Question 19 of 20 Advanced Scenario
How would you compare multiple imputation with complete-case analysis for missing measurements?
Short answer
Complete-case analysis is unbiased only under restrictive conditions and always loses power. Multiple imputation creates several plausible completed datasets, analyzes each and pools them with Rubin's rules. It is valid when data are missing at random given the variables in the imputation model, but neither method fixes data missing not at random.
Why
Sterne and colleagues describe three mechanisms. Missing completely at random means missingness is unrelated to anything, such as a broken freezer. Missing at random means missingness depends only on observed data, such as older patients missing a biomarker more often.
Missing not at random means it depends on the unseen value itself, such as the sickest patients skipping a visit. Complete-case analysis drops every subject with any missing value. Multiple imputation keeps them and carries the uncertainty of the filled values into the standard errors.
How to work it through
- Describe the amount and pattern of missingness by variable and by group.
- Look for observed variables that predict missingness, which tells you whether missing at random is plausible.
- Run the complete-case analysis as a reference.
- Build an imputation model that includes the outcome, all analysis variables and auxiliary variables that predict missingness, with transformations for skewed variables.
- Create enough imputed datasets, analyze each and pool with Rubin's rules.
- Compare results and run sensitivity analyses that shift imputed values to mimic missing not at random.
An example
A biomarker is missing in 100 of 400 patients, more often in older patients and at one site. Complete-case analysis on 300 patients gives a hazard ratio of 1.40 with a 95 percent interval of 0.98 to 2.00. Multiple imputation with 30 datasets, including age, site, the outcome and a correlated lab value, gives 1.35 with an interval of 1.01 to 1.80, narrower because 100 patients contribute again.
A sensitivity analysis that lowers imputed values by 20 percent moves the estimate to 1.30. The report presents imputation as primary and both alternatives as sensitivity analyses.
Where the answer stops holding
Multiple imputation assumes missing at random given the model, which cannot be verified from the data. Omitting the outcome from the imputation model biases associations toward zero. With very high missingness the results lean heavily on the model.
In confirmatory trials, the handling of missing data should be prespecified and tied to the estimand, as regulators expect.
A strong answer covers
- MCAR, MAR and MNAR defined
- Complete-case loses power and can be biased
- Imputation model includes the outcome and auxiliary variables
- Pool with Rubin's rules
- Sensitivity analysis for missing not at random
Common mistakes
- Using a single mean imputation, which shrinks variance and makes intervals too narrow
- Leaving the outcome out of the imputation model for a predictor
- Claiming multiple imputation corrects bias from data missing not at random
The interviewer might ask next
How would you design a sensitivity analysis for data that may be missing not at random?