A Missing-Data Audit Before You Choose a Method

Define what should have been observed

Start with one variable at one time point. Identify the records expected to contain that measurement. A blank caused by a missed visit differs from a field that was never applicable; a database code for “unknown” also differs from a measured zero.

Write these definitions into the data dictionary before recoding. Preserve the original data and make transformations reproducible. Resolve ambiguous codes with the collection team instead of inferring their meaning from a distribution.

Count without disguising the denominator

In a fictional example, 18 missing measurements among 120 expected measurements means 15% missing and 102 observed. Report the counts alongside the percentage. The Missing Data Summary Tool performs this arithmetic without uploading data.

Repeat summaries by relevant group and time point. Do not add missing counts from different variables to estimate incomplete participants: the same person may be missing several measurements. Participant-level patterns require a separate audit in your approved analysis environment.

Investigate reasons and patterns

Review whether missing measurements relate to visits, site processes, treatment discontinuation, or recorded participant characteristics. Document what is known and what remains unknown. An absence of an obvious observed pattern cannot prove that missingness is unrelated to unobserved outcomes.

A percentage alone cannot determine a safe analysis. Complete-case analysis and imputation require assumptions, and neither is automatically appropriate. Discuss plausible missingness mechanisms, the intended effect, and sensitivity analyses with the study team.

Produce an auditable handoff

Your handoff should include variable definitions, expected and observed counts, group summaries, reasons where available, and a list of unresolved coding questions. Preserve the distinction between descriptive findings and assumptions proposed for analysis.

Use the missing-data investigation worksheet. The Cochrane risk-of-bias chapter explains why the amount of missing outcome data alone is insufficient to judge bias.