A Bayesian Hierarchical Modeling Approach to Dietary Assessment via Food Frequency

Abstract

Previous likelihood-based linear modeling of nutritional data has been limited by the availability of software that allows flexible error structures in the data. We demonstrate the use of a Bayesian modeling approach to the analysis of such data. Our goal is to model the relationship between the energy intake derived from Food Frequency Questionnaires (FFQs) and the energy expenditure estimated from the doubly labeled water method. We consider models with different distributions for the FFQ energy intake. The models include previously identified covariates describing social desirability, education and their possible interaction that are felt to impact the reported FFQ. The models also include random effects to account for subject specific random variation (frailty) and also to account for the complex patterns of measurement error inherent in these data. Issues arising within the work relate both to the selection of relevant linear and non-linear models, the use of random effects, and the relevance of goodness-of-fit criteria such as DIC and PPL in assessing the most appropriate model.