Missing Data Handling via EM and Multiple Imputation in Network Analysis using Glasso and Atan Regularization.

Area of Science:

• Psychological network analysis
• Statistical methodology
• Missing data handling

Background:

• Current literature on missing data in psychological networks is limited to likelihood-based methods.
• Existing approaches often focus on convex regularization with varied missing data handling implementations.
• There is a need for standardized and comparative evaluations of different missing data handling techniques.

Purpose of the Study:

• To implement and evaluate a missing data handling approach using stacked multiple imputation.
• To compare stacked multiple imputation against direct and two-step EM methods.
• To assess performance under varying network conditions and regularization types.

Main Methods:

• Simulated cross-sectional psychological networks with varying network size, observations, and missingness.
• Implementation of stacked multiple imputation, direct EM, and two-step EM methods.
• Evaluation using convex (glasso) and nonconvex (atan) regularization with EBIC and BIC model selection.

Main Results:

• Missing data handling methods showed similar performance across many simulated conditions.
• Two-step EM with glasso and EBIC performed best overall, followed closely by stacked multiple imputation.
• Stacked multiple imputation was most consistent for atan regularization with BIC.

Conclusions:

• Stacked multiple imputation offers a viable and consistent alternative for missing data handling in psychological networks, particularly with nonconvex regularization.
• The choice of missing data handling method can depend on the regularization technique employed.
• Further research comparing imputation and EM methods in network analysis is warranted.