ABSTRACT
In this chapter we describe how hierarchical random-effects models can be applied to meta-analysis using a fully Bayesian approach. Using a metaanalysis of randomized trials of selective decontamination of the digestive tract as an example, inferences are made using Gibbs sampling via BUGS, a freely available software package. We illustrate the usefulness of graphical modeling techniques for expressing the conditional independence assumptions of the parameters in the model and show how specification of the model in BUGS leads naturally from the graph formulation. Problems with using a standard noninformative prior distribution for the population variance are discussed and suitable alternative prior distributions are derived and compared.
