Presentation (CTI)

OBJECTIVES: Health technology assessment in oncology often requires extrapolating overall survival beyond trial follow-up despite limited tail data and high decision uncertainty. External real-world controls can improve long-term precision, but inappropriate borrowing may distort survival tails and mislead value conclusions when trial and external populations differ. This study compared Bayesian borrowing strategies, including conflict-robust options, and assessed downstream cost-effectiveness impacts.
METHODS: A randomized trial setting (N=480; 1:1 allocation; 24-month censoring; 214 deaths) with an external real-world control cohort (N=1500; 60-month follow-up) was evaluated. Overall survival was modeled using Bayesian proportional hazards with flexible baseline hazards. External information entered baseline parameters via power priors with fixed borrowing weights (0.2, 0.5, 1.0), an estimated borrowing weight, and commensurate priors with a commensurability parameter controlling shrinkage. A robust mixture commensurate prior attenuated borrowing under conflict. Over 1000 replicates under concordant and conflicting external scenarios, bias, RMSE, and 95% coverage for five-year mean survival were assessed. Posterior draws informed a partitioned survival cost-effectiveness model with probabilistic sensitivity analysis.
RESULTS: Under concordance, borrowing reduced five-year mean survival uncertainty by 31-38% and RMSE by 12-18%, while maintaining 93-95% coverage. Under severe conflict, full borrowing introduced substantial bias (-0.41 life-years) and reduced coverage to 62%. Robust commensurate and estimated borrowing down-weighted external influence, limiting bias to -0.07 life-years and restoring coverage to 90-92%. In cost-effectiveness analysis, concordant borrowing narrowed incremental cost-effectiveness ratio (ICER) uncertainty by approximately 30%; under conflict, naïve full borrowing shifted mean ICER from $128,000 to $99,000/QALY, whereas robust approaches remained near baseline at $124,000-$130,000/QALY.
CONCLUSIONS: Bayesian borrowing can improve oncology extrapolation precision, but conflict-robust methods and transparent diagnostics are essential to avoid biased long-term tails and decision-relevant shifts in value assessments.