No single percentage of missing values determines the correct method. The response depends on the analysis, missingness process, observed predictors of missingness, model compatibility and sensitivity of the decision.
Plan before choosing a method
- Map missingness. Report variables, cases, patterns, timing and reasons where known.
- Protect the estimand. State the target quantity and which missing values threaten its interpretation.
- Make assumptions explicit. Explain why complete-case analysis, maximum likelihood, multiple imputation or another method is defensible.
- Align models. Include variables and relationships needed to make the imputation and analysis models compatible.
- Test sensitivity. Examine whether conclusions depend on plausible departures from the main missingness assumption.
Deletion is an analysis choice
Complete-case analysis can change the target population and reduce precision. It can be valid under particular conditions, but it should not be treated as a neutral default. Likewise, imputation does not recover unknown truth; it propagates uncertainty under a stated model.
Worked use case
Income is more often missing among self-employed respondents. A complete-case model may change the composition of the sample. Include observed predictors related to missingness, document the imputation model and compare the substantive conclusion with a complete-case and sensitivity analysis.
Evidence boundary
Missing-at-random assumptions cannot generally be proven from observed data alone. Diagnostics can reveal patterns and model problems, but sensitivity analysis is needed when departures could alter the decision.
Primary sources
- National Research Council. The Prevention and Treatment of Missing Data in Clinical Trials. Retrieved 5 August 2026.
- CONSORT. CONSORT guidance. Retrieved 5 August 2026.