Presentation (CTI)

OBJECTIVES: Multilevel network meta-regression (ML-NMR) is increasingly preferred by Health Technology Assessment (HTA) agencies, such as the National Institute for Health and Care Excellence (NICE), for anchored population-adjusted indirect comparisons. During a recent European HTA submission in a rare indication, an initial fixed-effects network meta-analysis (NMA) faced agency requests for an ML-NMR to address imbalances in baseline treatment effect modifiers across trial populations.
METHODS: The primary rate outcome was initially modelled using Poisson likelihoods for both the individual patient data (IPD) trials and the aggregate-level data (AgD) trials. However, severe overdispersion in the IPD trials caused convergence issues and uninterpretable credible intervals. The standard approach to handle overdispersion is to use negative binomial (NB) models. Because the R package multinma (v0.7.2) lacked native NB support, the underlying Stan code was manually modified. Two approaches were tested: 1) applying an NB likelihood to both IPD and AgD models (a.k.a “dual-NB model”), and 2) applying an NB likelihood to the IPD model alongside a Poisson likelihood for the AgD model (a.k.a “hybrid NB-Poisson model”).
RESULTS: The dual-NB model failed to yield meaningful results, likely due to insufficient information to inform the AgD overdispersion parameter. Conversely, the hybrid NB-Poisson model successfully converged to a sensible posterior that supported the original NMA conclusions, despite persistent Stan warnings regarding divergent transitions.
CONCLUSIONS: This case study underscores the need for continued methodological development in ML-NMR framework extensions, particularly for non-continuous metrics like count data. When complex computational issues arise, alternative population-adjustment methods (e.g., MAIC, STC) remain vital sensitivity analyses to support robust HTA submissions.