Data quality dimensions

In one line

The standard six dimensions — completeness, accuracy, timeliness, consistency, validity, uniqueness — are a good checklist. In health data there's a seventh that isn't on the list and outranks all of them: why is this missing?

The six, with health teeth

Completeness — are the fields populated? Easy to measure, and the measurement misleads. 100% of your patients have a recorded weight because the field is mandatory, and half of them weigh 70kg because that's what you type to get past the form.

Accuracy — does it match reality? Almost never checkable, because reality left. You have no gold standard for what the blood pressure actually was.

Timeliness — is it current, and was it recorded when it happened? Those are different questions, and most schemas conflate them. A vital signed at 22:00 for a 14:00 observation is timely by one definition and fiction by another.

Consistency — does it agree with itself? The discharge summary says the patient is on warfarin; the medication list doesn't. Both are in your database. Neither is flagged.

Validity — does it conform to the rules? A date of birth of 1900-01-01 is perfectly valid and almost certainly a placeholder someone typed to get through registration.

Uniqueness — one patient, one record? Duplicates are the most expensive data quality failure in healthcare and the least measured.

The seventh dimension: informative missingness

Here is the thing that makes health data different from every other domain, and it's why the generic frameworks fall short.

In healthcare, missing data is not random. Missingness is itself clinical information.

• A troponin is missing because nobody suspected a heart attack. The absence tells you what the clinician was thinking.
• A follow-up HbA1c is missing because the patient stopped coming — which may mean they got better, moved, or died.
• A pain score is missing because the ward was short-staffed that shift.

So the standard instruction — impute the missing values, or drop those rows — is actively dangerous here. Dropping rows with missing follow-up drops the patients who did worst. Your analysis of "our diabetics" silently becomes "our diabetics who kept attending", and their outcomes are better. The data didn't lie; your handling of it did. This is the same trap as population health's denominator.

The discipline: model the missingness, don't erase it. Ask what mechanism produced the gap before you decide what to do about it. "Is this MCAR, MAR, or MNAR?" is a genuinely clinical question wearing a statistical hat.

Kahn's framework — worth knowing

The Kahn et al. harmonised terminology is the reference for EHR data quality, and it organises the work usefully into three categories:

• Conformance — does it match the expected format, structure, and value set?
• Completeness — is it present?
• Plausibility — is it believable? Cross-checked against other facts, against time, against what's biologically possible.

Plausibility is the one people skip and the one that catches the real problems. A blood pressure of 400/300 is valid (numbers in fields), complete (present), conformant (correct type) — and impossible. A male patient with a pregnancy diagnosis is conformant nonsense. A patient whose medication started before their birth date is each-field-fine and jointly absurd.

OHDSI's Data Quality Dashboard implements these checks over OMOP and is worth running against any research dataset before you trust it.

The organisational truth

Data quality is not a technical problem you fix once. It is produced by workflow: the mandatory field that gets junk typed into it, the copy-forward that propagates a stale problem list, the clerk under pressure at 2am.

You cannot clean your way out of a collection problem. Every hour spent on the cleaning script is an hour not spent on the form that caused it — and the script has to run forever.