P50 Competing Risk and Multistate Model Compared to Partitioned Survival Model in Metastatic Non-Small Cell Lung Cancer


      Partitioned survival analysis (PartSA) is commonly used for economic evaluations in oncology. We compare PartSA to a competing risk semi-markov multi-state model (MSM) and investigate differences in estimated cost-effectiveness in metastatic non-small cell lung cancer from a French perspective.


      To populate both models, pooled data from Checkmate 017 and 057 was digitized to reconstruct patient-level data from overall survival, progression-free survival and post-progression survival Kaplan Meier curves. Models consisted of three states: progression-free, progressed disease and death. Statistical analysis was performed under R to fit parametric survival models for the PartSA and to estimate state transitions for the MSM (package “mstate”). Costs and utilities were integrated in the model using values from nivolumab’s French HTA submission. Outcomes were discounted at 2,5% per annum over a 7-year time horizon.


      Both models produced similar results during trial period (15 months) but showed discrepancies during extrapolation, inducing different costs and quality-adjusted life years (QALYs) over time. Incremental life years gained (LYG) of nivolumab versus docetaxel were 1,11 LYG (PartSA) and 0,88 LYG (MSM). However, both incremental cost effectiveness ratios (ICER) were similar, 88 979€/QALY (PartSA) and 90 148€/QALY (MSM). Scenarios assessing different parametric extrapolations produced an average of 112 496€/QALY (PartSA) and 113 085€/QALY (MSM) with coefficients of variation (CV) estimated at 19,0% (PartSA) and 16,3% (MSM). Probabilistic ICER results were similar between models with CV estimated at 21,8% (PartSA) and 23,2%(MSM).


      This study provides new evidence on structural uncertainty. The MSM produced similar results to the PartSA for a 15 months follow-up. Ideally, both models should be tested to validate extrapolations and to ensure that the structural uncertainty is limited. However, when the time horizon is short (<10 years), the impact of the model choice on the ICER seems limited.