ESTIMATION OF OVERALL SURVIVAL FROM INITIAL PROGRESSION-FREE SURVIVAL DATA: A BAYESIAN PARAMETRIC COPULA-BASED APPROACH LEVERAGING EARLIER-PHASE TRIAL DATA
Author(s)
Daniel J. Sharpe, PhD1, Ashley E. Tate, PhD2, Tuli De, PhD3, Jackie Vanderpuye-Orgle, MSc, PhD4.
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International, Cupertino, CA, USA, 4Parexel International Ltd, La Verne, CA, USA.
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International, Cupertino, CA, USA, 4Parexel International Ltd, La Verne, CA, USA.
OBJECTIVES: Reimbursement submissions based on interim data for a composite survival endpoint such as progression-free survival (PFS) present high decision risk since payers require lifetime overall survival (OS) estimates to evaluate cost-effectiveness. Here, we proposed that when trial OS data remain unavailable but earlier-phase trial data can provide a prior representation of the joint PFS-OS distribution, a parametric bivariate survival copula model estimated from these historical data can be combined with an extrapolated PFS distribution for the current trial to yield estimated OS probabilities.
METHODS: PFS-OS data were simulated for phase III (interim PFS and final analyses) and II trials of first-line immunochemotherapy in patients with advanced cervical cancer. Five candidate copula functions (namely: Clayton, Frank, Hougaard, Joe, and Plackett) encompassing archetypal association patterns were fitted to the phase II trial data. A parametric distribution for the PFS density in the phase III trial was then estimated, employing a Bayesian framework with informative multivariate normal prior distribution. At each Monte Carlo iteration, estimated long-term OS probabilities in the phase III trial were calculated as a weighted sum of conditional OS distributions for given PFS, according to the extrapolated PFS probability density.
RESULTS: The method accurately predicted OS probabilities from the interim phase III PFS trial data when certain criteria were met, including: appropriate parametric marginal distribution and copula function specification, strong prognostic relevance of progression events (e.g., Spearman’s rho >0.7), and highly similar dependence structure of OS on PFS between the studies. The approach was sensitive to the PFS extrapolations; incorporating prior information on longer-term PFS hazards into the phase III trial predictions enabled reasonable uncertainty levels in the corresponding OS estimates.
CONCLUSIONS: When earlier-phase trial data are available to characterize the treatment- and population-specific PFS-OS dependence pattern, the copula-based model provides an interpretable approach to generate plausible OS probabilities from initial PFS data.
METHODS: PFS-OS data were simulated for phase III (interim PFS and final analyses) and II trials of first-line immunochemotherapy in patients with advanced cervical cancer. Five candidate copula functions (namely: Clayton, Frank, Hougaard, Joe, and Plackett) encompassing archetypal association patterns were fitted to the phase II trial data. A parametric distribution for the PFS density in the phase III trial was then estimated, employing a Bayesian framework with informative multivariate normal prior distribution. At each Monte Carlo iteration, estimated long-term OS probabilities in the phase III trial were calculated as a weighted sum of conditional OS distributions for given PFS, according to the extrapolated PFS probability density.
RESULTS: The method accurately predicted OS probabilities from the interim phase III PFS trial data when certain criteria were met, including: appropriate parametric marginal distribution and copula function specification, strong prognostic relevance of progression events (e.g., Spearman’s rho >0.7), and highly similar dependence structure of OS on PFS between the studies. The approach was sensitive to the PFS extrapolations; incorporating prior information on longer-term PFS hazards into the phase III trial predictions enabled reasonable uncertainty levels in the corresponding OS estimates.
CONCLUSIONS: When earlier-phase trial data are available to characterize the treatment- and population-specific PFS-OS dependence pattern, the copula-based model provides an interpretable approach to generate plausible OS probabilities from initial PFS data.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR141
Topic
Methodological & Statistical Research
Disease
Oncology