OBJECTIVES: Extrapolation of immature overall survival (OS) or progression-free survival (PFS) trial data using standard parametric (SP) models based on prior distributions (e.g. Exponential, Weibull, Gompertz) is commonly used for economic modelling in HTA. Due to complex hazard functions, SP models may produce a poor fit for OS and PFS for immuno-oncology (I-O) treatments. This project aims to compare SP models with cubic spline (CS) methods, to assess the impact of extrapolation methods on estimated cost-effectiveness of nivolumab compared to dacarbazine in metastatic melanoma.

METHODS: OS and PFS data were extracted from digitised Kaplan-Meier curves reported in the CheckMate 066 trial. SP and CS models were fitted to two data cut-offs at 16 and 52 months to assess the effect of different follow-up periods on model choice. Incremental cost-effectiveness ratios (ICERs) per QALY gained were estimated using a partitioned survival model with a lifetime horizon.

RESULTS: ICERs for nivolumab vs. dacarbazine were similar for SP and CS models fitted to the 16-month data cut-off (£23,244, £24,026). Based on the 52-month cut-off, nivolumab was more cost-effective based on SP fit compared to CS (£22,269 vs. £34,850). CS models were a better fit to the KM data based on visual inspection. The choice of distribution for SP model of nivolumab 16-month OS presented a wide range of ICER estimates (£13,753-52,466). The Gompertz SP model often presented the lowest AIC, but also resulted in a significantly different predicted survival (and thus lower ICER) compared to other distributions.

CONCLUSIONS: The choice of statistical fit in SP models had a large impact on lifetime ICER with short-term trial data. Both models provided a good fit for the plateau in OS using the long data cut-off. CS models provide a robust fit and less uncertainty for HTA in the absence of long-term survival endpoints for I-O treatments.