Presentation (CTI)

OBJECTIVES: Multilevel network meta-regression (ML-NMR) allows prediction of relative and absolute effects in target populations, including those external to the evidence network. These are typically collapsed into static estimates before being incorporated into cost-effectiveness models (CEMs), limiting the ability to tailor estimates to the target population relevant for decision-making. We propose a theoretical framework for embedding ML-NMR posterior samples exported in Convergence Diagnosis and Output Analysis (CODA) format directly into an Excel-based CEM, enabling dynamic predictions of efficacy and safety inputs based on local population characteristics.
METHODS: The fitted ML-NMR is summarised in exportable components: posterior draws for model parameters including covariate beta coefficients, borrowed covariance structures, and marginal covariate distributions. These populate a parameter input sheet. The user specifies the local target population via two inputs: baseline risk (comparator response) and baseline characteristics (e.g. age ± variance). In the context of probabilistic CEMs, efficacy estimates can be tailored to the target population by drawing a parameter vector directly from the CODA samples and numerically integrating the inverse-link-transformed predictions over the local population's covariate distribution. This marginalization step yields a population-average (marginal) estimate on the natural scale, respecting the non-collapsibility of non-linear links. Individual draws can be directly used for the probabilistic sensitivity analysis.
RESULTS: Conceptually, the approach preserves joint parameter uncertainty, performs the marginalisation at the point of decision-making rather than at analysis stage, avoids the information loss of point-estimate transfer, and enables one fitted model serve heterogeneous decision contexts without refitting. It reflects effect-modification, prognostic effects & baseline risk. The Excel-resident CODA samples make adjustment accessible and transparent to HTA reviewers.
CONCLUSIONS: Direct CODA-to-CEM integration reframes ML-NMR as a live input layer rather than a one-off estimate, aligning population-adjusted indirect comparisons with the practical realities of HTA modelling. The computational costs associated with CODA-to-CEM merits further evaluation.