The Shapes a Data Integrity Failure Takes

Part of Clinical Data Management and Integrity: How Data Integrity Failures Are Found.

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

Why this matters

People imagine data integrity failure as fabrication: someone inventing results. That happens, and it is rare, and it is not what most findings are about.

Most data integrity failures are ordinary, committed by people who believe they are doing their jobs, and produce records that look entirely normal. They are identifiable by shape, and knowing the shapes is what lets you find them in your own data before somebody else does.

The shapes

Making the data agree. A discrepancy is resolved by changing one value to match another, rather than by determining which is right. The database becomes consistent and the evidence that two records ever disagreed is gone. This is the most common shape by a wide margin and it is almost always done helpfully.

Recording later as though contemporaneously. Work written up at the end of the day, the week or the visit, presented as a record made at the time. The content may be accurate; the record asserts something false about when it was made.

Selective retention. Keeping the reading that worked and discarding the ones that did not. A repeated measurement where only the accepted value survives. A sample re-drawn with no record of the first.

Completing the impossible. A required field with no true answer available, filled with something plausible. The form made it necessary and the record shows nothing to indicate it.

Attributing to the wrong person. Shared accounts, entries made on behalf of someone else, a signature applied by a colleague. Each produces a record asserting that a specific person did something they did not.

Deleting rather than correcting. A record removed instead of amended, so what happened is not that a value changed but that an event ceased to have occurred.

Creating the missing document. A document that should exist and does not, produced later to fill the gap. This is the one that converts a documentation finding into a credibility finding, because metadata identifies retrospectively created documents readily.

What they have in common

Each makes the record look better than the reality it describes, and each does so by removing information rather than by inventing it.

That is the diagnostic. Fabrication adds something false. These remove something true: the disagreement, the timing, the discarded reading, the uncertainty, the identity, the event, the gap.

Once you see that, you know what to look for. Not implausible data, which is what people search for. Data that is too clean: no disagreements, no missing values, no unusual readings, complete compliance, no exceptions. Real clinical data is messy, and a dataset without mess has usually had the mess removed.

Why good people do these

Every one of these has a sympathetic motivation.

The coordinator changing a value to match wants the query closed and believes the other record is right. The nurse writing up six visits at day's end was busy caring for participants. The technician discarding a failed reading believes the failure was technical. The site completing an impossible field cannot save the record otherwise. The colleague signing for an absent supervisor is keeping the study moving.

None of them is trying to mislead. The system created a situation where the honest action was harder than the dishonest one, and they took the available path.

This matters because it tells you where to intervene. Training people not to do these things works poorly. Removing the situations that produce them works well.

In practice

A data manager reviews a study for the shapes rather than for errors.

She asks: how many queries resolved with a value change and no explanation, how many visits have all their records created within a few minutes of each other, how many sites have zero missing values in optional fields, which required fields have no "unknown" option, how many accounts are shared, and how many records were deleted rather than amended.

Six questions. Four return nothing. Two return something: one site whose visit records are always created in a single burst three to five days after the visit date, and a required field for date of last menstrual period with no "not applicable" option in a study including post-menopausal participants.

Neither is misconduct. The first is a site batching its data entry, which is a contemporaneity question worth a conversation. The second is a form defect producing fabricated dates, and it is the form's fault.

What this looks like when it goes wrong

A site's data is exemplary. No missing values, no queries outstanding, complete compliance with the electronic diary, every visit within window.

It is the best-performing site in the study and is held up as such.

An inspection finds that one coordinator completes the electronic diary entries for participants who forget, using their devices during visits, because she believes an incomplete diary will reflect badly on the site.

Every entry is her recollection of what participants told her. The dataset records them as participant-reported daily entries.

The site's perfection was the signal. Real participants forget diary entries, and a site with none was not doing better than everyone else, it was doing something different. Nobody asked, because the metric was compliance and compliance was excellent.

Key takeaways