Multiple imputation with non-additively related variables: Joint-modeling and approximations.

Area of Science:

• Statistics
• Biostatistics
• Regression Analysis

Background:

• Multiple imputation is common for handling missing data in regression.
• Standard methods assume multivariate normality, which is violated by binary predictors and interactions.
• This violates assumptions for binary predictors and their interactions with continuous variables.

Purpose of the Study:

• To investigate multiple imputation methods for regression models with missing continuous predictors and interactions.
• To evaluate the performance of joint modeling and approximation methods for imputation.

Main Methods:

• Developed a joint modeling approach for multiply imputing missing covariates.
• Considered alternative imputation methods under multivariate normal assumptions as approximations.
• Evaluated methods via a simulation study and application to childhood trauma data.

Main Results:

• The joint modeling procedure performed well across various scenarios.
• Approximation methods incorporating interactions via stratification also showed good performance.
• Failure to include interactions in imputation models led to bias and low coverage.

Conclusions:

• Joint modeling and appropriate approximation methods are effective for multiple imputation with interactions.
• Including interactions in imputation models is critical for accurate regression analysis.
• The methods were successfully applied to analyze childhood trauma and gender interactions.