Bias and efficiency of multiple imputation compared with complete-case analysis for missing covariate values.

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Missing data in regression covariates is common.
• Multiple Imputation (MI) is often preferred over Complete-Case analysis (CC).

Purpose of the Study:

• Compare bias and efficiency of MI and CC under various missing data mechanisms.
• Evaluate the appropriateness of MI and CC for missing covariate data.

Main Methods:

• Theoretical arguments and simulation studies.
• Analysis of bias and efficiency under different missing data assumptions (MCAR, MAR, MNAR).

Main Results:

• When data are missing completely at random (MCAR), MI is more efficient than CC with negligible bias for both.
• Under missing at random (MAR), CC can be biased towards the null, while MI may be biased away from the null.
• Bias is generally smaller with MI than CC for more complex missing data mechanisms.

Conclusions:

• MI is not universally superior to CC for missing covariate problems.
• The choice between MI and CC should consider the specific missing data mechanism.
• Standard error comparisons alone are insufficient for method selection; consider the Fraction of Incomplete Cases among the Observed (FICO) for precision gains.