Improving the estimation of multiple correlated dietary effects on colon-rectum cancer in multicentric studies: a hierarchical Bayesian approach

The paper deals with the analysis of the effects of multiple exposures on the occurrence of a disease in observational case-control studies. We consider the case of multilevel data, with subjects nested in spatial clusters. As a result, we often face problems of small and sparse data, along with correlations among the exposures and the observations, which both invalidate the results from the ordinary analyses. A hierarchical Bayesian model is here proposed to manage the within-cluster dependence and the correlation among the exposures. We assign prior distributions on the crucial parameters by exploiting additional information at different levels and by making suitable assumptions according to the problem at hand. The model is conceived to be applied to a real multi-centric study aiming at investigating the association of dietary exposures with colon-rectum cancer occurrence. Compared with results obtained with conventional regressions, the hierarchical Bayesian model is shown to yield great gains in terms of more consistent and less biased estimates. Thanks to its flexibility, this approach represents a powerful statistical tool to be adopted in a wide range of applications. Moreover, the specification of more realistic priors may facilitate and extend the use of Bayesian solutions in the epidemiological field.