Diagnosing imputation models by applying target analyses to posterior replicates of completed data

Area of Science:

• Statistics
• Data Science
• Biostatistics

Background:

• Missing data is a common problem in statistical analyses.
• Multiple imputation is a widely used technique to handle missing data.
• Assessing the adequacy of imputation models is crucial for valid results.

Purpose of the Study:

• To develop and illustrate a method for assessing the adequacy of multiple imputation models.
• To provide a practical approach for quantifying evidence of imputation model inadequacy.
• To enhance the reliability of statistical analyses involving imputed data.

Main Methods:

• Comparing completed data with replicates simulated under the imputation model.
• Applying analyses of substantive interest to both datasets.
• Using posterior predictive checks of estimate differences to quantify model inadequacy.
• Integrating out imputed data and replicates to reduce comparison variance.

Main Results:

• The proposed posterior predictive check method effectively quantifies evidence of imputation model inadequacy.
• The checking procedure is implementable using standard imputation software.
• The method is applicable to both Bayesian and non-Bayesian imputation techniques.
• Real data applications and simulations demonstrate the method's utility and properties.

Conclusions:

• The developed method provides a statistically sound and practical approach for validating multiple imputation models.
• This facilitates more trustworthy data analysis when dealing with missing data.
• The approach promotes better understanding and application of imputation techniques in research.