BAYESIAN MULTI-PARAMETER EVIDENCE SYNTHESIS FOR INFORMED EXTRAPOLATIONS OF TIME-TO-EVENT OUTCOMES IN MULTISTATE MODELS
Author(s)
Daniel J. Sharpe, PhD1, Ashley E. Tate, PhD2, Tuli De, PhD3, Jackie Vanderpuye-Orgle, MSc, PhD3;
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International Ltd, Durham, NC, USA
1Parexel International Ltd, London, United Kingdom, 2Parexel International Ltd, Amsterdam, Netherlands, 3Parexel International Ltd, Durham, NC, USA
OBJECTIVES: Bayesian multi-parameter evidence synthesis (B-MPES) provides a flexible approach to incorporate historical trial data into extrapolations. Here, we adapted B-MPES survival models to the multistate problem, and conducted a simulation study to validate that the method can improve reliability and reduce uncertainty in joint extrapolations of overall and progression-free survival from immature data in oncology studies.
METHODS: We developed a continuous-time semi-Markov model based on parametric (Weibull) transition intensity functions for a three-state irreversible progressive illness-death model within a B-MPES framework. External information on state occupation probabilities at annual timepoints beyond the minimum follow-up period was incorporated via additional likelihood terms based on concentration parameters of Dirichlet distributions. We demonstrated the approach by fitting B-MPES and naïve models to synthetic interim (2-year) trial data emulating outcomes for patients with high-risk advanced ovarian cancer treated with PARP inhibitor plus bevacizumab (N=200). Prior expectations for longer-term (3-, 4-, and 5-year) state occupation probabilities were derived by applying 1-year transition probabilities, estimated from 5-year data for a synthetic historic trial of bevacizumab, to the 2-year observations. Model estimates were compared to the final (5-year) trial data.
RESULTS: The B-MPES model yielded estimates for the proportions of progression-free (16.0% [95% credible interval: 9.4-23.2%] and surviving (24.1% [16.6-32.6%]) patients at 5 years that were relatively accurate but somewhat conservative, compared to final observations (Aalen-Johansen estimates 24.7% [95% confidence interval: 19.0-32.2%] and 31.0% [23.7-37.6%], respectively). The B-MPES model was more accurate and much less uncertain than the naïve model, which overestimated the proportions of progression-free (35.3% [17.8-49.2%]) and surviving (47.9% [28.8-63.3%]) patients.
CONCLUSIONS: Incorporating prior knowledge on expected longer-term occupation probabilities of health states using B-MPES provides an intuitive approach to avoid speculation and reduce statistical uncertainty in extrapolations from multistate models fitted to early data cuts, which are a major source of decision risk in cost-effectiveness analyses.
METHODS: We developed a continuous-time semi-Markov model based on parametric (Weibull) transition intensity functions for a three-state irreversible progressive illness-death model within a B-MPES framework. External information on state occupation probabilities at annual timepoints beyond the minimum follow-up period was incorporated via additional likelihood terms based on concentration parameters of Dirichlet distributions. We demonstrated the approach by fitting B-MPES and naïve models to synthetic interim (2-year) trial data emulating outcomes for patients with high-risk advanced ovarian cancer treated with PARP inhibitor plus bevacizumab (N=200). Prior expectations for longer-term (3-, 4-, and 5-year) state occupation probabilities were derived by applying 1-year transition probabilities, estimated from 5-year data for a synthetic historic trial of bevacizumab, to the 2-year observations. Model estimates were compared to the final (5-year) trial data.
RESULTS: The B-MPES model yielded estimates for the proportions of progression-free (16.0% [95% credible interval: 9.4-23.2%] and surviving (24.1% [16.6-32.6%]) patients at 5 years that were relatively accurate but somewhat conservative, compared to final observations (Aalen-Johansen estimates 24.7% [95% confidence interval: 19.0-32.2%] and 31.0% [23.7-37.6%], respectively). The B-MPES model was more accurate and much less uncertain than the naïve model, which overestimated the proportions of progression-free (35.3% [17.8-49.2%]) and surviving (47.9% [28.8-63.3%]) patients.
CONCLUSIONS: Incorporating prior knowledge on expected longer-term occupation probabilities of health states using B-MPES provides an intuitive approach to avoid speculation and reduce statistical uncertainty in extrapolations from multistate models fitted to early data cuts, which are a major source of decision risk in cost-effectiveness analyses.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR115
Topic
Methodological & Statistical Research
Disease
SDC: Oncology