OBJECTIVES : When projecting long-term survival for cost-effectiveness models, standard parametric models (SPM) may not adequately capture the survival benefits of new treatments with varying hazards or other complex survival data. Consequently, alternative extrapolation methods, like mixture cure models (MCM), have been been successfully applied, and accepted by health technology assessment (HTA) agencies, to best capture the heterogeneity of patients and their complex survival outcomes. The objective of this study is to assess the performance of MCMs compared to SPMs for extrapolation purposes.

METHODS : Early data-cuts of three trials in melanoma, breast cancer (BC) and multiple myeloma (MM) were extrapolated and compared with later data-cuts of the same trials using SPM and MCM methods. The exponential, Gompertz, loglogistic, lognormal and Weibull MCMs were compared with SPMs. The best fitting distributions were selected based on leave-one-out cross-validation information criterion (LOOIC). To determine which method best predicted log-term outcomes, the observed (mature data) vs predicted (immature data) curves were evaluated based on ΔMean Absolute Deviation (MAD). Lower LOOIC and ΔMAD imply better predictions.

RESULTS : When comparing the predictions of the long-term data to the immature data in the melanoma dataset where cure is clinically possible, ΔMAD was 10.89 and 23.73 months for MCM and SPM, respectively. In the BC dataset where long-term survivors were observed, ΔMAD was 17.40 months for MCM and 44.07 months for SPM. In the MM dataset where no cure or long-term survivors were observed, ΔMAD was 8.70 months for MCM and 7.21 SPM months for SPM.

CONCLUSIONS : For two out of the three trials, the MCM method had better predictions than the SPMs as validated by mature data. Where cure or long-term survivors are expected, the MCM method outperforms the SPM method.