Borrowing from historical control data in a Bayesian time-to-event model with flexible baseline hazard function.

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Pharmacometrics

Background:

• Statistical methods leveraging historical trial data accelerate drug discovery and development.
• Bayesian methods offer dynamic borrowing, adjusting information incorporation based on data similarity.
• Existing Bayesian borrowing methods often approximate baseline hazard shapes, potentially limiting accuracy.

Purpose of the Study:

• To propose a novel Bayesian semiparametric model for borrowing information from one historical dataset.
• To improve the accuracy of modeling the baseline hazard in time-to-event analyses.
• To enhance statistical power and reduce bias in treatment effect estimation.

Main Methods:

• Developed a Bayesian semiparametric model allowing unrestricted baseline hazard forms via an ensemble average.
• Introduced smoothing priors for the posterior baseline hazard to enhance estimation and borrowing.
• Explored various prior borrowing structures, including a lump-and-smear prior.
• Proposed a principled method for selecting hyperparameters to control dynamic borrowing.

Main Results:

• The proposed model improves power and reduces bias in treatment effect estimation when exchangeability holds.
• Accurate baseline hazard modeling, rather than approximation, is key to effective borrowing.
• The model reduces type I error inflation in the presence of prior-data conflict.
• The lump-and-smear prior demonstrated benefits in improving type I error and power.

Conclusions:

• The novel Bayesian semiparametric model offers a more accurate and flexible approach to borrowing historical data.
• This method enhances clinical trial efficiency by improving statistical power and reliability.
• Accompanying R software facilitates the implementation of this advanced statistical approach in practice.