Adapting tree-based multiple imputation methods for multilevel data? A simulation study

Area of Science:

• Statistics
• Data Science
• Machine Learning

Background:

• Standard imputation methods assume data independence, limiting their use with hierarchical data.
• Multivariate Imputation by Chained Equations (MICE) is common for hierarchical data but has limitations.
• Tree-based methods are promising but underexplored for multilevel contexts.

Purpose of the Study:

• To evaluate novel tree-based imputation methods adapted for multilevel data.
• To compare their performance against traditional MICE imputation.
• To assess performance across various cluster sizes, missingness mechanisms, and rates.

Main Methods:

• Simulation study comparing chained random forests (missRanger) and extreme gradient boosting (mixgb) with MICE.
• Tree-based methods adapted with cluster membership dummy variables.
• Evaluated bias, type I error, and statistical power under random intercept and slope models.

Main Results:

• MICE offers robust inference for level 2 variables at low missingness (10%).
• Adapted boosting (mixgb) excels for level 1 variables at higher missingness (30%, 50%).
• Adapted boosting surpasses MICE for level 2 variables at high missingness (50%).

Conclusions:

• Adapted tree-based methods, particularly boosting with cluster dummies, are effective alternatives to MICE for multilevel data.
• These methods offer improved performance, especially under higher missingness rates.
• Appropriate adaptation is key for leveraging tree-based methods in complex data structures.