On the 12th January 2026 the FDA brought out recommended guidance on the use of Bayesian methods, this marks the first time that the FDA has brought out draft guidance on Bayesian methods outside of brief mentions within other guidelines (such as the rare disease, adaptive designs and medical devices) where it is often regulated to a short paragraph or section. They are applicable to both drugs and biologics development. [1]
This article, authored by Billy Amzal, PhD, Head of Strategic Consulting and Giles Partington, Consultant Statistician, summarizes:
- The FDA review on the ways that Bayesian trials have been used so far FDA submissions.
- Some key takeaways for drug developers
First, the document points out that the most common usages of Bayesian methods have been:
- borrowing information from previous trials/between subgroups,
- extrapolation (both paediatric or from similar disease types),
- utilizing external control data within a clinical trial,
- oncology dose finding.
Other elements of Bayesian usage such as generating missing values, or informing sample size calculations/trial characteristics within either Bayesian or frequentist trials are acknowledged as plausible but noting that they need to be defined up front when looking for agency approval.
The FDA also recalls the difference between frequentist and Bayesian settings with the shift of importance from the usual design and operating characteristics (Sample size, error rates and power) towards posterior probability of true treatment effect size. The FDA acknowledges that there can be several ways of calculating this dependent on the approach chosen. [2]
The FDA highlights, at length, the development and modelling of priors, and the importance of thoroughly reviewing the data and sources involved within developing a prior, especially if multiple sources are used and need to be combined.
Overall, the FDA clearly shows that there are strong reasons for using Bayesian methods however, due to the non-standard techniques and reliance on external information, it is important to clearly pre-define multiple extra elements within the protocol and study report. They also show a preference towards more standard methods where possible.
Below, we report some of the key take-aways for statisticians and practitioners:
- Success criteria can now be “Probability of Effect”:
Acknowledging the limitations related to type I errors e.g. in cases where trials cannot be ethically or practically implemented, FDA underlines that success thresholds (e.g., posterior-probability criteria) can be used as success criteria, provided supportive clinical justification and a discussion of whether criteria are calibrated to frequentist Type I error rate or to alternative Bayesian operating metrics after agreement with FDA.
An important reminder is that the estimand framework is still relevant in Bayesian methods with a few extra points to consider, such as the differences between the external data and the trial data in terms of estimands and estimators where they do not match directly in terms of timepoints and method of collection of variables.
- Priors need to be specified and justified vs operating characteristics via simulations:
The FDA understands that complete comparability between external data and trial data is not always possible and that priors specification and parameterization will affect trial data analyses. Hence, the FDA reports on the various critical ways to prospectively specify distribution priors in the analysis plan and to evaluate their impacts on trial outcomes and operating characteristics. The development of the prior should be stated, showing if it is an informed prior, where the information came from and its appropriateness (is it of good quality, is it a reliable source); and if it is a non-informative prior, how it reduces its influence on the posterior distribution.
In particular, scenario simulations evaluating operating characteristics under specific prior and model parameterizations or assumptions should guide the choice of priors and help controlling type I error or “probability of correct decisions”.
- Bayesian borrowing is recognized as a powerful approach also for extrapolation provided thorough evaluation:
The FDA highlights the role of borrowing even in the context of extrapolation e.g. from adults to pediatrics, provided a thorough evaluation of the prior-data conflict and proper discounting to address it, e.g. via dynamic borrowing.
Highly informative priors can overwhelm the trial data, but not utilizing the data sufficiently makes the trial less efficient and increases the needed sample size.
The FDA still wants to see that other methods that do not include borrowing have been considered and found as infeasible before borrowing information from other trials is suggested, where borrowing is determined to be the most sensible path forward, they want to be assured the information borrowed is truly relevant to how it will be used. They also want to see that the prior is not counterintuitive to the trial characteristics, conflicting with other data or too strong an opinion in one direction that it would overtly influence the posterior distribution given any observed outcome. Thorough sensitivity analyses are encouraged in all cases. As a complement, this recent EMA-sponsored simulation-based evaluation illustrates how model and prior parameterization influences performance and operating characteristics of Bayesian borrowing [3].
- Bayesian approaches require more effort in technical documentation and verification
Ensuring a clearly defined, justified and pre-specified study design, estimand framework and analysis is key, similarly to non-Bayesian approaches. However, a Bayesian implementation will include several additional steps and elements to test and document that would not be standard or applicable within the frequentist framework. For example, details toon the proposed prior and the underlying assumptions and rationale; justification of any borrowing and where it is from; prior influence and discounting approaches, diagnostics supporting MCMC convergence and reliability, computational settings reporting e.g. seed numbers, sensitivity analyses, codes and software documentation enabling full reproducibility and robustness of conclusions.
Bayesian methods can provide powerful opportunities in drug development, but only when designed, justified, and implemented correctly. With clearer expectations now set by the FDA, early statistical strategy and robust documentation are more critical than ever.
If you’re considering Bayesian approaches for borrowing, extrapolation, or decision making, now is the time to sense-check your assumptions and design choices.
References
2.Viele, K., Berry, S., Neuenschwander, B., Amzal, B., Chen, F., Enas, N., Hobbs, B., Ibrahim, J. G., Kinnersley, N., Lindborg, S., Micallef, S., Roychoudhury, S., & Thompson, L. (2014). Use of historical control data for assessing treatment effects in clinical trials. Pharmaceutical Statistics, 13(1), 41–54. https://doi.org/10.1002/pst.1589
3.Fauvel, T., Tanniou, J., Godbillot, P., Génin, M., & Amzal, B. (2025). Comparison of Bayesian methods for extrapolation of treatment effects: A large-scale simulation study. ArXiv. https://arxiv.org/pdf/2504.01949
4.https://www.valueinhealthjournal.com/article/S1098-3015(25)06001-2/abstract
