Presentation (CTI)

OBJECTIVES: Network meta-analyses can be challenging to interpret when the number of distinct nodes is large. Combining treatments/doses/routes (lumping) risks obscuring differences or introducing heterogeneity, while creating a distinct node for each combination (splitting) reduces precision and increases the risks for identifying spurious relationships, chance violations of intransitivity, and implausibly large treatment effects. Existing data-based solutions have only been provided via custom Bayesian samplers or two-step frequentist models, and both approaches lack consideration for heterogeneity, particularly baseline risk, and integration of expert knowledge. This study addresses these limitations, aiming for a more robust and practical approach to treatment node definition in NMA.
METHODS: We adapted and extended a Bayesian non-parametric approach based on a Dirichlet process prior, replacing the the spike-and-slab base measure with a regularized horseshoe prior. This facilitates implementation with generic samplers (e.g., JAGS), avoids mixing issues, and frames treatment effects in terms of regularization. We incorporated meta-regression on baseline risk to account for heterogeneity, allowing for clustering under varying baseline risks. We also developed methods to integrate domain knowledge by limiting the number of clusters or specifying informative priors. We illustrate these methods in an application to the certolizumab baseline risk adjustment example from NICE TSD 3.
RESULTS: Node lumping differed under the unadjusted and adjusted models, with two groups in the former and three in the latter. Model fit statistics were worst in the unadjusted clustering model (DIC: 165) and best under the splitting baseline risk adjusted model (DIC: 155) and clustered model with baseline risk adjustment model (DIC: 156) representing the ability to reduce dimensionality in treatment effects by 50% without sacrificing model fit.
CONCLUSIONS: Data-based lumping/splitting decisions informed by prior expectation of unique treatments is feasible to implement within existing NICE TSD code. Lumped treatments can differ significantly in adjusted compared to unadjusted models.