Bayesian Meta-Analysis

An approach combining evidence across studies using Bayesian methods, formally incorporating prior information to produce a posterior effect estimate distribution.

Source: Dias et al. 2013

Bayesian meta-analysis combines the evidence from several studies using Bayesian statistics, which start from a prior distribution expressing what was believed before the data and update it with the studies to yield a posterior estimate. This lets external information, such as findings from related research or expert judgement, be brought formally into the pooling rather than set aside. It also produces full probability distributions for the quantities of interest, which express uncertainty richly, at the cost of greater complexity and the need to justify the priors. Updating belief with data is its logic. Sutton and Abrams (2001) describe this approach.

Source: Sutton & Abrams 2001

Bayesian meta-analysis works by specifying prior distributions for the parameters, such as the pooled effect and the between-study variance, combining them with the likelihood of the observed study data, and computing the posterior distribution, often using simulation methods such as Markov chain Monte Carlo. The posterior summarises the effect and its uncertainty given the data and priors. It naturally accommodates random-effects and complex structures, including networks of treatments. So Bayesian meta-analysis combines priors with study data to produce posterior estimates, using computational methods to handle the models, and expressing results as distributions that convey uncertainty about the synthesised effect.

Source: Dias et al. 2013

The advantages of Bayesian meta-analysis include its natural handling of complex models, such as network meta-analysis and hierarchical structures; its full representation of uncertainty through posterior distributions, including for the between-study variance; its ability to incorporate external or prior evidence; and its coherent framework for probabilistic statements, such as the probability that one treatment is best. It integrates well with decision modelling. So Bayesian meta-analysis is advantageous for complex syntheses and for conveying uncertainty comprehensively, making it widely used in network meta-analysis and health technology assessment where these features are valuable.

Source: O'Hagan & Stevens 2001

The challenges of Bayesian meta-analysis include the need to specify prior distributions, which can influence results, especially with few studies, and must be justified; the computational demands of simulation methods and the need to check their convergence; the complexity of building and interpreting the models; and the requirement for expertise. Sensitivity to priors, particularly for the between-study variance, is a concern. These challenges mean Bayesian meta-analysis is conducted with careful choice and justification of priors, checks on computation, and sensitivity analyses, so that its flexibility and rich output are realised without the results being unduly driven by the prior assumptions.

Source: Dias et al. 2013

Bayesian meta-analysis combines prior information with the data to produce posterior distributions for the effect, expressing uncertainty probabilistically, while frequentist meta-analysis produces a point estimate and confidence interval without formally incorporating prior beliefs. Bayesian methods handle complex models and uncertainty in the between-study variance more naturally and allow probabilistic statements, but require priors and computation, whereas frequentist methods are simpler and more familiar but less flexible for complex syntheses. So the two differ in whether prior information is formally used and in how uncertainty is expressed, with Bayesian meta-analysis offering flexibility at the cost of greater complexity and prior specification.

Source: Dias et al. 2013