Choosing complete cases or multiple imputation
Choose complete cases or multiple imputation from the missingness assumption and the analysis, never from a percentage cutoff.
Fit
When it matters, and when it does not
It matters when
- MCAR is plausible, so complete-case analysis may be valid, though less precise.
- For linear regression, completeness depends only on observed covariates and not on the outcome given those covariates. The conditional slope may remain unbiased.
It matters less when
- Under plausible MAR, multiple imputation can use observed predictors of missingness and retain information from incomplete records. Read: Multiple imputation
- Under plausible MNAR, neither ordinary approach resolves the uncertainty. Add a delta or tipping-point sensitivity analysis. Read: Sensitivity analyses
Before you start
What to check in your study
- Which variables are missingMissing outcomes and missing predictors can need different arguments.
- The model and target quantitytargetsummaryMeasurePreserve interactions, nonlinear terms, clusters and follow-up structure.
- Useful observed informationInclude measured predictors of gaps and missing values. Check that they will be available for the intended use.
- Choose a method because of a plausible missingness assumption.How to check: Missing completely at random (MCAR) means gaps depend on neither observed nor unseen values. Missing at random (MAR) means observed information explains the gaps. Under MAR, multiple imputation can help, but it is not inherently better. Missing not at random (MNAR) remains plausible if unseen values predict gaps after conditioning on observed information. Under MNAR, add a delta or tipping-point sensitivity analysis.
- The imputation and analysis models agree.How to check: Check distributions, interactions and clusters. For prediction, fit imputation inside training resamples and use only information available at prediction time.
Plain language
How it works
Complete-case analysis uses only people with every needed value. MCAR is sufficient for many analyses, but is not always necessary. In correctly specified linear regression, selection independent of the outcome given covariates can preserve conditional slopes. It does not automatically preserve population means. Under MAR, multiple imputation draws missing values from models using observed information. It can recover information, but a poor model can add bias.
The choice also depends on what is missing and which quantity you estimate. Compare imputations with observed values, inspect model fitting and pooling, and vary assumptions where MNAR is plausible. A symptom score after death needs an outcome definition, not a routine imputation.
Worked example
A worked example
- Setting
- A study records an outcome for 100 adults, but some exposures are missing.
- Question
- Does the exposure predict the outcome?
- Approach
- 20 exposures are missing. Recorded severity predicts the gaps and the exposure, making MAR plausible conditional on severity.
- Result
- 80 of 100 records are complete, or 80 in 100. Complete-case analysis discards 20 records whose outcomes are known.
- What it means
- Multiple imputation may use their observed outcomes and severity. That does not guarantee less bias or better precision; the model must fit.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Multiple imputation does not make missing-at-random true, and complete-case analysis is not automatically biased under every mechanism. Justify the assumptions for the specific model and perform sensitivity analyses where they matter.
- Choosing imputation whenever missingness exceeds a fixed percentage.Instead: Patterns and causes matter. Explain why the assumption and model support your choice.
- Treating pooled imputed records as extra participants.Instead: Fit the analysis separately in each completed dataset, then pool estimates and their uncertainty.
- Dropping people whose exposure is missing without discussing selection.Instead: Dropping them risks selection bias. Consider multiple imputation under MAR, with sensitivity analysis for plausible MNAR.
Inside Mallard
How Mallard uses this
The comparison cue offers complete cases versus imputation for causal or associational linear or logistic regression with missing-data cues and explicitly independent, unrepeated observations.
concept.choosing_complete_case_mi, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls (opens in a new tab) · BMJ · doi 10.1136/bmj.b2393
- Missing Data in Clinical Research: A Tutorial on Multiple Imputation (opens in a new tab) · Can J Cardiol · doi 10.1016/j.cjca.2020.11.010