3  The Regulatory Framework

A drug is approved or not approved on a legal finding, and everything in this book is aimed at that finding. A trial can be elegant, efficient and ethically admirable, and it still fails its sponsor if the evidence it produces does not meet the standard the law sets. The standard does not say how the evidence must be produced, and that gap is where design choices, including Bayesian and adaptive ones, live. This chapter is the map. It says what the law requires, what international and national guidance recommends on top of it, which meetings a sponsor can ask for, and where the European and British regulators agree or differ.

Two cautions before starting. First, the regulatory text changes quickly. Several of the documents below were issued as drafts in 2025 and 2026, and a draft can be revised or withdrawn. The statuses in this chapter are as of November 2026, and the dates are those printed on the documents. Second, regulatory documents differ in force. A statute binds. A regulation binds. A guidance “contains nonbinding recommendations”. This chapter keeps those categories apart.

ImportantRegulatory anchor

This chapter is about the anchors themselves. The documents that carry the legal and procedural weight are:

• FD&C Act section 505(d) (United States Congress 1962), as amended in 1997 (United States Congress 1997): the drug cannot be approved without “substantial evidence” that it has the effect it claims, defined as adequate and well-controlled investigations. This is law.
• 21 CFR 314.126 (U.S. Food and Drug Administration 2024a): the regulation that lists the characteristics of an adequate and well-controlled study. This is regulation.
• FDA draft guidance on substantial evidence, June 2026 (U.S. Food and Drug Administration 2026a): how FDA reads the standard, including when one trial plus confirmatory evidence can suffice. Draft, nonbinding.
• ICH E8(R1), E9, E9(R1), E20 (draft), M15 (International Council for Harmonisation 2021, 1998, 2019, 2025a, 2026): the harmonized statistical and design principles that FDA, EMA and other regulators adopt as guidance.
• FDA Adaptive Designs (final, November 2019) (U.S. Food and Drug Administration 2019a) and Bayesian Methodology (draft, January 2026) (U.S. Food and Drug Administration 2026e): the two FDA documents that govern the designs in this book.
• FDA Formal Meetings guidance (August 2026) (U.S. Food and Drug Administration 2026b) and the Complex Innovative Trial Design guidance (December 2020) (U.S. Food and Drug Administration 2020): how a sponsor gets answers from FDA before the trial starts.

3.1 The legal standard

What the statute says

The central sentence of US drug regulation comes from the Kefauver-Harris Amendments of 1962, passed after the thalidomide tragedy, which for the first time required manufacturers to show that a drug works and not only that it is reasonably safe. Section 505(d) of the Federal Food, Drug, and Cosmetic Act directs FDA to refuse approval of a new drug application if there is a lack of substantial evidence that the drug will have the effect it purports or is represented to have under the conditions of use in the labeling. The Act then defines substantial evidence as “evidence consisting of adequate and well-controlled investigations”, including clinical investigations, by experts qualified by training and experience, on the basis of which it could fairly and responsibly be concluded by such experts that the drug will have the claimed effect (United States Congress 1962).

Two features of this definition matter for everything that follows. The first is the plural: the statute speaks of investigations, and for decades FDA read that as two independent adequate and well-controlled trials, each with a positive result. The second is that the standard is one of evidence and not of method: the statute names no test, level, sample size, prior or design, and asks only whether qualified experts could fairly and responsibly conclude that the drug works. Frequentist hypothesis testing at a one-sided 0.025 level became the customary way to show this, and the customary way hardened into expectation, but the law does not mention it. The Bayesian methodology draft says so in effect when it allows that its default calibration to 0.025 “may not be applicable or appropriate” in some circumstances (U.S. Food and Drug Administration 2026e).

The 1997 amendment: one trial and confirmatory evidence

The Food and Drug Administration Modernization Act of 1997 (United States Congress 1997) added a sentence to section 505(d): if the Secretary determines, based on relevant science, that “data from one adequate and well-controlled clinical investigation and confirmatory evidence” are sufficient to establish effectiveness, the Secretary may consider those data to constitute substantial evidence. The confirmatory evidence may come from before or after the trial. The power is discretionary: it is FDA’s determination, based on relevant science, not a right of the sponsor.

The amendment has been used more often over time, especially in oncology, rare diseases and pediatric settings where a second trial is impractical or unethical. FDA’s draft guidance on the topic was first issued in December 2019 (U.S. Food and Drug Administration 2019b) and was revised in June 2026 (U.S. Food and Drug Administration 2026a), which, when finalized, will replace the 1998 guidance Providing Clinical Evidence of Effectiveness for Human Drug and Biological Products. The revision was announced in the Federal Register on 24 June 2026 under docket FDA-2019-D-4964.

The regulation: what “adequate and well-controlled” means

The statute leaves “adequate and well-controlled” undefined. FDA’s regulation at 21 CFR 314.126 defines it, and it is the single most useful list for a trialist to carry in the head (U.S. Food and Drug Administration 2024a). The regulation says the purpose of conducting clinical investigations of a drug is to distinguish the effect of the drug from other influences, such as spontaneous change in the course of the disease, placebo effect, or biased observation. It then lists the characteristics of an adequate and well-controlled study. In paraphrase, the study:

1. has a clear statement of its objectives and of the method of analysis to be used;
2. uses a design that permits a valid comparison with a control, so that the effect of the drug can be quantified (the regulation names placebo concurrent control, dose-comparison, no treatment, active treatment and historical control as recognized types);
3. selects subjects by an adequate method, with reasonable assurance that they have the disease or condition being studied;
4. assigns subjects to treatment and control groups in a way that minimizes bias and assures comparability on baseline factors such as age, sex, severity and duration of disease (randomization is the usual way);
5. takes adequate measures to minimize bias on the part of subjects, observers and analysts of the data (blinding is the usual way);
6. uses well-defined and reliable methods of assessing response;
7. analyses the results adequately to assess the effects of the drug.

Two observations. Item 7 asks for an adequate analysis, and the regulation expects the protocol to state the proposed methods, and where it did not, the report to explain how the methods were chosen. This is the regulatory root of the pre-specification that every guidance in this chapter repeats, and it applies to Bayesian analyses as it applies to frequentist ones. Item 2 admits historical control as a recognized type, but the 2026 draft cautions that external controls have a weakness: differences between groups arise “because of the lack of randomized treatment assignment” (U.S. Food and Drug Administration 2026a). The hierarchy of controls will come back in Chapter 19.

The 2026 substantial-evidence draft

The June 2026 revised draft guidance, Demonstrating Substantial Evidence of Effectiveness for Human Drug and Biological Products (U.S. Food and Drug Administration 2026a), restates how FDA reads the statute. It is a draft, so it is not for implementation, and it revises the draft issued in December 2019 (U.S. Food and Drug Administration 2019b). What it says that matters here, in the order the document gives it:

• It describes the conventional route (two trials) and a single-trial route. The single-trial route “will involve a highly persuasive trial or a source of strong confirmatory evidence” (Section IV.A of the draft).
• It lists the factors FDA weighs in deciding whether one trial and confirmatory evidence suffice: the seriousness of the disease and the unmet need, the strength of the single trial, the strength and source of the confirmatory evidence, and “whether it is ethical and practicable to conduct more than one adequate and well-controlled trial”.
• It allows confirmatory evidence of types that are not trials. For instance, “mechanistic evidence of the drug’s treatment effect in a particular disease may be appropriate to use as confirmatory evidence”, within the regulatory flexibility the draft describes in its Section V for serious disease and unmet need.
• It says that the plan for one trial “should be discussed with FDA prior to initiating the single trial”.

The draft explicitly recognizes Bayesian analyses, and it matters for them: a Bayesian analysis that uses an informative prior built from earlier data is a close cousin of “one trial plus confirmatory evidence”, and the borrowing discussion in Chapter 19 returns to the connection. Prespecified borrowing formalizes how earlier evidence enters the analysis, but FDA still evaluates its relevance, weight and contribution to substantial evidence, including in discussions before the trial.

TipTry it in the lab

Open Module 11, What counts as proof (https://nytel.ai/testing/11-substantial-evidence.html). Before you touch a control, predict: a first trial just clears the win line with p = 0.04 (two-sided, counting luck in both directions). If the drug’s true effect is exactly what that trial saw, what is the chance that a second trial of the same size also wins? Then run the repeats. The lab’s answer, about half, is the statistical reason a regulator may prefer replication; its second panel shows the other side, two independent wins at one-sided 0.025 by luck about once in 1,600 times.

What the standard implies for the numbers

Two adequate and well-controlled trials, each tested at a two-sided 0.05 level (one-sided 0.025), independent and each required to succeed in the prespecified beneficial direction, imply that a useless drug passes both only with probability \(0.025 \times 0.025 = 1/1600\). A single trial at 0.025 passes a useless drug 1 time in 40. A single trial at a stricter level (the book’s companion online lab, module 11, https://nytel.ai/testing/11-substantial-evidence.html, uses 0.001) passes a useless drug 1 time in 1,000. In this sense, one very persuasive trial and two ordinary trials can offer comparable protection against a false positive, which is why FDA’s willingness to accept “a highly persuasive trial” is consistent with the statute. For trials with an overall frequentist statistical analysis plan, the Bayesian methodology draft says that the familywise type I error rate is almost always required to be no greater than 0.025 one-sided (U.S. Food and Drug Administration 2026e). The draft also describes direct interpretation of posterior probabilities and decision-theoretic or benefit-risk criteria, so calibration is one of its routes (the book’s default for confirmatory adaptive designs without borrowing), not a universal requirement. FDA’s REBYOTA review used two posterior thresholds chosen as analogues of one-sided type I error rates of 0.00125 and 0.025 without borrowing (Section 3.11). The notation of this book follows suit: one-sided \(\alpha = 0.025\) is the default, and we say so whenever we use two-sided.

3.2 What a traditional submission looks like

It is worth describing the traditional package accurately, because it is what a regulatory reviewer expects to see and what the Bayesian design is measured against.

The development path. A sponsor begins with nonclinical work, then files an Investigational New Drug application (IND). FDA has 30 days to object; if it does not, the first human trials can begin. Phase 1 examines tolerability and pharmacokinetics, phase 2 looks for efficacy and chooses a dose, and phase 3 is the pivotal comparison, usually randomized, controlled and blinded. The sponsor then files a New Drug Application (NDA) or, for biologics, a Biologics License Application (BLA), which FDA reviews, checking completeness for 60 days first. At the major junctures FDA offers meetings: a pre-IND meeting, an end-of-phase 2 meeting to agree the phase 3 plan, and a pre-NDA or pre-BLA meeting to confirm the file is complete. Module 13 of the companion lab draws this road.

The phases themselves, what each is designed to learn and how Bayesian methods enter each, are the subject of Chapter 2, the chapter before this one. Bioequivalence studies, which establish that a generic or reformulated product delivers the same exposure as a reference product and so replace a new efficacy trial, follow a different pathway from the phase 1 to 3 efficacy sequence when they support a generic (ANDA) approval. They also occur inside an innovator program, for example to bridge formulations, and FDA’s final bioequivalence guidance covers INDs and NDAs as well as ANDAs. Chapter 2 treats the Bayesian version of the bioequivalence question.

The pivotal trial protocol. The protocol fixes, before any patient is randomized, the primary endpoint and its definition, the estimand (the precise treatment effect the trial targets, with a stated strategy for intercurrent events such as rescue medication or treatment discontinuation, as set out in ICH E9(R1) (International Council for Harmonisation 2019)), the randomization scheme, blinding, the sample size computed from a power formula, the analysis populations, and the primary analysis. ICH E9 says the statistical methods must be described in the protocol and that “all interim analyses should be carefully planned in advance and described in the protocol” (International Council for Harmonisation 1998). The sample size is usually the number \(n\) that gives power \(1-\beta\) (often 80 or 90 percent) against a clinically relevant effect at one-sided \(\alpha = 0.025\). Chapter 9 derives that formula.

The statistical analysis plan (SAP). Written before database lock and typically before unblinding, the SAP states the primary analysis, the handling of missing data, the multiplicity strategy (a single primary endpoint, a gatekeeping hierarchy, a Bonferroni or graphical procedure; see the FDA multiple endpoints guidance of October 2022 (U.S. Food and Drug Administration 2022b) and Chapter 10), the sensitivity analyses, and the timing and boundaries of any interim analyses, usually a group sequential design with alpha spending (Chapter 11). An independent Data Monitoring Committee (DMC) sees the unblinded interim results under a charter, and the sponsor’s trial team does not.

The evidence. For the standard route to approval, the package contains two adequate and well-controlled trials, each with a statistically significant result on its pre-specified primary endpoint, or one such trial plus confirmatory evidence in the circumstances discussed above. The primary result is a point estimate, a confidence interval and a p-value. FDA’s statistical reviewers re-run the analysis from the datasets the sponsor submits, in standardized formats, and check that what was done matches what the protocol and SAP said would be done.

What the regulator is really asking. The questions a statistical reviewer asks of a traditional trial are worth listing, because the same questions reappear in new form for Bayesian and adaptive designs. Was the primary analysis specified before the data were seen? Was the type I error rate controlled for the family of hypotheses that supports the claim? Is the effect estimate credible, free of selection bias, and clinically meaningful? Was the trial conducted with integrity: blinded, with limited access to comparative data, with deviations documented? The answers are almost entirely about pre-specification and error control, which the traditional package makes simple: the type I error rate is \(\alpha\) by construction and the power is a formula.

Where the traditional design is the right choice. For a single-look trial with a well-understood endpoint and no usable prior information, the fixed-sample design is simple, transparent, and cheap to review. A sponsor with no meaningful prior information and no appetite for the added simulation work should not adopt a more elaborate design merely because it is available. Chapter 21 shows a case (REWIND) where the scheduled design wins.

3.3 What changes with a Bayesian adaptive design

The short answer is that the legal standard does not change, the evidentiary bar does not change, and the demands for pre-specification and error control do not change. What changes is how the operating characteristics are established, who has to be convinced, and what the sponsor must hand over. It is helpful to separate three kinds of change.

The design becomes a rule set, and the rule set is the thing that is pre-specified. In a traditional trial the protocol pre-specifies a test and a sample size. In an adaptive trial it also pre-specifies decision rules that depend on accumulating data: when to stop for efficacy or futility, when to drop an arm, how to change allocation, and how to re-estimate the sample size. FDA’s 2019 guidance defines an adaptive design as “a clinical trial design that allows for prospectively planned modifications to one or more aspects of the design based on accumulating data” (U.S. Food and Drug Administration 2019a), and the draft ICH E20 sharpens the point: “prospectively planned means that the potential trial adaptations are pre-specified in the clinical trial protocol prior to initiation of the trial” (International Council for Harmonisation 2025a).

Operating characteristics are established by simulation, not by formula. A group sequential design has analytic boundaries. A Bayesian adaptive design with predictive probability stopping, response-adaptive allocation and borrowing generally does not. The 2019 guidance says “a major use of simulations in adaptive trial design is to estimate trial operating characteristics” (U.S. Food and Drug Administration 2019a), and the Bayesian draft recommends “comprehensive” simulation covering the plausible range of assumptions, with the simulation code available to FDA for verification (U.S. Food and Drug Administration 2026e). Chapter 14 is devoted to this. This is the currency in which the book itemizes costs: each benefit of a Bayesian adaptive design is bought with a simulation burden, and the simulation report is how the sponsor pays.

The success criterion is stated on the posterior and calibrated to error. The Bayesian draft gives the standard form: success is declared when the posterior probability that the true treatment effect exceeds a threshold is high enough, \(\Pr(\delta > a \mid \text{data}) > c\), where \(a\) is the minimum benefit and \(c\) the minimum probability (U.S. Food and Drug Administration 2026e). Section IV.A describes three routes: calibration of the thresholds to a type I error rate (“typically 0.025 one-sided” for the overall familywise error rate), direct interpretation of posterior probabilities (for example with borrowing, by agreement with FDA), and decision-theoretic or benefit-risk criteria. Calibration is this book’s default for confirmatory adaptive designs without borrowing, and the draft says it may be appropriate for complex adaptive designs; it is one route, not a requirement. Against a single unadjusted analysis the calibrated Bayesian bar is not lower without an informative prior. The other routes require agreement with FDA, and their justification is heavier, not lighter (Chapter 8 and Chapter 19).

Table 3.1 lines the two submissions up.

Table 3.1: What changes between a traditional and a Bayesian adaptive submission.

Element	Traditional fixed-sample or group sequential	Bayesian adaptive	Where developed
Legal standard	Substantial evidence (505(d), 314.126)	Identical	This chapter
Primary inference	p-value and confidence interval	Posterior probability and credible interval, usually also with a calibrated frequentist error rate	Chapter 6, Chapter 7
Type I error	\(\alpha\) by construction; alpha spending for looks	Estimated by simulation; calibrate \(a\) and \(c\) to FWER 0.025 one-sided by default	Chapter 14
Prior information	None in the primary analysis	Pre-specified analysis prior; justified; sensitivity analyses	Chapter 8, Chapter 19
Sample size	Fixed from power formula (or event count)	Maximum plus expected size under scenarios; predictive stopping	Chapter 9, Chapter 15
Interim looks	Pre-planned group sequential, DMC	Pre-planned decision rules, DMC, firewalls, access limits	Chapter 11, Chapter 23
Allocation	Fixed randomization ratio	Fixed or response-adaptive (the latter needs strong justification)	Chapter 16
Documents beyond the protocol	SAP	SAP plus simulation report with code, prior documentation	Chapter 23
Meeting with FDA	Pre-IND, end-of-phase 2, pre-NDA	Same, plus CID paired meeting when design is novel	This chapter

The third row of the table is a regulatory choice and not a mathematical necessity: what makes the calibrated route the default is that it leaves FDA’s reviewers where they were: with a number they know how to interpret.

3.4 ICH: the harmonized principles

The International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) brings together regulators and industry to agree guidelines that member regulators then adopt. A guideline reaches Step 4 when the ICH Assembly adopts the final text, and regional regulators then implement it (FDA publishes it as guidance, EMA as a scientific guideline). Step 2 is a draft out for consultation. Six ICH documents shape this book.

E8(R1), General Considerations for Clinical Studies (Step 4, 6 October 2021) (International Council for Harmonisation 2021). It replaces E8 of 17 July 1997. E8(R1) is the document where “quality by design” enters clinical research: it says quality should be “driven proactively by designing quality into the study protocol and processes” (Section 3.1). It states that “the major purpose of a control group is to separate the effect of the treatment(s) from the effects of other factors” (Section 5.3). Its critical-to-quality factors, for a Bayesian adaptive trial, include the prior, the calibration of the success criterion and the integrity of the interim data, though E8(R1) does not say so.

E9, Statistical Principles for Clinical Trials (Step 4, 5 February 1998) (International Council for Harmonisation 1998). After more than 25 years E9 remains the foundation. It describes randomization, blinding, the analysis populations, the primary variable, interim analysis, and multiplicity. On randomization it says that the deliberate element of chance “tends to produce treatment groups in which the distributions of prognostic factors, known and unknown, are similar” (Section 2.3.2), and on analysis populations that “Preservation of the initial randomisation in analysis is important in preventing bias” (Section 5.2.1). On multiplicity it says that “the usual frequentist approach to the analysis of clinical trial data may necessitate an adjustment to the type I error” (Section 5.6). On interim analyses it asks for pre-planning: “all interim analyses should be carefully planned in advance and described in the protocol” (Section 4.5), and group sequential designs are “the usual way” to look repeatedly (Section 3.4). E9 mentions Bayesian approaches only briefly, in Section 1.2 and the Glossary, and does not elaborate them. Pre-specification applies to them as to any analysis; the point for this book is that E9 neither prohibits nor privileges them.

E9(R1), Addendum on Estimands and Sensitivity Analysis (Step 4, 20 November 2019) (International Council for Harmonisation 2019). FDA labels its own publication of this document May 2021. E9(R1) defines an estimand as “a precise description of the treatment effect reflecting the clinical question posed by a given clinical trial objective”, requires that “the targets of estimation are to be defined in advance of a clinical trial”, and introduces intercurrent events, events after treatment initiation that affect the interpretation or the existence of the measurements. It names five strategies for handling them: treatment policy, hypothetical, composite variable, while on treatment and principal stratum. A posterior probability is only as interpretable as the estimand it is posterior about. Estimands are developed in Chapter 5.

E20, Adaptive Designs for Clinical Trials (draft, Step 2, endorsed 25 June 2025) (International Council for Harmonisation 2025a). This is the first international guideline on adaptive designs. It remains a draft: the EMA consultation ran from 30 June to 30 November 2025, the EMA overview of comments was published on 16 February 2026, and the ICH work plan of July 2026 targets Step 3 and Step 4 sign-off in February 2027. Until final it binds no region, but it shows where regulators are heading. Its content: the definition quoted above; the expectation that a “proposed adaptive design requires a clear and compelling justification”; the expectation that in confirmatory trials “the number and complexity of adaptations should generally be limited”; the statement that “controlling the chances of false positive efficacy conclusions” is “an essential element of regulatory decision-making”; the warning that naive analyses can raise the type I error probability to “more than double” its nominal value; and a section on simulation (5.2). E20 also has a Bayesian section (5.3), which says that Bayesian adaptive designs “can achieve targeted operating characteristics” while “maintaining Type I error probability control” and defines Bayesian methods as those that “combine a prior probability distribution with current trial data to obtain a posterior probability distribution”. Where E20 and the FDA Bayesian draft meet, the expectation is calibration to type I error for confirmatory trials. Chapter 14 uses E20 section 5 for the contents of a simulation report. The adaptive combination methods behind confirmatory adaptive analyses go back to Bauer (1989) and Bauer and Köhne (1994), as the review of Bauer and colleagues recounts (Bauer et al. 2016); the interim-look machinery is in Chapter 11.

E11A, Pediatric Extrapolation (Step 4, August 2024) (International Council for Harmonisation 2024). E11A sets out when and how evidence from an adult or other source population can be extrapolated to children, and it promotes quantitative approaches including Bayesian borrowing and modeling and simulation to do so. It matters for this book because pediatric extrapolation is the clearest case where a regulator has asked for, rather than tolerated, a quantitative statement of how much prior information counts. It is taken up in Chapter 19, and belimumab in children is a worked case in Section 3.11. The Bayesian draft itself uses “adult data for children” as an example of borrowing (U.S. Food and Drug Administration 2026e).

M15, General Principles for Model-Informed Drug Development (Step 4, January 2026) (International Council for Harmonisation 2026). FDA published this document in June 2026. M15 defines model-informed drug development (MIDD) as “the use of computational modeling and simulation (M&S) methods that can include and integrate nonclinical data, clinical data, prior information” to inform decisions. Its structure is a credibility framework built on the question of interest, the context of use, model influence, the consequence of a wrong decision and model risk, where model risk is “derived by combining model influence and consequence of wrong decision” and “is essential for determining the requirements for model evaluation”. The more a decision rests on a model, and the worse a wrong decision would be, the more evidence the model needs. A Bayesian adaptive design is in part a model on which stopping decisions rest, and the simulation report is the evaluation, so the M15 vocabulary is a useful way to think about how much documentation FDA will want.

Three older ICH documents round out the picture and are referenced elsewhere: E3 (Step 4, November 1995; the structure of the clinical study report) (International Council for Harmonisation 1995), M4 (M4(R4), Step 4, 2016; the common technical document format of the submission) (International Council for Harmonisation 2016), and E6(R3) on good clinical practice (Step 4, January 2025) (International Council for Harmonisation 2025b). A sponsor writes the SAP and the simulation report to fit into the first two structures. E6(R3) governs conduct and applies to adaptive trials, including the integrity of interim data.

3.5 FDA design guidances

FDA guidance documents “contain nonbinding recommendations”. The agency’s own convention is that the word “should” in a guidance means “suggested or recommended, but not required”. A sponsor may depart from a guidance, but must be prepared to show that the alternative approach meets the statute and regulations. In practice a departure without prior discussion with the agency is a risk to the program, and that is why the meetings of the next section exist. A guidance is first issued as a draft, comments are taken (60 days after the Federal Register notice is usual), and a final follows, or the draft is revised or withdrawn. A draft is labeled as not for implementation, but sponsors often follow it because it tells them what reviewers will look for.

Adaptive Designs for Clinical Trials of Drugs and Biologics (final, November 2019)

This is the founding FDA document for the book (U.S. Food and Drug Administration 2019a). It is final. Its four principles for adaptive designs are the ones to remember: “the chance of erroneous conclusions should be adequately controlled, estimation of treatment effects should be sufficiently reliable, details of the design should be completely prespecified, and trial integrity should be appropriately maintained” (Section III). It says plainly that “adaptive features can inflate the Type I error probability of a trial”, with the simplest example a repeated look at accumulating data (Section III.A), and it works through the example of a single interim look at 50 percent of the information. It names simulation for operating characteristics (Section VI.A), limited access to comparative interim data (Sections III and VII), and an example of effectiveness rules based on a Bayesian posterior probability (Section V). It recommends that the four principles be addressed in the protocol and submission, and that complex designs be discussed with FDA beforehand. For frequentist and Bayesian adaptive designs alike, the 2019 document is the baseline, and the 2026 Bayesian draft builds on it.

Bayesian methods at the regulators before 2026

The January 2026 draft is not the first regulatory text on Bayesian trials. ICH E9 (1998) already allowed Bayesian approaches when the reasons for their use are clear and the conclusions are sufficiently robust (Section 3.4), and the device side went first. CDRH issued a draft Bayesian guidance on 23 May 2006 and the final version on 5 February 2010 (U.S. Food and Drug Administration 2010). The final text treats type I error as one operating characteristic among others and says that when prior information is used it may be appropriate to control it at a less stringent level, case by case (Section 4.8). The CDRH statisticians have written up the experience: Pennello and Thompson on reviewing Bayesian device trials (Pennello and Thompson 2008), Campbell on the Bayesian innovation the FDA sparked (Campbell 2011) and on the similarities between Bayesian and adaptive designs (Campbell 2013), and Campbell, Irony, Pennello and Thompson on progress since 2010 (Campbell et al. 2023).

On the drug side, the complex innovative trial design program gave Bayesian designs a forum for early dialogue (Section 3.7), and Ruberg and colleagues argued in 2023 that wider use of Bayesian approaches in drug development could start a virtuous cycle of experience and acceptance (Ruberg et al. 2023). The practical lesson is the same on both sides of the Agency: agree the prior and the operating characteristics with the reviewers before the trial starts.

Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products (draft, January 2026)

This is the document that most changes the conversation for a sponsor considering a Bayesian design (U.S. Food and Drug Administration 2026e). It is a draft guidance for industry, marked “Contains Nonbinding Recommendations”, issued by CDER and CBER under docket FDA-2025-D-3217 in January 2026. It does not apply to devices, which have their own Bayesian guidance (Section 3.5.2); two device approvals under it are in Section 3.11. Its scope is INDs, NDAs, BLAs and supplements, with a “primary focus” on “the use of Bayesian methods to support primary inference”. Its main elements, in the order a sponsor will meet them:

• Success criteria. The criterion is “most often based on the posterior probability that the true treatment effect size exceeds some threshold”, \(\Pr(\delta > a \mid \text{data}) > c\). For trials with an overall frequentist analysis plan, the thresholds \(a\) and \(c\) are almost always calibrated so that the familywise type I error rate across all primary and secondary estimands is no greater than 0.025 one-sided; the draft also describes posterior-probability and benefit-risk routes.
• Priors. “It is always critical for Bayesian analyses to have a prespecified prior.” The sponsor should “pre-specify and justify the full details of the proposed prior distribution in the protocol”. Skeptical priors “express skepticism about the presence of very large treatment effects”, and design priors that explore pessimistic assumptions should be considered. Where the data do not dominate the prior, “the accuracy of conclusions depends strongly on the choice of prior distribution”. Chapter 8 takes up these points.
• Borrowing. Static discounting fixes how much external information counts; dynamic methods “borrow less when the data are less similar and more when the data are more similar”. Borrowing needs “strong justification that considers feasibility … and the relevance of the available information”, information should typically come from patient-level data, and a prior that is too informative “risks overwhelming the data collected in the target population”. The draft says that type I error inflation is “not recommended” as the measure of borrowing, and prefers effective sample size and similar influence metrics. The draft’s named examples are summarized in Section 3.11. Chapter 19 develops borrowing.
• Operating characteristics and simulation. “Simulations should comprehensively cover the plausible range of assumptions”, and the plan should compare complex designs with simpler alternatives and request an FDA meeting before the trial starts.
• Documentation. The sponsor should submit protocol, SAP and simulation report “as early as possible” and “during the design stage”. The report on the trial should state the estimates and credible intervals, the success criteria, the sensitivity analyses, the computational details (software versions, code, the MCMC seed numbers). On software the draft says that “FDA does not require use of any specific software for statistical analyses”.

It requires nothing by itself, since it is a draft; it permits both the calibrated and the non-calibrated pathway, the latter with agreement. Chapter 23 turns the draft into a protocol and SAP checklist.

The response to the draft

Three commentaries on the draft appeared together in JAMA on 19 May 2026 (Lee et al. 2026; Evans et al. 2026; Gelman, Zwet, and Więcek 2026). The routes the draft leaves open are described in Chapter 3; this book takes calibration as the default for confirmatory adaptive designs without borrowing, and direct posterior criteria with FDA agreement when prior information is borrowed, as the draft allows. The older cautions about prior dependence, type I error and operational bias are collected in Section 1.9.

Master Protocols for Drug and Biological Product Development (revised draft, June 2026)

A first draft appeared in December 2023; a revised draft that “revises and replaces” it was announced in the Federal Register on 24 June 2026 (91 FR 37991), with comments closing on 24 August 2026 (U.S. Food and Drug Administration 2026d). FDA’s description of the changes is “additional recommendations on basket trials and minor changes for clarity on topics such as randomization, choice of control, and informed consent”. The document is issued under the Food and Drug Omnibus Reform Act of 2022 (U.S. Congress 2022). It defines a master protocol as “a protocol designed with multiple substudies, which may have different objectives and involve coordinated efforts to evaluate one or more drugs”; an umbrella trial “evaluate[s] multiple drugs concurrently for a single disease or condition”, a basket trial “evaluate[s] a drug for multiple diseases, conditions, or disease subtypes”, and a platform trial evaluates multiple drugs “in an ongoing manner, with drugs entering or leaving the platform”.

Its recommendations are specific. A shared control arm improves efficiency, and the draft discusses allocating more participants to the control arm than to each drug arm (the EMA platform-trials draft gives Dunnett’s square-root-of-\(k\) rule as one example). The primary comparison should use “only concurrently eligible control participants”; nonconcurrent controls are allowed only in rare circumstances agreed with FDA in advance, with concurrent-only sensitivity analyses. On multiplicity, FDA “generally does not recommend the use of multiplicity adjustments to strongly control the probability of making at least one type I error across the multiple comparisons of different drugs”, because the expected number of type I errors is the same whether the comparisons are made in independent trials or in one master protocol. Adjustment is still expected across doses or formulations of the same drug. For a new master protocol the draft points sponsors to early discussion with FDA, and a master protocol pre-IND meeting is requested as Type B (see the meetings section below). It complements the March 2022 oncology master protocols guidance (U.S. Food and Drug Administration 2022a). Chapter 20 develops all of this.

Multiple Endpoints in Clinical Trials (final, October 2022)

This final guidance (U.S. Food and Drug Administration 2022b) describes the multiplicity problem and the strategies for handling it when a trial has several endpoints, doses or comparisons, and the statistical procedures for strong control of the familywise error rate: Bonferroni, Holm, fixed-sequence and gatekeeping procedures, and graphical methods. It is the regulatory partner of ICH E9 section 5.6 and the basis of Chapter 10. For the Bayesian reader the guidance matters in two ways. It sets the expectation that a claim supported by several endpoints must come with a stated multiplicity strategy, which the Bayesian draft echoes by recommending that the FWER be controlled “across all primary and secondary estimands”. And it leaves room for hierarchical modeling as a complement, not as a replacement, for the confirmatory strategy.

AI and new approach methodologies

Two further FDA drafts share this book’s logic. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft, January 2025) (U.S. Food and Drug Administration 2025) sets a seven-step, risk-based credibility framework (question of interest, context of use, model risk, credibility plan, execution, documentation of results and deviations, adequacy determination), in which model risk combines “model influence” and “decision consequence”, the structure of M15. It covers AI that produces information supporting decisions on safety, effectiveness or quality, not AI used only for operational efficiency, so a machine-learned model that selects patients or imputes outcomes in a trial falls under it. General Considerations for the Use of New Approach Methodologies in Drug Development (draft, March 2026) (U.S. Food and Drug Administration 2026c) concerns nonclinical methods that replace or reduce animal testing. It says a NAM “does not necessarily need to be validated” and is judged in the weight of evidence.

Other FDA documents that recur

Several other FDA documents recur and are listed in Table 3.4: the final guidance Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases (August 2024, Project Optimus) (U.S. Food and Drug Administration 2024b) in Chapter 17, the February 2023 draft guidance on externally controlled trials (U.S. Food and Drug Administration 2023) in Chapter 19, the expedited programs guidance (final, May 2014) (U.S. Food and Drug Administration 2014), and the guidance on data monitoring committees (final, March 2006; revised draft, February 2024) (U.S. Food and Drug Administration 2006), the template for the DMC firewalls of Chapter 23. Expedited programs change the speed and intensity of FDA’s interaction, not the substantial-evidence standard: a faster review is not a lower bar.

3.6 Requires, recommends, permits

Because the force of a document varies, it helps to sort the documents above, and a book on design should say precisely what it means by “FDA requires”. Table 3.2 does the sorting.

Table 3.2: What each kind of regulatory source requires, recommends or permits.

Force	Source	Examples in this chapter	What it means for a sponsor
Requires (statute)	Act of Congress	FD&C Act 505(d) substantial evidence; FDAMA 1997 single-trial provision; FDORA 2022	Binding. The standard to be met
Requires (regulation)	Code of Federal Regulations	21 CFR 314.126 characteristics of adequate and well-controlled studies	Binding. Defines what the evidence must look like
Recommends (final guidance)	FDA guidance, final	Adaptive Designs 2019; Multiple Endpoints 2022; ICH E8(R1), E9, E9(R1) as adopted by FDA; Project Optimus 2024	Nonbinding, but FDA’s current thinking. Depart only with a reason and, ideally, with agreement
Recommends (draft guidance)	FDA guidance, draft	Bayesian Methodology (January 2026); Master Protocols (June 2026); Substantial Evidence (June 2026); AI (January 2025); NAMs (March 2026)	Nonbinding and subject to change. Shows where review is heading
Recommends (ICH Step 2 draft)	ICH draft, out for comment	E20 (June 2025)	Not yet adopted in any region
Recommends (ICH Step 4)	ICH final, adopted regionally	E11A (2024); M15 (January 2026)	Adopted by FDA as guidance; see the issuer’s page for the regional status
Permits	Statute or guidance giving discretion	One trial plus confirmatory evidence; non-calibrated Bayesian success criteria with FDA agreement; nonconcurrent controls in rare circumstances	Available only if the agency concurs, so the route goes through a meeting
Process	Procedural guidance, PDUFA commitments	Formal Meetings (August 2026); CID and MIDD paired meeting programs	Defines how to ask. Not a source of evidentiary standards

Two corollaries follow. A sponsor does not need a guidance to be final before using it; but a sponsor should never describe a draft as “FDA requirements” in a protocol or a meeting package. And a permission is not an entitlement. The permissions that matter most to Bayesian adaptive designs are all of the form “if FDA agrees”, and the practical meaning of “regulation as the spine” is that those agreements are sought in writing, early, and on the record.

3.7 Interacting with FDA

This section explains how a sponsor finds out whether FDA agrees. FDA’s meeting types and programs differ in what they can promise, and the differences matter.

Formal meetings under PDUFA

The procedural guidance Formal Meetings Between the FDA and Sponsors or Applicants of PDUFA Products (August 2026) is the current text; it states that it withdraws the December 2017 formal meetings guidance (U.S. Food and Drug Administration 2026b). PDUFA, the Prescription Drug User Fee Act, is the law, periodically reauthorized, under which sponsors pay fees and FDA commits to review goals and meeting timelines; the current commitments are in the PDUFA VII goals letter for fiscal years 2023 through 2027 (U.S. Food and Drug Administration 2021b). The guidance defines six meeting types (Section III): “Type A, Type B, Type B (end of phase (EOP)), Type C, Type D, and Initial Targeted Engagement for Regulatory Advice on CDER and CBER ProducTs (INTERACT)”. In plain terms:

• Type A is “necessary for an otherwise stalled drug or biological product development program to proceed or to address an important safety issue”. It is for a program that is stuck, for instance after a disagreement with the review division.
• Type B covers the milestone meetings: pre-IND, and pre-NDA or pre-BLA meetings. Type B (end of phase) covers certain end-of-phase 1 meetings (for products to be considered under 21 CFR part 312 subpart E or part 314 subpart H, or similar products) and end-of-phase 2 meetings, where the sponsor shows the phase 3 plan (Section III.C). A master protocol pre-IND meeting is requested as Type B.
• Type C is the residual category: any meeting other than A, B, B (EOP), D or INTERACT (Section III.D). It is the usual route for design questions between milestones, such as a novel endpoint or a Bayesian analysis plan.
• Type D is small and focused: “often one, but typically not more than two issues”. It is meant for a narrow question that does not need a multidisciplinary meeting.
• INTERACT is for very early engagement: it is “intended for novel drug and biological products and development programs that present unique challenges in early development”. It comes before the first IND.

The timelines in the guidance (Tables 1 and 2) are how a trialist plans a calendar. FDA responds to a meeting request within 14 calendar days of receipt for Type A, Type B (EOP) and Type D, and within 21 for Type B and Type C. The meeting or written response is scheduled within 30 days of receipt for Type A, 60 for Type B, 70 for Type B (EOP), 75 for Type C and 50 for Type D; the INTERACT clocks should be read from the guidance tables. Formats include face to face, video, teleconference and “written response only”, in which written answers replace a meeting (Section IV). Module 13 of the companion lab turns these tables into a clock, and the sensible planning conclusion is that a Type C meeting on a novel design takes roughly three months from request to answer, so it must be requested well before the protocol is final.

Meeting minutes matter. FDA’s written responses and the sponsor’s record of agreements are the main evidence at the time of review that a design was discussed. They are not a binding contract: a review division can change its position if the science changes or if the data raise new concerns. But an agreement documented in minutes is persuasive, and a sponsor that departs from it should expect a question.

The Complex Innovative Trial Design (CID) paired meeting program

The 21st Century Cures Act of 2016 (United States Congress 2016) directed FDA to promote novel trial designs, including complex adaptive and Bayesian designs, through public meetings and guidance. FDA’s response, under its PDUFA commitments, was the CID paired meeting program (U.S. Food and Drug Administration, n.d.a), and the guidance Interacting with the FDA on Complex Innovative Trial Designs for Drugs and Biological Products (final, December 2020) (U.S. Food and Drug Administration 2020) describes how it works. The program is the main route for a sponsor proposing a Bayesian or otherwise complex adaptive design. In outline: a sponsor applies with a description of the proposed design, and accepted sponsors receive a pair of meetings with extra FDA attention, the second after the sponsor has developed the simulation work discussed in the first. Admission is selective, and the program page is the place to check current eligibility and application windows. Two features matter for the reader. The meetings do not replace the ordinary milestones, and the program asks the sponsor to bring the thing that matters most to a Bayesian design, the simulations: a CID package normally includes the proposed design, the scenarios, the operating characteristics, and the comparison with a simpler alternative.

The MIDD paired meeting program

The model-informed drug development paired meeting program (U.S. Food and Drug Administration, n.d.d) offers paired meetings with FDA on the use of modeling and simulation approaches in a specific development program, for instance dose selection, trial design simulation, or use of an exposure-response model to support a labeling claim. It is the procedural counterpart of ICH M15 (International Council for Harmonisation 2026): M15 sets the principles for assessing model credibility, and the program is where a sponsor tests a particular modeling plan against FDA’s reading. A Bayesian adaptive design with a model-based dose-response component (see Chapter 17) can fall naturally under both the CID and the MIDD programs, and a sponsor should ask which fits the question: the CID program for the trial’s decision rules, the MIDD program for the models that feed them.

Critical Path Innovation Meetings (CPIM)

CPIM is a different kind of meeting (U.S. Food and Drug Administration, n.d.b). It is a nonbinding, informal discussion of methods or technologies that are not tied to a single product or an active application. FDA’s stated aim is to respond to a request for a CPIM within 14 days. For a group developing a general methodology, such as a class of Bayesian borrowing models, a CPIM is a low-cost way to get an informal view early.

Drug Development Tool (DDT) qualification

DDT qualification (U.S. Food and Drug Administration, n.d.c) is a formal process by which FDA qualifies a tool, such as a biomarker, a clinical outcome assessment or an animal model, for a stated context of use. Once qualified, the tool can be used by any sponsor within that context of use without FDA re-reviewing its suitability in each application. The context of use is the key idea: a tool is qualified for a purpose, not in the abstract. The 21st Century Cures Act (United States Congress 2016) put the process on a statutory footing. For trial design the relevance is endpoints: a qualified tool is the kind of “well-defined and reliable method of assessing response” that 21 CFR 314.126 asks for, and using it in its context removes one reviewer question. DDT qualification is slow and is typically pursued by consortia.

How to choose the meeting

Table 3.3 sorts the options by the question being asked. The “stuck, early, or small” rule of thumb in the takeaways of the companion lab (module 13) is a useful first filter.

Table 3.3: Which meeting fits which question.

Question	Meeting or program	Binding?	Chapter
Is this first-in-human plan acceptable for a novel product?	INTERACT, then pre-IND (Type B)	No (advice)	This chapter
Is the phase 3 plan acceptable?	Type B (end of phase 2)	No, but minutes matter	Chapter 23
One narrow question, one or two issues	Type D	No	This chapter
Our program is stalled, or there is a safety issue	Type A	No	This chapter
Our design is Bayesian, complex adaptive or simulation-driven	CID paired meeting program; Type C otherwise	No	Chapter 14, Chapter 23
Our dose or design depends on a model	MIDD paired meeting program	No	Chapter 17
A general method, not tied to a product	CPIM	No	This chapter
A biomarker or outcome assessment for many sponsors	DDT qualification	Qualified for stated context of use	This chapter
Is one trial plus confirmatory evidence acceptable?	Type B (EOP) or Type C, before the trial starts	No	Chapter 23
Our master protocol	Type B pre-IND, labeled as a master protocol	No	Chapter 20

TipTry it in the lab

Open Module 13, The regulator’s map (https://nytel.ai/testing/13-regulators-map.html). Before you click, write down which meeting you would request for each of the five scenarios in the lab (a new pill at phase 2, a gene therapy before the first IND, a stalled program, a new kind of trial design, and a dose picked by a computer model). Then pick each scenario and compare. Two scenarios (the new design and the model-picked dose) point to the paired-meeting programs, not to a standard meeting type. Then read the meeting clocks, which are as printed in Tables 1 and 2 of the August 2026 guidance.

3.8 EMA and MHRA counterparts

The European Medicines Agency (EMA) and the UK Medicines and Healthcare products Regulatory Agency (MHRA) apply the same statistical principles, since both adopt ICH guidelines, but their documents on adaptive and complex designs differ in emphasis and date. A sponsor seeking approval in several regions needs to read them together.

EMA adaptive-design reflection paper (2007) (European Medicines Agency, Committee for Medicinal Products for Human Use 2007). The reference paper on methodological issues in confirmatory trials planned with an adaptive design (CHMP/EWP/2459/02) was adopted in October 2007 and is still current until E20 reaches its final step in the EU. It is more conservative than the 2019 FDA guidance: adaptations need justification and seamless phase 2/3 designs are treated with caution.

EC/EMA/HMA Questions and Answers on complex clinical trials (2022) (European Commission and European Medicines Agency and Heads of Medicines Agencies 2022). This final document (EMA/298712/2022, version of 23 May 2022, published in June 2022) is the first European text to deal with complex clinical trials as a class: master protocols, platform, basket and umbrella designs, and trials using Bayesian methods. It is practical rather than statistical, covering the clinical trial application under the EU Clinical Trials Regulation, sub-protocols, transparency, and use in marketing authorization applications. On Bayesian methods it asks sponsors to justify and document the priors, to use prior-predictive simulation, and to engage in early dialogue.

EMA concept paper on Bayesian methods (2026) (European Medicines Agency, Methodology Working Party 2026). The Methodology Working Party agreed a concept paper for a future reflection paper on Bayesian methods in clinical development in November 2025 (EMA/CHMP/1813/2026); the CHMP released it for consultation on 19 January 2026 and the consultation ran from 30 January to 30 April 2026. A concept paper is not guidance: it proposes that a reflection paper be written and lists the questions it should answer, among them what defines technical success for a Bayesian trial, how to assess error control for primary and secondary endpoints without frequentist inference, how much weight the prior carries relative to the data, and which sensitivity analyses are required. The tentative timetable puts a draft reflection paper out for consultation in September 2027 and the final text in June 2028; an EMA workshop on the topic was held in June 2025. Until then a sponsor should read the 2022 Q&A and ICH E9 together.

EMA reflection paper on methodological aspects in confirmatory platform trials (draft, October 2026) (European Medicines Agency 2026). This is a draft (EMA/217509/2026) that follows the 2022 concept paper. The CHMP adopted it for consultation on 5 October 2026, and the consultation runs from 19 October 2026 to 31 January 2027. It concerns confirmatory platform trials only and deliberately avoids the broader term “master protocol”. It says that objectives, estimands and the overall testing strategy must be documented at the start “as far as possible”, warning that “lack of pre-specification could render a confirmatory platform trial as exploratory only”. It prefers concurrent controls, saying that “non-concurrent controls are usually not acceptable for the primary analysis of the primary endpoint”, and it holds the arm-to-control ratio for each comparison fixed, with the consequence stated plainly: “response adaptive randomisation (RAR) is not supported”. It does not discuss Bayesian methods. FDA’s June 2026 draft permits non-concurrent controls in rare cases by agreement, while the EMA draft is stricter on both controls and allocation, so a global platform should treat the stricter position as the design constraint until the paper is final. Chapter 20 and Chapter 16 discuss the consequences.

MHRA draft guideline on external control arms (2025) (Medicines and Healthcare products Regulatory Agency 2025). The MHRA consulted in 2025 on a draft guideline on the use of external control arms based on real-world data to support regulatory decisions. It is relevant here because external controls are a form of borrowing. It parallels FDA’s February 2023 draft guidance on externally controlled trials (U.S. Food and Drug Administration 2023) and is taken up in Chapter 19.

The three regulators share the principles: pre-specification, control of false positive conclusions, simulation, integrity. They differ in how much flexibility they will accept in the primary analysis of a confirmatory trial, and the EMA drafts are the more cautious. A global program should plan European scientific advice in parallel with the US meetings.

3.9 The document table

Table 3.4 collects the documents discussed in this chapter. “Status” is as of November 2026 and the dates are those printed on the documents or in the issuer’s notice. The last column gives the chapters that rely on the document; Appendix Appendix B lists the links.

Table 3.4: Regulatory documents in this book: issuer, date, status and use.

Document	Issuer	Date and status	Used in
FD&C Act 505(d), substantial evidence	US Congress	Law since 1962; amended 1997 (FDAMA)	Chapter 1, Chapter 23
21 CFR 314.126, adequate and well-controlled studies	FDA	Regulation	Chapter 5, Chapter 23
21st Century Cures Act	US Congress	Law, December 2016	This chapter
FDORA	US Congress	Law, December 2022	Chapter 20
ICH E8(R1), General Considerations for Clinical Studies	ICH	Step 4, 6 October 2021	Chapter 5, Chapter 23
ICH E9, Statistical Principles	ICH	Step 4, February 1998	Chapter 4, Chapter 5, Chapter 11
ICH E9(R1), Estimands addendum	ICH	Step 4, 20 November 2019	Chapter 5, Chapter 23
ICH E20, Adaptive Designs	ICH	Draft, Step 2, endorsed 25 June 2025; sign-off targeted February 2027	Chapter 14, Chapter 23
ICH E11A, Pediatric Extrapolation	ICH	Step 4, August 2024	Chapter 19
ICH E6(R3), Good Clinical Practice	ICH	Step 4, January 2025	Chapter 23
ICH M15, Model-Informed Drug Development	ICH (FDA publication)	Step 4, January 2026; FDA June 2026	Chapter 17, Chapter 14
Adaptive Designs for Clinical Trials of Drugs and Biologics	FDA	Final, November 2019	Chapter 11, Chapter 14, Chapter 16
Use of Bayesian Methodology in Clinical Trials	FDA	Draft, January 2026	Chapter 7, Chapter 8, Chapter 14, Chapter 19
Master Protocols for Drug and Biological Product Development	FDA	Revised draft, June 2026 (first draft December 2023)	Chapter 20
Master Protocols: Oncology Drugs and Biologics	FDA	Final, March 2022	Chapter 20
Demonstrating Substantial Evidence of Effectiveness	FDA	Revised draft, June 2026 (first draft December 2019); on finalization replaces the May 1998 guidance	Chapter 23, Chapter 19
ICH E3, Structure and Content of Clinical Study Reports	ICH	Step 4, November 1995	Chapter 23
ICH M4(R4), Common Technical Document	ICH	Step 4, 2016	Chapter 23
Bayesian Statistics in Medical Device Clinical Trials	FDA (CDRH, CBER)	Final, February 2010	Chapter 19, Section 3.11
PDUFA VII commitment letter	FDA and industry	Goals for FY2023 to FY2027	This chapter
Multiple Endpoints in Clinical Trials	FDA	Final, October 2022	Chapter 10
Optimizing the Dosage (Project Optimus)	FDA	Final, August 2024	Chapter 17
Considerations for Externally Controlled Trials	FDA	Draft, February 2023	Chapter 19
Expedited Programs for Serious Conditions	FDA	Final, May 2014	This chapter
Establishment and Operation of Clinical Trial Data Monitoring Committees	FDA	Final, March 2006; revised draft, February 2024	Chapter 23
Considerations for the Use of AI to Support Regulatory Decision-Making	FDA	Draft, January 2025	Chapter 23
General Considerations for the Use of NAMs	FDA	Draft, March 2026	This chapter
Formal Meetings Between FDA and Sponsors (PDUFA)	FDA	Final procedural guidance, August 2026	Chapter 23
Interacting with FDA on Complex Innovative Trial Designs	FDA	Final, December 2020	Chapter 14, Chapter 23
CID, MIDD paired meeting programs; CPIM; DDT qualification	FDA	Active programs; see program pages	Chapter 23, Chapter 17
Reflection paper on adaptive designs (CHMP/EWP/2459/02)	EMA	Adopted October 2007	Chapter 11
Complex clinical trials Q&A (EMA/298712/2022)	EC, EMA, HMA	Final, 2022	Chapter 20, Chapter 23
Concept paper, reflection paper on Bayesian methods (EMA/CHMP/1813/2026)	EMA	Concept paper, consultation January to April 2026	Section 3.8
Reflection paper on confirmatory platform trials (EMA/217509/2026)	EMA	Draft, October 2026; consultation to 31 January 2027	Chapter 20, Chapter 16
External control arms based on real-world data	MHRA	Draft, 2025	Chapter 19

3.10 Reading a regulatory document

Two habits prevent most misuse. Check the date and status on the document itself, and again before submission: FDA media URLs can serve a later revision, and the address of the 2019 substantial-evidence draft now serves the June 2026 revision. And read the context-of-use language of M15 and the AI draft as a rule of proportion: a design whose stopping rule decides approval is highly influential, so its simulation report should be proportionately thorough. None of this replaces the conversation: the review division that examines an application has the last word on what is adequate for that product, and the way to learn its view is to ask, with a concrete design and simulation report in hand.

3.11 In practice

The standard of this chapter has already been met with Bayesian arithmetic, in a vaccine, a biologic approved on one trial plus borrowing, a pediatric extension and two device approvals, and the January 2026 draft names further examples.

REBYOTA: one trial standing in for two

Ferring’s RBX2660 (REBYOTA) for recurrent C. difficile infection was approved on 30 November 2022 (US Food and Drug Administration, Center for Biologics Evaluation and Research 2022) on one phase 3 trial (PUNCH CD3) plus a Bayesian hierarchical model that borrowed from a phase 2b study, an approach FDA called “acceptable” (U.S. Food and Drug Administration, Center for Biologics Evaluation and Research 2022). Success was a posterior probability of superiority with two thresholds chosen as analogues of one-sided type I error without borrowing: 0.9993275 (analogue 0.00125, a stricter bar for one trial standing in for two) and 0.9750338 (analogue 0.025, one adequate and well-controlled trial). The observed posterior probability of superiority was 0.991, which clears the second threshold and misses the first, so the evidence counts as one adequate and well-controlled trial, not two. Approval rested on FDA’s assessment of the totality of evidence, and the thresholds were case-specific evidentiary criteria agreed with FDA, not numerical standards set by the 1997 amendment (Section 3.1). The 2026 draft cites it as its example of borrowing from an earlier trial of the same drug (U.S. Food and Drug Administration 2026e). The planning history, the borrowing model and the data-alignment lesson are in Section 19.17.1.

BNT162b2: a Bayesian criterion in an EUA and a BLA

The pivotal efficacy trial of the Pfizer and BioNTech vaccine (C4591001) had a Bayesian primary criterion written in its statistical analysis plan (Pfizer Inc. 2022): success was a posterior \(\Pr(\mathrm{VE} \ge 30\% \mid \text{data}) > 0.995\) at an interim analysis or \(> 0.986\) at the final analysis, thresholds calibrated so that the overall one-sided type I error is 0.025. The trial observed 8 cases on vaccine against 162 on placebo, vaccine efficacy (VE) 95.0%, 95% credible interval 90.3 to 97.6 (Polack et al. 2020). FDA issued the emergency use authorization on 11 December 2020 for ages 16 and older (U.S. Food and Drug Administration 2021a) and approved the BLA (Comirnaty) on 23 August 2021. The beta-binomial model and the prior are worked in Section 6.12. The lesson is that a pivotal trial can use a posterior criterion calibrated to 0.025 (calibration is one of the three routes in the January 2026 draft, and this book’s default), and that both an EUA and a BLA accepted it.

Belimumab in children: a tipping point on the borrowing weight

PLUTO (C1109), a small trial of belimumab in pediatric lupus, was not powered to stand alone (Brunner et al. 2020). At FDA’s request the sponsor borrowed from the adult trials through a robust mixture prior and reported a tipping point: the posterior probability of efficacy was at least 97.5% whenever the borrowing weight was 0.55 or more, a weight the review division judged reasonable, and FDA approved the pediatric indication (U.S. Food and Drug Administration, Center for Drug Evaluation and Research 2026; Pottackal et al. 2025). The analysis was post hoc, and FDA’s case study recommends prespecifying such borrowing with the agency. The full case is in Section 19.17.2.

Devices: Superion and OPTIMIZER Smart

CDRH has accepted Bayesian primary analyses for years under its 2010 guidance (U.S. Food and Drug Administration 2010). The Superion interspinous spacer (PMA P140004, approved 20 May 2015) was a randomized 1:1 non-inferiority trial against X-STOP with a Bayesian adaptive sample size, a 10% margin, independent Beta(1, 1) priors, and success declared if the posterior probability of non-inferiority was at least 0.958. The final analysis gave 0.9927 (U.S. Food and Drug Administration, CDRH 2015). The OPTIMIZER Smart system (P180036, approved 21 March 2019) analyzed FIX-HF-5C (\(n = 160\)) with a Bayesian model and 30% fixed borrowing from the earlier FIX-HF-5 study, with success if \(\Pr(\Delta > 0) > 0.975\) for the change in peak VO2. With borrowing the posterior probability was 0.977 to 0.989 across imputation models; the new trial alone gave 0.911 to 0.960 (the lower value when last-observation-carried-forward imputation is included), below the bar (U.S. Food and Drug Administration, Center for Devices and Radiological Health 2019). The lesson is that the borrowing weight can decide the outcome, which is why the draft asks for justification and sensitivity analysis.

The other precedents in the draft

The draft names the GBM AGILE (NCT03970447) and Precision Promise (NCT04229004) platforms, which use Bayesian models for temporal drift when augmenting concurrent controls (Saville et al. 2022); a pediatric multiple sclerosis non-inferiority study through the CID program, with a prespecified Bayesian analysis borrowing historical active-comparator data and modeling relapse rate by age (U.S. Food and Drug Administration, CDER, n.d.b); supportive Bayesian analyses for empagliflozin and linagliptin in pediatric type 2 diabetes, accepted because the pediatric pathophysiology resembles the adult; an epilepsy with myoclonic-atonic seizures (EMAS) placebo-controlled study proposed through the CID program, with a hierarchical model borrowing from other epilepsy types (U.S. Food and Drug Administration, CDER, n.d.a); and hierarchical subgroup borrowing in the LEADER trial of liraglutide and the Rinvoq Drug Trial Snapshot (U.S. Food and Drug Administration 2026e). For the last three we rely on the draft’s description.

Key points

• The legal standard is substantial evidence of effectiveness: adequate and well-controlled investigations (FD&C Act 505(d)), characterized by 21 CFR 314.126. The statute is a standard of evidence, not of method; it names no test, level or prior.
• Since FDAMA 1997, one adequate and well-controlled trial plus confirmatory evidence can suffice at FDA’s discretion. The June 2026 revised draft says this calls for a highly persuasive trial or strong confirmatory evidence, and that the plan should be discussed with FDA before the trial.
• A traditional submission rests on pre-specification and error control: a protocol with a fixed sample size, a SAP, a one-sided 0.025 level (two-sided 0.05), a stated multiplicity strategy, and a DMC for interim looks. Bayesian adaptive designs keep all of that and add a simulation report and a justified prior. For confirmatory adaptive designs without borrowing, this book uses calibration of the success criterion to one-sided 0.025; FDA’s draft also describes direct posterior interpretation and decision-theoretic or benefit-risk criteria (Section IV.A).
• The Bayesian draft (January 2026) states success as \(\Pr(\delta > a \mid \text{data}) > c\) and describes three routes, by agreement with FDA: calibration to a familywise error rate (typically 0.025 one-sided), direct posterior interpretation, and decision-theoretic or benefit-risk criteria. Calibration is this book’s default, not the draft’s. Against a single unadjusted analysis the calibrated bar is not lower without an informative prior.
• Documents differ in force: statutes and regulations bind; guidances recommend; drafts are subject to change; permissions depend on FDA’s agreement. Never describe a draft as a requirement.
• The statuses that matter as of November 2026: Adaptive Designs final (2019); Bayesian, Master Protocols and Substantial Evidence drafts (2026); ICH E20 at Step 2; ICH M15 at Step 4; EMA platform draft in consultation to 31 January 2027.
• Real approvals already use the arithmetic: REBYOTA’s two posterior thresholds (0.9993275 for one trial standing in for two, 0.9750338 for one adequate and well-controlled trial; observed 0.991), BNT162b2’s calibrated posterior criterion, belimumab’s tipping-point analysis, and CDRH’s Superion and OPTIMIZER Smart approvals (Section 3.11).
• FDA offers Type A, B, B (EOP), C, D and INTERACT meetings, the CID and MIDD paired meeting programs, CPIM, and DDT qualification. A novel design should go to FDA early, with simulations in hand.
• EMA’s drafts are more cautious than FDA’s on allocation and nonconcurrent controls; a global platform should design to the stricter position.

NoteDiscussion questions

1. The statute asks whether experts “could fairly and responsibly” conclude that the drug works, and does not mention a significance level. Why has 0.025 one-sided become so entrenched, and what would have to be true for a regulator to accept a different working standard in a given program?
2. The June 2026 substantial-evidence draft allows one trial with confirmatory evidence. In what ways is an informative Bayesian prior similar to confirmatory evidence, and in what ways is it different? Who decides the weight, and when?
3. Suppose your sponsor is considering a Bayesian adaptive phase 3 design with response-adaptive randomization across three arms, to be submitted in both the US and the EU. Using the FDA master protocol draft and the EMA platform draft, which design elements would you keep, and which would you change?
4. A colleague says “FDA requires a simulation report for any Bayesian design”. Is that accurate? Classify the statement using the requires, recommends, permits distinction, and name the document you would cite.
5. Pick one of the six formal meeting types and one of the programs (CID, MIDD, CPIM, DDT). For a trial of your own, say which you would request first, what you would bring, and what answer you would need to proceed.

Bauer, Peter, Frank Bretz, Vladimir Dragalin, Franz König, and Gernot Wassmer. 2016. “Twenty-Five Years of Confirmatory Adaptive Designs: Opportunities and Pitfalls.” Statistics in Medicine 35 (3): 325–47. https://doi.org/10.1002/sim.6472.

Brunner, Hermine I, Carlos Abud-Mendoza, Diego O Viola, Inmaculada Calvo Penades, Deborah Levy, Jordi Anton, Julia E Calderon, et al. 2020. “Safety and Efficacy of Intravenous Belimumab in Children with Systemic Lupus Erythematosus: Results from a Randomised, Placebo-Controlled Trial.” Annals of the Rheumatic Diseases 79 (10): 1340–48. https://doi.org/10.1136/annrheumdis-2020-217101.

Campbell, Gregory. 2011. “Bayesian Statistics in Medical Devices: Innovation Sparked by the FDA.” Journal of Biopharmaceutical Statistics 21 (5): 871–87. https://doi.org/10.1080/10543406.2011.589638.

———. 2013. “Similarities and Differences of Bayesian Designs and Adaptive Designs for Medical Devices: A Regulatory View.” Statistics in Biopharmaceutical Research 5 (4): 356–68. https://doi.org/10.1080/19466315.2013.846873.

Campbell, Gregory, Telba Irony, Gene Pennello, and Laura Thompson. 2023. “Bayesian Statistics for Medical Devices: Progress Since 2010.” Therapeutic Innovation & Regulatory Science 57 (3): 453–63. https://doi.org/10.1007/s43441-022-00495-w.

European Commission and European Medicines Agency and Heads of Medicines Agencies. 2022. “Complex Clinical Trials: Questions and Answers.” https://health.ec.europa.eu/system/files/2022-06/medicinal_qa_complex_clinical-trials_en.pdf.

European Medicines Agency. 2026. “Reflection Paper on Methodological Aspects in Confirmatory Platform Trials.” https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-methodological-aspects-confirmatory-platform-trials_en.pdf.

European Medicines Agency, Committee for Medicinal Products for Human Use. 2007. “Reflection Paper on Methodological Issues in Confirmatory Clinical Trials Planned with an Adaptive Design (CHMP/EWP/2459/02).” European Medicines Agency. https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-methodological-issues-confirmatory-clinical-trials-planned-adaptive-design_en.pdf.

European Medicines Agency, Methodology Working Party. 2026. “Concept Paper for the Development of a Reflection Paper on the Use of Bayesian Methods in Clinical Development.” https://www.ema.europa.eu/system/files/documents/scientific-guideline/concept-paper-development-reflection-paper-use-bayesian-methods-en.pdf.

Evans, Scott R., Thomas R. Fleming, Holly Janes, and Lori E. Dodd. 2026. “Reflections on FDA Draft Guidance on Bayesian Methods in Trials: Protecting Scientific Integrity and Evidentiary Standards.” JAMA 335 (19): 1659. https://doi.org/10.1001/jama.2026.4175.

Gelman, Andrew, Erik van Zwet, and Witold Więcek. 2026. “FDA Draft Guidance for the Use of Bayesian Methods in Clinical Trials.” JAMA 335 (19): 1662. https://doi.org/10.1001/jama.2026.4178.

International Council for Harmonisation. 1995. “ICH Harmonised Tripartite Guideline: Structure and Content of Clinical Study Reports E3.” https://database.ich.org/sites/default/files/E3_Guideline.pdf.

———. 1998. “ICH Harmonised Tripartite Guideline: Statistical Principles for Clinical Trials E9.” https://database.ich.org/sites/default/files/E9_Guideline.pdf.

———. 2016. “ICH M4: Organisation of the Common Technical Document for the Registration of Pharmaceuticals for Human Use.” https://www.ich.org/page/ctd.

———. 2019. “ICH E9(R1): Addendum on Estimands and Sensitivity Analysis in Clinical Trials to the Guideline on Statistical Principles for Clinical Trials.” https://database.ich.org/sites/default/files/E9-R1_Step4_Guideline_2019_1203.pdf.

———. 2021. “ICH E8(R1): General Considerations for Clinical Studies.” https://database.ich.org/sites/default/files/ICH_E8-R1_Guideline_Step4_2021_1006.pdf.

———. 2024. “ICH E11A: Pediatric Extrapolation.” https://database.ich.org/sites/default/files/ICH_E11A_Guideline_Step4_2024_0821.pdf.

———. 2025a. “E20 Adaptive Designs for Clinical Trials.” https://www.ema.europa.eu/en/ich-e20-adaptive-designs-clinical-trials-scientific-guideline.

———. 2025b. “ICH E6(R3): Guideline for Good Clinical Practice (GCP).” https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf.

———. 2026. “ICH M15: General Principles for Model-Informed Drug Development.” https://www.fda.gov/media/184747/download.

Lee, J. Jack, Frank E. Harrell, Lisa M. LaVange, and David J. Spiegelhalter. 2026. “Embracing Bayesian Methods in Clinical Trials.” JAMA 335 (19): 1664. https://doi.org/10.1001/jama.2026.4179.

Medicines and Healthcare products Regulatory Agency. 2025. “MHRA Draft Guideline on the Use of External Control Arms Based on Real-World Data to Support Regulatory Decisions.” https://www.gov.uk/government/consultations/mhra-draft-guideline-on-the-use-of-external-control-arms-based-on-real-world-data-to-support-regulatory-decisions.

Pennello, Gene, and Laura Thompson. 2008. “Experience with Reviewing Bayesian Medical Device Trials.” Journal of Biopharmaceutical Statistics 18 (1): 81–115. https://doi.org/10.1080/10543400701668274.

Pfizer Inc. 2022. “Protocol C4591001 (PF-07302048) Statistical Analysis Plan.” ClinicalTrials.gov NCT04368728, SAP_001, version 10. https://cdn.clinicaltrials.gov/large-docs/28/NCT04368728/SAP_001.pdf.

Polack, Fernando P., Stephen J. Thomas, Nicholas Kitchin, et al. 2020. “Safety and Efficacy of the BNT162b2 mRNA Covid-19 Vaccine.” New England Journal of Medicine. https://doi.org/10.1056/NEJMoa2034577.

Pottackal, Ginto, James Travis, Lei Nie, Rebecca Chiu, Rosemarie Neuner, Gregory Levin, Jing Niu, Anshu Marathe, Lily (Yeruk) Mulugeta, and Nikolay P. Nikolov. 2025. “Application of Bayesian Statistics to Support Approval of Intravenous Belimumab in Children with Systemic Lupus Erythematosus in the United States.” Lupus 34 (9): 943–52. https://doi.org/10.1177/09612033251349930.

Ruberg, Stephen J., Francois Beckers, Rob Hemmings, Peter Honig, Telba Irony, Lisa LaVange, Grazyna Lieberman, et al. 2023. “Application of Bayesian Approaches in Drug Development: Starting a Virtuous Cycle.” Nature Reviews Drug Discovery 22 (3): 235–50. https://doi.org/10.1038/s41573-023-00638-0.

Saville, Benjamin R., Donald A. Berry, Nicholas S. Berry, Kert Viele, and Scott M. Berry. 2022. “The Bayesian Time Machine: Accounting for Temporal Drift in Multi-Arm Platform Trials.” Clinical Trials 19 (5): 490–501. https://doi.org/10.1177/17407745221112013.

United States Congress. 1962. “Federal Food, Drug, and Cosmetic Act, Section 505(d) (21 u.s.c. 355(d)): Grounds for Refusing to Approve an Application.” https://www.govinfo.gov/content/pkg/USCODE-2023-title21/html/USCODE-2023-title21-chap9-subchapV-partA-sec355.htm.

———. 1997. “Food and Drug Administration Modernization Act of 1997.”

———. 2016. “21st Century Cures Act.” https://www.congress.gov/bill/114th-congress/house-bill/34.

U.S. Congress. 2022. “Food and Drug Omnibus Reform Act (FDORA) of 2022.” https://www.fda.gov/regulatory-information/selected-amendments-fdc-act/food-and-drug-omnibus-reform-act-fdora-2022.

U.S. Food and Drug Administration. 2006. “Guidance for Clinical Trial Sponsors: Establishment and Operation of Clinical Trial Data Monitoring Committees.” https://www.fda.gov/media/75398/download.

———. 2010. “Guidance for the Use of Bayesian Statistics in Medical Device Clinical Trials.” U.S. Food and Drug Administration. https://www.fda.gov/media/71512/download.

———. 2014. “Expedited Programs for Serious Conditions – Drugs and Biologics: Guidance for Industry.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/expedited-programs-serious-conditions-drugs-and-biologics.

———. 2019a. “Adaptive Designs for Clinical Trials of Drugs and Biologics: Guidance for Industry.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/adaptive-design-clinical-trials-drugs-and-biologics-guidance-industry.

———. 2019b. “Demonstrating Substantial Evidence of Effectiveness for Human Drug and Biological Products.” U.S. Food and Drug Administration. https://www.fda.gov/media/133660/download.

———. 2020. “Interacting with the FDA on Complex Innovative Trial Designs for Drugs and Biological Products: Guidance for Industry.” U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/interacting-fda-complex-innovative-trial-designs-drugs-and-biological-products.

———. 2021a. “Emergency Use Authorization Letter, Pfizer-BioNTech COVID-19 Vaccine (Reissued 10 May 2021; Original EUA 11 December 2020).” https://www.fda.gov/media/144412/download.

———. 2021b. “PDUFA Reauthorization Performance Goals and Procedures, Fiscal Years 2023 Through 2027.” https://www.fda.gov/media/151712/download.

———. 2022a. “Master Protocols: Efficient Clinical Trial Design Strategies to Expedite Development of Oncology Drugs and Biologics.” U.S. Food and Drug Administration. https://www.fda.gov/media/120721/download.

———. 2022b. “Multiple Endpoints in Clinical Trials: Guidance for Industry.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/multiple-endpoints-clinical-trials-guidance-industry.

———. 2023. “Considerations for the Design and Conduct of Externally Controlled Trials for Drug and Biological Products.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-design-and-conduct-externally-controlled-trials-drug-and-biological-products.

———. 2024a. “21 CFR 314.126: Adequate and Well-Controlled Studies.” https://www.ecfr.gov/current/title-21/chapter-I/subchapter-D/part-314/subpart-D/section-314.126.

———. 2024b. “Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases: Guidance for Industry.” Final guidance, Oncology Center of Excellence, CDER and CBER. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/optimizing-dosage-human-prescription-drugs-and-biological-products-treatment-oncologic-diseases.

———. 2025. “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological.

———. 2026a. “Demonstrating Substantial Evidence of Effectiveness for Human Drug and Biological Products.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/demonstrating-substantial-evidence-effectiveness-human-drug-and-biological-products.

———. 2026b. “Formal Meetings Between the FDA and Sponsors or Applicants of PDUFA Products: Guidance for Industry (Procedural).” https://www.fda.gov/media/172311/download.

———. 2026c. “General Considerations for the Use of New Approach Methodologies in Drug Development: Guidance for Industry.” https://www.fda.gov/media/191589/download.

———. 2026d. “Master Protocols for Drug and Biological Product Development: Guidance for Industry.” https://www.fda.gov/regulatory-information/search-fda-guidance-documents/master-protocols-drug-and-biological-product-development.

———. 2026e. “Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products.” U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/use-bayesian-methodology-clinical-trials-drug-and-biological-products.

———. n.d.a. “Complex Innovative Trial Design Meeting Program.” https://www.fda.gov/drugs/development-resources/complex-innovative-trial-design-meeting-program.

———. n.d.b. “Critical Path Innovation Meetings (CPIM).” https://www.fda.gov/drugs/novel-drug-approvals-fda/critical-path-innovation-meetings-cpim.

———. n.d.c. “Drug Development Tool (DDT) Qualification Programs.” https://www.fda.gov/drugs/development-approval-process-drugs/drug-development-tool-ddt-qualification-programs.

———. n.d.d. “Model-Informed Drug Development Paired Meeting Program.” https://www.fda.gov/drugs/development-resources/model-informed-drug-development-paired-meeting-program.

U.S. Food and Drug Administration, CDER. n.d.a. “Complex Innovative Trial Design Case Study: Epilepsy with Myoclonic-Atonic Seizures (EMAS) Study with a Bayesian Hierarchical Model.” https://www.fda.gov/media/172312/download.

———. n.d.b. “Complex Innovative Trial Design Case Study: Pediatric Multiple Sclerosis Non-Inferiority Study with a Bayesian Analysis.” https://www.fda.gov/media/172313/download.

U.S. Food and Drug Administration, CDRH. 2015. “Summary of Safety and Effectiveness Data, Superion InterSpinous Spacer, PMA P140004.” https://www.accessdata.fda.gov/cdrh_docs/pdf14/P140004B.pdf.

U.S. Food and Drug Administration, Center for Biologics Evaluation and Research. 2022. “FDA Briefing Document, Vaccines and Related Biological Products Advisory Committee Meeting, September 22, 2022: RBX2660 (Fecal Microbiota Suspension for Rectal Enema), Rebiotix, Inc., BLA STN 125739.” https://www.fda.gov/media/161679/download.

U.S. Food and Drug Administration, Center for Devices and Radiological Health. 2019. “Summary of Safety and Effectiveness Data, OPTIMIZER Smart System, PMA P180036.” https://www.accessdata.fda.gov/cdrh_docs/pdf18/P180036B.pdf.

U.S. Food and Drug Administration, Center for Drug Evaluation and Research. 2026. “LEADER 3D: Bayesian Statistics in Rare Disease Clinical Investigations, Belimumab (Benlysta).” https://www.fda.gov/media/193802/download.

US Food and Drug Administration, Center for Biologics Evaluation and Research. 2022. “Clinical Review Memorandum: REBYOTA (Fecal Microbiota, Live-jslm), BLA 125739.” US Food and Drug Administration. https://www.fda.gov/media/164040/download.