The use of dietary patterns empirically derived from principal components analysis and alternative strategies to identify associations between diet and disease

摘要

Dietary patterns derived empirically using principal components analysis (PCA) are widely employed for investigating diet-disease relationships. The aim of the study was to investigate whether PCA performed better at identifying associations between diet and disease than analysing each food on the FFQ separately, a process we refer to as exhaustive single food analysis (ESFA).udA systematic review of nutritional epidemiology literature relating to the use of PCA in identifying dietary patterns in observational and cohort studies from 2004-2009 was employed. Furthermore, we simulated diet and disease data using real food frequency questionnaire data and assuming that a number of foods or dietary pattern intakes were causally associated with disease. In each simulation, ESFA and PCA were employed to identify foods associated with disease using logistic regression, allowing for multiple testing and adjusting for energy intake. ESFA was further adjusted for principal components, foods which were significant in unadjusted ESFA, and propensity scores. For each method, we investigated the power, with which we could identify an association between diet and disease, and the power and false discovery rate (FDR) for identifying associations with specific food intakes. We apply our innovative methodology to a real dietary dataset (GA2LEN survey).udESFA had greater power to detect an association of diet with disease than PCA, and greater power and lower FDR for identifying associations with specific foods. FDR increased with increasing sample size using both methods. However, when ESFA was adjusted for foods that were significant in unadjusted ESFA, FDRs were controlled successfully at the desired level of 20%.udOur results raise questions about the use of PCA in nutritional epidemiology. Adjusted ESFA identifies foods that are causally linked to disease with a low rate of false discoveries, and surprisingly good power. These findings were not fully supported from the analysis of the GA2LEN data-set.