FDA’s January 2026 draft guidance on Bayesian methodology in drug and biologics trials signals a clear willingness to see Bayesian methods used for regulatory decisionmaking around clinical trials. The guidance lays out how sponsors (i.e., pharmaceutical manufacturers) should pre-specify priors, success criteria, and simulations so that Bayesian designs remain interpretable, control error rates when needed, and generate evidence that can contribute to substantial evidence of effectiveness. The document also emphasizes transparency around prior construction and operating characteristics, especially when borrowing external information like historical trials or real‑world data.
Example applications of Bayesian Statistics
The guidance highlights several recurring application areas where Bayesian methods have already appeared in submissions:
Borrowing from previous clinical trials: Under certain circumstances, an informative prior for a trial can be formed based on results from earlier studies of the same drug, such as the REBYOTA phase 3 study that formally incorporated data from a prior phase 2 trial in recurrent C. difficile infection.Augmenting a randomized concurrent control with external or nonconcurrent controls: In hard‑to‑recruit or ethically challenging settings, external or nonconcurrent control data can supplement a concurrent control arm, as proposed in GBM AGILE and Precision Promise platform trials and a pediatric multiple sclerosis non‑inferiority study.Pediatric extrapolation: When disease course and pharmacology are sufficiently similar between adults and children, adult data can inform informative priors for pediatric efficacy and dosing, as seen in empagliflozin and linagliptin programs for pediatric type 2 diabetes, with careful assessment of relevance.Borrowing across similar diseases or disease subtypes: For related conditions (e.g., tumors sharing a molecular alteration or related epilepsies), Bayesian methods—often via hierarchical models—can share information across indications or subtypes in basket‑style settings while still allowing heterogeneity.Borrowing between subgroups within a trial (subgroup analysis): Bayesian hierarchical models and shrinkage can stabilize subgroup estimates (e.g., regional effects) by partially pooling toward an overall effect, as illustrated by the liraglutide Cardiovascular Outcomes Results trial and several Drug Trial Snapshots including Rinvoq.Dose‑finding trials in oncology: Bayesian model‑based and model‑assisted dose‑escalation designs (CRM, BLRM, BOIN, mTPI/mTPI‑2, EWOC, and utility‑based approaches) are used to identify maximum tolerated or otherwise optimized doses more efficiently and flexibly than traditional 3+3 designs.
FDA’s perspective on priors
The guidance distinguishes three broad classes of priors and offers practical commentary on each:
Noninformative and minimally informative priors are framed as tools to encode general uncertainty, often appropriate when there is no relevant prior evidence, but FDA notes they can still have unintended influence in weak‑data settings and should be evaluated via simulation (including effective sample size and sensitivity checks).Skeptical priors are described as reasonable when large effect sizes should be viewed cautiously—such as after multiple failed programs in a therapeutic area or when only incremental benefit is plausible—and as potential tools in adaptive designs to temper early random highs while maintaining Type I error control.Informative priors that borrow external information (e.g., for pediatrics, rare diseases, or platform trials) are acceptable in principle but require “strong justification,” explicit modeling of relevance and uncertainty, formal evaluation of prior–data conflict, and careful calibration of how much borrowing occurs (often via dynamic discounting, hierarchical models, or mixture priors).
For data used to inform informative priors, FDA stresses that sources should be: high quality and reliable (with extra scrutiny for real‑world data), prespecified in a systematic evidence‑synthesis plan (inclusion criteria, analytic methods, and scope), relevant to the specific estimand and setting (population, endpoints, intercurrent event handling, recency, and evolving standard of care), ideally randomized and patient‑level where possible, and transparently documented including reasons for excluding any studies.
Other important topics in the guidance include success criteria and operating characteristics, quantifying prior influence (e.g., effective sample size [ESS]), estimands and missing data in a Bayesian setting, software and computation (including Markov Chain Monte Carlo [MCMC] diagnostics), and expectations for documenting and reporting Bayesian analyses.
Readers interested in the full detail should consult the draft guidance at https://www.fda.gov/media/190505/download.
