Presentation (CTI)

OBJECTIVES: Multilevel network meta-regression (ML-NMR) enables population-adjusted indirect treatment comparisons by combining individual patient data (IPD) with aggregate data. When individual level covariates are unavailable, ML-NMR marginalizes over the covariate distribution, effectively but discards any published subgroup specific treatment effects that may contain valuable information about effect modification. We propose an extension to ML-NMR using Bayesian Synthetic Likelihood (BSL) that leverages these routinely available subgroup summaries to improve estimation of effect modifier interactions.
METHODS: The proposed method generates synthetic datasets at each MCMC iteration, computes synthetic subgroup summaries, and matches them to observed summaries via a synthetic likelihood. We use a continuous relaxation approach to ensure smooth gradients for efficient Hamiltonian Monte Carlo sampling and implantation in stan. We evaluate the method using simulation studies with varying degrees of missing covariate data and compare performance to standard ML-NMR and oracle analysis (with complete IPD).
RESULTS: Simulation studies demonstrate that BSL-enhanced ML-NMR substantially improves upon standard ML-NMR, potentially approaching oracle performance despite incomplete covariate data. The method successfully recovers treatment effect modification parameters that ML-NMR cannot estimate precisely.
CONCLUSIONS: BSL provides a principled approach for incorporating published subgroup summaries that would otherwise be discarded in standard ML-NMR. Our approach may be particularly valuable in health technology assessment where outcome data and subgroup analyses are routinely reported but individual covariate data remain unavailable due to privacy or proprietary concerns.