OBJECTIVES:

Survival estimates used in cost-effectiveness analyses are often generated by applying network meta-analysis (NMA)-derived hazard ratios (HRs) to extrapolated trial data for a chosen ‘reference’ treatment. Limited guidance is available for selecting this reference treatment yet survival estimates are sensitive to the choice. This research investigated which reference treatment might be the most appropriate.

METHODS:

A published NMA for non-small cell lung cancer served as a case study to estimate survival for erlotinib. Reference treatments were chosen from trials in the NMA based on: most aligned population, longest follow-up, data maturity, NMA reference treatment, and most recent data. Survival functions for which the proportional hazards assumption holds (exponential, Weibull, Gompertz) were fitted to progression-free survival (PFS) data from each relevant trial to derive reference curves. NMA-derived HRs were applied to estimate PFS for erlotinib and validated against a real-world evidence (RWE) study with long-term follow-up of a similar patient population.

RESULTS:

Across the survival functions, the reference treatment with the most aligned population (gefitinib) produced mean PFS estimates closest to RWE (17.1% average deviation; 3.0 months). The NMA reference treatment (chemotherapy) had the poorest estimate (28.8% average deviation; 5.1 months), with a maximum deviation of 39.5% (7.0 months; Weibull). Data maturity (dacomitinib) estimated the most consistent results across survival functions (range=0.93 months), however, was less consistent with RWE (21.2% average deviation; 3.8 months).

CONCLUSIONS:

There are several factors to consider when choosing the most appropriate reference treatment for NMA-derived HRs. Similarity in patient population is expected to be an important factor. Importantly, the reference treatment of an NMA might not be an appropriate reference treatment for extrapolation, and data maturity can minimize uncertainty in extrapolations but not necessarily have external validity. However, this research considered only one NMA and RWE validation as a case study. Furthermore, conclusions might be endpoint specific.