Reproducibility of a food frequency questionnaire used in the New York University Women's Health Study: Effect of self-selection by study subjects

Transcription

1 European Journal of Clinical Nutrition (1997) 51, 437±442 ß 1997 Stockton Press. All rights reserved 0954±3007/97 $12.00 Reproducibility of a food frequency questionnaire used in the New York University Women's Health Study: Effect of self-selection by study subjects E Riboli 1, P Toniolo 2, R Kaaks 1, RE Shore 2, C Casagrande 1 and BS Pasternack 2 1 Unit of Nutrition and Cancer, International Agency for Research on Cancer, 150 cours Albert-Thomas, Lyon Cedex 08, France; and 2 Nelson Institute of Environmental Medicine and Kaplan Comprehensive Cancer Center, New York University School of Medicine, 341 East 25th Street, New York, NY , USA Objective: The aim of the study was to investigate the reproducibility of a food frequency questionnaire (FFQ) used in a cohort of women enrolled in the NYU Women's Health Study. Design: A subset of 474 cohort subjects who completed the dietary questionnaire at baseline (FFQ-1) were approached again on the occasion of a second visit to the mammography study centre and asked to complete the questionnaire a second time (FFQ-2) two to four years after FFQ-1. Two to three months later the questionnaire was mailed to the subjects, and they were asked to complete it a third time (FFQ-3). Setting: A breast cancer screening clinic. Subjects: Of the 474 subjects selected, 100% completed the second questionnaire while only 56% completed and mailed back FFQ-3. This made it possible to compare the long-term reproducibility of dietary intake measurements and baseline dietary habits between the two groups of subjects: `respondents', who agreed to complete the questionnaire a third time, and `non-respondents', who did not. Results: Among respondents (56% of study subjects), energy-adjusted correlation coef cients for short-term reproducibility between FFQ-2 and FFQ-3 ranged from 0.50±0.64 for nutrients, and from 0.44±0.67 for foods. The long-term reproducibility was lower, ranging from 0.36±0.53 for nutrients, and from 0.31±0.48 for speci c food groups. Among those who did not respond to FFQ-3, crude correlations for long-term reproducibility, unadjusted for energy intake, were generally lower than among respondents. Nevertheless, after adjustment for energy intake, correlations for long-term reproducibility (FFQ-2 to FFQ-1) were of similar magnitude in both groups. In addition, `non-respondents' reported lower intake of fruit and vegetables and higher intake of meat. Conclusions: The results of this study suggest that subjects who volunteer to participate in substudies on the validity or reproducibility of dietary questionnaire measurements may tend to provide more accurate responses to the questionnaire. The phenomenon seems related more to accuracy of reporting of absolute intake levels than of the relative composition of diet. In addition, self-selection may be associated with differences in dietary habits. Sponsorship: This study was supported by Public Health Service (PHS) grants CA 34588, CA and CA from the National Cancer Institute, National Institutes of Health (NIH), Department of Health and Human Services (DHHS); and by PHS ES from the National Institute of Environmental Health Sciences, NIH, DHHS. Descriptors: dietary measurements; reproducibility; selection bias; dietary questionnaires; cohort studies; dietary patterns Introduction In nutritional epidemiology studies, in which disease risk is estimated as a function of an individual's consumption levels of speci c foods or nutrients, it is essential that measurement errors leading to incorrect ranking of individuals by levels of dietary intake be kept to a minimum, so as to preserve the highest possible statistical power to demonstrate the presence of relevant diet-disease associations. As some measurement error will, however, always occur, additional substudies may be conducted to estimate the magnitude of statistical power loss or bias in estimates of diet-disease associations, incurred by the inaccuracy of dietary questionnaire measurements. These substudies can take the form either of a reproducibility study, based on successive administrations of the questionnaire, usually Correspondence: Dr E Riboli. Received 24 May 1996; revised 10 March 1997; accepted 14 March 1997 over a six- to twelve-month interval (Hankin et al, 1983; Wu et al, 1988; Jain et al, 1989; JaÈrvinen et al, 1993) or a `validity' study, using a comparison with records of actual food consumption during a number of days as well (Willett et al, 1985; Pietinen et al, 1988a, 1988b; Munger et al, 1992; Rimm et al, 1992; Callmer et al, 1993). In both types of substudy, with or without additional reference measurements, there is the potential problem that participation rates may be lower than 100% (in some studies, participation rates as low as 20% have been reported (Munger et al, 1992)), and that there may be self-selection of volunteers whose capacity or motivation to respond accurately to the dietary questionnaires is above average. In some of the articles reporting the results of validation studies, the main argument supporting the absence of bias due to `selfselection' was that nutrient intakes did not appreciably differ between participants in the validation study and the overall study population (Munger et al, 1992). We are not aware, however, of any study that has evaluated the

2 438 potential magnitude of over- or under-estimation of the validity or reproducibility of questionnaire measurements as a result of such self-selection. In this paper, we report results on the reproducibility of a dietary questionnaire used to collect baseline exposure data in the New York University Women's Health Study, a prospective cohort study designed to investigate the relations between diet, endogenous hormones, and incidence of cancers of the breast and other organs. The reproducibility study, which included 475 women, was designed to compare the concordance between repeat measurements of usual diet over a longer (2±5 y) as well as over a shorter (2±3 months) period of time. The rst two questionnaire measurements were obtained from a random subsample of the cohort, under conditions which ensured near perfect participation. The third measurement required an additional effort on the part of the subjects as they were asked to complete the questionnaire at home and mail it back. This simple procedure meant that the subjects could be separated a posteriori into two subgroups: those who lled out the questionnaire and mailed it back (respondents), and those who failed to do so (non-respondents). By analyzing results for respondents and non-respondents separately, we also give an estimate of the potential effect of self-selection of study participants on the evaluation of the longer-term reproducibility of questionnaire measurements. Materials and methods Study subjects The subjects were participants in a prospective cohort study on hormones, diet and cancer, the New York University Women's Health Study (Toniolo et al, 1991, 1995). Healthy women between 35 and 65 y of age, attending a large breast cancer screening clinic in New York City, were invited to enroll in the prospective study. Those who agreed were asked to donate 30 ml of peripheral venous blood and to answer a demographic and medical history questionnaire as well as a self-administered food-frequency questionnaire (FFQ-1). The dietary questionnaire was similar to that developed at the US National Cancer Institute (Block, 1989; Block & Hartman, 1989), with minor modi cations to the food list. Between March 1985 and June 1991, women were recruited. In an effort to obtain as many repeated blood specimens as possible from study subjects, all cohort members were contacted again each time they returned to the clinic for mammographic screening. The study design is summarized in Table 1. In brief, during a two-month period (April±May 1990) all 474 cohort members returning to the screening clinic one to ve years after being enrolled in the study were asked to complete the dietary questionnaire a second time (FFQ-2) while waiting for a mammographic examination in the clinic. In this situation, all of them complied. Two months later, the dietary questionnaire was mailed to the same 474 women, and they were asked to complete it for the third time (FFQ-3) and send it back in a prepaid return envelope. Without further solicitation, 267 women (56.3%) responded to the FFQ-3. We excluded six subjects who either completed FFQ-2 less than one year after FFQ-1 or returned FFQ-3 more than six months after FFQ-2. The mean time between FFQ-1 and FFQ-2 was 47.5 months ( 11.4) for those who responded to FFQ-3, and 46.2 months ( 12.9) for those who did not. The mean time between FFQ-2 and FFQ-3 was 3.4 months ( 0.9). Food consumption reported by the study subjects was rst checked for completeness and missing values. The number of lines in the questionnaire with missing values was used as an indicator of carefulness. In FFQ-1 and FFQ- 2, approximately 10% of the subjects had four or more missing values, and 3% were missing 10 or more. The frequency of blank lines was higher in FFQ-3, for which about 20% had four or more missing values, and 8.7% had 10 or more missing. Further analyses were restricted to 411 subjects (223 `respondents' and 188 `non-respondents') who had no more than seven missing values in any of the questionnaires. Thus, subjects in the upper 5% of the distribution by number of missing values in any of the three questionnaires were excluded, which is in line with rules applied in analyses of data on diet and breast cancer in this and other cohort studies (Willett et al, 1987; Toniolo et al, 1994). There were no differences in the proportions of excluded subjects between respondents and non-respondents. Daily intake of foods and nutrients was calculated on the basis of food lists and food composition tables elaborated at the NCI (Block & Hartman, 1989), with minor modi cations. Statistical analysis Mean differences and 95% con dence intervals were computed for the differences between FFQ-3 minus FFQ-1, Table 1 Study design for the evaluation of the effect of subjects' self-selection on the estimation of the reproducibility of a food frequency questionnaire (FFQ). The New York University Women's Health Study (NYU-WHS) Period FFQ Where FFQ was lled in How FFQ was returned No. of subjects who participated Phase ±1990 Phase 2 April±May 1990 Phase 3 June±July 1990 FFQ-1: Administered to all women enrolled in the NYU-WHS FFQ-2: Administered to all women returning to the clinic during a 2- month period FFQ-3: Mailed to the women who had completed FFQ-2 Screening clinic Handed to a nurse Screening clinic Handed to a nurse 474 Subject's home Mailed back to investigators 267 out of 474 (207 did not return the FFQ) Time between FFQ-1 and FFQ-2: Subjects who returned FFQ-3 (267): 3 years and 11.5 months ( 11.4 months) Subjects who did not return FFQ-3 (207): 3 years and 10.2 months ( 12.9 months) Time between FFQ-2 and FFQ-3 (267): 3.35 months ( 0.89)

3 FFQ-2 minus FFQ-1, and FFQ-3 minus FFQ-2. To estimate agreement in relative ranking, Spearman correlations were computed for the absolute intake variables as well as for the energy-adjusted variables, using the linear regression `residual' approach described by Rosner & Willett (1988). Missing answers for speci c food items were treated as zero consumption. In a few cases, where the frequency of consumption was reported but the portion size was missing, the `medium' portion was used for calculation of total quantities of a given type of food consumed. Results Means and standard deviations of daily intake of energy and main nutrients are shown in Table 2 for the entire cohort and for the subjects included in the reproducibility study. We compared the 411 subjects included in the present study to the entire cohort, and found no signi cant differences in the nutrient intake in FFQ-1. Regarding the nutrient pattern, it is worth noting that energy intake is somewhat lower than would be expected on the basis of the subjects' age and weight. This is typical of the underassessment of food consumption observed regularly in studies conducted with similar food frequency questionnaires (Willett et al, 1985; Munger et al, 1992). Nevertheless, the nutrient intake pattern seems characteristic of middle-aged north American women, about 40% of the energy being provided by fat and 17% by protein. Intake of other nutrients is also similar to what has been measured with food frequency questionnaires in similar populations. In particular, the estimated average intake of ber is relatively low (about 14 g/d) while the average dietary intake of vitamin C (supplements excluded) is relatively high (about 167 mg/d). The relatively low intake of vitamin E (about 8 mg/d) may be due to an underestimation of vegetable oil consumption (one of the main sources of vitamin E) as the questionnaire included speci ed questions on seasoning fats but not on oil used for cooking. When the means and standard deviations of daily nutrient intake obtained on the three occasions among respondents are compared (Table 3), no meaningful changes are observed over time for the intake of most nutrients; assessment of consumption of energy, fat, carbohydrates and protein remained remarkably stable. An exception is cholesterol, for which there was a 20% decrease from FFQ- 1 to both FFQ-2 and FFQ-3 but no difference between FFQ-2 and FFQ-3. Similar considerations apply to the data for food groups reported in Table 4. The fewer speci c foods for which there is a suggestion of a difference between FFQ-1 and FFQ-2 or FFQ-3 (not shown in Table 4) are beef, eggs and apples which decreased by about 20, 45 and 28%, respectively, and spaghetti and fruit juice which increased by 54 and 32%. Comparison between respondents and non-respondents suggested that total energy intake was slightly lower for non-respondents (761 Kcal per day). However, no meaningful differences in intake of any of the other nutrients reported in Table 2 were observed between respondents and non-respondents. As for foods, respondents reported slightly higher intakes of vegetables and our products and lower meat intake. Mean differences were statistically signi cant for meat (P ˆ 0.01) or close to statistical signi cance for our products (P ˆ 0.07) in FFQ-1 and for vegetables (P ˆ 0.09) in FFQ-2. Respondents and nonrespondents did not differ in age, height or weight. No information was available on study subjects' physical activity. Spearman rank correlations between repeat measurements are given in Tables 5 and 6 for nutrients and foods. Crude correlations for nutrients and foods were always higher for short-term (FFQ-2 vs FFQ-3) than for long-term reproducibility (). Crude correlations for energy were 0.54 for FFQ-1 vs FFQ-3 and 0.45 for. Crude correlations for energy-providing nutrients ranged between 0.39 and 0.49 for FFQ-1 vs FFQ-2 and between 0.54 and 0.59 for FFQ-2 vs FFQ-3. Adjustment for energy intake led to a modest decrease in the correlation coef cients for all nutrients except saturated fat and cholesterol, for which the correlations were slightly improved. After energy adjustment, the short-term correlations (average 0.54) remained higher than long-term correlations (average 0.37). Results for foods and food groups followed a similar pattern, but with correlations slightly higher than for nutrients. 439 Table 2 Average daily intake of selected nutrients estimated in three repeated administrations of the same food frequency questionnaire (FFQ) among 223 subjects who agreed to complete the questionnaire three times and 188 subjects who refused to participate the third time Respondents to FFQ-3 (No. ˆ 223) Non-respondents to FFQ-3 (No. ˆ 188) Total Cohort (No. ˆ ) Nutrients FFQ-1 FFQ-2 FFQ-3 FFQ-1 FFQ-2 FFQ ±1988 April±May 1990 June±July ±1988 April±May ±1991 Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d. Energy (Kcal Protein (g) Carbohydrates (g) Total fat (g) Saturated fat (g) Oleic acid (g) Linoleic acid (g) Cholesterol (mg) Dietary bre (g) Carotene (mg) Retinol (mg) Vitamin C (mg) Vitamin E (mg) Calcium (mg) Iron (mg) Sodium (mg) Potassium (mg)

4 440 Table 3 Means and 95% con dence intervals of the individual differences in nutrient intakes from questionnaires 3, 2 and 1. Analyses restricted to 223 subjects who responded to all three questionnaires Long-term reproducibility (2±5 years) Short-term reproducibility (2±3 months) Difference FFQ-2±FFQ-1 Difference FFQ-3±FFQ-1 Difference FFQ-3±FFQ-2 95% CI 95% CI 95% CI Mean Lower Upper Mean Lower Upper Mean Lower Upper Total calories (Kcal) Protein (g) Carbohdyrate (g) Total fat (g) Saturated fat (g) Oleic acid (g) Linoleic acid (g) Cholesterol (mg) Dietary bre (g) Carotene (mg) Retinol (mg) Vitamin C (mg) Vitamin E (mg) Calcium (mg) Iron (mg) Sodium (mg) Potassium (mg) Table 4 Average daily intake of main groups of foods estimated in three repeated administrations of the same food frequency questionnaire (FFQ) among 223 subjects who agreed to complete the questionnaire three times and 188 subjects who refused to participate the third time Respondents to FFQ-3 (No. ˆ 223) Non-respondents to FFQ-3 (No. ˆ 188) FFQ-1 FFQ-2 FFQ-3 FFQ-1 FFQ-2 Food groups 1985±1988 April±May 1990 June±July ±1988 April±May 1990 Mean s.d. Mean s.d. Mean s.d. Mean s.d. Mean s.d. Fruit Fruit juices Vegetables Cereals Pastries Potatoes Meat Poultry Fish Eggs Ice Cream Dairy products Spaghetti & Pizza Table 5 Spearman rank correlations between nutrient intake estimated in repeated administrations of the same food frequency questionnaire (FFQ) within a short time period (2±3 months) for 223 `respondents' and within a long time period (2±5 y) for respondents and `non-respondents' Respondents Non-respondents Short-term (2±3 months) FFQ-2 vs FFQ-3 Crude Energy-adjusted Crude Energy-adjusted Crude Energy-adjusted Energy 0.54 Ð 0.45 Ð 0.33 Ð Protein Carbohydrates Total fat Saturated fat Oleic acid Linoleic acid Cholesterol Fibre Carotene Retinol Vitamin C Vitamin E Calcium Average correlation

5 Table 6 Spearman rank correlation between food groups estimated in repeated administrations of the same food frequency questionnaire (FFQ) within a short time period (2±3 months) for 223 `respondents' and within a long time period (2±5 y) for respondents and `non-respondents' 441 Respondents Non-respondents Short-term (2±3 months) FFQ-2 vs FFQ-3 Crude Energy-adjusted Crude Energy-adjusted Crude Energy-adjusted Fruit Fruit juices Vegetables Cereals Pastries Potatoes Meat Poultry Fish Eggs Ice cream Dairy products Spaghetti & pizza Average correlation Without adjustment for total energy intake, correlation coef cients for nutrient intake were slightly higher for respondents than for non-respondents, for all nutrients except carotene (Table 5). When adjusted for energy intake, however, the correlation coef cients were very similar for both groups. As for foods, the correlations were higher among respondents for a few food groups, particularly fruit and vegetables, and quite similar for the others. Adjustment for energy lowered the correlation coef cients and generally made them more similar in both groups. Discussion We investigated the long- and short-term reproducibility of the dietary measurements obtained with the food frequency questionnaire used in the New York University Women's Health Study. Reproducibility was assessed at the group level by comparison of group means, and by calculation of correlation coef cients between the individuals' repeat measurements. Mean estimated intake levels remained relatively stable over time, although there was a small decrease in the reporting of food items that might be perceived as `bad for health', such as beef and eggs, and an increase in food items that may be seen as `healthy' foods such as fruit juice and spaghetti. There was, however, no meaningful change in the reported consumption of fresh fruit and vegetables. The energy-adjusted correlation coef- cients obtained for short-term reproducibility in the present study are similar to (or just slightly lower than) those reported in several other studies of similar design (Pietinen et al, 1988a, 1988b; Rimm et al, 1992) where correlations ranged between 0.60 and 0.70 for most of the nutrients considered. The short-term reproducibility, over a period of 2±3 months, was somewhat higher than the long-term reproducibility, after 2±5 y. One possible explanation for this difference is that the long-term reproducibility may have been in uenced more by changes in true dietary habits over time. Another view is that the short-term reproducibility may have been favoured by the fact that measurements were obtained during the same season (assuming that the measurements more strongly re ected dietary habits during the more recent months), or that some women still recalled some of their previous answers and did not try to give a second, more independent evaluation of their dietary habits. While we have no data for discriminating between different explanations, the higher reproducibility over short than over long periods is consistent with the ndings of previous studies (Friedenreich et al, 1992). A potentially important issue in the extrapolation of the results of a methodological study to the actual epidemiological investigation, in which a diet questionnaire is used to estimate relative risk of disease, is whether the subjects included in the methodological study are representative of those in the main etiological investigation. Methodological studies to evaluate the reproducibility and validity of dietary questionnaire measurements have so far been based mostly on self-selected volunteers. In some cases the subjects were a relatively small subgroup of very large cohort studies (Willett et al, 1985; Rimm et al, 1992; Munger et al, 1992), while in others the subjects were drawn independently of any on-going epidemiological study, even though the source population was that targeted for future studies on diet and cancer (Pietinen et al, 1988a, 1988b; Callmer et al, 1993). Concordance, either between repeat questionnaire measurements (in a reproducibility study) or between questionnaire and reference measurements (in a validity study), depends not only on the characteristics of the dietary questionnaire method being evaluated and the true dietary habits of the study subjects but also on the accuracy and dedication with which the subjects ll in the questionnaires. One might therefore suspect that the reproducibility and relative validity derived from studies on self-selected volunteers will be overestimated as compared to those which would actually be found in a large, less selective study population. Overestimation of the reproducibility or validity of questionnaire measurements may have consequences for the evaluation of the true statistical power of studies on diet and disease to detect an association, or may lead to an under-evaluation of attenuation biases in relative risk estimates. In our study, 56% of the women contacted agreed to ll out the FFQ for a third time. This participation rate is similar to that obtained in previous studies, and would in practice be considered as quite satisfactory in this type of methodological study. However, a comparison between the results of subjects who provided full participation and those who did not

6 442 respond to FFQ-3 suggests that self-selection may play a role, namely that volunteers may provide more reliable information than subjects less strongly motivated to participate in repeat dietary measurements. In addition, subjects who complied with the mailed questionnaire had reported somewhat higher vegetable intake, and a lower meat intake at baseline. These differences suggest that people who are willing to participate in a study on diet in relation to health may be more sensitive to dietary guidelines, which currently recommend eating more vegetables and reducing meat intake. The results of our study also suggest, therefore, that those involved in a reproducibility or validity study may not be a truly representative sample of the targeted study population when the rate of non-participation in the substudy is high. In our view, in order to ensure high compliance, future studies on the accuracy of dietary intake measurements should preferably be based on a reference method which does not require a large number of repeat measurements (for example several weeks of food recording) from the study participants. To achieve this, one approach is to include more subjects in the substudy, but with a smaller possible number of repeat reference measurements (weighed food records or 24 h recalls) per person. If the aim is to estimate the correlation between dietary questionnaire measurements and individuals' intake levels, a minimum of two additional measurements with independent random errors per person is required (Kaaks et al, 1995). If the aim is limited to estimating the variation in true intake levels actually distinguished by questionnaire measurements (as can be the case in on-going prospective cohort studies) then a `calibration' approach requiring only a single daily intake record or 24 h diet recall per person can be used (Kaaks et al, 1994, 1995). Such an approach is now being implemented in the `European Prospective Investigation into Cancer and Nutrition', a prospective cohort study in progress in nine European countries (Riboli & Kaaks, 1997). However, as the reference measurement is based on a single day's diet measurement per subject and it is not possible to estimate the correlation between dietary questionnaire measurements and individuals' true intake levels, this method may be less suitable when the prime objective is to develop a new questionnaire for which estimates of validity of measurement of speci c foods, food groups and nutrients are needed for subsequent improvement of the questionnaire itself. References Block G (1989): Health Habits and History Questionnaire: Diet History and Other Risk Factors. Personal Computer System Packet. Bethesda, MD: National Cancer Institute, Division of Cancer Prevention and Control. Block G & Hartman AM (1989): Issues in reproducibility and validity of dietary studies. Am. J. Clin. Nutr. 50, 1133±1138. Callmer E, Riboli E, Saracci R, aêkesson B & LindgaÈrde F (1993): Dietary assessment methods evaluated in the MalmoÈ food study. J. Intern. Med. 233, 53±57. Friedenreich CM, Slimani N & Riboli E (1992): Measurement of past diet: Review of previous and proposed methods. Epidemiol. Rev. 14, 177± 186. Hankin JH, Nomura AMY, Lee J, Hirohata T & Kolonel LN (1983): Reproducibility of a diet history questionnaire in a case-control study of breast cancer. Am. J. Clin. Nutr. 37, 981±985. Jain M, Howe GR, Harrison L & Miller AB (1989): A study of repeatability of dietary data over a seven-year period. Am. J. Epidemiol. 129, 422±429. JaÈrvinen R, SeppaÈnen R & Knekt P (1993): Short-term and long-term reproducibility of dietary history interview data. Int. J. Epidemiol. 22, 520±527. Kaaks R, Plummer M, Riboli E, EsteÁve J & van Staveren W (1994): Adjustment for bias due to errors in exposure assessments in multicenter cohort studies on diet and cancer: calibration approach. Am. J. Clin. Nutr. 59, 245s±250s. Kaaks R, Riboli E & van Staveren WA (1995): Calibration of dietary intake measurements in prospective cohort studies. Am. J. Epidemiol. 142, 548±556. Munger RG, Folson AR, Kushi LH, Kaye SA & Sellers TA (1992): Dietary assessment of older Iowa women with a food frequency questionnaire: nutrient intake, reproducibility, and comparison with 24-hour dietary recall interviews. Am. J. Epidemiol. 136, 192±200. Pietinen P, Hartman AM, Haapa E, RaÈsaÈnen L, Haapakoski J, Palmgren J, Albanes D, Virtamo J & Huttunen JK (1988a): Reproducibility and validity of dietary assessment instruments. I. A self-administered food use questionnaire with a portion size picture booklet. Am. J. Epidemiol. 128, 667±676. Pietinen P, Hartman AM, Haapa E, RaÈsaÈnen L, Haapakoski J, Palmgren J, Albanes D, Virtamo J & Huttunen JK (1988b): Reproducibility and validity of dietary assessment instruments. II. A qualitative food frequency questionnaire. Am. J. Epidemiol. 128, 655±666. Riboli E & Kaaks R (1997): The EPIC Project: Rationale and study design. Int. J. Epidemiol. 26, Suppl. 1, S6±S14. Rimm EB, Giovannucci EL, Stampfer MJ, Colditz GA, Litin LB & Willett WC (1992): Reproducibility and validity of an expanded self-administered semiquantitative food frequency questionnaire among male health professionals. Am. J. Epidemiol. 135, 1114±1126. Rosner R & Willett WC (1988): Interval estimates for correlation coef cients corrected for within-person variation: Implications for study design and hypothesis testing. Am. J. Epidemiol. 127, 377±386. Toniolo PG, Pasternack BS, Shore RE, Sonnenschein E, Koenig KL, Rosenberg C, Strax P & Strax S (1991): Endogenous hormones and breast cancer: a prospective cohort study. Breast Cancer Res. Treat. 18, 23s±26s. Toniolo P, Riboli E, Shore RE & Pasternack BS (1994): Consumption of meat, animal products, protein and fat and risk of breast cancer: a prospective cohort study in New York. Epidemiology 5, 391±397. Toniolo PG, Levitz M, Zeleniuch-Jacquotte A, Banerjee S, Koenig KL, Shore RE, Strax P & Pasternack BS (1995): A prospective study of endogenous estrogens and breast cancer in postmenopausal women. J. Natl. Cancer Inst. 87, 190±197. Willett WC, Sampson L, Stampfer MJ, Rossner B, Bain C, Witschi J, Hennekens CH & Speizer FE (1985): Reproducibility and validity of a semiquantitative food frequency questionnaire. Am. J. Epidemiol. 122, 51±65. Willett WC, Stampfer MJ, Colditz GA, Rosner BA, Hennekens CH & Speizer FE (1987): Dietary fat and the risk of breast cancer. N. Engl. J. Med. 316, 22±28. Wu ML, Whittemore AS & Jung DL (1988): Errors in reported dietary intakes. II. Long-term recall. Am. J. Epidemiol. 128, 1137±1145.