A Bayesian Hierarchical Model For Dietary Exposure And Biomarker Measurements

Marta PITTAVINO, International Agency for Research on Cancer (IARC), France
CHAJES V. 1 , JOHANSSON M. 2 , PLUMMER M. 3 , FERRARI P. 1

1 Nutritional and Metabolism Section, International Agency for Research on Cancer (IARC), Lyon, France
2 Genetics Section, International Agency for Research on Cancer (IARC), Lyon, France
3 Infections Section, International Agency for Research on Cancer (IARC), Lyon, France

Purpose
In nutritional epidemiology, self-reported assessments of dietary exposure are prone to random and systematic measurement errors. As a result, estimates of the association between dietary factors and risk of disease can be biased. To partially account for exposure misclassification, it has been suggested to complement self-reported dietary assessments with objective measurements, such as dietary biomarkers. A holistic approach which uses all available information is still missing. In this work, dietary and biomarker measurements were integrated in a Bayesian model, which was used in two nested case-control studies within EPIC.
 
Methods
A Bayesian latent factor hierarchical model with three structural components was developed: 1) an exposure model, to define the distribution of unknown true exposure (the latent factors), 2) a measurement model, to disclose the relationship between observed measurements (dietary questionnaires, 24-hour recalls and biomarkers) and the true exposures, 3) a disease model, to estimate the relationship between dietary exposures and disease status. Hierarchical models are used to build complex models through the specification of simpler conditional independence relationships, for which each variable in the model is conditionally related to only a few other variables. The marginal posterior distribution of model parameters is obtained from the joint posterior distribution, using Markov Chain Monte Carlo (MCMC) sampling techniques.  Analyses were carried out using JAGS.  
 
Results
The work focused on two applications. Firstly, the association between dietary fat and risk of breast cancer was complemented by gas-chromatography plasma phospholipids from 2,982 breast cancer cases matched to 2,982 controls. Secondly, the relationship between B-vitamins with kidney cancer risk was estimated by integrating dietary and blood level measurements.
 
Conclusions
Bayesian models make it possible to integrate complex problems into modular components with simpler structure, allowing the complex nature of dietary measurements to be accounted for.
 
Funding
WCRF Grant (GR-IARC-2012-10-10-03).