FDA's draft guidance on the use of Bayesian methodology in clinical trials was issued in January 2026 as a PDUFA VII commitment, one of the undertakings the next reauthorisation cycle will build on, and the agency is running a webinar on it on 14 October. Most of the discussion of it will be about the statistics, but section VIII deals with documentation, and that is the part which changes the work.

The guidance highlights settings such as pediatric and rare disease trials, where these methods are most useful, which is precisely where borrowing external information is most attractive and where the evidence base is thinnest. Those are also the programmes already carrying a disproportionate writing load against compressed reporting calendars.

Where the documentation lives

The guidance is direct about the requirement and about which document carries it: "Clear documentation is necessary for FDA to review Bayesian proposals. The study design, estimands, and analyses should be pre-specified and justified in a protocol."

The word "justified" is doing the work in that sentence. A conventional protocol also justifies its design, its estimands and its handling of missing data, but a Bayesian protocol carries an additional burden, which is to justify the prior and any external borrowing as part of the evidentiary argument. The guidance asks that the protocol "describe and justify the design and the planned statistical analysis methods," and for Bayesian methods that includes "detailed information to support the proposed prior distribution and any external information borrowing, likelihood function, success criteria, and trial operating characteristics."

Supporting technical detail may sit in "a statistical analysis plan or simulation report," but the protocol carries the core argument. Timing is stated too: all relevant information should reach FDA "during the design stage and as early as possible to ensure sufficient time for FDA feedback prior to initiation of the trial." In practice that feedback is sought through a formal meeting, which means the argument for the prior has to be written well enough to survive a meeting package capped at ten questions.

What the prior has to be accompanied by

For any proposed prior, the guidance asks for three things: a detailed description "including explicit specification of prior parameterization and underlying assumptions," with a description of induced priors where a function of parameters is of interest; "a rationale for the plausibility of the assumptions"; and "any data or other information that informed the prior distribution."

Where the trial uses an informative prior to borrow external information for the primary estimand, the list becomes considerably longer, and it is worth reading as a writing specification rather than a statistical one:

  • "A strong justification for borrowing that considers feasibility (e.g., of alternative approaches that do not involve borrowing) and the relevance of the available information."
  • "A thorough evaluation of all the available relevant evidence for informing the prior, including evidence that may suggest skepticism. This should include a discussion of steps taken to ensure information was not selectively obtained or used."
  • "A detailed description of each external information source used to inform the prior, as well as any source that was excluded and why," covering data quality and reliability, relevance, whether the data are patient-level or summary-level, completeness, and a rationale for the degree of borrowing.
  • Where multiple sources are used, how they will be synthesised and the modelling approach for that synthesis.
  • Where a discounting method is used, a description and justification including a rationale for the parameters that govern how much is borrowed.
  • A discussion of how the analysis will handle prior-data conflict, supported by simulation.

The clause that changes the drafting

Two requirements in that list are unusual enough to be worth reading twice, because neither has an obvious equivalent in a conventional protocol.

The first is the instruction to evaluate all relevant evidence "including evidence that may suggest skepticism," together with a discussion of the steps taken to ensure information "was not selectively obtained or used." A protocol is being asked to document its own search discipline and to include the findings that argue against the position it is taking. Pre-specification of that kind is what separates a defensible analysis from a contested one, as the record shows whenever a statistical analysis plan changes after the fact or an analysis population moves late.

The second is the requirement to describe "any source that was excluded and why." That is a provenance discipline written into a protocol. It asks the writer to account not only for what is in the document but for what was considered and left out, which is a materially different standard from citing what you used.

Together they turn the prior from a modelling choice into an evidenced argument with a documented search behind it. That is the same discipline the industry keeps arriving at from other directions, whether through writer-defined provenance in AI-assisted drafting, through the reconciliation problem when sources move underneath a document, or through the recognition that an absence of citations is not automatically the problem worth flagging.

The clinical study report grows too

Section VIII.B extends the reporting requirement into the CSR. Beyond "the typical content," a report of a trial with Bayesian analyses should carry the principal aspects of the design and analysis plan, treatment effect estimates with their uncertainty expressed through credible intervals, whether pre-specified success criteria were met, and the marginal posterior distributions of the estimands including measures of location and variation.

It should also carry sensitivity analyses quantifying how far results and conclusions move under "alternative reasonable choices for the prior distribution." Where an informative prior was used to borrow external information in the primary analysis, "the posterior results should be shown for other prior choices with different degrees of borrowing."

Then model checking, including an assessment of prior-data conflict and comparisons of model predictions against observed data; sampling convergence diagnostics, with simulation settings described as implemented for non-direct sampling algorithms; "any deviations from the planned implementation and a rationale for such deviations"; and the software used with version details, with "documented code" provided for all primary and key secondary analyses and any sensitivity analyses described in the report.

Several of those items have no natural home in many E3-derived CSR templates as they are commonly implemented, which means the template has to make space for them deliberately rather than leaving the writer to improvise a location after the analysis is done. E3 sits alongside the other guidelines a writer works to, and none of them yet anticipates a posterior distribution, which is a gap that will matter as CSR timelines compress.

What this asks of the writing organisation

The practical consequence is that a Bayesian trial front-loads writing into the design stage. The argument for the prior, the evidence evaluation behind it, the record of what was excluded and the simulation work supporting the operating characteristics all have to exist before the trial starts, and they have to exist in a protocol FDA can review in time to give feedback.

That is a different sequencing from the usual one, where the statistical detail firms up as the programme matures and the heavy writing happens at the reporting end. It is another instance of the submission argument moving earlier, into design-stage documents rather than end-of-study reporting, in the way that endpoints have had to carry more argument than they once did. It also puts a medical writer and a statistician on the same document much earlier than either is used to, on material where the justification is prose and the substance is quantitative.

For programmes in rare disease and pediatrics, where these methods are most attractive and phase-appropriate expectations are already difficult to hold, that shift lands on teams least likely to have spare capacity for it.

What the guidance asks for, underneath the statistics, is that a prior be written as an argument somebody else can audit: what informed it, why those sources were relevant, what was left out, and what evidence would have counted against it.

The draft guidance is at fda.gov/media/190505/download.