AbstractObjective:To assess the relative ability of principal components analysis (PCA)-derived dietary patterns to correctly identify cases and controls compared with other methods of characterising food intake.Subjects:Participants in this study were 232 endometrial cancer cases and 639 controls from the Western New York Diet Study, 1986–1991, frequency-matched to cases on age and county of residence.Design:Usual intake in the year preceding interview of 190 foods and beverages was collected during a personal interview using a detailed food-frequency questionnaire. Principal components analysis identified two major dietary patterns which we labelled ‘healthy’ and ‘high fat’. Classification on disease status was assessed with separate discriminant analyses (DAs) for four different characterisation schemes: stepwise DA of 168 food items to identify the subset of foods that best discriminated between cases and controls; foods associated with each PCA-derived dietary pattern; fruits and vegetables (47 items); and stepwise DA of USDA-defined food groups (fresh fruit, canned/frozen fruit, raw vegetables, cooked vegetables, red meat, poultry, fish and seafood, processed meats, snacks and sweets, grain products, dairy, and fats).Results:In general, classification of disease status was somewhat better among cases (54.7% to 67.7%) than controls (54.0% to 63.1%). Correct classification was highest for fruits and vegetables (67.7% and 62.9%, respectively) but comparable to that of the other schemes (49.5% to 66.8%).Conclusions:Our results suggest that the use of principal components analysis to characterise dietary behaviour may not provide substantial advantages over more commonly used, less sophisticated methods of characterising diet.
Is principal components analysis necessary to characterise dietary behaviour in studies of diet and disease?
