OBJECTIVES: Manufacturers often seek reimbursement of new treatments based on early clinical trial results. Extrapolation of overall survival (OS) trial data to a lifetime is often required to estimate cost-effectiveness. In oncology, this involves fitting parametric models to trial Kaplan-Meier data, however, different models will lead to different long-term survival estimates. For renal cell carcinoma, previous reimbursement submissions for nivolumab used OS data from a 1-year data cut of the CheckMate 025 (CM-025) trial. We assessed predicted survival from parametric models fitted to these data against a 3-year survival update from CM-025.

METHODS: We used a previous reimbursement dossier to fit a spline-based survival model to 1-year OS data for nivolumab and everolimus from CM-025. Function selection was based on criteria including visual inspection, statistical model fit, and validation versus longer-term data from the phase I CA209-003 study. A number of other parametric survival models were also evaluated.

RESULTS: Our base-case spline function fitted to early Kaplan-Meier data from CM-025 predicted 3-year OS rates of 36.1% and 27.3% for nivolumab and everolimus, respectively. Analysis of the updated 3-year data shows observed 3-year OS rates of 39.1% and 29.5%, respectively. A lognormal distribution, which gave closer predictions to the long-term data from CA209-003, estimated 3-year OS rates of 41.6% and 32.0 %, and the best fitting non-spline-model, log-logistic, provided rates of 39.1% and 29.5% for nivolumab and everolimus, respectively.

CONCLUSIONS: Three-year OS rate predictions using base-case parametric models originally fitted to 1-year survival data from CM-025 were conservative and underestimated the actual, observed 3-year survival rates. The observed incremental difference in 3-year OS between nivolumab and everolimus was also greater than the predicted estimates, with implications for cost-effectiveness analyses. These results show the importance of scenario testing with different OS curves in economic analyses and retrospective validation of the original results.