Bayesian Trials at the FDA: What the Draft Guidance Asks

The FDA now has a written playbook for using Bayesian statistics as the main evidence in pivotal drug and biologic trials ^[1][2]. The message is clear: sponsors can borrow data from earlier studies, but only with a prior that is prespecified, justified and stress tested. The draft is still open to change, yet FDA statisticians walked industry through it in a public webinar on October 14, 2026, so it already shapes design talks ^[3].

Why the FDA wrote this guidance

The draft is titled "Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products." CDER and CBER issued it together in January 2026 ^[1]. The notice ran in the Federal Register on January 12, 2026, under docket FDA-2025-D-3217 ^[2]. It is written for sponsors filing INDs, NDAs, BLAs and supplements ^[2].

The FDA said the document meets a commitment under the sixth reauthorization of the Prescription Drug User Fee Act, known as PDUFA VII ^[2]. That commitment aims to speed development through complex adaptive, Bayesian and other innovative trial designs ^[4]. It builds on the Complex Innovative Trial Design initiative, which began under the previous user fee agreement ^[4].

FDA Commissioner Marty Makary framed the goal in plain terms. He said Bayesian methods "help address two of the biggest problems of drug development: high costs and long timelines" ^[5]. The agency also pointed to limited patient availability as a problem these methods can ease ^[5].

The guidance is not binding. It creates no rights and does not bind the FDA or the public, and other approaches are allowed if they meet the law ^[2]. Comments closed on March 13, 2026 ^[2]. In October, the FDA's Small Business and Industry Assistance office hosted a session on the draft with statisticians from CDER and CBER ^[3].

How a Bayesian trial differs, in plain terms

Traditional statistical decisions in drug development rely only on the data from the new trial ^[4]. A Bayesian analysis adds a second input. It starts with a prior distribution, which sums up what is already known, and combines it with the new data through Bayes' theorem ^[1]. The result is a posterior distribution, an updated picture of how large the treatment effect is likely to be ^[1].

From the posterior, a sponsor can report a posterior mean, a credible interval and a posterior probability, such as the chance that the effect is above zero ^[1]. A credible interval is the Bayesian cousin of a confidence interval. It gives a range that likely holds the true effect, given the prior and the data.

Writing in JAMA, statisticians Andrew Gelman, Erik van Zwet and Witold Więcek welcomed the FDA's choice to describe the prior as "prestudy information" rather than personal belief ^[6]. They noted that prior knowledge already shapes conventional trials, for example when teams plan sample sizes ^[6]. In their view, the main gain from the Bayesian approach is that it uses prior knowledge formally and in the open ^[6].

The guidance lists three broad roles for these methods. They can set the timing and rules for interim analyses, inform design choices such as dose selection for later trials, or support the primary inference itself ^[1]. The draft focuses on that last role, where a Bayesian analysis carries the main claim of effectiveness or safety ^[1].

What the evidence shows

The FDA did not write this guidance from theory alone. It lists real cases where Bayesian methods have already supported decisions, and that record is the strongest sign of where the agency is comfortable.

Borrowing from an earlier trial

The clearest example is REBYOTA, a fecal microbiota product used to prevent recurrent Clostridioides difficile infection ^[4]. Its phase 3 primary analysis used a Bayesian model to formally include data from an earlier placebo controlled phase 2 study ^[4]. The product was approved in 2022 ^[4].

Platform trials and shared controls

The draft cites GBM AGILE and Precision Promise as platform trials that proposed Bayesian methods to use nonconcurrent control data ^[1]. Nonconcurrent controls are patients enrolled at an earlier point in a long running trial. The guidance flags the risk that outcomes drift over time, which can make older control data less comparable ^[1].

Children, related diseases and subgroups

Adult data have been borrowed through informative priors to support pediatric results ^[1]. The guidance names supportive analyses for empagliflozin and linagliptin in children with type 2 diabetes ^[1]. It also describes borrowing across disease subtypes in basket trials, and a proposed epilepsy study that uses a Bayesian hierarchical model ^[1]. For subgroups, it points to shrinkage of regional estimates in the LEADER trial ^[1].

Dose finding in oncology

Model based and model assisted designs, such as the continual reassessment method and BOIN, are already used to find the maximum tolerated dose in cancer trials ^[1]. The guidance adds that for targeted therapies, an optimized dosage may be more suitable than the maximum tolerated dose ^[1].

The FDA's own event page says pediatric and rare disease trials are the most common settings for Bayesian methods ^[3]. Kert Viele of Berry Consultants, a firm that designs adaptive trials, reads the list as proof that the FDA has real experience with Bayesian work ^[7]. He says the draft largely formalizes two decades of progress ^[7].

Success criteria: the type I error question

This is the section most statisticians will read first. In a conventional pivotal trial, the success rule is almost always set so the familywise type I error rate is no more than 0.025, one sided ^[1]. Type I error is the chance of declaring a drug effective when it is not.

A Bayesian success rule usually takes the form Pr(d > a) > c ^[1]. Here d is the treatment effect, a is the smallest effect worth having, and c is the posterior probability required ^[1]. The guidance warns that the usual frequentist rule may not suit Bayesian primary inference, especially when data are borrowed ^[1].

The draft sets out three ways to pick a and c ^[1]. The table below summarizes them.

Approach	How the threshold is set	Where it fits
Calibration to type I error	a is zero for superiority, and c is derived by algebra or simulation to keep the error rate in check [1]	Adaptive designs or noninformative priors [1]
Direct reading of the posterior	The posterior probability is taken at face value, so 0.98 for effectiveness means 0.02 for no effect [1]	Needs a prior that accurately reflects existing information [1]
Benefit and risk	The threshold reflects product risk, disease seriousness, alternatives or expected loss [1]	Decisions that weigh benefit against harm [1]

Viele notes that for a simple trial with a noninformative prior, the calibrated rule looks like a posterior probability above 0.975 ^[7]. He adds that most Bayesian trials at the FDA have used this calibrated route ^[7]. The trouble starts with borrowing. With an informative prior, the type I error rate depends on how well the old data match the new trial, and it can rise sharply when they diverge ^[7].

The guidance allows a way out. Designs that are not calibrated to type I error are possible when the sponsor and the FDA agree ^[1]. Those designs are judged on other measures, such as Bayesian power averaged over a design prior and the probability of a correct decision ^[1]. The FDA wants design priors to include pessimistic scenarios, not only ones where the drug works ^[1].

Gelman and his co-authors called the type I error approach "one of the most notable features" of the draft ^[6]. They also named the risk. Flexibility on type I error could let optimistic assumptions slip in through a favourable prior ^[6]. Their answer is that success criteria must be agreed in advance ^[6].

Borrowing external data: the rules on priors

The longest part of the guidance deals with priors. The FDA wants every prior prespecified and justified in the protocol, along with its influence on the result ^[1].

An informative prior that borrows external data needs strong justification ^[1]. The justification should weigh feasibility, including alternatives that do not borrow, and show that the external data are relevant ^[1]. Such priors are most often proposed in pediatrics and rare diseases, and other areas are reviewed case by case ^[1]. Viele warns that sponsors often read "case by case" as a default no ^[7].

The bar for source data is high. The guidance asks sponsors to check data quality, prespecify inclusion rules and judge relevance ^[1]. Relevance covers the population, endpoint, treatment, measurement, recency and changes in standard of care ^[1]. Randomized comparisons and patient level data are preferred ^[1].

The synthesis must also be planned in advance, so a sponsor cannot pick only the favourable studies ^[1]. Discounting is usually needed. Without it, adult data could swamp pediatric data whatever the trial shows ^[1].

The draft describes two families of discounting. Static methods, such as a power prior, apply a fixed weight between zero, meaning no borrowing, and one, meaning equal weight ^[1]. Dynamic methods, such as commensurate and elastic priors, borrow less when the old and new data conflict ^[1]. They add parameters that must be justified ^[1].

To show how much a prior matters, the FDA wants measures such as effective sample size, compared against a noninformative prior ^[1]. Effective sample size expresses the prior as an equivalent number of patients. The guidance explicitly does not recommend type I error inflation as a measure of borrowing ^[1]. Sensitivity analyses should vary the prior, including the strength of borrowing ^[1]. Viele suggests a tipping point analysis that shows how much borrowing a positive result needs ^[7].

Lawyers at Hogan Lovells summed up the stance: "borrowing must be disciplined, transparent, and proportionate to the strength and relevance of the external data" ^[8]. They also noted that CDRH, the FDA's device centre, has not adopted this draft ^[8].

The rules reach into computing too. The FDA does not require any specific software ^[1]. For MCMC, the sampling method most Bayesian analyses use, sponsors should report settings, convergence checks, software versions, documented code and seed numbers ^[1].

What industry and experts asked the FDA to change

Commenters broadly supported the draft but asked for clarity and more examples, according to RAPS ^[9]. Several themes stand out.

• Do not make simple Bayes harder. PhRMA warned the recommendations "could disincentivize sponsors from using noninformative priors" ^[9]. It asked the FDA to set different expectations for each type of prior ^[9].
• Do not treat Bayes as a niche tool. BIO said parts of the draft could be seen as too burdensome ^[9]. It argued that "the scientific value of Bayesian inference extends well beyond feasibility considerations" ^[9].
• Be clear on calibration. BIO asked when calibration is expected, optional or replaceable by other criteria ^[9]. EMD Serono asked how interim decisions should work without a type I error rule ^[9].
• Treat children as more than one group. The American Academy of Pediatrics said adult data may be more relevant for adolescents about 13 and older, or children over 40 kg ^[9].
• Look abroad. The Duke Margolis Institute noted that China's NMPA and the EMA are the only other regulators with Bayesian guidance or reflection papers ^[9].

Viele, writing for Berry Consultants, called the draft a "dramatic leap forward" ^[7]. His suggested edits were mostly about clarity. He wants a plain statement that Bayesian methods suit ordinary trials, not only those with informative priors ^[7].

What it means for people in the industry

For biostatisticians, the work moves up front. The protocol must describe and justify the prior, the borrowing, the likelihood, the success criteria and the operating characteristics ^[1]. Operating characteristics need enough detail for the FDA to verify them ^[1]. Simulation scenarios should be prespecified, and the simulation report should include the code ^[1]. In practice, that makes the simulation report a core deliverable rather than an appendix.

For clinical development leads, the guidance changes what is worth asking for. In pediatric and rare disease programs, borrowing adult or earlier trial data now has a documented path ^[1]. Outside those areas, proposals are reviewed case by case, and early FDA discussion is advised ^[1].

For regulatory affairs, timing matters most. The FDA recommends that sponsors request a meeting before a complex Bayesian trial starts ^[1]. RAPS reported that the agency wants all relevant information during the design phase, so it has time to give feedback ^[4]. Viele notes the guidance warns that complex reviews may take longer ^[7].

For business development, the draft is a factor in valuing small population assets. A program that can credibly borrow strength from earlier data may need fewer new patients. That benefit only holds if the prior survives FDA review, so diligence teams should ask to see it.

For anyone in clinical operations or data management, the reporting rules are concrete. The final study report should include sensitivity analyses, prior data conflict checks, convergence diagnostics and the code used ^[1]. Those tasks need planning and staffing well before database lock.

Watch two things next. One is the final version, which may respond to calls from PhRMA and BIO for lighter expectations on simple designs ^[9]. The other is whether CDRH and other regulators move toward the same standards ^[9][8]. For ongoing coverage of FDA policy, see PharmoniQ news.

Key takeaways

• The FDA's January 2026 draft explains how Bayesian analysis can support the main claim in drug and biologic trials ^[1][2].
• It meets a PDUFA VII commitment, and the comment period closed on March 13, 2026 ^[2].
• Success can be judged by calibration to type I error, a direct posterior probability, or a benefit and risk threshold ^[1].
• Borrowing external data is allowed but needs a prespecified, justified prior, discounting and sensitivity analyses ^[1].
• Prior influence should be shown with measures like effective sample size, not type I error inflation ^[1].
• Industry groups backed the draft but asked the FDA not to add burden to simple Bayesian designs ^[9].
• Pediatric and rare disease programs have the clearest path today, and early FDA meetings are advised ^[1][3].

References

1. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products (Draft Guidance for Industry). U.S. Food and Drug Administration, Jan 2026.
2. Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products; Draft Guidance for Industry; Availability. Federal Register, Jan 12, 2026.
3. Navigating the FDA Bayesian Draft Guidance: Design, Analysis, and Review Considerations - 10/14/2026. U.S. Food and Drug Administration, Oct 14, 2026.
4. FDA issues guidance on use of Bayesian methods to support drug development. RAPS Regulatory Focus, Jan 9, 2026.
5. FDA Issues Draft Guidance to Advance Bayesian Methods in Clinical Trials. Applied Clinical Trials, Jan 14, 2026.
6. FDA Draft Guidance for the Use of Bayesian Methods in Clinical Trials. JAMA, Mar 23, 2026.
7. Guide to the Draft FDA Bayesian Guidance 2026. Berry Consultants, Jan 30, 2026.
8. FDA signals long-awaited embrace of Bayesian methods in drug and biologic trials. Hogan Lovells, 2026.
9. Industry asks FDA to affirm Bayesian methods guidance applies broadly to drugs. RAPS Regulatory Focus, Mar 18, 2026.