A Simulation and Case Study to Evaluate the Extrapolation Performance of Flexible Bayesian Survival Models when Incorporating Real-World Data

Full abstract

Read the full abstract

BackgroundAssessment of long-term survival for health technology assessment often necessitates extrapolation beyond the duration of a clinical trial. Flexible Bayesian survival models that incorporate longer-term data sources, including registry data and population mortality, have been proposed as an alternative to using standard parametric models with trial data alone.MethodsThe performance of extrapolations from the survextrap Bayesian M-spline survival model was evaluated through simulation, with 5 years of trial data follow-up for our primary analysis. We assessed long-term survival and incremental effect estimates under a range of settings and assumed long-term real-world data informed the control arm. Comparisons were made with standard and flexible parametric models.ResultsWhen relevant long-term external data on the control arm were available, a flexible Bayesian approach substantially improved the precision and accuracy of extrapolations for that arm. Improvements in estimates of extrapolated treatment effects were more variable and sometimes deteriorated, reflecting sensitivity to assumptions about how hazards relate across arms. Compared with exponential and Weibull models, the survextrap Bayesian model had better within-trial fit and more plausible extrapolations, although a Royston-Parmar spline often provided comparable accuracy of incremental effects in many settings. We noted further that the survextrap Bayesian approach has the potential strength of explicitly quantifying the structural uncertainty, making transparent the strong assumptions required by any survival modeling approach to extrapolate.ConclusionsFlexible Bayesian survival models can potentially improve long-term survival extrapolation in the control arm when well-matched external data are available, but improvements in estimates of incremental treatment effects were often quite variable across scenarios and models. This highlights the need for careful consideration of appropriate modeling assumptions and real-world data quality when using this approach.HighlightsFlexible Bayesian models that incorporate both clinical trial and real-world data can improve the accuracy of control-arm survival extrapolations compared with using trial data alone.Estimates of incremental long-term treatment benefit were sensitive to assumptions about treatment-effect structure, highlighting the need for clinically motivated modeling choices and sensitivity analyses.Our simulation study guides users on the important modeling choices, sensitivity analyses, and software settings to consider when implementing these flexible Bayesian survival models.By allowing the hazard to vary smoothly outside the data, flexible Bayesian methods can enable better quantification of structural uncertainty in survival extrapolations.