The Discrepancy That Cannot Be Reconciled

Part of Clinical Data Management and Integrity: Queries and Discrepancy Management.

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

Why this matters

Most discrepancies resolve. One record was wrong, the source settles it, the data is corrected and the trail records why.

Some do not. Two records of the same fact, both contemporaneous, both attributable, both apparently reliable, and they disagree. There is no further evidence and no principled way to determine which reflects what happened.

How an organization handles that case is the clearest available signal of whether its data integrity culture is real, because it is the case where the honest answer is uncomfortable and the dishonest one is easy, invisible and feels like tidying up.

What must not happen

Do not choose. Selecting the more plausible value, the one that agrees with the rest of the data, or the one from the system you trust more, is a judgment presented as a fact. Nothing in the record will show that a choice was made.

Do not average, adjust or interpolate. Creating a third value that appeared in neither record invents data.

Do not ask the site to re-record it. A site asked to confirm a value it recorded eighteen months ago is being asked to recall something it cannot recall, and it will produce an answer, and the answer will be whichever record it is shown.

Do not close the query as resolved. An unresolvable discrepancy that is closed looks identical to one that was resolved, and that is the whole of the problem.

What must happen

Preserve both values with their provenance. Both records, both sources, both timestamps, both authors.

Document the investigation: what was checked, what was found, and why it could not be determined.

Escalate to whoever can adjudicate. Usually the investigator for a clinical fact, sometimes the statistician for an analysis consequence, sometimes both.

Record the disposition explicitly. Which value, if any, the analysis uses, and on what stated basis. "The site-recorded value is used as the primary source for this variable, with the vendor value retained and the discrepancy documented" is a defensible position. Silence is not.

Consider whether it is a signal. One unresolvable discrepancy is an incident. A pattern of them, at a site or with a vendor or in a field, is evidence of a process problem, and the pattern is visible only if each one was recorded rather than closed.

Why the honest handling is also the safer one

The discomfort is that documenting an unresolvable discrepancy looks like admitting a data quality failure, and there is a strong pull toward making it go away.

Consider how each version reads to an inspector.

The documented version: a discrepancy identified, investigated, found unresolvable, escalated, adjudicated, and disposed of on a stated basis. That is a functioning quality system detecting and handling a problem, which is exactly what the system is for.

The resolved version: a value silently changed to agree with another record, with an audit trail entry saying "per query." That is indistinguishable, from the outside, from altering data to make it agree with expectations. There is no way for the sponsor to demonstrate it was not, and the burden of that demonstration falls on them.

The honest handling produces a better record and a better defense.

In practice

A participant's serious adverse event is recorded in the safety database with an onset of 14 March and in the clinical database with an onset of 11 March. Both entries were made within days of the event by different site staff, and both are attributable and contemporaneous.

The source records do not settle it: the clinical note describes symptoms beginning "a few days ago" at a visit on 14 March.

The data manager does not choose. She documents both values with their provenance, records that the source does not permit determination, and escalates to the investigator, who reviews and states that the onset cannot be determined more precisely than the week of 11 to 14 March.

The disposition: the analysis uses the clinical database value, with the discrepancy and the investigator's assessment documented. The event's reporting timeline is assessed against the earlier date, which is the conservative position for a regulatory clock.

Three records where there might have been one silent change, and every one of them is defensible.

What this looks like when it goes wrong

A vendor's laboratory result and the site's recorded value differ for a potassium measurement, 5.8 against 3.8.

The data manager queries the site. The site, seeing the vendor value, changes its record to 5.8, and the query closes.

Two things were destroyed. First, the evidence that the two records ever disagreed, which was the signal that something in the process had failed and would fail again. Second, and more seriously, a potassium of 5.8 may have required clinical action at the time, and the question of whether the site acted on 3.8 because that is what it believed the value to be is now unaskable.

The database is consistent. A participant may have had a clinically significant result that nobody responded to, and the record no longer contains the information that would reveal it.

Key takeaways