Regularized Cross-Sectional Network Modeling with Missing Data: A Comparison of Methods

Area of Science:

• Network Science
• Psychological Measurement
• Statistical Modeling

Background:

• Network modeling is crucial for analyzing psychological variables, often using the regularized Gaussian graphical model (GGM) with the graphical lasso (glasso).
• Existing methods for handling missing data with glasso are underdeveloped, limiting the use of efficient data collection designs.
• Planned missing data designs can reduce participant burden but require robust missing data handling techniques.

Purpose of the Study:

• To compare three distinct approaches for handling missing data within the graphical lasso framework.
• To evaluate the performance of these methods under varying simulation conditions, including sample size and missing data proportions.
• To provide practical guidance for researchers analyzing psychological network data with missing observations.

Main Methods:

• A two-stage estimation approach using a saturated covariance matrix prior to glasso.
• Single-stage approaches combining glasso with the expectation-maximization (EM) algorithm, utilizing either EBIC or cross-validation for tuning parameter selection.
• A simulation study assessing performance across different sample sizes, missing data proportions, and network structures, supplemented by a real-world data example.

Main Results:

• The expectation-maximization (EM) algorithm combined with cross-validation for tuning parameter selection performed optimally among the evaluated methods.
• All three compared methods showed viability, particularly in scenarios with larger sample sizes and lower proportions of missing data.
• The study identified practical considerations for selecting appropriate missing data handling techniques in psychological network analysis.

Conclusions:

• The EM algorithm with cross-validation offers a promising strategy for addressing missing data in graphical lasso network analyses.
• Researchers should consider sample size and missing data prevalence when choosing a method for psychological network analysis.
• Further methodological development is warranted to enhance the handling of missing data in complex network models.