What if your clinical trial could formally learn from the past, without sacrificing regulatory rigor? What if probability itself became part of the evidence FDA evaluates? With its new draft guidance, FDA explains how Bayesian methods, that were previous described for medical devices, can be used for drugs and biologics as well, but with guardrails. In its new draft guidance on the Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products, the FDA explains its position on Bayesian inferences for drug and biologics product development. While Bayesian statistics are not new to FDA, this document clarifies when, how, and under what conditions Bayesian methods may support primary inference for drug and biologic approvals.
First, FDA explicitly acknowledges that Bayesian methods may be used not only for trial design and adaptation, but also to support primary evidence of effectiveness, provided the approach is well-justified, prespecified, and transparent. This is a meaningful evolution for drug developers.
Second, the guidance places heavy emphasis on prior distributions. Sponsors must justify the source, relevance, and influence of priors, especially when borrowing external information from historical trials, real-world data, pediatric extrapolation, or related disease populations. FDA clearly expects quantitative assessments of prior influence, prior–data conflict, and robust sensitivity analyses. Third, FDA reframes success criteria. Rather than defaulting to a frequentist p-value, Bayesian trials often rely on posterior probabilities (e.g., Pr(effect > threshold) > c). However, FDA requires that these criteria be carefully calibrated, either to traditional Type I error control or justified through decision-theoretic or benefit–risk frameworks.
FDA presents its real-world experience with Bayesian analysis in drug applications into concrete scenarios, reflecting situations where traditional frequentist approaches alone were insufficient or inefficient. One of the most prominent examples cited is REBYOTA, a fecal microbiota product approved in 2022. FDA accepted a Bayesian primary analysis that formally borrowed data from a prior Phase 2 study to support effectiveness in a Phase 3 randomized trial. This example signals FDA’s willingness to accept informative priors derived from earlier trials of the same product when relevance and consistency are well justified.
In oncology trials such as GBM AGILE and Precision Promise, borrowing nonconcurrent control data improved feasibility in rare or rapidly evolving diseases. Bayesian methods were used for pediatric indications for empagliflozin and linagliptin, where disease biology and treatment response were sufficiently similar between adults and pediatric populations thereby reducing pediatric trial burden. FDA explicitly endorses Bayesian hierarchical models in basket trials, where treatment effects may be shared across biologically related diseases. Bayesian shrinkage approaches are cited as tools FDA has already used to improve precision in subgroup analyses, such as regional effects in large cardiovascular outcomes trials. These methods help prevent overinterpretation of noisy subgroup signals while preserving interpretability.
FDA recognizes Bayesian dose-escalation and dose-optimization designs (e.g., CRM, BLRM, BOIN) as increasingly important for modern oncology drugs, particularly targeted therapies where maximum tolerated dose is no longer the optimal endpoint. The guidance documents FDA past acceptance of Bayesian analysis for drug applications to describe how such acceptance depends on pre-specification, relevance of borrowed data, simulation-based operating characteristics, and transparent justification.
FDA’s approach here is philosophically consistent with its long-standing Bayesian framework for medical devices, where Bayesian methods have been routinely accepted, particularly for PMAs, adaptive designs, and borrowing from prior device iterations. In both contexts, FDA emphasizes pre-specification of priors and decision rules, simulation-based evaluation of operating characteristics, and transparency in documentation and reporting. So, the underlying Bayesian mechanics are aligned. The regulatory posture, however, is not identical.
The most important distinction between applying Bayesian analysis to devices versus drugs and biologics is evidentiary tolerance. For medical devices, FDA has historically allowed greater flexibility in borrowing prior data, reflecting iterative design changes and engineering-based risk mitigation. For drugs and biologics, FDA is more cautious, primarily because systemic exposure, population heterogeneity, and benefit–risk irreversibility demand stronger confirmatory evidence.
Unlike devices, drug and biologic trials must still convincingly demonstrate substantial evidence of effectiveness, often across multiple endpoints or studies. As a result, FDA places tighter expectations on relevance and discounting of external data, control of false positives when borrowing information, adequacy of prospective safety data, regardless of prior strength. In short, Bayesian methods for drugs are acceptable, but not a shortcut.
FDA’s new Bayesian guidance is a structured pathway. It rewards rigor, transparency, and disciplined borrowing, while rejecting black-box probability. For sponsors who get it right, Bayesian methods can be a regulatory asset.
