Meta-Analysis of Bayesian Analyses

Meta-analysis aims to combine results from multiple related statistical analyses. While the natural outcome of a Bayesian analysis is a posterior distribution, Bayesian meta-analyses traditionally combine analyses summarized as point estimates, often limiting distributional assumptions. In this paper, we develop a framework for combining posterior distributions, which builds on standard Bayesian inference, but using distributions instead of data points as observations. We show that the resulting framework preserves basic theoretical properties, such as order-invariance in successive updates and posterior concentration. In addition to providing a consensus analysis for multiple Bayesian analyses, we highlight the benefit of being able to reuse posteriors from computationally costly analyses and update them post-hoc without having to rerun the analyses themselves. The wide applicability of the framework is illustrated with examples of combining results from likelihood-free Bayesian analyses, which would be difficult to carry out using standard methodology.