Unlocking, and What It Costs

Part of Clinical Data Management and Integrity: Database Lock.

About 5 minutes. Free, in your language, no account needed.

Why this matters

Sometimes a locked database must be changed. An error is found that materially affects the analysis, and leaving it would produce a wrong result.

Unlocking is possible and it is an event. It will be described in the study report, it will be asked about at inspection, and the question will be why the error was not found before lock. That question has good answers and bad ones, and which one you have is determined long before the unlock.

When unlocking is justified

A material error affecting the analysis. A systematic problem in a variable the study depends on.

Missing data that has been recovered and materially affects the result, such as a vendor transfer found to have been incomplete.

A regulatory requirement, where an authority requires a change.

And essentially nothing else. In particular, not a preference for how the data looks, not a late-arriving improvement, and not a discrepancy that could equally be documented rather than corrected.

What an unlock requires

A documented rationale stating what the error is, how it was found, what it affects, and why correcting it is necessary rather than merely preferable.

Approval at a level above the study team, because the study team is now unblinded and is not the right body to authorize a change to its own data.

A defined scope. Which records, which fields, which participants. An unlock authorizing "necessary corrections" is an unlock with no boundary.

Controlled execution, with the changes made by named people, within the scope, with the audit trail recording each one.

A re-lock with the same checks the original lock had for the affected scope.

Disclosure. The study report describes the unlock, its reason and its effect. An undisclosed unlock discovered at inspection is a different and much more serious matter than a disclosed one.

Why the timing damages credibility

The difficulty is not the correction. It is that by the time a database is unlocked, the team is unblinded, and a change made to data by people who know the result cannot be shown to be independent of it.

This holds even when the error is unambiguous and the correction is obviously right. The reviewer's question is not whether this correction is justified. It is how the sponsor knows there are not other errors of similar size that were not found, given that this one was found only after the result was visible.

That question is unanswerable except by pointing to the quality of the pre-lock process, which is why the answer is determined before the lock rather than during the unlock.

Avoiding the need

Everything else in this course is, in one sense, about this.

The specific practices that prevent unlocks: reconciliations run at transfer frequency rather than when convenient; derivations checked against the collection instrument before the instrument is finalized; audit trail review as a routine rather than a pre-lock activity; the blind data review given enough time to actually review; and a lock checklist with a name against every line.

None of these is dramatic. Each removes a category of error that would otherwise be found afterwards.

In practice

A study locks, unblinds, and begins analysis. A statistician finds that one site's laboratory values for a key safety parameter are implausible as a group, and the investigation shows the site's local laboratory reported in different units from the protocol's specification for four months.

The error is real, systematic, affects 31 participants, and materially affects the safety analysis.

The unlock proceeds properly: a documented rationale, approval above the study team, a scope limited to that parameter for that site in that period, a controlled conversion with each change recorded, a re-lock with the relevant checks re-run, and disclosure in the study report.

The inspection question, when it comes, is not about the conversion. It is why a unit inconsistency persisted for four months undetected.

The answer is that the reconciliation compared the received file against itself rather than against an expected range, and no cross-site comparison of the parameter's distribution was performed until analysis.

That answer is a process finding, and it is accepted, because it is specific and the sponsor found and stated it. The answer "we do not know" would not have been.

Key takeaways