Changing Trial Endpoints: When Is an Adaptation Defensible?

If a team wants to replace, promote, or reweight an endpoint during an ongoing trial, first reconstruct when the idea arose and what information the decision-makers had. A formal adaptive design anticipates the possible modification before the trial starts, with an analysis that accounts for the selection. A new idea arising during conduct is an unplanned change, even if an independent committee proposes it. That distinction determines the review needed; it does not by itself prove that every unplanned change is unusable.

The FDA adaptive-design guidance addresses endpoint selection in section V.F and recommends early discussion with the review division. For unplanned changes motivated by comparative interim results, section VI.H strongly discourages implementation before meeting with FDA, provided patient safety is not compromised. The decision therefore needs both statistical assessment and attention to applicable amendment procedures.

The ICH E20 draft defines an adaptive design through prospectively planned modifications based on interim analysis of accumulating trial data. Its section 1 puts potential adaptations in the protocol before trial initiation. The same section excludes unplanned modifications, including an amendment proposed by an independent data monitoring committee (IDMC) after unexpected interim results, and changes based entirely on information external to the trial. These are scope boundaries, not declarations that necessary safety changes are illegitimate.

Pallmann and colleagues similarly distinguish planned adaptations from ad hoc modifications. Their methods review separates validity, including appropriate tests and estimates, from integrity, including protection against information leakage. Both matter: a sound calculation cannot remove changes in recruitment or assessment behavior caused by revealing interim treatment comparisons.

A practical first record is a dated comparison of the original protocol, proposed endpoint definition, reason for changing it, and data-access history. Identify whether the proposal was triggered by a prespecified interim rule, blinded aggregate information, new external evidence, or unplanned comparative results. These situations warrant different assessments. The record should preserve uncertainty where access or timing cannot be reconstructed.

Status checked October 22, 2026: the FDA guidance page lists final guidance; its linked document is dated November 2019, while the page heading says December 2019. The FDA E20 page lists a September 2025 draft, not for implementation. The EMA E20 record lists Step 2b with consultation closed, covering June 30 to November 30, 2025, and a comments overview first published February 16, 2026. Do not describe that consultation as still open.

FDA guidance describes agency thinking and permits alternative approaches that satisfy applicable law. The regulation has a different legal role. The discussion below concerns drug and biologic trials, particularly confirmatory settings under a U.S. IND; it does not describe device or EU amendment procedures. E20 recommendations are identified as draft recommendations throughout.

Locate the proposed adaptation in the protocol, statistical analysis plan (SAP), and associated planning documents. The record should identify the potential changes, number and timing of interim analyses, information used, anticipated decision rule, and resulting analysis. If the interim decision selects a different estimand, define every candidate precisely and explain its clinical relevance. Listing several endpoint names without explaining how selection affects the final test leaves the central question unanswered.

The E20 draft section 6.1 allows detailed rule and access arrangements in restricted documents when disclosure could reveal interim results. Core design elements and rationale still belong in the protocol. Its population- and treatment-selection discussions also recognize methods that permit flexibility around anticipated rules. Do not generalize that discussion into permission for arbitrary endpoint switching: any flexibility needs suitable statistical support and transparent documentation.

A report should establish how the whole design behaves, including selection and subsequent testing, rather than evaluate only the test performed after a favorable endpoint has been chosen. FDA section VI.A discusses analytic theory for established designs and simulations when relevant distributions are unavailable or computationally intractable. Combining adaptive features can make simulations particularly important. The absence of a large simulation program is not, on its own, proof that an established design lacks error control.

• False positive risk: evaluate the relevant null configurations and assumptions, including dependence between candidate endpoints. A result under one favorable correlation assumption does not characterize the entire design.

• Power and selection: examine clinically plausible effects, uncertainty in nuisance parameters, and whether the interim information can support a reliable decision.

• Sample size and duration: report expected, minimum, and maximum sample size where relevant, and how recruitment and outcome timing affect execution.

• Estimation: assess treatment-effect bias and uncertainty, including confidence-interval behavior after selection or early stopping. Testing validity and estimation reliability are separate questions.

When simulations are used, preserve the scenarios, code and software versions, replication count, and the precision of estimated operating characteristics. Explain why the scenario range is adequate, identify assumptions that need sensitivity checks, and report Monte Carlo uncertainty. A simulation estimate is not an exact error probability, and a single estimate just below the target is not sufficient evidence of control across all relevant scenarios.

Blinded, non-comparative adaptations generally have no effect or a limited effect on Type I error according to FDA section IV. That does not make every blinded endpoint change automatically acceptable: clinical meaning, measurement, and the resulting analysis still need review. Record the information actually used rather than classifying every interim decision as equally risky.

Under the FDA DMC guidance and FDA adaptive-design guidance section VII, comparative interim results should be protected from people conducting or managing the trial. An IDMC and a separated statistical reporting team are common arrangements. FDA also describes a dedicated independent adaptation body, and limited justified sponsor access with strong separation. An IDMC is therefore not the sole possible holder of all unblinded information.

Document named access roles, confidentiality agreements, data-transfer procedures, meeting records, and the content of recommendations sent to the sponsor. The E20 draft section 3.5 suggests that sensitive adaptation details may sit in a confidential charter appendix accessible to designated sponsor personnel separated from trial conduct. This is an option for protecting information, not a universal requirement to place every procedure in that appendix.

Plan stopping boundaries and error allocation before the relevant unblinded analysis. Consider what the adaptation itself might reveal, even if no treatment-effect table is shared. If access deviates from the plan, document the information, recipients, timing, and potential effect on conduct. Such exposure can seriously threaten interpretability; its consequences require assessment rather than an automatic declaration that every observation is worthless.

The following hypothetical example concerns a confirmatory trial that prospectively considers candidate estimands A and B. It illustrates records to examine, not a validated trial design. A and B are placeholders for fully specified, clinically justified estimands. No trial data or simulations were generated, and no numerical operating-characteristic results are claimed. The comparison shows where a retrospective claim of prespecification would fail.

A real review should connect each row to dated records and responsible roles. A late SAP version cannot establish that an earlier decision was planned; a confidential charter cannot by itself establish error control. Conversely, executing an already described adaptation does not automatically require a new amendment merely because the decision occurs mid-trial. Assess the actual protocol change and applicable requirements.

For covered Phase 2 or 3 changes under a U.S. IND, 21 CFR 312.30(b) requires an amendment when safety, scope, or scientific quality is significantly affected. FDA submission and responsible IRB approval precede implementation; either may occur first. This provision does not impose a new 30-day FDA-approval waiting period for a change in protocol.

A change intended to eliminate an apparent immediate hazard may be implemented immediately, followed by FDA notification by amendment and notification of the reviewing IRB. The exception is tied to an apparent immediate hazard, not simply an inconvenient endpoint. Safety action and subsequent assessment of what conclusions remain supportable are separate responsibilities.

Registration and reporting should preserve the history of the decision. In Jones and colleagues' 2015 systematic review of 27 studies, the median proportion of trials with a registered-versus-published primary-outcome discrepancy was 31% (IQR 17–45%). In the six-study subset comparing prospectively registered outcomes with publications, the median was 41% (IQR 33–48%). These are study-level medians, not pooled trial-level rates. The heterogeneous older evidence had moderate-to-high risk of bias; discrepancies do not all imply misconduct and do not estimate the frequency of invalid adaptive designs.

As a practical transparency step, keep the original outcome definition, changed definition, reason, dates, and information-access history together. Review the applicable registry updates and explain changes in reporting, using the site's guides to ClinicalTrials.gov outcome records and CONSORT-Outcomes reconciliation. A current registry entry alone may conceal the sequence that matters for interpretation.

Under ICH E9(R1), naming a measurement is insufficient to specify a treatment effect. For each candidate, identify treatment, population, endpoint, strategies for relevant intercurrent events, and population-level summary. Then align the estimator and data collection. Death, treatment discontinuation, or rescue medication may change the clinical question; missing observations create a different analysis problem. Use the estimand guide for the full framework.

The FDA multiple-endpoints guidance distinguishes success on any one of several endpoints from success requiring every co-primary endpoint. The latter does not itself create the same multiplicity problem, although it reduces joint power. Testing one composite also does not automatically mean multiple primary tests. Adaptive selection adds the question of how the chosen endpoint and all subsequent claims were selected and tested.

A fixed testing hierarchy cannot simply be assumed to remain valid after data-driven selection. Evaluate the combined interim and final procedure, including effects on correlated secondary endpoints. The site's multiplicity article covers baseline testing structures, and the SAP-lock article covers document ownership. Here the additional decision is whether the entire selection process was planned and its operating characteristics remain supportable.

FDA section VI.I recognizes well-motivated design changes based entirely on external information. Its DMC guidance sections 3.2 and 4.4.1.4 explain why organizers insulated from comparative interim data can consider external findings without the same concern that the change was motivated by internal results. This is distinct from a prospectively planned adaptation, and it does not establish that every externally motivated endpoint is clinically suitable.

• External trigger: preserve the publication or communication, its date, the proposed change, and why it changes the clinical question.

• Independence: reconstruct who formulated the proposal and whether internal treatment comparisons could have informed them, using access and meeting records.

• Analysis impact: reassess the estimand, available observations, measurement comparability, testing plan, and consequences for previously collected data.

• Implementation: document applicable protocol, SAP, regulatory, IRB, registry, and reporting changes without overwriting the earlier versions.

These are practical documentation checks derived from the cited principles, not an additional regulator-issued checklist. If the evidence was mixed with internal comparative results, do not describe the rationale as entirely external. Conversely, an urgent safety-driven amendment remains possible even when it falls outside the planned adaptive framework.

This editorial routing table identifies what needs assessment. No row certifies a trial design or grants regulatory acceptance. Safety needs take priority, and missing statistical evidence is distinct from demonstrated statistical failure. The actions draw on FDA adaptive-design guidance and 21 CFR 312.30.

For the proposed endpoint change, the usable output is a decision record that connects clinical rationale, planning dates, analysis support, and information access. Where those records are incomplete, state the unresolved inference question before implementation. This article reviews general methodology and source boundaries; it does not run simulations or validate a particular trial's design.