Enhancing Phase 3 Trials Through Bayesian Borrowing
Integrating Prior Evidence in Clinical Trial Design
Bayesian borrowing enables Phase 3 clinical trials to quantify and incorporate all available relevant prior information—such as previous trial results—into the statistical evaluation of current studies. Rather than limiting inference to only new trial data, the Bayesian framework specifies a prior probability distribution that encodes information about treatment efficacy. This approach is especially relevant when strong biological rationale, supportive comprehensive data, or prior results in closely related disease populations exist.
For example, using data on the efficacy of a treatment in studies on adults to estimate the effect in pediatrics, or from one rare disease indication to another with a shared mechanism, often makes use of prior data for improved inference and efficient new trials. In circumstances where a prior probability of efficacy is high, the new trial may justifiably be sized or analyzed differently than for a treatment with no supporting evidence.
The scientific validity of Bayesian borrowing depends on the comprehensiveness, relevance, and integrity of the chosen prior. The use of comprehensive empirical results improves the credibility and acceptance of the Bayesian approach. Selective use of favorable subgroups, end points, or post hoc populations undermines both the credibility of the analysis and regulatory and scientific acceptance.
Statistical and Regulatory Considerations
The integration of prior information alters the statistical properties of confirmatory trials, with a particular focus on Type 1 error. When the conclusions of a trial are constrained to just that trial, as is typical in a phase 3 trial, the probability of incorrectly concluding efficacy (Type 1 error) is tightly controlled, often at a one-sided 2.5%. By introducing an informative prior distribution, for example, using Bayesian borrowing, permits a trial to reach regulatory probability thresholds (e.g., a combined 97.5% posterior probability) even if the new data alone would not meet this standard. The effective new trial Type 1 error, with respect to the incremental evidence of the current trial, will be increased above the 2.5% level.
Regulatory authorities assess both the statistical foundation and the clinical suitability of the borrowed evidence. They may agree to higher Type 1 error rates in the new trial when evidence already exists supporting efficacy. This existing data could be from previous trials, related indications, or related populations.
Critical to regulatory acceptance is the avoidance of data that have been cherry-picked or have a high risk of bias. For the same reasons drawing inferences from cherry-picked analyses is problematic – using these same cherry-picked data in isolation to form a prior distribution is problematic. This principle is fundamental and appropriate for the FDA's evaluation. Any prior must be discussed, justified, and accepted in advance with regulators.
To address potential mismatch between past and new data, Bayesian analyses frequently utilize dynamic borrowing models (the FDA draft guidance refers to this as Bayesian discounting). If the new trial outcomes diverge from historical trends, dynamic models algorithmically down-weight the prior data. This adaptive mechanism provides a safeguard against continuing to include historical results that have accruing evidence they are not reflective of the new results. These models and approach can reduce the risk of errors compared to static borrowing, which applies a pre-set discount regardless of observed compatibility.
Practically, the Bayesian approach involves openness in showing operating characteristics to regulators—displaying power, expected error rates, and example trials to clarify decision rules. Simulated trial results are presented to demonstrate both the probability of success under true effects and the risk of false positives. A key presentation of the simulations is to provide single example outcomes of the new trial and the resulting Bayesian analysis result. This allows a clear vision for regulators to understand the behavior and face validity of the modeling approach.
Case Studies and Operational Lessons
Numerous Phase 3 programs have utilized Bayesian borrowing. In the WATCHMAN left atrial appendage closure device program, the analysis of the PREVAIL trial utilized Bayesian borrowing from the PROTECT-AF trial, using a 50% static discounting. The FDA approved this approach following comprehensive simulation of trial operating characteristics and direct illustration of successful and failed outcomes.
The REBYOTA approval for recurrent C. difficile infection offers further clarity. The Phase 2 trial had an observed positive result, but did not demonstrate classical significant on its own. The Phase 3 trial in the rare disease was plagued by challenging enrollment due to the randomized control. A Bayesian analytic plan that combined the phase 2 and phase 3 trial sources of evidence was agreed upon with FDA prior to unblinding the Phase 3 trial. The combined Bayesian dynamic borrowing demonstrated a 99.1% posterior probability of benefit of the REBYOTA treatment compared to the control. The FDA incorporated this Bayesian probability along with the posterior mean effect as the primary metrics in the drug label.
Conclusion
Bayesian borrowing provides a methodologically robust and explicit pathway for integrating prior evidence into Phase 3 clinical trial designs. This approach allows for more data-efficient trials and can accelerate development all while meeting regulatory requirements, but demands exacting attention to prior integrity, pre-specification, and statistically robust modeling. There is no recipe or algorithm for how to create a Bayesian analysis plan – which is good, medical and regulatory decision making is hard. Science is hard. The FDA draft guidance, and the described examples, demonstrates the feasibility and scientific validity of Bayesian borrowing when applied with transparency, dynamic safeguards, and clear communication with regulatory agencies.