Outcome-sensitive multiple imputation: a simulation study.

Area of Science:

• Biostatistics
• Health Services Research
• Data Science

Background:

• Missing data is a common challenge in healthcare research.
• Optimal strategies for multiple imputation (MI) are not fully established, particularly regarding outcome inclusion and handling missing not at random data.

Purpose of the Study:

• To evaluate the performance of various multiple imputation methods under different missing data scenarios.
• To provide clear recommendations on best practices for multiple imputation in healthcare research.

Main Methods:

• Simulated thousands of datasets with varying missingness mechanisms (MCAR, MAR, MNAR), missing data extents (20-80%), and sample sizes (1,000-10,000).
• Compared complete case analysis with seven MI methods differing in outcome inclusion/imputation and secondary outcome inclusion.
• Assessed performance using mean absolute error, bias, coverage, and power.

Main Results:

• Multiple imputation methods showed minimal performance differences when the outcome was included in the imputation model.
• All MI approaches performed well, even with extensive missingness.
• Secondary outcome inclusion had minimal impact; dataset size and missingness extent influenced performance as expected.
• MI offered partial protection against missing not at random data.

Conclusions:

• Including the outcome in the imputation model is key; minor differences exist between imputing, deleting, or not imputing the outcome.
• All informative covariates, regardless of missingness, should be included in the imputation model.
• Multiple imputation provides some robustness against simple missing not at random mechanisms.