FDA Draft Bayesian Guidance

• Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products Guidance for Industry
• Released January 2026
• Longstanding efforts of many biostatisticians in FDA CDER & CBER

Immediate Benefits

• Satisfies a long-time need at FDA
• CDRH published a Bayesian guidance 20 years ago under the leadership of Greg Campbell
• Drugs and Biologics have never had one
• Further legitimizes Bayes in general, not just for special cases
• Provides a pathway for Bayesian designs and analyses
• Communicates to industry that Bayesian proposals are welcomed
• awareness of Bayes and Bayes training at pharma & FDA (statisticians + reviewers)

What Problems Needed To Be Addressed?

• Reluctance of sponsors to use Bayesian methods in Phase III trials
• flexibility of RCT designs
• Better actionability of efficacy evidence measures
• Formal incorporation of extra-study information
• Pathway to less arbitrary handling of multiplicities
• Paradigm simplification

Paradigm Simplification

• All applications of Bayes were previously required to compute
• requires complex simulations that violated the likelihood principle
• Had to factor in intentions

Example Need for Paradigm Simplification

• Asthmatx Alair bronchial thermoplasty for refractory asthma
• Placebo bronchial ablation - strong control
• Bayesian decision criterion, flat prior:
  • P(benefit) > 0.99 at interim look
  • P(benefit) > 0.964 at final look
• Controls overall at 0.05

Example, continued

• Enrollment far exceeded expectations
• No time for interim look
• Final posterior prob. 0.96
• Failure of primary endpoint, which required 0.964
• Failed because of an intended look that never happened
• Hybrid Bayesian/frequentist procedures create complexity and confusion and violate the likelihood principle (data that could be obtained but were not are irrelevant)

fharrell.com/post/hybrid

Evidence Quantification and Actionability

• Longstanding problem in statistics: was never an error probability
• = probability of making an assertion
• Cannot call something an error by assuming the very fact that would make it one
• Translating P(getting data even more impressive if treatment doesn't work) to P(treatment works) is a leap
• Regulator's regret cannot be computed from or -values
• It is 1 - P(efficacy > 0 | data, prior)

Asking a Bayesian to Compute is Like ...

• telling a patient the specificity of the test he already underwent (P(test - | no disease))
• asking a poker player winning $10M/year to justify his ranking by how often he places bets in games he didn't win
• judging a politician by how often he speaks when he's honest instead how often he lies when he speaks

Historical Note: Illusion of Objectivity

Fisher touted his approach as being objective by

• sneakily changing the question (does the treatment work?)
• to a different question that could be answered without computers and without any prior knowledge:
  • If the treatment doesn't work, how surprised should we be by data more impressive than ours?

Paradigm Contrast in a Nutshell

• Do you want the probability of a positive result when there is nothing?
• Or the probability that a positive result turns out to be nothing?

Most Important Aspects of the Guidance

• Documents that Bayes has its own operating characteristics
• These OCs having nothing to do with
• They are P(correct decisions from data)
• Bayesian OCs can only be poor when study generating prior (simulation/sampling prior) is at odds with analysis prior

OC Simulations of Interest

• Simulate P(correct decision) under mismatch of priors and
• Example decision rule:
• Primary OC is
• Simulate 1000 RCTs with a universe of effects =
• Find the subset of the 1000 for which is triggered (under prior )
• Accuracy = proportion of this subset for which generating the data (from ) was actually

hbiostat.org/bayes/design

Guidance Provides Two Paths

• Not thinking hard about the prior
  • Noninformative prior (allows for huge treatment effects)
  • Demonstrate is conserved
• Thinking hard about the prior
  • Demonstrate reliability of the efficacy decision rule
  • excellent pure Bayesian OCs

Multiplicity

• Frequentist: Multiplicity comes from chances you give data to be extreme (e.g., multiple data looks)
  • And from detaching clinical and statistical significance
• Bayesian: No multiplicity from sequential looks
  • New looks merely make previous looks obsolete
  • No other real multiplicities, just assertions of varying stringency

fharrell.com/post/nfl

Challenges of Multiplicity Adjustments

• Not seen as logical in fields outside statistics
  • E.g. why should evidence for A vs B be discounted because we also examined C vs D?
• Arbitrary with no general principle to dictate choices
• Done on the randomness scale; should be on the clinical scale

Bayes Provides Logical Clinical Ways to Handle Multiplicities

• Multiplicity has been miscast as a randomness issue all along
• Multiplicity setting the bar too low in the assertions
• E.g. easy for P(treatment benefit on of endpoints) to be high
• Intersection assertions: Asserting that 4 things are simultaneously true is a high bar

Dealing with Union Assertions

• Make the bar high enough by demanding evidence for more than trivial effects
• E.g. P(any mortality effect or 10% reduction in nonfatal endpoint) > 0.95

fharrell.com/bayes/bet/multiplicity

Where Do Frequentists Need to Spend Time?

• data model specification
• approximations to complex sampling distributions and use of method (with poor confidence coverage)
• asymptotics
• sufficient statistics and ancillarity
• simulating accounting for intentions
• multiplicity adjustments
• gauging evidence for compound assertions
• ...

Frequentists Don't Spend Enough Time on Sampling Distributions

• Aside from simple cases, exact p-values and CLs are seldom available
• When interim analyses are done
  • sampling distributions are extremely complex (can be bimodal)
  • no RCT publications include correct point estimates and CLs
• Even commonly used models such as binary logistic regression and nonlinear mixed effect models have p-values and CLs of uncertain accuracy

Where Do Bayesians Need to Spend Time?

• data model specification
• choice of prior
• compute time
• diagnostics for convergence of posterior samples

Summary

The FDA draft Bayesian guidance

• establishes Bayesian methods as viable options in all contexts
• demonstrates the practicality of Bayes
• establishes that Bayes has its own OCs that are distinct from
• justifies use of pure Bayesian OCs if the prior is well justified or simulations show that reasonable vs. conflicts still result in excellent decision accuracy

Summary, continued

The guidance

• aids in understanding that is not an error probability
• outlines approaches for prior specification
• generates more exposure, interest, and demand for training in Bayesian methods

More Information

• hbiostat.org/bayes
• Lee, Harrell, LaVange, Spiegelhalter, JAMA 2026
• hbiostat.org/bayes/design - simulation of Bayesian sequential trial and OCs
• fharrell.com/post/bthink - Bayesian thinking
• fharrell.com/post/journal - My Bayesian journey
• hbiostat.org/bbr/alpha - vs. decision errors
• fda.gov/media/190505/download