Presentation (CTI)

OBJECTIVES: Indirect treatment comparisons (ITCs) are increasingly used in health technology assessment (HTA) when direct comparative evidence is unavailable. Anchored matching-adjusted indirect comparison (MAIC) and multilevel network meta-regression (ML-NMR) represent two established ITC approaches, differing in their approach to population adjustment. MAIC reweights individual patient data (IPD) to match aggregate comparator characteristics prior to fitting parametric survival models; ML-NMR jointly models baseline hazard and covariate-adjusted treatment effect within a single Bayesian framework using both IPD and aggregate data. This study compares parametric survival extrapolations from each method and their implications for long-term survival estimation in HTA.
METHODS: IPD for the index treatment and summary-level comparator data were simulated across scenarios varying population imbalance. MAIC was applied to reweight IPD, followed by fitting a range of parametric models. ML-NMR was implemented using multinma R package, with parametric survival likelihoods and covariate-adjusted treatment effects. Model performance was assessed using AIC/BIC for MAIC and LOOIC for ML-NMR, alongside visual inspection of extrapolated survival curves.
RESULTS: MAIC and ML-NMR produced differing patterns of model fit and survival extrapolation under population imbalance. Model selection based on information criteria identified different best-fitting distributions across methods and scenarios. MAIC results were sensitive to effective sample size after reweighting, with flexible distributions sometimes favoured. ML-NMR showed more consistent model performance across specifications. Visual comparisons indicated that best-fitting models could yield different long-term survival estimates, particularly with greater covariate imbalance.
CONCLUSIONS: MAIC and ML-NMR may yield different extrapolated survival estimates, particularly when populations are imbalanced. Differences are driven by both population adjustment approaches and the choice of parametric distribution. ML-NMR may offer greater stability when effective sample size is reduced under MAIC and allows estimation in alternative target populations through explicit covariate modelling. These findings support the use of multiple modelling approaches and fit diagnostics are recommended in HTA sensitivity analyses.