WHEN THE TAIL WAGS THE DOG - AND WHEN A POPULATION FLOOR REINS IT IN: EXTRAPOLATION TRANSPARENCY AND RELATIVE SURVIVAL ACROSS FIVE ONCOLOGY TRIALS
Author(s)
Andrew Briggs, DPhil1, Alex Hill, PhD1, Anna Forsythe, MBA, MSc, PharmD2.
1London School of Hygiene & Tropical Medicine, London, United Kingdom, 2Founder, Oncoscope, Miami, FL, USA.
1London School of Hygiene & Tropical Medicine, London, United Kingdom, 2Founder, Oncoscope, Miami, FL, USA.
OBJECTIVES: Modelled lifetime benefit in oncology rests heavily on survival extrapolation, yet how much of it is assumed, where the benefit sits, and whether a population-mortality floor should discipline it are seldom reported. We applied one framework - the Extrapolation Benefit Share (EBS), the Survival Ratio Shift (SRS), and a relative-survival (excess-hazard) floor - across five recent trials in breast, pancreatic, and bladder cancer.
METHODS: For each trial, pseudo-individual patient data were reconstructed from the published Kaplan-Meier curves; eight parametric distributions were fitted jointly; mean survival was decomposed into observed (RMST) and extrapolated (EMST) components over a lifetime horizon; and overall survival (OS) was additionally modelled as relative survival with a population-mortality floor. Trials were ordered by whether the excess hazard fades towards a long-term-survivor pattern or persists.
RESULTS: EBS tracked maturity, not statistical significance: near-total for immature OS (HER2CLIMB-05 95-100%; KEYNOTE-905 75-83%), still high where medians were mature (TROPION-Breast02 65-82%), and low only where follow-up was genuinely deep (PANOVA-3 13-40%) - so mature medians did not guarantee a trustworthy extrapolation. SRS exposed tail-loaded benefit (RASolute302 1.9-3.7). The floor's value was conditional and predictable: it did almost nothing where disease-specific mortality dominated (pancreatic, triple-negative breast), and earned its keep in the curative-intent, high-cure-fraction setting (KEYNOTE-905), clipping an implausible OS extrapolation (Gompertz 104→38 months; range 47-104→20-38).
CONCLUSIONS: One transparent spine - EBS, SRS, and a relative-survival floor - separates observed from assumed benefit, locates where it sits, and shows when a population anchor is essential rather than cosmetic. Because the floor is costless where excess mortality dominates and decisive where a cure fraction exists, it should be applied as standard - the survival-extrapolation counterpart to the background mortality already routine in Markov models. Its action, or inaction, is itself diagnostic of how a cancer's excess mortality behaves.
METHODS: For each trial, pseudo-individual patient data were reconstructed from the published Kaplan-Meier curves; eight parametric distributions were fitted jointly; mean survival was decomposed into observed (RMST) and extrapolated (EMST) components over a lifetime horizon; and overall survival (OS) was additionally modelled as relative survival with a population-mortality floor. Trials were ordered by whether the excess hazard fades towards a long-term-survivor pattern or persists.
RESULTS: EBS tracked maturity, not statistical significance: near-total for immature OS (HER2CLIMB-05 95-100%; KEYNOTE-905 75-83%), still high where medians were mature (TROPION-Breast02 65-82%), and low only where follow-up was genuinely deep (PANOVA-3 13-40%) - so mature medians did not guarantee a trustworthy extrapolation. SRS exposed tail-loaded benefit (RASolute302 1.9-3.7). The floor's value was conditional and predictable: it did almost nothing where disease-specific mortality dominated (pancreatic, triple-negative breast), and earned its keep in the curative-intent, high-cure-fraction setting (KEYNOTE-905), clipping an implausible OS extrapolation (Gompertz 104→38 months; range 47-104→20-38).
CONCLUSIONS: One transparent spine - EBS, SRS, and a relative-survival floor - separates observed from assumed benefit, locates where it sits, and shows when a population anchor is essential rather than cosmetic. Because the floor is costless where excess mortality dominates and decisive where a cure fraction exists, it should be applied as standard - the survival-extrapolation counterpart to the background mortality already routine in Markov models. Its action, or inaction, is itself diagnostic of how a cancer's excess mortality behaves.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
EE517
Topic
Clinical Outcomes, Economic Evaluation, Methodological & Statistical Research
Topic Subcategory
Trial-Based Economic Evaluation
Disease
Oncology