Bayesian non-parametric lumping and splitting of nodes in Network Meta-Analysis under heterogeneity

TL;DR

This paper introduces a Bayesian non-parametric approach for network meta-analysis that clusters treatments based on effects, accounts for heterogeneity, and propagates clustering uncertainty, improving decision-making in treatment comparisons.

Contribution

It presents a novel Bayesian non-parametric framework using a Dirichlet process prior with a horseshoe base, enabling treatment clustering with uncertainty propagation and meta-regression under heterogeneity.

Findings

Adjusting for baseline risk heterogeneity significantly alters treatment clustering.

The method effectively propagates clustering uncertainty in NMA.

Application to case studies demonstrates practical utility.

Abstract

Network meta-analysis (NMA) synthesizes evidence for multiple treatments, but decisions on node formation can have important statistical implications including bias or inflated uncertainty. Existing data-driven methods often lack flexibility or fail to fully account for node uncertainty and adjust for between-trial heterogeneity simultaneously. We introduce a Bayesian non-parametric framework using a Dirichlet process prior with a regularized horseshoe base measure. This data-driven approach allows treatments to cluster based on their effects while formally propagating uncertainty about the clustering structure itself. We extend this method to incorporate baseline risk meta-regression, enabling clustering even under heterogeneity, and demonstrate implementation using standard MCMC software. We apply the method to case studies in rheumatology and pain and find adjusting for baseline risk heterogeneity can substantially change which treatments are clustered together, highlighting the importance of methods to allow for meta-regression.