ORIGINAL ARTICLE. 1 Department of Epidemiology, German Institute of Human Nutrition, Potsdam-Rehbrücke, Germany; 2 Dietary Exposure Assessment

Transcription

1 (2009) 63, S263 S274 & 2009 Macmillan Publishers Limited All rights reserved /09 $ ORIGINAL ARTICLE Specific food group combinations explaining the variation in intakes of nutrients and other important food components in the European Prospective Investigation into Cancer and Nutrition: an application of the reduced rank regression method J Kröger 1, P Ferrari 2,27, M Jenab 3, C Bamia 4, M Touvier 5,6, HB Bueno-de-Mesquita 7, MT Fahey 8, V Benetou 4, M Schulz 1, E Wirfält 9, H Boeing 1, K Hoffmann 1,{, MB Schulze 1,28, P Orfanos 4, E Oikonomou 4, I Huybrechts 2, S Rohrmann 10, T Pischon 1, J Manjer 11, A Agren 12, C Navarro 13, P Jakszyn 14, MC Boutron-Ruault 5, M Niravong 5, KT Khaw 15, F Crowe 16, MC Ocké 7, YT van der Schouw 17, A Mattiello 18, M Bellegotti 19, D Engeset 20, A Hjartåker 21, R Egeberg 22, K Overvad 23, E Riboli 24, S Bingham 25,26,{ and N Slimani 2 1 Department of Epidemiology, German Institute of Human Nutrition, Potsdam-Rehbrücke, Germany; 2 Dietary Exposure Assessment Group, International Agency for Research on Cancer, Lyon, France; 3 Lifestyle and Cancer Group, International Agency for Research on Cancer, Lyon, France; 4 Department of Hygiene, Epidemiology and Medical Statistics, University of Athens, Medical School, Athens, Greece; 5 Inserm, ERI 20, Institut Gustave Roussy, Villejuif, France; 6 AFSSA (French Food Safety Agency), DERNS/PASER, Maisons- Alfort, France; 7 National Institute for Public Health and the Environment (RIVM), Bilthoven, The Netherlands; 8 Biostatistics Unit, Medical Research Council and University of Cambridge, Cambridge, UK; 9 Department of Clinical Sciences, Lund University, Malmö, Sweden; 10 Division of Cancer Epidemiology, German Cancer Research Center, Heidelberg, Germany; 11 Department of Surgery, Malmö University Hospital, Malmö, Sweden; 12 Department of Public Health and Clinical Medicine, Nutritional Research, University of Umeå, Umeå, Sweden; 13 Epidemiology Department, Murcia Health Council, Murcia and CIBER en Epidemiología y Salud Pública (CIBERESP), Spain; 14 Unit of Nutrition, Environment and Cancer, Cancer Epidemiology Research Programme, Catalan Institute of Oncology (ICO), Barcelona, Spain; 15 University of Cambridge School of Clinical Medicine, Addenbrookes Hospital, Cambridge, UK; 16 Cancer Epidemiology Unit, University of Oxford, Oxford, UK; 17 Julius Center for Health Sciences and Primary Care, University Medical Center, Utrecht, The Netherlands; 18 Department of Clinical and Experimental Medicine, University of Naples, Federico II, Naples, Italy; 19 Nutritional Epidemiology Unit, Department of Preventive & Predictive Medicine, Fondazione IRCCS Istituto Nazionale dei Tumori, Milano, Italy; 20 Institute of Community Medicine, University of Tromsø, Tromsø, Norway; 21 Cancer Registry of Norway, Oslo, Norway; 22 Institute of Cancer Epidemiology, Danish Cancer Society, Copenhagen, Denmark; 23 Department of Clinical Epidemiology, Aalborg Hospital, Aarhus University Hospital, Aalborg, Denmark; 24 Department of Epidemiology, Public Health and Primary Care, Imperial College, London, UK; 25 Diet and Cancer Group, MRC Mitochondrial Biology Unit, Cambridge, UK and 26 Department of Public Health and Primary Care, MRC Centre for Nutritional Epidemiology in Cancer Prevention and Survival, University of Cambridge, Cambridge, UK Correspondence: J Kröger, Department of Epidemiology, German Institute of Human Nutrition, Potsdam-Rehbrücke, Arthur-Scheunert-Allee , Nuthetal, Germany. kroeger@dife.de { The authors are deceased. 27 Current address: Data Collection and Exposure Unit (DATEX), European Food Safety Authority, Parma, Italy. 28 Current address: Public Health Nutrition Unit, Technische Universität München, Freising, Germany. Guarantor: JKröger. Contributors: JK carried out the statistical analyses, prepared the tables and figures, and wrote the paper, taking into account comments from all co-authors. NS was the overall coordinator of this project and of the EPIC Nutrient DataBase (ENDB) project. JK, PF, MJ, CB, MT, HBB-d-M, MTF, VB, MS, EW, HB, KH and MBS were members of the writing group and gave input on the statistical analyses, drafting of the manuscript and interpretation of results. The other co-authors were local EPIC collaborators who participated in the collection of dietary and other data, and in the ENDB project. ER is the overall coordinator of the EPIC study. All co-authors provided comments and suggestions on the manuscript and approved the final version.

2 S264 Objective: To identify combinations of food groups that explain as much variation in absolute intakes of 23 key nutrients and food components as possible within the country-specific populations of the European Prospective Investigation into Cancer and Nutrition (EPIC). Subjects/Methods: The analysis covered single 24-h dietary recalls (24-HDR) from subjects ( men and women), aged years, from all 10 countries participating in the EPIC study. In a set of 39 food groups, reduced rank regression (RRR) was used to identify those combinations (RRR factors) that explain the largest proportion of variation in intake of 23 key nutrients and food components, namely, proteins, saturated fatty acids, monounsaturated fatty acids, polyunsaturated fatty acids, cholesterol, sugars (sum of mono- and disaccharides), starch, fibre, alcohol, calcium, iron, potassium, phosphorus, magnesium, vitamin D, b-carotene, retinol and vitamins E, B1, B2, B6, B12 and C (RRR responses). Analyses were performed at the country level and for all countries combined. Results: In the country-specific analyses, the first RRR factor explained a considerable proportion of the total nutrient intake variation in all 10 countries ( %). The subsequent RRR factors were much less important in explaining the variation (p6%). Strong similarities were observed for the first country-specific RRR factor between the individual countries, largely characterized by consumption of bread, vegetable oils,, milk, cheese, potatoes, and meat. The highest explained variation was seen for protein, potassium, phosphorus and magnesium (50 70%), whereas sugars, b- carotene, retinol and alcohol were only marginally explained (p5%). The explained proportion of the other nutrients ranged between these extremes. Conclusions: A combination of food groups was identified that explained a considerable proportion of the nutrient intake variation in 24-HDRs in every country-specific EPIC population in a similar manner. This indicates that, despite the large variability in food and nutrient intakes reported in the EPIC, the variance of intake of important nutrients is explained, to a large extent, by similar food group combinations across countries. (2009) 63, S263 S274; doi: /ejcn Keywords: food group combinations; food component intakes; nutrient intakes; reduced rank regression; EPIC; 24-h dietary recall Introduction High correlations among foods and subsequently among nutrients are a particular challenge in nutritional epidemiology. When evaluating the relationship between dietary factors and occurrence of chronic diseases, adjustment for other foods or nutrients that are correlated with the dietary factor and the disease is one method of accounting for confounding. This approach is, however, cumbersome if there are many potential confounders, and may be ineffective owing to measurement error in self-reported dietary intakes. Furthermore, adjustment for many correlated variables can obscure relationships, and attenuate or eliminate any effect. Alternative strategies for dealing with nutritional variables, which are usually correlated, are therefore a challenge for nutritional epidemiology. The evaluation of food group combinations or nutrient profiles, instead of single foods or nutrients, may offer such an alternative approach (Hu, 2002). The variation in nutrient intake between individuals is determined by the variation in both food intake and food composition. It is, however, less clear which combinations of foods are responsible for the total nutrient variations at the population level. Descriptive analyses of this kind would complement studies on food sources of single nutrients or food components, as described in this supplement (Cust et al., 2009; Halkjær et al., 2009; Jenab et al., 2009; Linseisen et al., 2009; Olsen et al., 2009; Sieri et al., 2009; Welch et al., 2009), by investigating a group of nutrients simultaneously in a multivariate analysis and allowing for correlations among nutrients. Recently, researchers have begun identifying statistical tools capable of addressing such complex methodological questions. In particular, multidimensional approaches, such as partial least square and reduced rank regression (RRR), provide a suitable statistical framework (Hoffmann et al., 2004). Using RRR, for example, the relationship between two kinds of variable groups can be analysed by constructing linear combinations of one group of variables (for example, food intakes) that explain as much variation as possible in the other variable group (for example, nutrient intakes) (Hoffmann et al., 2004). To shed light on the complex relationship between food and nutrient intake, approximately h dietary recalls (24-HDRs) collected in the multicentre setting of the European Prospective Investigation into Cancer and Nutrition (EPIC) (Riboli et al., 2002) were analysed by the RRR method to determine which combinations of 39 food groups explain most effectively the absolute intake of 23 nutrients and food components that are of current major epidemiological research interest and for which concentration values have been standardized across countries. The main objectives of this paper are to describe and understand how food combinations and nutrient profiles are linked, to determine whether there are food combinations that can considerably explain the variation in total nutrient intakes and to explore how much country-specific heterogeneity actually exists. Thus, this paper should help to understand

3 the complex relationship between food and nutrient intake, which is of great importance for investigations on dietary components in relation to the risk of chronic diseases. Further, under the assumption that nutrients are the bioactive components that affect the risk of chronic diseases, a food combination that explains the variation in nutrient intakes may also have a role in chronic disease aetiology. Thus, this study provides hints for potential disease-related food group combinations, which could be evaluated in future studies. Materials and methods Study sample The EPIC study is a multicentre prospective cohort study designed to investigate the relationship of diet with the risk of cancer and other chronic diseases, using the data from 23 study centres in 10 European countries (Greece, Spain, Italy, France, Germany, Netherlands, UK, Denmark, Sweden and Norway). In France, Norway, Utrecht (Netherlands) and Naples (Italy) only women were recruited, whereas all the other cohorts included both men and women. Detailed information on the EPIC study, including descriptions of the source populations, can be found elsewhere (Riboli et al., 2002; Bingham and Riboli, 2004). Within a calibration substudy designed to correct for measurement error in baseline dietary intake measurements, a single 24-HDR interview was performed between 1995 and 2000 in a subsample of the EPIC cohort. The calibration samples represent between 5 and 12% (1.5% in the UK) of each EPIC cohort population (Slimani et al., 2002b). This calibration study provided the data for the present analyses. The initial dataset included 24-HDRs from subjects. After a systematic exclusion of the participants outside the age range of years, HDRs (obtained from men and women) were included in the present study. Dietary assessment Information on dietary intake was collected by means of one computerized 24-HDR interview per subject using the computer programme EPIC-SOFT (Slimani et al., 1999), which was specially developed to ensure standardization of the interviews (Slimani et al., 2000) within and between countries. All 24-HDRs were collected during a face-to-face interview, except in Norway, where interviews were conducted by telephone (Brustad et al., 2003). During the dietary interview, each reported food item was searched for, described, quantified and checked according to standard rules. Methods used to estimate the portion size included photographs, household measures and standard units. Standardized food component data were compiled within the framework of the ENDB (EPIC nutrient database) project (Slimani et al., 2007). The aim of the ENDB was to improve the comparability of nutrient databases across the 10 countries participating in EPIC through collaboration with national compilers of nutrient databases, and food and nutrient experts from each country. The nutrient contents for each food were calculated by following a standardized procedure, drawing upon the nutrient information available in the national food composition tables. Other common rules and calculations were adopted for computing the nutrient values of mixed recipes and foods that were missing from the national food composition tables. In total, 26 important nutrients and food components were standardized within the ENDB (Slimani et al., 2007). The present study considered all these ENDB nutrients and food components, except water, total fat and glycaemic carbohydrates. The latter were excluded in order to avoid using intakes of these components twice in RRR, because the constituents of fat (saturated fatty acids (SFA), monounsaturated fatty acids (MUFA) and polyunsaturated fatty acids (PUFA)) and glycaemic carbohydrates (starch and sugars) were used in the analysis. Thus, 23 key nutrients and food components were finally included in this study, namely, protein, SFA, MUFA, PUFA, cholesterol, sugars (comprising mono- and disaccharides), starch, fibre, alcohol, calcium, iron, potassium, phosphorus, magnesium, vitamin D, b-carotene, retinol, and vitamins E, B1, B2, B6, B12 and C. Intakes from dietary supplements were not considered in the present study, but this topic is addressed in another paper of this special issue (Skeie et al., 2009). For the sake of simplicity and clarity regarding the wording, the term nutrient is often used in this paper to refer to the 23 nutrients and food components. However, it should be borne in mind that many scientists do not consider alcohol and fibre to be nutrients. Classification of food groups A special food classification, which is a slightly modified version of the food classification available in the EPIC-SOFT program, was adopted in the EPIC for this supplement (Slimani et al., 2000). Similar to the EPIC-SOFT classification, this modified food classification contains 17 major groups, each of which are further divided into several sub-groups. For the present analysis, all food groups were considered, either at the main group or at the subgroup level, and finally 39 food groups (defined in Table 1) were selected, in line with previous studies on dietary patterns (Newby and Tucker, 2004). Non-dietary variables Data on other lifestyle factors, including education level, total physical activity and smoking history, considered in this analysis were collected at baseline through standardized questionnaires and clinical examinations, and have been described for the calibration sampling elsewhere (Riboli et al., 2002; Slimani et al., 2002b). Data on age, as well as S265

4 S266 Table 1 Definitions and contents of the 39 food groups considered in the analysis Food group Potatoes Leafy vegetables Fruiting and root vegetables Cabbage Other vegetables Legumes Fruits Nuts Other fruits Milk and dairy products Cheese Pasta, rice Bread Breakfast cereals Other cereals Red meat and game Poultry Processed meat Fish Eggs Margarine Vegetable oils Butter Other fats Sugar and confectionery Cake, cookies Fruit and vegetable juices Soft drinks Tea Coffee Water Wine Beer Spirits Other alcoholic beverages Sauces Condiments and yeast Soups Miscellaneous Definition and content Potatoes and other tubers e.g., spinach, iceberg lettuce e.g., carrots, tomatoes, sweet pepper and avocados Broccoli and other cabbages Mushrooms, grain and pod vegetables, onion, garlic, stalk vegetables, sprouts and mixed salads Dried peas, lentils and beans Fruits, e.g., apples, citrus fruits, grapes and stone fruits Nuts and seeds Mixed fruits and olives Milk and milk beverages, including fermented ones, yoghurt, curd, cream desserts and puddings (milk based), dairy creams, milk for coffee and creamers (dairy) All cheese, including cream cheese and ricotta Pasta, rice and other grains Bread, crispbread and rusks e.g., muesli, cornflakes, puffed rice and porridge based on water Flour, flakes, starches, semolina, salty biscuits, crackers, dough and pastry Beef, veal, pork, mutton/lamb, horse, goat, game and offal Chicken, turkey, duck, goose and rabbit Processed meat from or poultry (e.g., ham, bacon, sausages, etc.) Fish and fish products, crustaceans and molluscs Eggs and egg products Pure vegetable s, s of mixed origin and non-specified e.g., olive oil, soya oil, sunflower oil, rapeseed oil and walnut oil Butter Deep frying fats, other animal fats and fat non-specified Sugar, honey, jam, syrup, chocolate, confectionery, ice cream and water ice Cakes, sweet pies, pastries, puddings (non-milk based), dry cakes, biscuits Fruit and vegetable juices Carbonated/soft/isotonic drinks, diluted syrups and non-specified non alcoholic beverages Tea and herbal tea Coffee, chicory and coffee substitutes Tap water, ice cubes, mineral water and other water products Red/white/rosé wine and non-specified wines Beer, cider, fruit wines and drinks based on beer/cider Spirits, brandy, liqueurs and aniseed drinks Fortified wines, cocktails and non-specified alcoholic beverages Tomato sauces, dressing sauces, mayonnaises, dessert sauces and other sauces Yeast, spices, herbs, flavourings and condiments Soups and bouillons Vegetarian foods, soya products, dietetic products (e.g., artificial sweeteners), non-dairy creams, amphibians and snacks body weight and height were self-reported by the participants during the 24-HDR interview. The mean time interval between these baseline questionnaire measures and the 24-HDR interview varied by country, from 1 day to 3 years later (Slimani et al., 2002b). Data analysis All statistical analyses were performed using SAS release 9.1 (SAS Institute, Cary, NC, USA). Analyses were carried out separately for men and women and stratified by country. In the United Kingdom, the health-conscious subjects and those recruited from the general population were considered as two separate cohorts (Slimani et al., 2002b). We applied RRR using the partial least square procedure in SAS to identify combinations of food groups that explain the largest proportions of variation in the intake of the 23 nutrients and food components. RRR can be considered as a generalization of a simple regression to the multivariate case, but with a dimensionality reduction aim. This technique was recently introduced in nutritional epidemiology to derive dietary patterns (Hoffmann et al., 2004). Briefly, RRR extracts linear combinations (the so-called factors) of predictor variables (for example, food group intakes) that explain as much variation in response variables (for example, nutrient intakes) as possible. One property of the successive extracted factors is that they are uncorrelated, and hence the variation in the responses can be broken down into fractions of variation explained by the factors obtained. The 24-HDR data (g/day) on 39 food groups were used as predictors, and those on 23 nutrients as responses. RRR analyses were performed both for the whole calibration sample to obtain overall factors and for each country separately. To improve the comparability between countries, the data on food groups and nutrients were adjusted for several covariates using linear regression and by entering the

5 residuals into the RRR analysis. The following variables were taken into account as covariates: age, weight and height (continuous variables); day of the week of interview (Monday/Tuesday/Wednesday/Thursday versus Friday/Saturday/ Sunday); season of interview (spring, summer, autumn and winter); study centre (the 23 EPIC study centres were redefined into 27 centres and geographical regions (Slimani et al., 2002b) and 26 dummy variables were coded, with the Malmö cohort as reference centre); education (none/primary school, technical/professional/secondary school, university and missing data); smoking status (never smoker, former smoker, smoker and missing data); and physical activity (inactive, moderately inactive, moderately active, active and missing data). An RRR factor is a complex construct, because it is a linear combination of the z-standardized intakes (AM (arithmetic mean) ¼ 0, s.d. (standard deviation) ¼ 1) of 39 food groups, with each group multiplied by an individual weight. To make the results easier to interpret, the first RRR factor was also examined in a shortened, simplified form. For this purpose a new, simplified factor was calculated, as the unweighted sum of the z-standardized intakes (AM ¼ 0, s.d. ¼ 1, standardization performed within the whole calibration sample) of only the indicator food groups of the first RRR factor. The indicator food groups were considered to be groups with a factor loading (statistical measure that indicates the relationship between the food group and the derived RRR factor) X0.2 of the absolute value. This selection criterion was based on the assumption that a high factor loading gives a variable much more weight in constructing a factor (Hatcher, 1994). A simplified food factor is easier to interpret and to communicate, and better reflects individual food consumption, but still has a very similar meaning compared to the original factor. More details on this method of simplification and its applicability have been presented previously (Schulze et al., 2003). As the list of food groups that met the criterion of a factor loading X0.2 was very similar for men and women, a single simplified factor was developed for both. Thus, RRR was applied to the whole calibration sample without splitting by gender, and food groups with factor loadings X0.2 of the absolute value were used for the joint simplified factor. The creation of the simplified factor was the only part of the analysis that was not performed using stratification for gender. Percentages of total nutrient variation, explained by the simplified pattern, were calculated for men and women separately. To compare the original (gender-specific) RRR factors with the simplified (joint) factor, Pearson s correlation coefficients were calculated. Results As the number of extracted RRR factors equals the number of selected responses, 23 factors were extracted during RRR analysis. When RRR was performed on the whole calibration sample, the first RRR factor explained substantially more of the total nutrient variation (31.1% for men, 29.6% for women) than the following factors: factor 2, 6.2% for men, 6.3% for women; factors 3 23, consecutively less. This observation was also made when RRR was performed on single country-specific populations. It was therefore decided to focus the analysis on the first factor. Descriptive characteristics of the first RRR factor derived for men, stratified by country, are shown in Table 2. The amount of variation in total nutrient intake explained by the first RRR factor was similar between countries, ranging from 28.3% for the UK general population cohort to 37.1% for the Netherlands. Strong similarities between the EPIC countries were also observed for the proportion of explained variance for single nutrients. In every country, relatively large proportions of protein, potassium, phosphorus and magnesium intake were explained by the first factor, which seemed to be mainly driven by the food group bread, which often constitutes an important part of the overall diet. The explained variation for protein ranged from B65% up to B78%. For potassium, the explained variability ranged from 60 to B70% in all populations except for the UK general population, in which only 53.1% of the potassium intake was accounted for. The explained variation for phosphorus was in general from 70 to B80%, except in the Greece population, in which only 50.1% of the phosphorus variation was accounted for. For magnesium, the explained variation ranged between B50 and B70%. In contrast, the explained variation for sugars, alcohol, b-carotene and retinol intake by the first factor was modest (p10%) in the country-specific models. The proportion of explained variance for the other nutrients (SFA, MUFA, PUFA, cholesterol, starch, fibre, calcium, iron, and vitamins D, E, B1, B2, B6, B12 and C) ranged between these two extremes. Strong similarities between countries were also detected regarding the food groups that contributed most to the first countryspecific RRR factor (with a factor loading X0.2 of the absolute value). Bread was the food group with the highest factor loading in most countries. Other important food groups with factor loadings X0.2 in most countries were red meat, vegetable oils, milk, cheese, potatoes, meat and. Table 3 shows the descriptive characteristics of the first RRR factor obtained for women and stratified by country. In general, the results for women were comparable to those for men, again with strong similarities across countries. The country-specific proportion of explained nutrient variation by the first RRR factor ranged from 27.4% for the UK healthconscious cohort to 33.9% for Sweden. As observed in men, the largest proportions of variation were explained for protein (B65 75% explained variance, but with lower values in the UK general population (59%) and health-conscious cohorts (50%)), potassium (B60% explained variance), phosphorus (range B70 80%, but with lower values in Greece (45%) and the UK health-conscious cohort (55%)) and magnesium (range B50 70%, with lower values in S267

6 S268 Table 2 Percentage of nutrient-specific and total nutrient variation explained by the first (country-specific and overall) RRR factor and ranking of food groups with factor loadings X0.2 (absolute value) in the European Prospective Investigation into Cancer and Nutrition (EPIC) calibration study (n ¼ ), stratified by country Men Greece Spain Italy Germany The Netherlands UK general population UK healthconscious cohort Denmark Sweden Overall n Protein SFA MUFA PUFA Cholesterol Sugars o o Starch Fibre Alcohol Calcium Iron Potassium Phosphorus Magnesium Vitamin D b-carotene Retinol Vitamin E Vitamin B Vitamin B Vitamin B Vitamin B Vitamin C Total Food groups with factor loading X0.2 Bread (0.51), veg. oils (0.42), cheese (0.32), fish (0.24), root veg. (0.22), (0.21) Veg. oils (0.41), bread (0.39), (0.30), meat (0.26), milk (0.25), potatoes (0.21) Veg. oils (0.38), bread (0.34), fruits (0.28), wine (0.25), milk (0.24), (0.24), cheese (0.23), pasta (0.22), potato (0.22), cake (0.21) Bread (0.36), (0.35), meat (0.32), milk (0.28), potato (0.27), cheese (0.20), cake (0.20), fruits (0.20) Bread (0.36), (0.34), meat (0.31), (0.30), potatoes (0.29), cheese (0.27), milk (0.26), sauce (0.25) Milk (0.36), bread (0.34), (0.32), potatoes (0.28), (0.25), meat (0.23), veg. oils (0.21), cheese (0.21) Bread (0.40), potatoes (0.37), milk (0.33), (0.29), root veg. (0.23), other vegetables (0.20) Bread (0.41), milk (0.36), cheese (0.27), (0.27), (0.25), potatoes (0.25), meat (0.23), fish (0.23), sauce (0.23) Bread (0.41), (0.39), milk (0.31), cheese (0.26), potatoes (0.26), (0.24), meat (0.21) Bread (0.42), (0.29), veg. oils (0.29), milk (0.28), cheese (0.26), potatoes (0.25), meat (0.23), (0.23) Abbreviations: MUFA, monounsaturated fatty acids; PUFA, polyunsaturated fatty acids; RRR, reduced rank regression; SFA, saturated fatty acids; veg., vegetable.

7 S269 Table 3 Percentage of nutrient-specific and total nutrient variation explained by the first (country-specific and overall) RRR factor and ranking of food groups with factor loadings X0.2 (absolute value) in the European Prospective Investigation into Cancer and Nutrition (EPIC) calibration study (n ¼ ), stratified by country Women Greece Spain Italy France Germany The Netherlands UK general population UK healthconscious cohort Denmark Sweden Norway Overall n Protein SFA MUFA PUFA Cholesterol Sugars o Starch Fibre Alcohol Calcium Iron Potassium Phosphorus Magnesium Vitamin D b-carotene Retinol Vitamin E Vitamin B Vitamin B Vitamin B Vitamin B Vitamin C Total Food groups with factor loading X0.2 Bread (0.47), veg. oils (0.44), cheese (0.31), leafy veg. (0.22), fish (0.20) Veg. oils (0.43), bread (0.36), other veg. (0.28), root veg. (0.26), (0.23), milk (0.21), meat (0.20) Veg. oils (0.38), bread (0.35), fruits (0.25), milk (0.25), (0.22), potatoes (0.21), cheese (0.21), cake (0.21), wine (0.20), root vegetables (0.20) Bread (0.35), cheese (0.32), (0.30), milk (0.26), veg. oils (0.23), wine (0.23), root veg. (0.22), cake (0.21) Milk (0.30), bread (0.29), cheese (0.28), potatoes (0.28), (0.25), root veg. (0.25), fruits (0.23), veg. oils (0.23), meat (0.22) Bread (0.37), milk (0.34), (0.31), potatoes (0.29), cheese (0.28), (0.27), sauce (0.23), meat (0.23), nuts (0.20) Milk (0.38), cake (0.33), bread (0.32), veg. oils (0.30), breakfast cereals (0.28), (0.24) Potatoes (0.34), misc. (0.31), bread (0.29), cake (0.28), root veg. (0.27), leafy veg. (0.27), nuts (0.23), butter (0.22), fruits (0.22), other veg. (0.22), cabbage (0.20) Bread (0.39), milk (0.35), (0.28), potato (0.24), fish (0.24), cake (0.22), eggs (0.21), root veg. (0.21), cheese (0.20) Bread (0.37), (0.32), milk (0.31), cheese (0.28), potatoes (0.24), (0.24), fish (0.22), cake (0.21) Bread (0.45), milk (0.37), (0.28), cheese (0.26), nuts (0.25), fish (0.23), potatoes (0.22), root veg. (0.21), (0.20) Bread (0.39), milk (0.29), veg. oils (0.29), cheese (0.28), (0.26), potatoes (0.24), root veg. (0.21) Abbreviations: misc., miscellaneous; MUFA, monounsaturated fatty acids; PUFA, polyunsaturated fatty acids; RRR, reduced rank regression; SFA, saturated fatty acids; veg., vegetable.

8 S270 France (38%) and the UK general population (46%)). In contrast, sugars, alcohol, b-carotene and retinol were only marginally accounted for (approximately p10% explained variation). Examination of those food groups that contributed most to the first RRR factor revealed that bread again emerged as a very important food item, having the highest factor loading in 6 of 11 cohorts. Other important food groups with a factor loading X0.2 in most countries were milk, vegetable oils, cheese,, potatoes, root vegetables, and cake. The characteristics of the second, third and fourth overall RRR factors derived for men and women separately are presented in Table 4. In general, the characteristics were similar between men and women. In both genders, the second factor explained much less nutrient variation than the first factor (B6% compared with B30% for the first factor). The second factor accounted for large proportions of fibre, vitamin C and cholesterol variation, and also for alcohol and SFA, and in men, potassium was also explained in considerable proportions. This RRR factor was mainly characterized by fruits, root vegetables, eggs, wine, beer and spirits. The third factor explained large amounts of alcohol variation, especially in men (465%). In women, this factor also accounted for considerable proportions of calcium and vitamin B2. Consequently, this factor was mainly driven by alcoholic beverages, that is, wine, beer and spirits, but milk and cheese also had important roles, especially for women. The fourth factor showed differences between men and women. Among men, calcium, vitamin E, MUFA and PUFA were best explained by this factor, and foods such as vegetable oils,, milk and cheese had factor loadings X0.2. Among women, the fourth RRR factor explained alcohol intake in particular, but also starch and fatty acid intake. Consequently, alcoholic beverages, such as Table 4 Percentage of nutrient-specific and total nutrient variation explained by the second, third and fourth overall RRR factor and ranking of food groups with factor loadings X0.2 (absolute value) in men (n ¼ ) and women (n ¼ ) in the European Prospective Investigation into Cancer and Nutrition calibration study RRR factor Men RRR factor Women Protein 5.2 o Protein SFA SFA MUFA MUFA PUFA PUFA Cholesterol Cholesterol 21.9 o0.1 o0.1 Sugars o o0.1 Sugars 0.1 o Starch Starch Fibre 33.6 o Fibre Alcohol Alcohol Calcium Calcium Iron o Iron o Potassium Potassium Phosphorus Phosphorus Magnesium Magnesium Vitamin D Vitamin D 3.3 o0.1 o0.1 b-carotene b-carotene Retinol Retinol 1.1 o0.1 o0.1 Vitamin E Vitamin E Vitamin B1 0.4 o0.1 o0.1 Vitamin B1 0.9 o0.1 o0.1 Vitamin B Vitamin B Vitamin B o0.1 Vitamin B Vitamin B Vitamin B Vitamin C Vitamin C Total Total Food groups with factor loading X0.2 Eggs (0.28), beer (0.27), spirits (0.23), (0.22), fruits ( 0.48), fruiting & root vegetables ( 0.29), bread ( 0.23) Wine (0.57), beer (0.43), spirits (0.35), milk ( 0.27), ( 0.25), cheese ( 0.23) Vegetable oils (0.41), (0.37), milk ( 0.61), cheese ( 0.34) Food groups with factor loading X0.2 Eggs (0.29), wine (0.27), cake (0.23), fruits ( 0.53), fruiting & root vegetables ( 0.31) Wine (0.47), vegetable oils (0.26), milk ( 0.63), cheese ( 0.35) Wine (0.55), milk (0.24), spirits (0.23), bread ( 0.32), ( 0.31), cake ( 0.28), vegetable oils ( 0.23) Abbreviations: MUFA, monounsaturated fatty acids; PUFA, polyunsaturated fatty acids; RRR, reduced rank regression; SFA, saturated fatty acids.

9 Explained % of total nutrient variation Greece Spain Italy Germany Netherlands UK General population UK "Health-conscious" Denmark Sweden Figure 1 Percentage of country-specific total nutrient variation explained by the simplified pattern in the European Prospective Investigation into Cancer and Nutrition (EPIC) calibration study (n ¼ ) Men. (Calculated as the sum of the standardized intakes of bread, vegetable oils,, milk, cheese, potatoes, and meat.) Explained % of total nutrient variation Greece Spain Italy France Germany Netherlands UK General population UK "Health-conscious" Denmark Sweden Norway Figure 2 Percentage of country-specific total nutrient variation explained by the simplified pattern in the European Prospective Investigation into Cancer and Nutrition (EPIC) calibration study (n ¼ ) Women. (Calculated as the sum of the standardized intakes of bread, vegetable oils,, milk, cheese, potatoes, and meat.) S271 wine and spirits, and foods rich in starch or fat, such as bread, cake, and vegetable oils, contributed most to this factor. As mentioned before, a simplified food factor was formed by considering those food groups with a factor loading X0.2 (absolute value) for the first overall RRR factor in the whole calibration sample, jointly for men and women. The food groups bread, vegetable oils,, milk, cheese, potatoes, and meat met this criterion and were used to form the simplified factor. Pearson s correlation coefficients between the originally derived first factor (gender-specific) and the simplified (joint) food factor were 0.78 for men and 0.73 for women. Vertical bar diagrams were used to illustrate the amount of explained country-specific nutrient variation by the joint simplified food factor stratified by gender. For men, the amount of explained total nutrient variation ranged from 15.1% for the UK general population to 24.2% for the Netherlands (Figure 1). In five of the nine cohorts, the combination of bread, vegetable oils,, milk, cheese, potatoes, and meat explained more than 20% of the nutrient variation (Netherlands, Greece, Sweden, the UK health-conscious cohort and Denmark). For women, the explained country-specific variations in nutrient intake using the simplified food factor were mostly lower than those for men (Figure 2). The explained nutrient variations ranged from only 8 13% for the UK cohorts to 24.6% for Greece, with most values between 15 and 20%. Discussion This study shows that a considerable proportion of the variation in intake of 23 key nutrients can be explained using a specific combination of food groups represented by an RRR food factor. Strong similarities were seen across the EPIC countries for this food factor (both in terms of food groups contributing the most to this factor and in terms of the explained proportions of nutrient variances), which was unexpected given the wide variations in dietary intakes and patterns across centres described previously (Slimani et al., 2002a). The RRR food factor s ability to predict nutrient variations largely persisted when a simple food factor was adopted, comprising bread, vegetable oils,, milk, cheese, potatoes, and meat, indicating that these foods make the most important contribution to the intake variation of key nutrients in all EPIC countries. A major strength of this study is that it covers populations from 10 European countries, thus making it possible to study and compare the relationship between food intake and nutrient provision in populations from various geographical regions in Europe with multifaceted cultural backgrounds. It should, however, be noted that for many of these countries the population studied is not representative of the general population (Slimani et al., 2002b); therefore, the results presented here should not be considered to be representative of a given country. An important advantage of this study lies in the method used to collect the dietary data, which were obtained by means of a highly standardized dietary assessment tool (EPIC-SOFT) used by all participating countries. Furthermore, nutrient data from the ENDB were used, which greatly improves the comparability of nutrient intakes across countries when compared with the use of different countryspecific nutrient databases (Slimani et al., 2007). However, when interpreting our results, some methodological issues related to the use of a single 24-HDR as the data basis must be taken into account. First, one 24-HDR cannot reliably reflect an individual s long-term dietary intake. Thus, the results of this study cannot contribute directly to research on dietary patterns, which are supposed to reflect the habitual dietary behaviour. Rather, these results describe the combinations of foods consumed on one particular day, which explain the largest proportions of variation in intake of important nutrients on that day. Second, the large intra-individual day-to-day

10 S272 variability in dietary intake should not have affected the reproducibility and stability of the results, because the 24- HDRs were not used to draw conclusions about individual dietary intake, but were seen as single measurements of the simultaneous intake of foods and nutrients. In other words, if the study participants had completed their recall interview on a different day from the actual interview day, the results of this study would, most likely, not have changed very much. In fact, the results did not change appreciably when the HDRs were randomly allocated to two groups and the analyses were repeated (data not shown). Third, as the data were based on a single 24-HDR, the frequency distribution of food intakes in the study population may be very different between common foods and less regularly eaten foods, which may have many zero values. This might have contributed to the observation that commonly eaten foods, such as bread,, milk, cheese and potatoes, were the most important constituents of the first RRR factor. In analyses of habitual dietary intake, this observation might not be expected, because there would be far fewer zero intakes among the less regularly eaten foods. This issue needs to be evaluated in future studies. It should be noted that differences in food groups intake variances per se could not have affected our findings because, by default, predictors and responses are z-standardized (mean ¼ 0, variance ¼ 1) in the course of an RRR analysis. This study was not intended at generating food factors that represent a particular type of diet, such as one with high intakes of healthy nutrients. In contrast, it was aimed at understanding the complex nutrient profile generated by a certain combination of foods. The RRR method optimizes the combination of foods according to the explained variation in nutrients. The rationale for using this method is the assumption that variation in nutrient intake is the component that affects the risk of chronic diseases. Thus, in future studies, it might be worth examining whether the retrieved food group combination is related to the risk of chronic diseases. The nutrients selected for this study comprise those that are currently the focus of epidemiological research, and for which solid information is available in the common EPIC Nutrient Database (Slimani et al., 2007). It should be noted that a change in the response variables would influence the selection of foods and that the results of this study are valid only for the response variables used in this analysis. However, in a sensitivity analysis using only 16 nutrients (protein, sucrose, starch, SFA, MUFA, PUFA, alcohol, fibre, potassium, calcium, iron, b-carotene, and vitamins C, B1, B2 and B6) as responses, the results did not change appreciably (data not shown). This indicates that the results of this study may not be very sensitive to a change in the response variables. It was decided to present the results only for the first four overall RRR factors. The first four factors together explain a large proportion of the total nutrient variation accounted for by all 23 RRR factors (45.7% for the first four factors compared with 58.2% for all 23 factors in men, 44.5% compared with 56.3% in women). Thus, it is unlikely that important information was overlooked in the presentation of results. Indeed, most nutrients were not importantly explained by any of the remaining 19 factors (data not shown). However, in both genders, a considerable proportion of cholesterol was explained by the eighth factor (men 17.0% and women 13.1%), of starch by the fifth factor (men 37.5%, women 27.5%) and of vitamin D by the seventh factor (men 16.6%, women 13.0%). Furthermore, vitamin C was explained in relatively large amounts by the fifth factor in men (15.2%). It was a deliberate choice not to use energy-adjusted intakes of foods and nutrients in the RRR analyses, because the aim of this study was to identify combinations of food groups that explain the largest proportions of absolute nutrient intakes. Nevertheless, we also performed RRR analyses using energyadjusted intakes of foods and nutrients (that is, residuals from the regression of the respective dietary factor on total energy intake) to understand the impact of energy adjustment on the results. The analyses using energy-adjusted intakes focus on the variation in relative intakes, which is a different concept compared with the analyses that were presented in this paper. When using energy-adjusted intakes, the explained total nutrient variation by the first RRR factor was only % in individual countries (compared with % when intakes were not adjusted for energy). Furthermore, the first country-specific RRR factors differed substantially between the countries, both in terms of the nutrients that were best explained by these factors and in terms of the food groups that contributed most to these factors (data not shown). These findings indicate that energy adjustment may have a great impact on RRR food patterns that are derived to explain nutrient variation. Some nutrients, especially sugars, were not substantially explained by any of the 23 RRR factors. For sugars, this is probably related to the fact that most variability in intake is explained by the food group sugar and confectionery. As this group comprises foods mainly composed of sugars, it cannot explain substantial variability for any of the 22 remaining nutrients. The broad food groups in this study may have contributed to the similarity of the first food factor across countries. This may be especially the case for bread, which was the most important food group for the first factor in many countries and which is a composite of many kinds of bread in our study, with different fibre and nutrient contents and prepared using different manufacturing techniques. One would expect different types of bread to be important for explaining the nutrient variation in different geographical and cultural regions. Unfortunately, it was not possible to investigate this issue in detail, because the data did not allow a distinction between bread subtypes that is relevant for this study (for example, wholemeal versus white bread), owing to the different naming and composition of breads in individual countries. The food factors revealed in the RRR analysis are highly complex constructs because of their large number of

11 components (39 food groups), each of which is weighted differently. It is obvious that such a comprehensive factor is difficult to communicate and cannot, in practice, be related directly to eating habits. Thus, the findings for the first factor were translated into an observable food profile by constructing a simple food factor, which is calculated as the simple (unweighted) sum of the indicator food groups (food groups with a factor loading X0.2 of the absolute value). By doing so, a particular food consumption profile was identified, comprising the food groups bread, vegetable oils,, milk, cheese, potatoes, and meat. This simple combination of food groups (without the use of weighting factors) still explained an important proportion of variation in the 23 key nutrients. One concern regarding simplification is that it might lead to a nonnegligible loss of information (and thus loss of power to predict the food component variance). In fact, Pearson s correlation coefficient between the original and the simplified factor was relatively high for both men (0.78) and women (0.73), indicating that information loss during simplification was limited. However, compared with the initial food factors, the simplified food factor was characterized by a loss of 10 15% of the explained variance in total nutrient intake in men and women. Nevertheless, the simplification procedure made it possible to obtain a clear list of food groups that, in combination, were shown to be powerful for explaining the nutrient variation. In conclusion, a combination of food groups was identified that explained a considerable proportion of the nutrient intake variation in every country-specific EPIC population in a similar manner. In addition to descriptive purposes, the results of this study may have direct implications for future scientific research, including the evaluation of the retrieved food group combination in relation to chronic disease risk. Conflict of interest M Jenab has received grant support from the World Cancer Research Fund. S Bingham has received grant support from MRC Centre. The remaining authors have declared no financial interests. Acknowledgements The work described in the paper was carried out with the financial support of the European Commission: Public Health and Consumer Protection Directorate ; Research Directorate-General 2005; Ligue contre le Cancer (France); Société 3M (France); Mutuelle Générale de l Education Nationale; Institut National de la Santé et de la Recherche Médicale (INSERM); Institut Gustave Roussy; German Cancer Aid; German Cancer Research Center; German Federal Ministry of Education and Research; Danish Cancer Society; Health Research Fund (FIS) of the Spanish Ministry of Health; Spanish Regional Governments of Andalucía, Asturias, Basque Country, Murcia and Navarra and the Catalan Institute of Oncology; and ISCIII RETIC (RD06/0020), Spain; Cancer Research UK; Medical Research Council, UK; the Stroke Association, UK; British Heart Foundation; Department of Health, UK; Food Standards Agency, UK; the Wellcome Trust, UK; Greek Ministry of Health; Hellenic Health Foundation; Italian Association for Research on Cancer; Italian National Research Council, Regione Sicilia (Sicilian government); Associazione Iblea per la Ricerca Epidemiologica ONLUS (Hyblean association for epidemiological research, NPO); Dutch Ministry of Health, Welfare and Sport; Dutch Prevention Funds; LK Research Funds; Dutch ZON (Zorg Onderzoek Nederland); World Cancer Research Fund (WCRF); Swedish Cancer Society; Swedish Research Council; Regional Government of Skane and the County Council of Vasterbotten, Sweden; Norwegian Cancer Society; the Norwegian Research Council and the Norwegian Foundation for Health and Rehabilitation. We thank Sarah Somerville, Nicole Suty and Karima Abdedayem for their assistance with editing, and Kimberley Bouckaert and Heinz Freisling for their technical assistance. References Bingham S, Riboli E (2004). Diet and cancer the European Prospective Investigation into Cancer and Nutrition. Nat Rev Cancer 4, Brustad M, Skeie G, Braaten T, Slimani N, Lund E (2003). Comparison of telephone vs face-to-face interviews in the assessment of dietary intake by the 24 h recall EPIC SOFT program the Norwegian calibration study. Eur J Clin Nutr 57, Cust AE, Skilton MR, van Bakel MME, Halkjær J, Olsen A, Agnoli C et al. (2009). Total dietary carbohydrate, sugar, starch and fibre intakes in the European Prospective Investigation into Cancer and Nutrition. Eur J Clin Nutr 63(Suppl 4), S37 S60. Halkjær J, Olsen A, Bjerregaard LJ, Deharveng G, Tjønneland A, Welch AA et al. (2009). Intake of total, animal and plant proteins, and their food sources in 10 countries in the European Prospective Investigation into Cancer and Nutrition. Eur J Clin Nutr 63(Suppl 4), S16 S36. Hatcher L (1994). A Step-by-Step Approach to Using the SAS system for Factor Analysis and Structural Equation Modeling. SAS Institute Inc: Cary, North Carolina. Hoffmann K, Schulze MB, Schienkiewitz A, Nothlings U, Boeing H (2004). Application of a new statistical method to derive dietary patterns in nutritional epidemiology. Am J Epidemiol 159, Hu FB (2002). Dietary pattern analysis: a new direction in nutritional epidemiology. Curr Opin Lipidol 13, 3 9. Jenab M, Salvini S, van Gils CH, Brustad M, Shakya-Shrestha S, Buijsse B et al. (2009). Dietary intakes of retinol, b-carotene, vitamin D and vitamin E in the European Prospective Investigation into Cancer and Nutrition cohort. Eur J Clin Nutr 63(Suppl 4), S150 S178. Linseisen J, Welch AA, Ocké M, Amiano P, Agnoli C, Ferrari P (2009). Dietary fat intake in the European Prospective Investigation into Cancer and Nutrition: results from the 24-hour dietary recalls. Eur J Clin Nutr 63(Suppl 4), S61 S80. Newby PK, Tucker KL (2004). Empirically derived eating patterns using factor or cluster analysis: a review. Nutr Rev 62, S273