Why a Deviation Needs Its Own Record
An adaptive trial depends on decisions being made according to rules specified in advance. A departure from a rule can be consequential even when the decision appears reasonable, produces no obvious harm, or is later judged unlikely to change the trial’s conclusion. A clear record allows reviewers to distinguish what the design prescribed from what happened, identify which decisions and participants were affected, and assess the resulting evidence without relying on memory or a polished summary alone.
The purpose of deviation documentation is not to declare a trial invalid whenever an operational error occurs. Nor is it to defend every departure as harmless. It is to provide a traceable account of the facts and a reasoned assessment of their implications. Regulatory guidance for adaptive trial designs and the earlier material on prospective plans versus post hoc adaptations establish why the planned rule matters; this tutorial focuses on what to record when actual conduct departs from it.
This definition requires a comparison between two accounts: the applicable planned rule and the documented event. A rule may have more than one linked requirement. For instance, a rule could specify which participants count at an interim cutoff, how an interim statistic is calculated, who reviews it, and what action follows each threshold. Recording only that “the adaptation was implemented” is not enough to establish whether each requirement was followed.
Reconstruct the Event Before Assessing It
A reliable record begins with chronology and source evidence. Preserve the version of the protocol, statistical analysis plan, adaptation charter, or other controlled document that governed at the time. Record the relevant rule accurately, identify the event that differed from it, and cite the records that support the account, such as dated data extracts, system logs, meeting minutes, decision forms, and correspondence. Do not replace the original planned rule with wording that makes the actual event appear compliant.
The account should distinguish facts from explanations and judgments. “The interim extract included four additional outcomes entered after the cutoff” is a factual statement if confirmed by the data and timestamp records. “The team misunderstood the cutoff instruction” is a proposed cause that should be supported by evidence, such as the instruction available to staff and the training record. “The extra outcomes could not have changed the adaptation” is an impact conclusion requiring a documented reconstruction, not a substitute for reporting what happened.
Useful records commonly identify the trial and adaptation, the applicable document version, the planned rule, the date and time of the event, the people and roles involved, the information available when the decision was made, the actual action, and the evidence sources. They also state the suspected or confirmed cause, the scope of affected decisions or participants, immediate containment, subsequent corrective or preventive actions, and the status of the impact assessment. A site or organization may define severity categories, but a label such as “minor” or “major” should never replace this factual account.
Worked Example: An Interim Cutoff Includes Late Data
Hypothetical scenario. The adaptation plan states that the interim review will use outcomes recorded by 17:00 on 30 June and that no later records will enter the locked extract. The data manager creates the extract on 2 July. A timestamp check finds that four of 60 records in it were entered on 1 July. The decision committee used the extract to select the prespecified sample-size increase.
What the record says. It quotes or references the applicable cutoff rule, identifies the extract and its creation time, and reports that 4 of 60 included records (4 ÷ 60 = 0.0667, or 6.67%) were entered after the specified cutoff. It identifies the committee decision and the records that establish which version of the extract the committee received. The percentage describes the extract, not the proportion of all trial participants affected unless the denominator is known to be that population.
Cause and impact assessment. Suppose the audit trail shows that the extract query used the entry date rather than the prespecified cutoff timestamp. Record that as the supported operational cause. Then reconstruct the decision using the 56 records that met the cutoff and the same prespecified calculation. If the rule yields the same action on the corrected extract, document the result and the calculation; do not conclude that the deviation never occurred. If it yields a different action, identify the altered decision and the participants or trial operations affected. In either case, assess any implications for the planned inferential procedure with the appropriate statistical and governance reviewers rather than assuming that a matching action proves error control is unaffected.
Actions. Containment could include preserving both extracts and preventing further use of the incorrect one. Corrective action could include issuing a verified cutoff-based extract and documenting whether any decision can still be appropriately reconsidered under the governing plan. Preventive action could include validating the timestamp field in the extraction process. The record should state who completed each action and when.
Link the Departure to Decisions and Evidence
An impact assessment follows a sequence. First identify the exact departure; next establish what information and decision it affected; then determine whether the departure could have changed trial conduct or the planned interpretation of results. This order matters because a deviation’s seriousness is not determined by its label or by whether the final result looks favorable. It depends on the design, the facts, and the consequences.
Identify the governing version and compare its requirement with the dated action or data actually used.
Identify which calculation, review, authorization, allocation, enrollment action, or analysis relied on the departure, and who received the relevant information.
Establish affected participants, time periods, and decisions; evaluate whether the departure could alter operating characteristics, analysis, or interpretability.
Describe containment and follow-up, assign responsibility, and state which questions remain unresolved.
The justification for this sequence is straightforward. A decision’s consequences depend on the information entering it and on the rule mapping that information to action. If either differs from the plan, one must trace whether the difference changes the action or its evidential basis. In compact form:
Each arrow requires evidence. A departure may have no demonstrated effect on the selected action but still affect the documented conduct of the trial. Conversely, a change in conduct does not by itself establish the magnitude or direction of any inferential effect. If the consequences cannot be determined, say so and explain what analysis or review is needed. Avoid unsupported claims that a departure “does not affect validity” or “invalidates the trial.”
Worked Example: A Threshold Is Applied to the Wrong Statistic
Hypothetical scenario. A prespecified rule says to increase the target sample size if the interim conditional power is below 40%; otherwise, retain the target. The authorized analyst calculates 38% from the specified model, which calls for an increase. A spreadsheet error displays 43% to the committee, which retains the target. The discrepancy is found before the final participant is enrolled.
Reconstruct the comparison. The record states the threshold and direction exactly: 38% is below 40%, so the planned action is to increase the target. The value shown to the committee was 43%, which is not below 40%, so the action taken was to retain the target. The arithmetic comparison is 38% < 40% and 43% > 40%. The deviation therefore changed the rule-based recommendation and the trial’s target sample size.
Document the evidence and response. Preserve the original calculation, displayed spreadsheet, version history, and committee record. If the spreadsheet formula was incorrectly linked to a different cell, document that as the confirmed technical cause only after verifying the file history. State how many participants had already been enrolled against the retained target and whether the planned increase can still be implemented consistently with the governing documents and oversight decisions. Do not simply overwrite the spreadsheet or record a new committee decision as if it were the original one.
Interpretation. Because the rule-based recommendation differed from the action taken, assess the implications for the planned sample size, interim decision process, and inferential properties. The numbers 38% and 43% alone do not quantify any change in Type I error or power. That assessment needs design-specific statistical evaluation. The deviation record reports the known facts and conclusions, while the supporting evaluation supplies the basis for any inferential claim.
Causes, Corrective Actions, and Transparent Wording
Cause analysis should seek a plausible, evidence-supported explanation rather than assign blame by default. Potential causes include ambiguous wording, an incorrect data transformation, a system configuration error, an access-control failure, a timing problem, or a misunderstanding of delegated authority. Several causes may contribute. If evidence is incomplete, distinguish a confirmed cause from a hypothesis and record what remains to be checked.
Actions also need precise labels. Containment limits further use or spread of the immediate error, such as restricting access to an incorrect extract while preserving it for audit. Correction addresses the specific error where possible, such as creating and validating a corrected calculation. Preventive action reduces the chance of recurrence, such as adding an independent verification or testing a calculation before the next interim review. None of these actions erases the historical event; the record should retain both the event and its response.
A useful narrative is factual, bounded, and explicit about uncertainty. For example: “The committee received a display of 43%, although the archived calculation using the specified model was 38%. The prespecified threshold was 40%; therefore the displayed value led to a different action from the action indicated by the rule. The spreadsheet formula error is under review. The target sample size was not increased at that decision. Statistical assessment of implications for the planned analysis is ongoing.” This wording does not conceal the departure or claim an unverified cause or consequence.
Worked Example: A Decision Is Made by an Unauthorized Role
Hypothetical scenario. The adaptation plan assigns the final recommendation to an independent committee and requires sponsor approval before implementation. After a scheduled review, a sponsor statistician emails the trial operations team instructing it to switch recruitment to the larger allocation ratio. The independent committee has not issued its recommendation and sponsor approval is not yet documented. The operations team implements the change for 24 participants.
Deviation and scope. The departure is not merely a change in allocation ratio; it is a departure from the specified decision authority and authorization sequence. The record identifies the relevant roles and requirements, the email and system timestamps, the 24 participants assigned after implementation, and the allocation ratio actually used. It distinguishes the statistician’s recommendation from an authorized decision.
Assessment and actions. The responsible team should establish whether allocation concealment, the intended ratio, or any other trial operation was affected, using system records and the randomization specification. The record should not infer that participant outcomes are biased simply because the authorization sequence was not followed. It should describe immediate steps to prevent further unauthorized changes, who reviewed the event, and what decisions remain. Any adjustment to documents or future conduct is handled through the applicable governance and change-control processes; the historical record remains unchanged.
Diagnostics, Alternatives, and Limits
Several warning signs indicate that a deviation account is incomplete: it gives no version of the governing rule; uses dates without time zones or source records when timing determines eligibility; reports the final action but not the data or calculation used; calls a departure harmless without showing a reconstruction; or describes training as corrective action without checking whether the process itself contributed. A further warning is a retrospective “clarification” that changes the apparent meaning of the original rule. Preserve the original wording and explain any interpretation separately.
A deviation log and a study report serve related but different purposes. A log is the traceable event-level record, with evidence, chronology, responsibilities, and status. The clinical study report presents a study-level account of adaptations and their implications, as covered in the earlier material on reporting adaptations in clinical study reports. The report can summarize relevant deviations, but it should not be the only location where the underlying facts and supporting records exist. A change-control record addresses governance of controlled documents; it does not replace the record of an event that departed from the then-applicable rule.
| Approach | What it contributes | Why it is not a substitute for a deviation record |
|---|---|---|
| Deviation record or log | Event-level facts, source evidence, impact assessment, and actions | This is the primary traceable account of what departed and what followed |
| Study-report summary | Study-level presentation and interpretation for readers | A concise summary may omit operational detail needed to reconstruct the event |
| Change-control documentation | Controlled review and approval of document changes | It records document governance, not necessarily the event that occurred under the earlier version |
| Informal explanation | May help locate context or evidence | Memory or correspondence alone may be incomplete and may not establish an approved decision |
The limits are important. Documentation can establish what happened and organize an impact assessment, but it cannot recover information that was never recorded or guarantee that a design’s statistical properties are preserved. Some effects require statistical re-evaluation, and some operational consequences may remain uncertain. Record the uncertainty, supporting evidence, and responsible follow-up rather than presenting a categorical conclusion that the available facts do not support.
Check Your Understanding
Answer each question using the principles of event-level, evidence-based documentation.
- A decision follows the prespecified threshold exactly, but it changes the target sample size. Is this a deviation? Explain.
- An interim extract contains 3 records entered after the cutoff, out of 75 total records. What proportion of that extract is late, and what must still be established before interpreting the impact?
- A team discovers that the wrong threshold was displayed, but has not yet found the cause. How should it describe the cause and the immediate response?
- Why should a report that summarizes the deviation not replace the underlying event record?
Solutions. (1) No. The trial adapted, but the decision conformed to the prespecified rule; an adaptation is not itself a deviation. (2) The proportion is 3 ÷ 75 = 0.04, or 4.00%. The team must establish which data and decision used the extract, whether removing those records changes the rule-based action, and what trial conduct or inference may be affected. (3) State the confirmed facts, identify the cause as under investigation rather than asserting a guess, preserve the relevant records, and describe containment to prevent further use of the incorrect display. Assign follow-up for cause and impact assessment. (4) A report-level summary is designed to communicate the study-level account and may not preserve the detailed chronology, source evidence, roles, calculations, and actions needed to reconstruct and review the event.