FDA Guidance on Adaptive Trial Designs

Here, \(N_k\) represents the information level or sample size associated with interim look \(k\).

The key regulatory issue is that the repeated opportunities for decision-making must be incorporated into the statistical design.

2. Sample-Size Adaptation

A sample-size adaptation changes the planned sample size based on accumulating information.

For example, a trial could have:

• Minimum sample size = 300
• Maximum sample size = 500
• One interim assessment after 200 participants
• A prespecified rule that permits increasing the final sample size when information is insufficient

The purpose is not to “look at the p-value and add patients until it becomes significant.” The adaptation should be governed by a rule whose statistical properties have been evaluated before the trial begins.

3. Adaptive Enrichment

An adaptive enrichment design changes the population being enrolled based on accumulating information.

For example, suppose patients are eligible across biomarker-defined groups A and B. Interim evidence might indicate that the treatment effect is concentrated in group A. A prospectively specified design could allow subsequent enrollment to focus on group A.

This can be particularly useful when treatment effect heterogeneity is scientifically plausible but the appropriate target population is uncertain before the trial starts.

4. Treatment-Arm Selection

A multi-arm study can prospectively allow less promising doses or treatment arms to be dropped.

For example:

The exact selection rule must be specified and its operating characteristics evaluated.

5. Patient Allocation

Some designs modify randomization probabilities based on accumulating data. These approaches can have ethical or efficiency motivations, but they can also introduce operational and statistical complexity.

If investigators can predict which arm will receive more future patients based on interim results, knowledge of the adaptation can influence behavior. Maintaining appropriate confidentiality and operational separation therefore becomes especially important.

6. Endpoint Adaptation

FDA also discusses adaptations involving endpoint selection. These designs require particular care because changing the primary endpoint after seeing comparative results can directly affect the credibility of the confirmatory claim.

If endpoint adaptation is contemplated, the candidate endpoints, timing, selection rule, and final inferential procedure should be defined in advance.

Multiple Adaptive Features

Some innovative trials combine several adaptations:

• Drop an ineffective treatment arm.
• Change the sample size.
• Enrich the population.
• Modify randomization probabilities.
• Stop early for efficacy or futility.

Combining adaptations can produce a powerful and efficient design, but the number of possible decision paths grows rapidly.

Simulation Is Often Central to Adaptive Design

For a simple fixed design, operating characteristics can sometimes be derived analytically. Adaptive designs frequently require simulation because the sample size, treatment allocation, stopping decision, or analysis method can depend on interim results.

A generic simulation proceeds as follows:

Type I Error Simulation

To evaluate Type I error, simulations are run under scenarios in which the null hypothesis is true.

If \(M\) simulated trials are generated and \(R\) incorrectly declare success, the estimated Type I error is:

For example, if 9,870 of 400,000 simulated null trials reject the null:

The simulation estimate itself has Monte Carlo uncertainty, so the number of simulations should be sufficiently large for the intended precision.

Power Simulation

Power is evaluated under clinically relevant alternative scenarios.

A strong simulation program does not evaluate only one optimistic alternative. It typically explores a range of plausible treatment effects and nuisance parameters.

Simulation Scenarios

What FDA Expects From a Simulation Report

FDA's adaptive-design guidance recommends substantial detail when simulations are the primary or sole technique used to evaluate operating characteristics. The guidance identifies an overall design description, example hypothetical trials with different conclusions, parameter configurations and their justification, the number of simulated trials, and detailed operating characteristics among the important elements. citeturn1search21

Example: Adaptive Sample-Size Reassessment

Consider a two-arm trial with a binary endpoint. The sponsor anticipates an absolute treatment difference of 10 percentage points but is uncertain about the control event rate and variability.

A simplified adaptive concept might be:

The point is not to make the decision after seeing whether the treatment is “significant.” The adaptation should be linked to the statistical information needed to answer the study question.

Conditional Power

Conditional power is one possible concept in adaptive sample-size planning. It asks how likely the trial is to meet its success criterion given the information observed so far and specified assumptions about future data.

Conditional-power-based rules must be defined carefully because different definitions and assumptions can lead to very different adaptation behavior. The operating characteristics of the complete rule—not just the conditional power calculation—must be evaluated.

Adaptive Enrichment Example

Suppose a trial enrolls biomarker-positive and biomarker-negative patients. The treatment effect may differ between the two groups.

An adaptive enrichment design could allow a prespecified interim decision to continue both populations or concentrate future enrollment in one population.

Important questions include:

• What evidence is sufficient to change the enrollment population?
• Does the design preserve the intended Type I error?
• What happens if the subgroup effect is uncertain?
• How will the treatment effect be estimated after enrichment?
• What happens to patients enrolled before the population changes?

Adaptive Dose Selection

Dose-selection adaptations are common candidates for adaptive designs because early development often involves uncertainty about the dose-response relationship.

For example:

The design can use interim information to discontinue doses that are sufficiently unpromising while retaining more informative doses.

The challenge is ensuring that the selection process does not create misleading final estimates or inflate the chance of an erroneous conclusion.

Bayesian Adaptive Designs

FDA's guidance specifically addresses Bayesian adaptive designs. Bayesian methods can be useful when the design naturally incorporates posterior probabilities, predictive quantities, or other Bayesian decision criteria.

A generic Bayesian update can be written as:

where \(p(\theta)\) is the prior distribution, \(p(D\mid\theta)\) is the likelihood, and \(p(\theta\mid D)\) is the posterior distribution.

A Bayesian adaptive design might use a posterior or predictive probability to decide whether to continue, stop, change enrollment, or modify treatment allocation.

Prior Distributions Need Justification

When prior information affects an adaptive decision, the sponsor should be able to explain:

• Where the prior information came from.
• Why it is relevant to the current population.
• How sensitive conclusions are to alternative priors.
• How much the prior can influence adaptation decisions.
• How the prior affects operating characteristics.

Time-to-Event Adaptive Designs

Adaptive designs become particularly interesting when the primary endpoint is time to an event such as progression or death.

In these settings, the relevant information may be the number of observed events rather than the number of enrolled patients.

This creates an operational challenge: the interim analysis may be triggered by an event count, while enrollment continues and event rates can differ from the assumptions used during planning.

Intermediate Endpoints

An adaptive decision may be based on an intermediate endpoint while the final trial objective is based on a later or more definitive endpoint.

Examples might include:

• Early biomarker response used to guide enrollment.
• Short-term efficacy used for dose selection.
• Intermediate event information used to assess whether continued enrollment is justified.

The relationship between the intermediate endpoint and the final endpoint should be scientifically justified and reflected in the simulation strategy.

Safety Adaptations

Adaptive designs must also account for safety. A design that is statistically efficient for efficacy can be inappropriate if safety information is too immature or if the adaptation exposes participants to unacceptable risk.

The adaptive framework should therefore address:

• What safety information is reviewed.
• Who reviews it.
• Whether the review is comparative.
• What safety stopping rules apply.
• How safety decisions interact with efficacy adaptations.

Maintaining Trial Integrity

Trial integrity is one of the most important themes in FDA's guidance. Adaptive designs create opportunities for operational bias if people who can influence trial conduct learn comparative interim results.

Potential biases include:

• Selection bias
• Investigator behavior changes
• Enrollment changes
• Changes in patient or site behavior
• Changes in concomitant treatment or assessment practices
• Unintentional disclosure of treatment differences

The goal is to allow the adaptation to occur while limiting unnecessary access to comparative information.

Firewalls and Independent Decision Functions

Depending on the design, a Data Monitoring Committee or dedicated adaptation committee may receive interim information and implement the prespecified decision rule.

FDA emphasizes the importance of limiting access to comparative interim results when such access could affect trial conduct. The exact operational structure depends on the design, but the statistical algorithm and responsibilities should be clearly defined. citeturn1view0

Why Blinding Alone Is Not the Whole Story

FDA's guidance deliberately distinguishes between comparative and non-comparative interim analyses rather than relying exclusively on the words “blinded” and “unblinded.” An analysis can be comparative even if individual treatment assignments are coded as A and B.

The important question is whether the analysis uses information that distinguishes treatment groups in a way that can influence the adaptation.

Non-Comparative Adaptations

Some adaptations can be based on information that does not require treatment comparisons. For example, a blinded sample-size reassessment may use pooled variance information without examining the treatment difference.

Such adaptations can sometimes be operationally easier to protect from bias, although the exact statistical and regulatory properties still need to be evaluated.

Comparative Adaptations

Comparative adaptations use treatment-group information to determine the modification.

Examples include:

• Dropping an inferior treatment arm.
• Enriching a subgroup with a larger apparent treatment effect.
• Changing randomization probabilities based on comparative outcomes.
• Stopping for efficacy or futility based on between-group evidence.

These adaptations require particularly careful control of access to interim comparative results.

Unplanned Design Changes

FDA separately discusses design changes that were not prospectively planned. These are not equivalent to a prespecified adaptive design.

An unplanned change may sometimes be unavoidable—for example, because of an unexpected operational or scientific issue—but it requires a different regulatory and statistical assessment.

Changes Based on External Information

FDA also discusses adaptations based on information external to the trial. Examples can include emerging evidence from other studies or changes in the external clinical environment.

External information creates additional questions:

• Was the external source prespecified?
• How reliable and relevant is it?
• Who has access to it?
• Could it reveal the comparative performance of the current trial?
• How does the external information affect the final inference?

Adaptive Designs in Exploratory Trials

The regulatory burden and statistical objectives may differ between exploratory trials and trials intended to provide substantial evidence of effectiveness. FDA's guidance recognizes this distinction.

Exploratory studies can often tolerate greater flexibility because their goal may be learning rather than providing the definitive evidence for a regulatory claim.

However, exploratory does not mean “anything goes.” The design should still protect participants, maintain scientific credibility, and clearly distinguish exploratory findings from confirmatory evidence.

Estimating Treatment Effects After Adaptation

A frequent mistake is to focus on the decision rule but neglect estimation.

Suppose an adaptive trial selects the most promising treatment arm. The observed treatment effect among selected arms can be biased upward because the arm was selected precisely because it looked favorable.

Therefore, the final analysis should consider the adaptive selection mechanism.

FDA explicitly includes bias and confidence-interval coverage among the operating characteristics that may need evaluation. citeturn1search21

Selection Bias in Treatment-Arm Selection

Imagine three doses with true effects:

Even when the true effects are identical, random sampling variation will cause one dose to look best at an interim analysis.

If that apparently best dose is selected, its observed interim effect will tend to be larger than its true effect.

This is a classic winner's-curse problem and is one reason why adaptive selection designs require careful simulation and estimation planning.

Confidence-Interval Coverage

Suppose a nominal 95% confidence interval is reported:

Under an adaptive design, the usual formula may not automatically have 95% coverage if the selection or stopping process has been ignored.

The relevant question is whether the complete adaptive procedure produces the intended coverage.

Multiplicity

Adaptive trials can involve multiplicity from:

• Multiple treatment arms
• Multiple interim analyses
• Multiple populations
• Multiple endpoints
• Multiple doses
• Multiple adaptive pathways

Multiplicity is not solved simply by calling a design adaptive. The final inferential framework must account for the opportunities to make an erroneous claim.

Adaptive Designs and the Null Hypothesis

Consider a trial with multiple doses:

If the design selects one dose based on interim evidence and then tests it, the complete selection-and-testing process determines the Type I error—not just the final test applied to the selected dose.

Protocol Specification

An adaptive protocol should be explicit enough that an independent reader can understand the decision process without reconstructing it from programming code.

Statistical Analysis Plan

The SAP should go beyond saying that an adaptive design will be used. It should define how the data are transformed into interim decisions and how the final analysis is performed.

FDA's guidance recommends prespecification of statistical methods used to produce interim results, guide adaptation decisions, conduct hypothesis tests, estimate treatment effects, and quantify uncertainty. FDA also recommends prespecifying software used for interim and final analyses and providing sufficient information about novel or custom software when applicable. citeturn1search21

Programming Considerations

For statistical programmers, an adaptive trial introduces an additional validation layer. The programming task is not simply to produce a final TLF. It is to implement a decision algorithm whose behavior may alter the future trial.

At minimum, programming specifications should define:

• Input datasets and data cuts
• Interim-analysis population
• Endpoint derivations
• Eligibility rules for adaptation
• Decision statistics
• Thresholds
• Randomization updates, if applicable
• Stopping rules
• Sample-size rules
• Final-analysis rules
• Simulation code and random seeds
• Independent validation procedures

Illustrative Adaptive Algorithm

if interim_information >= required_information:

    calculate interim statistic

    if efficacy_boundary_crossed:
        stop_for_success()

    else if futility_boundary_crossed:
        stop_for_futility()

    else if sample_size_rule_triggered:
        increase_sample_size(
            to = prespecified_value
        )

    else:
        continue_as_planned()

This pseudocode is intentionally generic. A real trial requires a complete statistical definition, including treatment-effect estimation, missing-data handling, multiplicity, and final inference.

Simulation in R: Skeleton

set.seed(20260911)

simulate_trial <- function(
  true_effect,
  n_initial = 200,
  n_max = 400
) {

  ## 1. Generate initial data
  ## 2. Perform interim analysis
  ## 3. Apply the prespecified adaptation
  ## 4. Generate additional observations if required
  ## 5. Perform final analysis
  ## 6. Return success, sample size, estimate, CI, etc.

  list(
    success = NA,
    n_total = NA,
    estimate = NA_real_,
    lower = NA_real_,
    upper = NA_real_
  )
}

results <- replicate(
  10000,
  simulate_trial(
    true_effect = 0
  ),
  simplify = FALSE
)

The important point is that the simulation should reproduce the actual trial algorithm, not a simplified approximation that omits the adaptive decision.

Scenario Grid in R

scenarios <- expand.grid(
  true_effect = c(0, 0.05, 0.10, 0.15),
  sd = c(1.0, 1.25, 1.5),
  accrual_rate = c(5, 10, 20)
)

scenario_results <- lapply(
  seq_len(nrow(scenarios)),
  function(i) {

    s <- scenarios[i, ]

    ## Run many simulated trials under scenario s

    data.frame(
      true_effect = s$true_effect,
      sd = s$sd,
      accrual_rate = s$accrual_rate,
      type1_or_power = NA_real_,
      expected_n = NA_real_
    )
  }
)

Why Random Seeds Matter

Simulation results should be reproducible. Record:

• Random-number seed
• Simulation code version
• Software version
• Number of iterations
• Scenario definitions
• Input assumptions

For a regulatory submission, reproducibility is especially important because the simulation may be part of the evidence supporting the trial design.

Adaptive Trial Documentation Package

A strong submission package should make it possible for FDA reviewers to understand the proposed design, reproduce its operating-characteristic evaluation, and determine how the adaptive procedure affects inference.

Interacting With FDA Early

Adaptive designs can benefit from early regulatory interaction because seemingly small design choices can have major statistical consequences.

FDA has a dedicated Complex Innovative Trial Design (CID) meeting program for eligible innovative designs where simulation may be necessary to establish operating characteristics. FDA's current program information requests a description of the design, adaptive features, analysis plan, decision criteria, simulation scenarios, preliminary operating characteristics, and specific questions for discussion. citeturn1search1

This is particularly useful when the design combines multiple complex features or when analytical derivation of operating characteristics is difficult.

What to Bring to an FDA Discussion

2025 ICH E20 Draft: Why It Matters

FDA published a draft guidance entitled E20 Adaptive Designs for Clinical Trials in September 2025 under the ICH process. FDA describes the draft as a transparent and harmonized set of recommendations focused on planning, conduct, analysis, and interpretation of adaptive clinical trials intended to confirm efficacy and support benefit-risk assessment. citeturn0search2

Because the document is currently identified by FDA as a draft and “Not for Implementation,” it should be distinguished from the final 2019 FDA guidance. It is nevertheless important to monitor because it represents movement toward international harmonization of adaptive-design principles.

Adaptive Designs for Medical Devices Are Different Guidance

FDA also has a separate 2016 guidance for adaptive designs in medical-device clinical studies. That guidance addresses device clinical investigations and has its own scope and regulatory context. citeturn0search1

Therefore, a drug or biologic development team should not automatically use the medical-device guidance as its primary regulatory reference.

Example End-to-End Adaptive Design

Consider a hypothetical randomized trial with two experimental doses and a control.

The design would then be evaluated under each plausible scenario before the trial begins.

Decision Tree for Reviewing an Adaptive Proposal

Common Mistakes

1. Calling an unplanned protocol change an adaptive design. Adaptive features should generally be planned before comparative interim information is examined.
2. Optimizing only for power. Type I error, bias, confidence-interval coverage, sample size, duration, and other operating characteristics matter.
3. Using a single simulation scenario. Adaptive designs should be stress-tested across a range of plausible assumptions.
4. Ignoring selection bias. Selecting the best-looking treatment can bias naive treatment-effect estimates.
5. Ignoring interim-data confidentiality. Operational access to comparative results can influence enrollment and trial conduct.
6. Leaving the adaptation rule vague. “The sample size may be adjusted based on interim data” is not an adequate statistical specification by itself.
7. Failing to define the final analysis. The final estimator and confidence interval must reflect the adaptive design.
8. Using simulation code that does not reproduce the actual design. Every decision path should be represented.
9. Forgetting operational feasibility. An adaptation that occurs after all patients are enrolled is not practically useful.
10. Overcomplicating the design. Every additional adaptive feature creates additional implementation and validation burden.

Adaptive Design Quality-Control Checklist

How FDA Guidance Translates Into a Statistical Programmer's Workflow

What Should Be Reported After the Trial?

A completed adaptive trial should clearly document what happened—not merely what was planned.

Useful reporting elements include:

• Actual timing of interim analyses.
• Actual information available at each interim analysis.
• Whether an adaptation occurred.
• The rule that triggered the adaptation.
• Actual sample size before and after adaptation.
• Treatment arms or populations retained or dropped.
• Stopping decisions and timing.
• Final analysis method.
• Simulation-based expectations compared with observed behavior where informative.
• Any deviations from the prespecified adaptive algorithm.

Regulatory Interpretation: Guidance Is Not a Regulation

FDA states that guidance documents generally represent the Agency's current thinking and do not establish legally enforceable responsibilities unless specific statutory or regulatory requirements are cited. The 2019 adaptive design guidance itself is labeled as containing nonbinding recommendations. citeturn1view0

This distinction matters when translating the guidance into a study protocol. A sponsor should understand which elements are regulatory requirements and which are FDA recommendations reflecting current statistical and scientific thinking.

Reference Implementation Pattern

# Pseudocode for an adaptive confirmatory design

define_design()

specify(
    interim_times,
    adaptation_features,
    decision_rules,
    stopping_rules,
    maximum_sample_size,
    final_analysis
)

simulate(
    null_scenarios,
    alternative_scenarios,
    nuisance_parameter_scenarios
)

evaluate(
    type_I_error,
    power,
    expected_sample_size,
    min_sample_size,
    max_sample_size,
    bias,
    confidence_interval_coverage,
    duration
)

validate(
    simulation_code,
    interim_analysis,
    adaptation_algorithm,
    final_analysis
)

document(
    protocol,
    statistical_analysis_plan,
    simulation_report,
    regulatory_submission
)

The Most Important Concept

The defining feature of an FDA-compatible adaptive design is not complexity. It is disciplined flexibility.

The trial can change because the accumulating data provide information that was not available at the beginning. But the way the trial is allowed to change should itself be designed in advance.

That principle explains why FDA places so much emphasis on operating characteristics, simulations, interim-data confidentiality, treatment-effect estimation, and regulatory interaction.

References

U.S. Food and Drug Administration. Adaptive Design Clinical Trials for Drugs and Biologics: Guidance for Industry. November 2019. Final Level 1 Guidance. CDER/CBER.

U.S. Food and Drug Administration. E20 Adaptive Designs for Clinical Trials. September 2025. Draft Level 1 Guidance — Not for Implementation.

U.S. Food and Drug Administration. Interacting with the FDA on Complex Innovative Trial Designs for Drugs and Biological Products. Guidance for Industry.

U.S. Food and Drug Administration. Adaptive Designs for Medical Device Clinical Studies: Guidance for Industry and Food and Drug Administration Staff. July 2016.

Official FDA resources: 2019 Adaptive Design Guidance, ICH E20 Draft Guidance, and FDA Complex Innovative Trial Design Meeting Program.