Appropriate inclusion of interactions was needed to avoid bias in multiple imputation

Area of Science:

• Epidemiology
• Biostatistics
• Data Science

Background:

• Missing data are a common issue in research, potentially causing bias in complete case analysis (CCA).
• Multiple imputation (MI) is a method to handle missing data, but its application with interaction terms requires careful consideration.

Purpose of the Study:

• To evaluate the performance of different multiple imputation (MI) strategies in the presence of interaction terms.
• To determine the conditions under which complete case analysis (CCA) and MI yield unbiased estimates.

Main Methods:

• Simulated data with binary explanatory variables (X, Z) and their interaction (XZ), with continuous and binary outcomes (Y).
• Assessed six scenarios with varying interaction strengths and five missing data mechanisms.
• Compared CCA, MI without interactions, MI with interactions, and stratified imputation using directed acyclic graphs.

Main Results:

• Multiple imputation (MI) excluding interaction terms resulted in biased estimates and low coverage.
• MI including interactions and stratified imputation provided valid and equivalent inference across all simulated scenarios.
• Complete case analysis (CCA) was unbiased only when the interaction term was absent in the population model.

Conclusions:

• Epidemiologists must ensure their multiple imputation (MI) models are consistent with their analysis models.
• Accurate handling of interactions in MI is crucial for valid epidemiological research.
• Inappropriate MI models can lead to erroneous conclusions about important effect modifications.