Food patterns defined by cluster analysis and their utility as dietary exposure variables: a report from the Malmo Diet and Cancer Study

Objective: To explore the utility of cluster analysis in defining complex dietary exposures, separately with two types of variables. Design: A modified diet history method, combining a 7-day menu book and a 168-item questionnaire, assessed dietary habits. A standardized questionnaire collected information on sociodemographics, lifestyle and health history. Anthropometric information was obtained through direct measurements. The dietary information was collapsed into 43 generic food groups, and converted into variables indicating the percent contribution of specific food groups to total energy intake. Food patterns were identified by the QUICK CLUSTER procedure in SPSS, in two separate analytical steps using unstandardized and standardized (Z-scores) clustering variables. Setting: The Malmo Diet and Cancer (MDC) Study, a prospective study in the third largest city of Sweden, with baseline examinations from March 1991 to October 1996. Subjects: A random sample of 2206 men and 3151 women from the MDC cohort (n=28 098). Results: Both variable types produced conceptually well separated clusters, confirmed with discriminant analysis. 'Healthy' and 'less healthy' food patterns were also identified with both types of variables. However, nutrient intake differences across clusters were greater, and the distribution of the number of individuals more even, with the unstandardized variables. Logistic regression indicated higher risks of past food habit change, underreporting of energy and higher body mass index (BMI) for individuals falling into 'healthy' food pattern clusters. Conclusions: The utility in discriminating dietary exposures appears greater for unstandardized food group variables. Future studies on diet and cancer need to recognize the confounding factors associated with 'healthy' food patterns.