Use of multiple correspondence analysis and cluster analysis to study dietary behaviour: Food consumption questionnaire in the SU.VI.MAX. cohort C. Guinot 1 , J. Latreille 1 , D. Malvy 2 , P. Preziosi 3 , P. Galan 3 , S. Hercberg 3 & M. Tenenhaus 4 1 CE.R.I.E.S., Neuilly Sur Seine; 2 INSERM U330, University Victor Segalen Bordeaux 2, Bordeaux; 3 SU.VI.MAX. co-ordination, ISTNA, CNAM, Paris; 4 HEC School of management – Paris, Department SIAD, Jouy en Josas, France Accepted in revised form 11 September 2001 Abstract. Although the effects of individual foods or nutrients on the development of diseases and their risk factors have been investigated in many studies, little attention has been given to the effect of overall dietary patterns. The main objectives of this study were to identify dietary patterns and groups of sub- jects with similar food consumption habits, i.e. ‘di- etaryprofiles’,usingmultiplecorrespondenceanalysis and cluster analysis. A food frequency questionnaire was sent to a large population-based sample (2923 women and 2180 men), recruited among the ‘SUp- pl ementation en VItamines et Min eraux AntioXy- dants’ (SU.VI.MAX.) cohort participants in France. The food items were dichotomised in order to focus the study on the highest levels of consumption. Multiple correspondence analysis allows the con- struction of principal components, which optimally summarise the data, and enables the construction of graphical displays. An interesting property of these graphical displays is that associations between food items can be observed on various projection planes, each category of each food item being located at the centre of gravity of the subjects corresponding to this category. An ascending hierarchical classification was unsuccessfully tried in order to determine clusters from these principal components. Therefore, a ‘dis- section’ of the cloud of points was performed ac- cording to the orientation of the axes, providing a readily interpretable eight-dietary profiles typology for each sex. This statistical approach allows identi- fication of particular dietary patterns and dietary profiles, which might be more appropriate in studies of diet-disease associations than the single food or nutrient approach that has dominated past epidemi- ological research. Keywords: Clusteranalysis,Dietarypatterns,Foodconsumptionhabits,Foodfrequencyquestionnaire,Multiple correspondence analysis, Nutritional epidemiology Introduction The hypothesis, that diseases not caused by severe deficiency malnutrition could nevertheless be linked to nutritional factors emerged in the 1960s. These diseases, which today comprise major public health problems in industrialised countries (notably cancers, cardiovascular diseases, obesity, osteoporosis, etc.), are clearly of multifactorial origin. Among the poten- tially determinant factors, dietary behaviour seems to play a important role, even more so because it is possible to act on it and thus there is the hope of reducingtheriskofdisease[1,2].Sincethefeaturesof dietary behaviour are of a multidimensional nature, its description and an understanding of its impact on the possible development of chronic diseases com- prise a major factor in terms of scientific knowledge and public health. In the majority of works performed to date, the analyses of links between diet and health have taken onlyone‘isolated’foodornutrientintoconsideration. Thedietarydataaremostoftencollectedbymeansof aquestionnaireonthefrequencyofconsumptionona qualitative or ‘semi-quantitative’ basis and relating to a few dozen groups of foods [3]. In addition, the in- formationaboutphysicalactivitythatisindispensable for understanding the nutritional balance and the re- lationshipbetweendietandhealthisgenerallymissing ofverybrief[3].Inepidemiologicalobservationstudies (ecologicalstudies,case–controlstudiesorprospective studies),therelationshipsbetweentheconsumptionof foods [4] (e.g. meats, fish, cereals, fruit, sugar, fats, wine, etc.) or nutrients [5] (fats, carbohydrates, pro- teins, trace elements) on the one hand and either the value of nutritional status markers [6, 7] (blood levels ofcholesterol,triglyceridesorglucose,bodyfatindex, markers of vitamin and mineral status) or health in- dicators [8, 9] (cancer, cardiovascular disease, diabe- tes, obesity, osteoporosis) on the other hand are analysed. European Journal of Epidemiology 17: 505–516, 2001. Ó 2002 Kluwer Academic Publishers. Printed in the Netherlands.