Bayesian Parametric Mixture Cure Survival Models With Delayed Treatment Effects and Application to Targeted Therapies in Advanced Ovarian Cancer
Author(s)
Daniel J. Sharpe, PhD1, Georgia Yates, MSc1, Ashley E. Tate, PhD2, Tuli De, PhD3, Jacqueline Vanderpuye-Orgle, MSc, PhD3.
1Advanced Analytics, Access Consulting, Parexel, London, United Kingdom, 2Advanced Analytics, Access Consulting, Parexel, Amsterdam, Netherlands, 3Advanced Analytics, Access Consulting, Parexel, Durham, NC, USA.
1Advanced Analytics, Access Consulting, Parexel, London, United Kingdom, 2Advanced Analytics, Access Consulting, Parexel, Amsterdam, Netherlands, 3Advanced Analytics, Access Consulting, Parexel, Durham, NC, USA.
OBJECTIVES: Mixture cure models (MCMs) are commonly employed to analyze heterogeneity in survival outcomes but assume cure at randomization. This feature poses challenges for clinical interpretation of MCMs in studies of advanced cancers, where responses to novel therapies are typically delayed. To avoid long-term outcomes being misrepresented, we extended conventional parametric (Weibull) MCMs to allow an initial period during which all patients are “uncured”, by incorporating a cure delay time parameter.
METHODS: The proposed representation yields a change-point model in which a proportion of surviving patients become latently classified as statistically cured at the delay time threshold, while the remaining patients continue to follow the uncured survival function. We analyzed reconstructed progression-free survival (PFS) data for first-line maintenance olaparib or placebo plus bevacizumab in homologous recombination deficiency-positive advanced ovarian cancer from 5-year data of the phase III PAOLO-1 study. The model was estimated in a Bayesian framework with marginalization of the cure delay time parameter in monthly intervals and a corresponding weakly informative beta-binomial prior distribution.
RESULTS: The most probable (86.0%) cure delay time in the olaparib arm was 23 [90% credible interval (CrI): 23-24] months. The estimated cure fraction was 42.1% [95% CrI: 35.7-48.6%] (vs 43.3% [95% CrI: 36.1-50.1% in the conventional MCM) and 3-year PFS in the uncured subpopulation was 40.8% [95% CrI: 32.4-50.2%] (vs 25.5% [95% CrI: 18.4-34.4%] in the conventional MCM). In the placebo arm, the estimated cure delay time was 25 [90% CrI: 8-26] months and cure fraction was 19.3% [95% CrI: 15.7-24.6%].
CONCLUSIONS: Extension of MCMs to relax the baseline cure assumption provides a more clinically plausible survival model for advanced cancers, and yields slightly more conservative estimates for the cure fraction. Reasonable uncertainty intervals and similarity of the posterior distribution for the cure delay time between the treatment arms of PAOLO-1 support robustness of the proposed method.
METHODS: The proposed representation yields a change-point model in which a proportion of surviving patients become latently classified as statistically cured at the delay time threshold, while the remaining patients continue to follow the uncured survival function. We analyzed reconstructed progression-free survival (PFS) data for first-line maintenance olaparib or placebo plus bevacizumab in homologous recombination deficiency-positive advanced ovarian cancer from 5-year data of the phase III PAOLO-1 study. The model was estimated in a Bayesian framework with marginalization of the cure delay time parameter in monthly intervals and a corresponding weakly informative beta-binomial prior distribution.
RESULTS: The most probable (86.0%) cure delay time in the olaparib arm was 23 [90% credible interval (CrI): 23-24] months. The estimated cure fraction was 42.1% [95% CrI: 35.7-48.6%] (vs 43.3% [95% CrI: 36.1-50.1% in the conventional MCM) and 3-year PFS in the uncured subpopulation was 40.8% [95% CrI: 32.4-50.2%] (vs 25.5% [95% CrI: 18.4-34.4%] in the conventional MCM). In the placebo arm, the estimated cure delay time was 25 [90% CrI: 8-26] months and cure fraction was 19.3% [95% CrI: 15.7-24.6%].
CONCLUSIONS: Extension of MCMs to relax the baseline cure assumption provides a more clinically plausible survival model for advanced cancers, and yields slightly more conservative estimates for the cure fraction. Reasonable uncertainty intervals and similarity of the posterior distribution for the cure delay time between the treatment arms of PAOLO-1 support robustness of the proposed method.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR45
Topic
Methodological & Statistical Research
Disease
Oncology