Bayesian modelling of long-term dietary intakes from multiple sources

Human exposure to a specific pesticide or other chemical can occur from a combination of food and drink products. Probabilistic risk assessments are used to quantify the distribution of mean total daily exposures in the population, from the available data on residues and consumptions. We present a new statistical method for estimating this distribution, based on dietary survey data for multiple food types and residue monitoring data. The model allows for between-food correlations in both frequency and amounts of consumption. Three case studies are presented based on consumption data for UK children, considering the distribution of daily intakes of pyrimethanil, captan and chlorpyrifos aggregated over 4, 6 and 10 food types, respectively. We compared three alternative approaches, each using a Bayesian approach to quantify uncertainty: (i) a multivariate model that explicitly includes correlation parameters; (ii) separate independent parametric models for individual food types and (iii) a single parametric model applied to intakes aggregated directly from the data. The results demonstrate the importance of accounting for correlations between foods, using model (i) or (iii), for example, but also show that model (iii) can produce very different results when the aggregated intakes distribution is bimodal. The influence of residue uncertainty is also demonstrated.