TITLE: JOINT VERSUS INDEPENDENT MODELING OF PROGRESSION-FREE AND OVERALL SURVIVAL IN NICE ONCOLOGY APPRAISALS: A REVIEW AND SIMULATION STUDY OF THE IMPACT ON DECISION UNCERTAINTY
Author(s)
Emma Hawe, BSc, MSc1, Andrea Berardi, PhD2.
1SVP, Managing Director, Precision AQ, London, United Kingdom, 2Precision AQ, London, United Kingdom.
1SVP, Managing Director, Precision AQ, London, United Kingdom, 2Precision AQ, London, United Kingdom.
OBJECTIVES: Cost-effectiveness modelling in oncology typically extrapolate progression‑free survival (PFS) and overall survival (OS) independently. Because PFS and OS are correlated, this can generate implausible extrapolations. This study assessed the extent to which NICE oncology appraisals account for PFS-OS dependence and evaluated implications for uncertainty and decision‑making.
METHODS: Candidate joint frameworks (copula‑based bivariate survival, multi‑state illness-death, and shared‑frailty models) were reviewed. Twenty most recent NICE oncology TAs were examined, extracting survival‑modelling choices, treatment of PFS-OS dependence, structural constraints, and EAG critique. Correlated PFS and OS data were simulated across dependence structures (Kendall’s τ 0-0.8), comparing independently fitted models with a copula‑based approach. Outcomes included bias and uncertainty in time‑in‑state, QALYs, and cost‑effectiveness acceptability curves, implemented in an R Shiny application.
RESULTS: None of the appraisals used a formal joint model for PFS and OS. Of the 17 that modelled survival, independent partitioned survival models were used in 88% despite repeated committee concerns regarding implausible extrapolations (e.g. PFS and OS curves crossing in 18%, or long post‑progression survival). In practice, these were addressed through ad hoc distributional adjustments, structural constraints, or occasional multi‑state models, rather than formal modelling of PFS-OS dependence (e.g. copula‑based or shared‑frailty models). NICE and DSU guidance provide no recommendation for joint extrapolation despite assuming structural relationships, highlighting inconsistency in how dependence is treated. Simulations showed that ignoring correlation mischaracterised uncertainty, increasing with dependence. Independent modelling understated uncertainty in overall survival, while impacts on net benefit varied with utility assumptions, shifting acceptability curves despite similar extrapolations. The R Shiny tool identified when PFS-OS dependence would materially affect extrapolations and decision‑relevant conclusions.
CONCLUSIONS: Independent survival modelling, the NICE standard, ignores PFS‑OS correlation and may mischaracterise uncertainty. Joint models (e.g. copulas) impose coherent structure without restricting model shape and generate uncertainty that propagates appropriately to QALYs, particularly in probabilistic analyses, and with immature data.
METHODS: Candidate joint frameworks (copula‑based bivariate survival, multi‑state illness-death, and shared‑frailty models) were reviewed. Twenty most recent NICE oncology TAs were examined, extracting survival‑modelling choices, treatment of PFS-OS dependence, structural constraints, and EAG critique. Correlated PFS and OS data were simulated across dependence structures (Kendall’s τ 0-0.8), comparing independently fitted models with a copula‑based approach. Outcomes included bias and uncertainty in time‑in‑state, QALYs, and cost‑effectiveness acceptability curves, implemented in an R Shiny application.
RESULTS: None of the appraisals used a formal joint model for PFS and OS. Of the 17 that modelled survival, independent partitioned survival models were used in 88% despite repeated committee concerns regarding implausible extrapolations (e.g. PFS and OS curves crossing in 18%, or long post‑progression survival). In practice, these were addressed through ad hoc distributional adjustments, structural constraints, or occasional multi‑state models, rather than formal modelling of PFS-OS dependence (e.g. copula‑based or shared‑frailty models). NICE and DSU guidance provide no recommendation for joint extrapolation despite assuming structural relationships, highlighting inconsistency in how dependence is treated. Simulations showed that ignoring correlation mischaracterised uncertainty, increasing with dependence. Independent modelling understated uncertainty in overall survival, while impacts on net benefit varied with utility assumptions, shifting acceptability curves despite similar extrapolations. The R Shiny tool identified when PFS-OS dependence would materially affect extrapolations and decision‑relevant conclusions.
CONCLUSIONS: Independent survival modelling, the NICE standard, ignores PFS‑OS correlation and may mischaracterise uncertainty. Joint models (e.g. copulas) impose coherent structure without restricting model shape and generate uncertainty that propagates appropriately to QALYs, particularly in probabilistic analyses, and with immature data.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR197
Topic
Health Technology Assessment, Methodological & Statistical Research
Disease
Oncology