Missing Is Not a Single Category in Medical Data

An empty cell can represent several different events. A laboratory test may have been scheduled but missed, never ordered because it was not clinically indicated, completed at another facility but unavailable, suppressed by a technical interface, or impossible because the participant died. Treating all of these as the same “missing” value discards information about study conduct and can distort the analysis.

The first missing-data task is therefore not imputation. It is classification at collection. A study should preserve why a value is absent whenever that reason is knowable, using controlled codes that are distinct from legitimate observed values.

01Distinguish operational states

At minimum, teams should decide how to represent:

• Unknown: the value exists in principle, but the respondent or reviewer does not know it.
• Not collected: the study intended to collect the value, but collection did not occur.
• Not applicable: the question does not logically apply because of an earlier answer or participant state.
• Not done: a planned procedure or measurement was not performed.
• Pending or not yet available: collection may be complete, but the result has not entered the dataset.
• Unavailable from source: the value may exist elsewhere, but it cannot be obtained under the current workflow.
• Structurally nonexistent: the measurement cannot exist, as with an outcome after death in some settings.

These categories should be adapted to the study rather than copied indiscriminately. Too many codes can overwhelm data entry and reduce reliability. Too few erase distinctions needed for quality control and interpretation. The data dictionary should define every code, specify where it is permitted, and keep missingness codes outside the valid measurement range. A value such as `999` is dangerous if 999 can also be a real result or if software treats it as numeric.

02Do not confuse intercurrent events with missing data

ICH E9(R1) draws a critical distinction. Events after treatment starts—such as rescue medication, discontinuation, treatment switching, or death—may affect the meaning or existence of an outcome. These intercurrent events require a scientific strategy tied to the estimand. They are not automatically missingness.

For example, a glucose measurement after rescue therapy may be fully observed, yet answer a different treatment question from a measurement without rescue. Conversely, an outcome scheduled after death may not be merely unrecorded; the variable may be undefined. The protocol should identify important intercurrent events, while the collection system records both the event and the reason a planned value is absent.

03Preserve reasons before choosing an analysis

Statistical labels such as missing completely at random, missing at random, and missing not at random describe assumptions about the missingness mechanism. They cannot be read directly from a code. Still, operational reasons help analysts judge which assumptions are credible and design sensitivity analyses.

If follow-up is absent because participants with worsening symptoms are more likely to stop responding, a complete-case analysis may overstate favorable outcomes. If a laboratory feed failed for one calendar week independently of patient condition, the implications differ. A single null field hides that difference; reason codes, timestamps, and workflow logs make it assessable.

04Build quality checks around patterns

Missingness should be summarized by variable, site, time period, participant group, and study stage—not only as one overall percentage. Sudden increases can reveal a form revision, staff turnover, interface failure, or misunderstood instruction. Contradictory patterns deserve queries: “not applicable” without the required skip condition, “test not done” paired with a result, or “unknown” for a derived value whose source fields are present.

CONSORT 2025 requires transparent reporting of how missing data were handled, and its explanation recommends reporting strategies used to reduce missingness. That reporting is more credible when the study has retained operational reasons rather than attempting to reconstruct them at manuscript time.

05Use a missingness specification

For each critical variable, document whether it is required, conditionally required, or optional; the allowed absence codes; the triggering conditions; the person or system that assigns them; edit checks; and the downstream analytical treatment. Keep the raw code, even if an analysis dataset later maps several categories to a common value.

Before database lock, compare every generic blank with the study’s permitted reasons. Some blanks will remain genuinely unexplained, and they should be labeled as such. The goal is not to manufacture certainty. It is to avoid throwing away distinctions the research process already knows.

Missingness is data about the data. Record it with the same discipline used for the outcome itself.

This article is educational and intended for research purposes. It does not provide individual medical advice, diagnosis, or treatment. No patient data were used.