Complete cases or multiple imputation?
The decision
Describe which variables are missing, for whom, when and why. Distinguish an unavailable measurement from an outcome prevented by death or an ineligible observation; those may be estimand questions rather than values to impute. Consider what observed information predicts missingness and the missing values.
When complete-case analysis may be defensible
Restricting analysis to complete records is simple and can be valid under suitable conditions. Missing completely at random is sufficient for many analyses, but some regression settings allow weaker conditions. The relevant conditions depend on which variables are missing and the model. Loss of information and a changed represented population can still matter even when a coefficient is unbiased.
When multiple imputation may help
Multiple imputation propagates uncertainty about missing values through several completed datasets and pooled estimates. Under an appropriate missing-at-random assumption conditional on included information, and a suitable imputation model, it can use more of the observed data. Include relevant outcome and auxiliary information for explanatory analyses, preserve variable types and clustering, and make the imputation model compatible with interactions and nonlinear terms in the analysis.
What a choice needs
- Do not use one missingness percentage as an automatic switch between methods. Pattern, causes and information content matter.
- Single mean imputation understates uncertainty and is not multiple imputation. Pool estimates with appropriate rules instead of treating imputed datasets as extra participants.
- Inspect convergence, plausibility of imputations and Monte Carlo stability. More imputations cannot repair an implausible model.
- Missing-not-at-random sensitivity analyses examine departures from the primary assumptions; the observed data alone cannot verify MAR.
- Prediction evaluation has a separate information boundary: learn imputation inside training resamples and use only information available at intended prediction time.
Record in the workspace
For this aim, state the primary missingness assumption, analysis population, variables and structure in the imputation model if used, pooling procedure and sensitivity strategy. Retain the reason for choosing complete cases when that is the justified primary approach.
What to watch out for
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.
- 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