, Bueno-de-Mesquita H B, Smeets F W M, Runia S, Hulshof K F A M. The reproducibility of a food frequency questionnaire among controls participating in a case-control study on cancer. Nutr Cancer 1992; 18: Rimm E J, Giovannucci E L, Stampfer M J, Colditz G A, Litin L B, Willett W C. Reproducibility and validity of an expanded self-administered semiquantitative food frequency questionnaire among male health professionals. Am J Epidemiol 1992; 135: Martin-Moreno J M, Boyle P, Gorgojo L et al. Development and validation of a food frequency questionnaire in Spain. Int J Epidemiol 1993; 22: Longnecker M P, Lissner L, Holden J M et al. The reproducibility and validity of a self-administered semiquantitative food frequency questionnaire in subjects from South Dakota and Wyoming. Epidemiology 1993; 4: Gnardellis C H, Trichopoulou A, Katsouyanni K, Polychronopoulos E, Rimm E B, Trichopoulos D. Reproducibility and validity of an extensive semiquantitative food frequency questionnaire among Greek school teachers. Epidemiology 1994; 6: Bittoni M A, Wikins J R. Assessment of reliability of a diet history questionnaire. Nutr Cancer 1994; 21: Ajani U A, Willett W C, Seddom J M. Reproducibility of a food frequency questionnaire for use in ocular research. Invest Ophthalmol Vis Sci 1994; 35: Goldbohm R A, van t Veer P, van den Brandt P A et al. Reproducibility of a food frequency questionnaire and stability of dietary habits determined from five annually repeated measurements. Eur J Clin Nutr 1995; 49: Bingham S. The dietary assessment of individuals: methods, accuracy, new techniques and recommendations. Nutr Abstr Rev 1987; 57: Black A E, Goldberg G R, Jebb S A, Livingstone M B E, Cole T J, Prentice A M. Critical evaluation of energy intake data using fundamental principles of energy physiology: 2. Evaluating the results of published surveys. Eur J Clin Nutr 1991; 45: Goldberg G R, Black A E, Jebb S A et al. Critical evaluation of energy intake data using fundamental principles of energy physiology: 1. Derivation of cut-off limits to identify under-recording. Eur J Clin Nutr 1991; 45: Callmer E, Riboli E, Saracci R, Akesson B, Lindgarde F. Dietary assessment methods evaluated in the Malmö food study. J Intern Med 1993; 233: Goldbohm R A, Van den Brandt P A, Brants H A M et al. Validation of a dietary questionnaire used in a large-scale prospective cohort study on diet and cancer. Eur J Clin Nutr 1994; 48: