Comparing and Selecting Network Meta-Analysis Models for Cost-Effectiveness Analysis of Anticancer Drugs

OBJECTIVES: To systematically evaluate how survival estimates from network meta-analyses (NMA) models affect oncology cost-effectiveness analysis (CEA) results, particularly under different proportional hazards (PH) assumptions and extrapolation strategies. METHODS: A total of 19 time-to-event NMA models were evaluated in this analysis, encompassing Cox proportional hazards (Cox-PH), fractional polynomial (FP), Royston-Parmar (RP) models, piecewise exponential (PWE), generalized gamma (Gengamma), and parametric survival models (PSM), with 2 extrapolation strategies for non-PH models: a constant-tail hazard ratio (Fixed-HR) and a parametric extrapolation approach (Varying-HR). NMA model performance was assessed via relative error (RE) by comparing CEA outcomes derived from NMA-based survival estimates with those from head-to-head clinical data. RESULTS: Fifteen CEA models were conducted. The selection of the NMA model exerted a greater influence on CEA outcomes than traditional model parameters (ie, costs or utilities). RP models appeared to perform best, while PWE, second-order FP, and Gengamma models tended to perform comparatively worse. Fixed-HR extrapolation approaches generally yielded more accurate long-term projections than Varying-HR methods did, particularly under non-PH conditions. Violations of the PH assumption substantially increased uncertainty, especially for non-PH models. Notably, the choice of NMA model led to reversals in cost-effectiveness conclusions in up to one-third of the evaluated CEA cases. CONCLUSIONS: NMA model choice substantially influences oncology CEA outcomes. Underutilized but well-performing RP models show strong potential for broader application. Fixed-HR extrapolation was preferred in our study. These findings are context specific, and further validation is needed. When PH assumption is violated, the selection of non-PH models requires greater caution. Our findings may offer insights to support future refinements in CEA methodological guidance.