OBJECTIVES: Survival extrapolation is needed to assess the long-term costs and benefits of new treatments for health technology assessment (HTA) decision making. We use both case studies and simulation approaches to quantify the accuracy and improvement in precision in long-term estimates of patient survival when incorporating external data.

METHODS: We use the survextrap R package to implement flexible Bayesian M-spline survival models for estimating overall survival in patients treated with radiotherapy for head and neck cancer (N = 213, 5-years follow-up), using models with and without external data from SEER population registry (data available from 6 to 25 years). We assess the impact of incorporating SEER data on reducing the posterior variance of restricted mean-survival time (RMST) estimates and survival probabilities at long-term timepoints up to 40 years. We use a simulation study to further assess the potential impact on bias when incorporating imperfect external data, where we generate real-world data based approximately on SEER data, where the hazard rates deviate by up to 20% from an underlying true model.

RESULTS: For the radiotherapy control arm, we observed that at landmark times of 20 and 40 years, the inclusion of SEER data reduced the posterior standard deviation in estimates of RMST by 64% and 81%, respectively. In simulations, when incorporating imperfect external data we found the relative bias in estimates of RMST at 20 and 40 years was 4% and 6%, respectively. This compared with relative bias of 7% and 18%, respectively, when extrapolating using trial data alone.

CONCLUSIONS: The Bayesian model demonstrated marked improvements in both the accuracy and precision of mean survival estimates when real-world data is incorporated. Moreover, while utilizing imperfect real-world data comes with an increase in bias, there are still notable gains in accuracy compared with a constant hazard extrapolation relying solely on trial data.