Presentation (CTI)

OBJECTIVES: Partitioned survival models rely on extrapolated curves yet seldom report where the modelled benefit falls: inside the observed data, or in the assumed tail. Using RASolute302 in pancreatic cancer, where a large overall-survival (OS) effect remains immature, we applied the Extrapolation Benefit Share (EBS: the share of incremental mean survival arising beyond the observed data) and the Survival Ratio Shift (SRS: whether the advantage is tail-concentrated), with a relative-survival floor for OS.
METHODS: Pseudo-individual patient data were reconstructed (SurvdigitizeR) from the published progression-free survival (PFS) and OS Kaplan-Meier curves (daraxonrasib vs investigator's choice of chemotherapy; metastatic pancreatic cancer). Eight parametric distributions were fitted jointly (shared distribution, treatment covariate). Mean survival was decomposed into observed and extrapolated components over a 40-year horizon, with boundary determined by reverse-Kaplan-Meier diagnostic. OS was additionally modelled as relative survival (excess hazard plus ONS UK general-population mortality).
RESULTS: With nine months' median follow-up, both endpoints were extrapolation-dominated, and the modelled OS gain exceeded the PFS gain. For PFS, EBS was 37-69% with incremental survival 3-5 months; for OS, EBS was 65-86% with incremental survival 5-13 months. The population floor barely altered OS (relative survival ≈ all-cause; excess mortality dominates). In the extrapolation period distributions split sharply: increasing-hazard fits keep the excess hazard rising (persist), whilst long-tailed fits turn it over (fade) - a data-unresolved choice swinging the OS gain ~2.5-fold. Crucially, SRS exceeded 1 under every distribution (≈1.9-3.7): the OS advantage is loaded into extrapolated tail whichever extrapolation is chosen.
CONCLUSIONS: A large, significant OS effect can rest almost entirely on extrapolation. It’s lifetime magnitude hinging on an open persist-versus-fade choice that a population-mortality floor cannot resolve. Reporting EBS with SRS makes explicit not only how much of the benefit is assumed, but that it is structurally a tail phenomenon.