OBJECTIVES: Especially for cancers developing late in life, the reported all-cause mortality hazards (ACM) reflect a mixture of general population mortality hazards (GPM) and disease-specific mortality hazards (DSM). This study aims to analyze standard and novel approaches that can be used to incorporate GPM in survival extrapolations.

METHODS: Six approaches were explored; (1) GPM not considered (“standard”), (2) GPM considered once parametric hazards < GPM hazards (“converging hazards”), (3) GPM added to parametric hazards after the fit (“additive outside the fit”), (4) GPM added to parametric hazards in the fit (“additive in the fit”), (5) a “mixture cure model” and (6) proportional hazards compared to GPM. The approaches were applied to two datacuts (35 and 78 months) of a multiple myeloma (MM) dataset and to a breast cancer (BC) dataset. The fit, hazards over time, mean (incremental) survival, and corresponding uncertainty of some of the most common parametric models (e.g. exponential, Weibull, and log-logistic) were compared for the approaches.

RESULTS: The differences between the approaches were largest in the 35 months MM dataset where decreasing hazards within the trial were observed. The log-logistic incremental survival estimates were 5.1 (0.7 - 9.3), 2.9 (0.5 - 5.0), 2.1 (0.4 - 3.6), 3.6 (1.3 - 5.5), and 3.7 (0.5 - 6.1) years for the first five approaches, respectively. For approach 6 (proportional hazards), the incremental survival estimate was 1.6 (0.4 - 3.0) years.

CONCLUSIONS: Especially, in case of decreasing hazards and immature data, the six tested approaches resulted in different incremental mean survival and corresponding uncertainty estimates. Clinical plausibility of the extrapolations of the tested approaches will differ based on the underlying data Therefore, all extrapolations should be compared to historical data or validated by clinical experts.