The proportion of missing data should not be used to guide decisions on multiple imputation

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Missing data presents challenges in statistical analysis, with researchers debating between multiple imputation (MI) and complete case analysis.
• A large proportion of missing data can significantly impact the validity and reliability of study findings.

Purpose of the Study:

• To provide guidance on selecting appropriate methods for analyzing data with substantial missingness.
• To investigate the performance of multiple imputation under various missing data scenarios.

Main Methods:

• Simulations were conducted to assess the impact of missing data proportion, fraction of missing information (FMI), and auxiliary variables on MI performance.
• Outcome data were simulated under missing completely at random and missing at random (MAR) assumptions.

Main Results:

• Multiple imputation (MI) demonstrated benefits in reducing bias when sufficient auxiliary information was available, without compromising efficiency.
• Fraction of missing information (FMI) was a more reliable indicator of efficiency gains from MI than the proportion of missing data when bias was absent.
• Precision of effect estimates was comparable for models with similar FMI, irrespective of the proportion of missing data.

Conclusions:

• Valid multiple imputation (MI) effectively reduces bias in data with missing at random (MAR) mechanisms, even with large proportions of missingness.
• Researchers should utilize FMI to select auxiliary variables for maximizing efficiency in imputation analyses.
• Sensitivity analyses with varying imputation models may be necessary when the number of complete cases is limited.