FDA's Draft Bayesian Guidance: What It Means for Medical Device Sponsors

The Guidance You Didn’t Know You Needed

In January 2026, FDA's Center for Drug Evaluation and Research (CDER) and Center for Biologics Evaluation and Research (CBER) released a long-anticipated draft guidance on the use of Bayesian methodology in clinical trials. Technically, it’s addressed to drug and biologics sponsors. Practically, device sponsors and their statisticians should read every page.

Here's why: CDRH has operated under its own 2010 Bayesian guidance for over a decade. That document was pioneering for its time, but the field has moved substantially since then. The methods FDA is now codifying for drugs — dynamic borrowing, hierarchical models, elastic priors, rigorous simulation-based operating characteristics — are the same methods increasingly appearing in IDE applications, PMA submissions, and De Novo requests. The January 2026 guidance signals where FDA's Bayesian literacy has arrived. Sponsors who don't meet that bar will find out the hard way.

We work in this space every day. This post walks through what the guidance actually says, where it matters most operationally, and what device sponsors should be thinking about right now.

What Bayesian Methods Are Being Used For

The guidance catalogs six settings where Bayesian methods have appeared in submissions:

• Borrowing from prior trials — using results from a completed study to construct an informative prior for a new one

• Augmenting concurrent controls — supplementing a randomized control arm with external or nonconcurrent control data

• Pediatric extrapolation — borrowing adult trial data to support pediatric inference

• Cross-disease borrowing — leveraging related disease data in basket trials or hierarchical models

• Subgroup analysis — shrinkage estimation via Bayesian hierarchical models to improve subgroup precision

• Dose-finding — CRM, BLRM, BOIN, and related designs for oncology phase I

Device sponsors will recognize analogs immediately: single-arm IDE studies leveraging objective performance criteria or historical controls, pediatric device trials with limited enrollment, spine or orthopedic indications where patient-level data from prior generations of a device are available. The regulatory contexts differ, but the underlying statistical logic is shared. In our experience supporting Bayesian designs across device submissions, the documentation expectations FDA is articulating here are broadly consistent with what we encounter from CDRH reviewers.

Success Criteria: Don’t Default to Frequentist Thinking

One of the most practically important sections concerns how to define trial success in a Bayesian framework. FDA identifies three approaches:

• 1. Calibration to Type I error rate. Bayesian success criteria are set so that the familywise Type I error rate is controlled at 0.025 one-sided. This is the safest path for sponsors uncertain about FDA's appetite for a purely Bayesian inferential framework. It's also what most adaptive device trials have done — use Bayesian machinery internally while preserving a frequentist success threshold.

• 2. Direct posterior probability interpretation. If the prior accurately reflects pre-study belief, then Pr(δ > a) > c can be interpreted directly. A posterior probability of 0.98 means a 2% chance the treatment is ineffective — and FDA says that can be the basis for a regulatory conclusion, if the prior is sufficiently well-justified. This is a higher bar than it sounds; the prior justification requirements are substantial (see below).

• 3. Benefit-risk or decision-theoretic approaches. Success thresholds incorporate disease severity, availability of alternatives, and formal loss functions. This is intellectually elegant and relatively uncommon in device submissions so far. There may be some interesting use cases for HDE applications or rare disease contexts.

The key operational point: whichever framework you choose, it must be pre-specified, justified, and agreed upon with FDA before the trial starts. Sponsors who treat success criteria as an afterthought create unnecessary risk. For complex Bayesian designs, this conversation belongs at the pre-submission or IDE stage. Which framework is appropriate depends heavily on the specific design, the nature of the prior, and the regulatory context.

Priors: Where Most Proposals Fall Short

This is the heart of the guidance, and the section most likely to generate deficiency letters for sponsors who underestimate the documentation burden.

Noninformative and Minimally Informative Priors

These express general uncertainty, are overwhelmed by the data, and require less justification. Even here, the FDA flags a nontrivial issue: noninformative priors are not invariant under parameter transformations. For example, the prior you specify on a risk difference may induce an unintended distribution on the log odds ratio. This needs to be evaluated during design, not discovered during review.

Skeptical Priors

FDA explicitly endorses skeptical priors, distributions that express skepticism about large treatment effects, in settings where prior programs have failed or where only incremental improvement over an existing therapy is plausible. In device terms: if you’re pursuing a submission in a mature indication with a well-characterized comparator, a skeptical prior may be both defensible and scientifically appropriate. Whether it’s the right choice depends on the specific situation and requires careful judgment.

Informative Priors: The High-Bar Case

This is where the guidance concentrates its most detailed requirements, and where the operational lift is greatest.

FDA's expectation for an informative prior that borrows external data is essentially a miniature systematic review:

• Pre-specified inclusion/exclusion criteria for data sources

• Evaluation of data quality and reliability for each source

• Assessment of relevance across estimand attributes (population, endpoint, treatment conditions, intercurrent event handling, recency)

• Preference for patient-level data over summary-level

• Formal evidence synthesis if multiple sources are used

• Documentation of sources considered but excluded, and why

We've gone through this process for device submissions involving historical control data, OPC derivation, and cross-trial borrowing. The rigor FDA is articulating here is not theoretical; reviewers can and do ask these questions. The difference between a smooth review and a substantive deficiency often comes down to how systematically this work was documented in the protocol.

Discounting: A Decision That Demands Expert Judgment

When you borrow external data, proper care is essential. Discounting adjusts the weight the prior places on external information relative to the prospective trial data. FDA walks through the main approaches:

Each of these methods has legitimate applications, and each carries tradeoffs in terms of operating characteristics, implementation complexity, communication to reviewers, and behavior under prior-data conflict. Static approaches tend to be simpler to implement and explain; dynamic approaches offer protection against conflict scenarios but introduce additional parameters that must be justified in turn. The guidance is explicit that whichever approach is chosen, the similarity measure, borrowing rate, and parameter values must be pre-specified and justified; "we used an MAP prior” is not sufficient documentation.

Selecting the right discounting approach requires deep familiarity with the methods, the regulatory context, and the specific data being borrowed. This is not a decision that should be made by analogy to what another sponsor did in a different indication. It requires statistical expertise and regulatory judgment working together, ideally in close consultation with FDA through a pre-submission meeting.

Operating Characteristics: Simulation Is Not Optional

For calibrated designs, the requirements mirror frequentist standards: Type I error, power, bias, MSE, coverage. These are estimated via simulation, and the simulation report is a required deliverable.

For non-calibrated designs, the requirements shift to Bayesian analogs: Bayesian power (probability of meeting the success criterion, averaged over a design prior), expected bias, expected MSE, expected credible interval width, and probability of a correct decision. FDA makes an important distinction here between the design prior and the analysis prior — these are separate objects and should be treated as such. The design prior is what you use to characterize operating characteristics during planning; the analysis prior is what goes into the actual posterior calculation. Conflating them is a design error.

A few practical implications we see repeatedly in device submissions:

• Simulations should cover prior-data conflict scenarios explicitly, including the null. Sponsors often simulate only under the anticipated effect size.

• Trial characteristics are most sensitive to the analysis prior when sample size is small or when there's an early interim. Device trials with small N, rare disease, pediatric, and HDE, need particularly careful sensitivity coverage.

• The simulation report must include code, scenarios, and assumptions, and must be pre-specified. A post-hoc simulation is not acceptable.

Documentation: The Burden Is Real

FDA's reporting requirements for Bayesian analyses are extensive. The protocol must pre-specify the prior, likelihood, success criteria, and operating characteristics. The clinical study report must include posterior distributions with measures of location and variation, sensitivity analyses across alternative prior specifications, prior-data conflict assessment, convergence diagnostics, and reproducible code with seed numbers.

For MCMC-based analyses specifically, FDA wants: number of chains, warmup/burn-in length, number of iterations, convergence diagnostics, and the seed used for chain initiation.

This is not boilerplate. Sponsors who treat the Bayesian reporting section as an afterthought will receive deficiency letters. Building simulation reports and SAP Bayesian annexes as standard deliverables is increasingly essential.

What Device Sponsors Should Take Away

This guidance is non-binding and addressed to drug sponsors — device sponsors remain governed by CDRH expectations, including the 2010 Bayesian guidance. That said, the methodological principles here are not center-specific, and the rigor FDA is articulating for drugs reflects a broader evolution in how the agency approaches Bayesian submissions. It is worth understanding.

A few observations:

FDA's Bayesian expectations have matured. The documentation bar reflected in this guidance is high, higher than many sponsors have historically (no pun intended) met. Prior justification, discounting methodology, and simulation coverage all require systematic, pre-specified treatment.

Early FDA engagement matters for complex designs. The guidance explicitly recommends pre-submission meetings for sponsors considering informative priors and external borrowing. For device sponsors, CDRH's Q-submission process serves the same function. These conversations are worth having before a design is locked.

The design prior and analysis prior are distinct objects. They serve different purposes and should be treated separately in any statistical analysis plan.

Every methodological choice in a Bayesian design requires justification. Prior type, discounting method, similarity measure, ESS, sensitivity analyses — none of these can be treated as defaults. Each must be reasoned through in the context of the specific trial, the available external data, and the regulatory standard being applied. These are not decisions that lend themselves to off-the-shelf answers. They require statistical expertise and regulatory experience working together — and ideally, familiarity with how FDA reviewers have approached similar questions in prior submissions.

BSC® has designed and analyzed Bayesian clinical trials across IDE, PMA, De Novo, and HDE pathways. Contact us to discuss your design.