OBJECTIVE: In many oncology trials, especially those of innovative immunotherapies, many patients remain alive at the end of the trial resulting in the censoring of survival times. This introduces uncertainty over long-term survival curve extrapolation when estimating life-years gained (LYG), and consequently impacts decision making at the time of health technology assessment. The objective of this study is to assess, under a number of clinically plausible scenarios, how the amount of censoring impacts the uncertainty arising from survival curve extrapolations.

METHODS: To reflect a real-world scenario, we assumed staggered entry of patients (N=125, 250, 500, 1000) into a study. A simulation study with 1,000 iterations with survival curve extrapolation using exponential, Weibull, Gompertz, log-logistic, log normal and generalised gamma time to event data was conducted under different censoring scenarios (0%, 20%, 40%, 60% and 80%). Area under the curve (AUC) was used to estimate the LYG and its variability using three approaches: on the Kaplan-Meier estimate (0% censoring), parametric extrapolations, and a combination of the two.

RESULTS: In general, under all simulated scenarios and AUC estimation approaches, the immature data did not introduce any estimation bias. However, it is apparent that a reduction in sample size and an increase in censoring (i.e., more immature data) rapidly increases the variability of AUC, thus the LYG. Specifically, the AUC variability of the exponential model with constant hazard was approximately proportional to the amount of censoring. while the Weibull, Gompertz and log-logistic models performed better than the exponential when censoring was low (<60%). The generalised gamma and the log normal models were most susceptible to AUC variability.

CONCLUSIONS: As there is no universally accepted definition in the literature on a minimum required level of survival data maturity, the results are useful in informing payers on uncertainty this can introduce in decision making.