The FDA’s January 2026 draft guidance, “Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products,” clarifies how the Agency expects sponsors to justify Bayesian approaches, especially when an informative prior borrows external information to support primary inference. As a draft guidance, it is nonbinding and not for implementation.

This blog highlights strategic considerations that should inform development planning, protocol/SAP design, and FDA engagement.

Type I error control is not the only path

The guidance notes that calibrating Bayesian success criteria to a Type I error rate “may not be applicable or appropriate” when borrowing external information. In those settings, sponsors may instead define success using posterior probability criteria (e.g., Pr(d>a)>c) and, where appropriate, benefit-risk or decision-theoretic frameworks.

At the same time, the draft guidance also recognizes that Bayesian methods are often used within an overall frequentist framework (e.g., to facilitate complex adaptive designs), where Type I error calibration can remain appropriate. Regardless of the framework, success criteria should be pre-specified and justified.

Strategic implication:

When the FDA and sponsor agree that a design does not need to be calibrated to the Type I error rate (often discussed in pediatrics and rare diseases), the draft guidance describes alternative operating characteristics such as Bayesian power (probability of success averaged over a prior) and the probability of a correct decision (akin to positive predictive value). That flexibility increases the premium on a well-justified analysis prior, credible simulations, and early FDA alignment.

Prior specification is now a regulatory deliverable

The draft guidance recommends that sponsors pre-specify and justify the prior in the protocol, document external information sources (including exclusions), and quantify prior influence metrics. For informative priors, the FDA emphasizes a systematic, transparent review of the totality of relevant evidence — effectively bringing evidence-synthesis discipline into prior construction.

Key expectations:

• Pre-defined source selection criteria before searching for external data
• Patient-level data preferred over published summary statistics
• Randomized controlled evidence is generally preferred over single-arm or observational sources
• Documentation of sources considered and excluded, with rationale

Strategic implication:

Prior construction cannot be a post-hoc exercise. Build the evidence base for your prior prospectively — ideally while Phase 2 is ongoing — and then plan early for patient-level data access and any needed re-analyses to align primary estimand/estimators/strategies for handling intercurrent events. If patient-level data from prior studies are not accessible, negotiate data-sharing early or design natural history studies with Bayesian use in mind.

Dynamic discounting provides protection — with complexity

The draft guidance discusses both static and dynamic discounting approaches for borrowing external information. Dynamic approaches (e.g., commensurate/supervised power priors, mixture priors, Bayesian hierarchical models, elastic priors) can reduce borrowing when prior-data conflict emerges. These approaches can improve robustness but introduce additional parameters and assumptions that need justification. The FDA also notes the applicability of discounting methods is case-by-case and should be discussed with the Agency.

Strategic implication:

For rare diseases with uncertain external data relevance, dynamic discounting is often an important safeguard. For common diseases with robust and highly relevant prior data, simpler (static) discounting may suffice and can simplify the regulatory narrative. Either way, determine the discounting approach while still blinded to the results of the trials that will be borrowed — per the guidance’s explicit recommendation — and support the choice with simulations that span plausible degrees of prior data conflict.

Effective sample size is a central metric — Not Type I error inflation

The draft guidance recommends against using Type I error inflation to measure prior influence, calling it “philosophically inconsistent.” Instead, it highlights Effective Sample Size (ESS) and other metrics (e.g., the prior-only estimate) as more interpretable ways to quantify borrowing. The guidance also notes that multiple ESS calculation methods exist, and that ESS can exceed the source-study sample size when the variability in the target population variability is higher.

Strategic implication:

Quantify and present ESS across a plausible range of outcomes, including summary statistics such as maximum and mean values. For dynamic methods, show how ESS changes with different degrees of prior-data agreement. Be prepared to explain why ESS may differ from the original study’s nominal samples size and reassess influence after trial completion when dynamic priors are used.

Simulation standards are now explicit

The draft guidance recommends providing a comprehensive simulation report (including code, implementation details, and results) across pre-specified, plausible scenarios, including pessimistic assumptions about treatment effect. Simulations should address statistical parameters (e.g., variance, background rate, intercurrent events) as well as operational assumptions such as accrual rate. For MCMC-based analyses, computational settings (warmup/burn-in, iterations, chains, convergence diagnostics) and any other important algorithm-specific settings should be documented for reproducibility.

Strategic implication:

Treat simulations and computational reproducibility as submission-grade deliverables, not just internal design exploration. Establish reproducible computational workflows from the start. Pre-specify scenarios and decision rules, and define contingency procedures for implementation issues (e.g., MCMC non-convergence) before the first interim look and before the final analysis.

Early FDA engagement is essential

The draft guidance states that “the time needed for FDA and the sponsor to align on an appropriate prior should be considered in the development of the intended trial” and recommends submitting information “as early as possible to ensure sufficient time for FDA feedback prior to initiation.” The draft guidance also states that sponsors should have early discussions with the Agency about the planned estimands, estimators, and approaches for handling missing data in the analyses of external data that will be borrowed, and any differences relative to the approaches planned for the prospective trial data.

Strategic implication:

Use early interactions (e.g., Pre-IND or End-of-Phase 2 meetings and, where applicable, the Complex Innovative Trial Design (CID) program) to align on prior specification, success criteria, operating characteristics, and simulation strategy before protocol finalization. Include detailed design comparisons in meeting packages — the draft guidance explicitly recommends comparing proposed Bayesian designs against an alternative, including simpler alternatives.

Interim analyses: Design the decision points upfront

The guidance emphasizes that in trials with interim decision-making (e.g., group sequential designs), success criteria should be specified for each decision point. When Bayesian success criteria are calibrated to Type I error rate, interim criteria can be constructed to preserve overall control of the family-wise error rate across looks.

For designs not calibrated to Type I error rate, operating characteristics are calculated relative to the prior and can be especially sensitive when the sample size is small — or when an early interim look makes the effective sample size small. The guidance also notes that skeptical (or enthusiastic) priors can be used in adaptive settings to temper early stopping behavior for efficacy (or futility), but the resulting decision framework should be demonstrated via simulation.

Key interim analysis considerations:

• Pre-specify what decisions can be made at each look (e.g., stop for efficacy, stop for futility, adapt) and the exact posterior or predictive criteria that trigger each action.
• Simulate interim timing under realistic accrual, endpoint maturation, and missing data patterns — not just idealized information fractions.
• Plan prior sensitivity and robustness checks targeted at early looks, where prior influence is greatest (e.g., alternative priors and alternative borrowing strengths).
• Operationalize Bayesian computation for interim timelines: reproducible pipelines, diagnostic thresholds, locked code/versioning, and contingency plans for non-convergence.
• Protect safety and benefit-risk interpretability: consider minimum exposure or follow-up requirements even if an early efficacy threshold is met.

Strategic implication:

Treat interim analyses as part of the regulatory-facing Bayesian package, with pre-specified decision rules, simulations that stress-test early looks, and an execution plan that can be reproduced under tight timelines.

Rare vs. Common Disease Considerations

The bottom line

This draft guidance provides a clearer regulatory pathway for Bayesian methods, but that pathway requires substantial upfront investment in prior construction, estimand definition, and strategies for handling intercurrent events, simulations, and documentation at submission quality. The strategic question is not whether Bayesian methods are acceptable in principle — it is whether the efficiency gains justify the additional complexity and review burden for your specific program.

For rare diseases, the answer is often yes. Bayesian borrowing may be the only viable path to interpretable and approvable evidence. For common diseases, the calculus is more nuanced; borrowing typically needs a stronger relevance argument and may be most defensible when embedded in a Type I calibrated framework. Either way, the strategic decisions about prior specification, discounting method, and operating characteristics should be made early, documented thoroughly, and aligned with FDA before the pivotal trial initiation.

What’s clear is that biostatisticians must now be prepared to operate in both paradigms:

• To calibrate Bayesian designs to Type I error when appropriate, and
• To construct and defend fully Bayesian alternatives (including borrowing) when circumstances warrant.

The January 2026 draft guidance does not eliminate the traditional framework; it expands the toolkit. Using that expanded toolkit effectively will require new skills, new conversations, and new ways of thinking about evidence.

The statistical methodology exists. FDA expectations are clearer. The challenge is execution.

Share Via