Bayesian networks in public health

Uncertainty, complexity, and causality are common issues in statistical modeling in public health. Sometimes, we need to calculate the likelihood of one of many possible causes when we observe certain evidence. For example, we would like to calculate the probability of an infection source after we observe symptoms in the population. However, it is much easier in practice to obtain the reversed conditional probability, i.e. probability of observing evidence given its cause. Despite their strengths, Bayesian networks have not received much attention in public health so far. The method uses the reversed conditional probability distribution to represent the joint probability distribution of a large complex set of random variables with possible mutual causal relationship. The main objective of the method is to obtain the posterior conditional probability distribution of the outcome variable(s) after observing new evidence. Bayesian networks may be constructed either manually with a priori knowledge of the underlying domain, or may be learned automatically from databases with appropriate software. Moreover, intuitive and appealing visual interface of Bayesian networks is helpful to facilitate communication between statisticians and policy-makers. This method has been successfully applied in studies of health outcomes monitoring, Spiegelhalter (1998), or modeling of socio-demographic factors and obesity related behavior, Harding (2011).