How They Are Actually Found

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

Data integrity problems are not usually found by looking at individual records. They are found by looking at distributions, patterns and the shape of a dataset as a whole, because the individual records look fine.

The techniques are not secret and they are not difficult. They are simply not applied, because nobody is assigned to apply them and no finding triggers them.

Digit and rounding patterns

Human-generated numbers have characteristic signatures. People rounding measurements produce an excess of values ending in zero and five. People reading an analog scale produce a distribution with peaks at the graduations.

Comparing the terminal digit distribution across sites is a cheap check that requires no clinical knowledge. A site whose blood pressure values end in zero 70 percent of the time is recording differently from one where the proportion is 20 percent, and the difference is about measurement practice rather than about participants.

This does not establish misconduct. It establishes that something differs, and that is enough to ask.

Variability that is too low

Real biological measurements vary. Repeated measurements on the same participant vary, and measurements across participants vary more.

A site whose values are markedly less variable than the study as a whole is worth examining, because implausibly low variability is a signature of values that were not independently measured: carried forward, estimated, or entered from memory.

The same applies to participant-reported data, where a diary with unusually smooth day-to-day variation suggests batch completion rather than daily reporting.

Timing patterns

Entry timestamps carry a great deal of information that nobody looks at.

Visits recorded in a single burst days after the visit date. Diary entries for a week entered in four minutes. Records created outside the site's working hours. Assessments whose recorded duration is implausibly short for the procedure described.

None of these proves anything. Each identifies a place where the record and the reality may have diverged.

Distributions that are too good

Enrollment that exactly meets targets. Compliance at or near 100 percent. Zero missing values. No protocol deviations. Every laboratory value within range.

Each of these is possible and each becomes less likely as the study grows. A site with 40 participants and no protocol deviations has either extraordinary practice or an unreported deviation process, and the second is more common.

Cross-site comparison

Most of the above works by comparison rather than by absolute threshold, which is why it is available to a multi-site study and not to a single site.

The comparison is the tool. The study's own distribution defines what normal looks like, and the question is always which site differs and why, rather than which value is wrong.

In practice

A data manager builds a standing monthly report with six site-level comparisons: terminal digit distribution for the main vital signs, within-participant variability for the primary endpoint, the gap between visit date and record creation, the deviation rate, the missing value rate, and diary compliance.

None of these requires clinical knowledge and all of them are computed from data she already holds.

In month four, one site appears as an outlier on three of the six. Its digit distribution is skewed to zero, its within-participant variability is about half the study's, and its record creation gap averages four days.

The monitoring visit that follows finds the cause and it is not misconduct. One nurse measures blood pressure manually with a device the site prefers, records it on a worksheet, and enters a batch of records weekly. The values are real measurements, rounded by a human reading a dial, and entered late.

The finding is a contemporaneity issue and a measurement standardization issue, both fixable, both found in month four rather than at the closeout inspection.

What this looks like when it goes wrong

A study runs for three years and performs no cross-site comparison of any kind.

An inspection at one site finds that a coordinator has been completing participant diaries. The authority then asks the sponsor what its monitoring of data patterns had shown.

The sponsor has performed none. Its monitoring was risk-based and focused on source data verification of key fields, which verified that the diary entries in the database matched the diary entries in the device, which they did, because the coordinator entered them into the device.

The verification was performed correctly and could not have detected this, because it compared the database to the device and the falsification happened before the device.

The finding against the sponsor is not the coordinator's conduct. It is that a study of that size had no mechanism capable of detecting a site whose data differed from every other site, and the patterns were present in the data throughout.

Key takeaways