Principal components analysis of diet is not good at identifying foods that are causally linked to disease: a simulation study

Introduction Dietary patterns derived empirically from food frequency questionnaire (FFQ) data using principal components analysis (PCA) are widely employed for the investigation of diet-disease relationships. The aim of the study was to investigate whether PCA performed better at identifying associations between diet and disease than an analysis of each individual food in the FFQ separately after adjusting for multiple testing, a process we refer to as exhaustive single food analysis (ESFA).

Methods Using simulated data employing a known model for the associations between food intakes and disease, and a realistic joint distribution of food intakes, we investigated the performance of PCA and ESFA in correctly identifying associations between diet and disease. Performance was assessed in terms of the power with which we could identify at least one association between a food intake and disease, and the power and false discovery rate (FDR) for identifying specific food intakes that were causally linked to disease in the model.

Results ESFA had greater power than PCA to detect an association of at least one food with disease, and greater power and lower FDR for identifying specific foods causally linked to disease. With both methods FDRs increased with sample size, even using an FDR-controlling adjustment. However, when we adjusted the ESFA for foods that were significant in univariate analyses, FDRs were controlled at the specified level.

Conclusions An exhaustive analysis of single foods out-performed PCA in identifying associations between diet and disease using FFQ data.