Presentation (CTI)

OBJECTIVES: Multilevel network meta-regression (ML-NMR) is increasingly used in indirect treatment comparisons to adjust for cross-study differences in effect modifiers and estimate effects in target populations. Guidance on incorporating outputs into cost-effectiveness models, particularly probabilistic sensitivity analysis (PSA), remains limited. Directly applying conditional effects to marginal baseline survival curves mixes incompatible estimands, introducing aggregation bias. Considering marginal curves is recommended but computationally intensive. This study compares alternative approaches for including ML-NMR outputs into PSA.
METHODS: Our case study used the newly diagnosed multiple myeloma network from multinma package. Progression-free survival (PFS) was informed by three individual patient data (IPD) trials and two aggregate-data (AgD) trials evaluating lenalidomide, thalidomide, and placebo. Overall survival (OS) was reconstructed using digitised AgD publications and simulated curves for the IPD trials, calibrated to published outcomes. A partitioned survival model was considered. Three PSA approaches were compared to the marginal survival curves across all posterior draws: (1) marginal curves from a random subset of draws; (2) marginal-curve reconstruction by sampling survival-model coefficients from their multivariate normal distribution with a patient covariate sample; (3) conditional model evaluated at mean covariates ("average patient"). Approaches were assessed on computational efficiency and ability to reproduce model outcome estimates
RESULTS: Approaches 1-2 produced similar estimates of progression-free, post-progression, and total life-years, with overlapping distributions and comparable means to the reference. Approach 1-2 produced similar uncertainty to the reference, whereas approach 3 showed greater divergence. Approach 3 produced biased estimates for both PFS and OS outcomes. Runtime savings were negligible with approach 1, whereas approach 2 reduced runtime by approximately 80% relative to the reference. Approach 3 was the fastest method.
CONCLUSIONS: PSA implementation to incorporating ML-NMR outputs can influence results. Conditional approaches may reduce computational burden but risk bias. Marginal-curve reconstruction reproduced reference results while reducing runtime, offering an efficient approach.