- Prepare a missingness map, method rationale and sensitivity table.
Describe the pattern first
Summarize missing values by variable and group. Distinguish zero, unknown, not applicable and recording failures. Preserve known reasons and improve collection where missingness is preventable.
Treat mechanisms as assumptions
MCAR, MAR and MNAR concern relationships between missingness and observed or unobserved information. A pattern or test cannot establish all these assumptions. Use subject knowledge, collection procedures and sensitivity analyses.
Avoid automatic deletion or mean replacement
Complete-case analysis changes the analytical sample and may be biased under some mechanisms. Mean replacement commonly distorts variability and relationships. Neither should be selected by default without checking design and analytical goals.
Multiple imputation requires a full workflow
The R package mice creates multiple imputed datasets; analyze and pool them appropriately. Check variable types, temporal or clustered structure, auxiliary information, diagnostics and plausible imputed values. Do not present imputed values as actually measured observations.
Report decisions and their consequences
Document sample changes, patterns, assumptions, methods and settings; compare plausible alternatives. Explain material dependence of conclusions on these decisions. A missingness percentage alone does not provide a universal rule for selecting the method.
Practical research checklist
- Separate missing codes from zero and inapplicability.
- Justify mechanisms and methods.
- Document analysis and pooling.
- Report sample changes and sensitivity.
Worked case and implementation decisions
The following is a fictional teaching case. Do not use its numbers or wording as actual study findings.
In the fictional sample, 12 of 100 outcomes are missing and nine missing cases have low baseline scores. This pattern warrants investigation but alone establishes neither MAR nor MNAR. Count missingness by variable and time and distinguish known reasons from guesses. Do not default to complete-case analysis or mean substitution without rationale. If multiple imputation is chosen, document its model, auxiliary variables, number of datasets and pooling, and assess compatibility with the design and sensitivity to untestable assumptions.
| Stage | Teaching example | Verification question |
|---|---|---|
| Count | 12 of 100 in the example | Logged by variable and time? |
| Reason | Known, unknown or hypothesized | Are guesses labeled? |
| Assumption | Defensible missingness mechanism | Are evidence and limits stated? |
| Method | Justified deletion or imputation | Compatible with design and analysis? |
| Sensitivity | Defensible alternative assumption | Are changes in results reported? |
Exercise output: Prepare a missingness map, method rationale and sensitivity table.
Sources and further reading
Official sources for verification and further reading