Fields That Produce Unusable Data

Part of Clinical Data Management and Integrity: Designing the Data Collection Instrument.

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

Why this matters

Some fields are complete, internally consistent, free of queries, and cannot be analyzed. They pass every quality measure the study applies and contribute nothing.

They are worth studying as a category because they are invisible to every automated check, they are identifiable in advance by anyone who knows what to look for, and they are usually cheap to fix before enrollment and impossible to fix after.

The free-text field where a coded field belonged

Free text is the reflex when the options are not obvious. It is easy to build and it converts an analysis problem into a much larger one.

A study collecting "reason for discontinuation" as free text receives several hundred distinct phrasings for perhaps a dozen actual reasons, in multiple languages, with spelling variation. Making that analyzable means coding it retrospectively, which means a human reading each entry and assigning a category, which is expensive, slow, and introduces interpretation that nobody planned or validated.

Free text is right when the content is genuinely unbounded and is intended to be read rather than counted: a narrative of an adverse event, a description of a protocol deviation. It is wrong whenever the analysis will count things.

The test: if the analysis plan says "number of participants who," the field must be coded.

The compound field

A field collecting more than one fact cannot be analyzed for either.

"Dose and frequency" as a single text field. "Site and severity" of a finding. "Date and time of onset" where one is known and the other is not. Each requires the analysis to parse the value, and parsing free-form human entry is unreliable in exactly the cases that matter.

The field whose meaning depends on the person completing it

"Is this clinically significant?" means different things to different clinicians and the variation is not noise; it is systematic by site, by specialty and by individual.

Such a field is legitimate when the judgment is the thing being collected, made by a defined role, against a defined standard, and recorded as a judgment. It is not legitimate when it is used as though it were a fact.

The same applies to any field using a term the protocol does not define. "Adequate," "stable," "improved" and "tolerated" all mean whatever the person completing the form thinks they mean.

The ambiguous blank

A field left empty carries no information about why. Not assessed, assessed and absent, assessed and zero, and forgotten are four different facts sharing one representation.

The fix is structural: either the field is required with an explicit option covering each case, or there is a companion field recording why it is empty. Attempting to recover the distinction later from other evidence is guesswork presented as data cleaning.

The field collected at the wrong grain

A value collected once per visit when the analysis needs it per assessment. A value per participant when the analysis needs it per event. Concomitant medications recorded as a list per visit rather than as records with start and stop dates, so exposure cannot be reconstructed.

This one is easy to miss in review, because the field looks correct in isolation. It is only wrong relative to a question, which is why the analysis plan must be in hand during form design.

In practice

A data manager reviews a draft form against the analysis plan and flags five fields.

Reason for treatment discontinuation, free text, feeding an analysis that counts participants by reason. Coded, with an "other, specify" for the genuine remainder.

Concomitant medication dose, a single text field holding amount, unit and frequency. Split into three.

"Was the assessment adequate?" with no definition of adequate. Replaced with the specific objective criteria the clinical team actually meant.

Laboratory abnormality present as a checkbox with no way to record that the test was not performed. Given a not-done option with a reason.

Vital signs collected once per visit, where the analysis compares pre-dose and post-dose. Given a timepoint field.

Five changes, about three days of build. Each one, left in place, produces data that is complete and unusable, and the discovery would come at analysis.

What this looks like when it goes wrong

A study collects protocol deviations in a free-text field, with a plan to categorize them later.

Four thousand deviations are recorded across 60 sites and three years, in five languages, written by several hundred people, with no controlled vocabulary.

At analysis, the study must report deviations by category. Categorizing four thousand free-text entries retrospectively takes two people six weeks, requires clinical judgment on ambiguous entries, and produces a categorization nobody can validate because the original intent of each entry is not recoverable.

The reported deviation categories are, in the end, one team's retrospective interpretation of what several hundred people meant. It is the best that can be done and it is not a measurement.

A coded field with an "other, specify" option would have cost two days during design.

Key takeaways