FDA Holds a Webinar Oct. 14 on Its Bayesian Trial Guidance. Here Is How to Read the Trials It Covers.
The Center for Drug Evaluation and Research holds a one-hour webinar Oct. 14 on its January draft guidance for Bayesian trials. The guidance is nonbinding, but it sets out how to weigh a prior, a posterior probability and borrowed data.
Idea in Brief
The News
FDA's drug center holds a one-hour virtual webinar on Oct. 14, noon to 1 p.m. Eastern, on its draft guidance for Bayesian methods in drug and biologic trials.
What the Guidance Says
Success is usually judged by a posterior probability, and the prior, including any borrowed adult or earlier-trial data, must be justified and tested.
The Caveat
The guidance was published in January as a nonbinding draft, and the webinar is aimed at sponsors, not clinicians.
On Oct. 14, the Food and Drug Administration will hold a one-hour virtual webinar on its draft guidance for Bayesian methods in clinical trials of drugs and biologics. The Center for Drug Evaluation and Research, through its Small Business and Industry Assistance office, lists the session for noon to 1 p.m. Eastern. Attendees are asked to watch a pre-recorded overview before the live panel.
The guidance itself is not new. FDA posted it in January, and the Federal Register notice of Jan. 12 set a comment deadline of March 13. It remains a draft with nonbinding recommendations. The webinar is pitched at sponsors and statisticians, but the guidance shapes how some pivotal trials are designed and judged, and those are the trials physicians are asked to interpret.
What the webinar covers
FDA says the session will cover success criteria, operating characteristics, prior distributions, estimands and missing data, software, and documentation and reporting. The presenters are statisticians from CDER and the Center for Biologics Evaluation and Research, and the panel includes a CDER biostatistics division director and a deputy director of the Division of Pediatric and Maternal Health. The page names pediatric and rare disease trials as the most common settings for Bayesian methods.
What a posterior probability tells you
A p-value and a confidence interval describe how trial data would behave under fixed assumptions about the true effect; the guidance describes frequentist inference as based on the conditional probability of observing certain data. A Bayesian analysis starts from a prior distribution, combines it with the trial data and reports a posterior distribution. From that come a posterior mean, a credible interval and posterior probabilities, such as the probability that the effect exceeds zero.
The guidance gives a plain example. If the posterior probability of effectiveness is 0.98, the posterior probability that the treatment is ineffective is 0.02. It attaches a condition: that reading holds when the prior was chosen to accurately summarize what was known before the trial. A credible interval can look like a confidence interval on a forest plot, but it depends on that prior.
Success criteria change too. In conventional pivotal trials, the guidance says the familywise Type I error rate is almost always held to no more than 0.025, one-sided. In a Bayesian trial the criterion is most often a posterior probability that the effect exceeds a threshold, so a reader has two numbers to check: how large a benefit counts and how probable it must be. Some designs, such as complex adaptive trials, calibrate the Bayesian criterion back to the 0.025 standard. When a trial borrows outside information, the guidance says that default may not apply.
A posterior probability of 0.98 is only as good as the prior behind it. Read the prior before you read the probability.
Clinical Pearls
- A Bayesian analysis combines a prior distribution with the trial data to produce a posterior distribution, from which credible intervals and posterior probabilities are drawn.
- Conventional pivotal trials almost always hold the familywise Type I error rate to no more than 0.025, one-sided; with borrowing, that default may not apply.
- A Bayesian success criterion is most often a posterior probability that the true effect exceeds a prespecified threshold.
- FDA says informative priors have most often been proposed in pediatric and rare disease trials and require strong justification.
- Without discounting, adult data used as a pediatric prior would often overwhelm the pediatric results regardless of what the trial shows.
- The guidance recommends sensitivity analyses that use a range of reasonable alternative priors, such as varying the amount of borrowing.
Borrowing in pediatric and rare disease trials
The guidance says informative priors have been most often proposed in pediatrics and rare diseases, and it asks sponsors for strong justification, including why an approach without borrowing is not feasible and how relevant the outside data are. In pediatric extrapolation, adult results can form the prior for a pediatric trial. FDA warns that an undiscounted adult prior would often meet typical success criteria on its own and overwhelm the pediatric data, so a prior centered on adult-like benefit but with more uncertainty is often more reasonable.
The guidance cites examples already in use. The phase 3 trial of Rebyota, a fecal microbiota product for recurrent Clostridioides difficile infection, used a Bayesian primary analysis that incorporated a prior phase 2 placebo-controlled study; the product was approved in 2022. Supportive Bayesian analyses were also used in pediatric type 2 diabetes supplements for empagliflozin and linagliptin, where reviewers judged the adult information relevant enough to borrow.
The main hazard is prior-data conflict, where the trial results are inconsistent with the prior. The guidance notes that borrowing typically inflates the Type I error rate above the nominal 0.025. It distinguishes static discounting, which borrows a fixed amount, from dynamic discounting, which borrows less when the data diverge from the prior.
What to check when you read one
These questions are drawn from the draft guidance as a reading aid, not an FDA checklist. First, find out whether the Bayesian analysis was primary or supportive; the guidance notes Bayesian calculations can also govern interim analyses and inform dose selection. Second, find the prior and its source: an earlier trial of the same drug, adult data or a related population. The guidance calls a prespecified prior critical.
Third, ask how much the prior counts. The guidance points to effective sample size, which expresses borrowed information in patient-equivalents, and to discounting parameters. Fourth, look at the success criterion: the threshold, the required probability and whether both were prespecified. Fifth, look for sensitivity analyses that vary the prior, for example by changing the amount of borrowing. A conclusion that survives only under the sponsor's preferred prior deserves caution.
Last, apply the usual questions about estimands and missing data, which the guidance says apply to a Bayesian trial as to any other. The same habits that help in reading guidelines apply here: find the assumptions before weighing the conclusion. The guidance is still a draft, so details could change in a final version.
Reported, fact-checked and edited by The Script Pad's editorial staff.
This article is for professional education and does not replace clinical judgment. Treatment decisions should be based on the individual patient and current guidelines.
