Covariate Selection for Multilevel Models with Missing Data.

Area of Science:

• Statistics
• Biostatistics
• Health Research Methodology

Background:

• Missing covariate data presents significant challenges for variable selection in multilevel regression.
• Existing methods for multiply-imputed data often use problematic techniques like list-wise deletion or stepwise selection.
• Standard variable selection methods are typically designed for independent models, not multilevel mixed-effects models with incomplete data.

Purpose of the Study:

• To develop a novel methodology for covariate selection in multilevel random effects models with missing data.
• To enable accurate variable selection across multiply-imputed datasets when predictors are incomplete.
• To address the limitations of current variable selection techniques in complex data structures.

Main Methods:

• Proposed a method to stack multiply-imputed datasets.
• Applied a group variable selection procedure using group lasso regularization.
• Assessed the overall impact of predictors across imputed datasets for multilevel models.

Main Results:

• Simulations demonstrated superior performance of the proposed method compared to existing approaches.
• The novel methodology effectively handles missing covariate data in multilevel settings.
• The technique successfully performed covariate selection across multiply-imputed data.

Conclusions:

• The developed methodology offers an effective solution for variable selection in multilevel models with missing data.
• This approach improves upon traditional methods by accommodating complex data structures and missingness.
• The method was successfully applied to a real-world cancer prevention study dataset.