OBJECTIVES

: It is often not feasible to run controlled trials in oncology, especially for rare cancers or in heavily pre-treated patient populations. Unfortunately, single-arm trials and lack of direct comparator data pose challenges for Health Technology Appraisals with regards to comparative clinical assessments and cost effectiveness modelling (CEM). Determining which parametric survival distribution to use for extrapolations is one of the main issues for CEM. There are several methods to simulate survival curves. The aim of this study is to compare choice of distribution in the event of simulating data (SimD) versus estimation from individual patient data (IPD).

METHODS

: IPD from a clinical trial CheckMate025 (minimum follow-up 14 months) investigating efficacy and safety of an immuno-oncology agent (nivolumab) in patients with renal cell carcinoma was the main source of data used for this study. The survival data were simulated using a published algorithm. Distributions for overall survival were estimated for both SimD and IPD. Five distributions were examined: exponential, gamma, loglogistic, lognormal, and Weibull distribution. Goodness of fit was assessed graphically and by comparing several statistics such as Akaike’s Information Criteria (AIC), Bayesian Information Criteria, and loglikelihood.

RESULTS

: For both SimD and IPD, the best fit by statistical criterion, resulted from the loglogistic distributions. Differences in parameter estimates were observed with the shape parameters (SimD=1.50; IPD=1.48) and scale parameters (SimD=24.90; IPD=25.46). Graphical inspection revealed that differences between survival curves were less than 1% over a 20-year time horizon.

CONCLUSIONS

: Estimated distribution parameters varied little between the SimD and IPD. This study demonstrates that simulated survival data can be a good proxy for IPD when determining the optimal distribution for extrapolation in oncology modelling when there is a lack of direct comparator data.