Handling missing values when using neighborhood selection for network analysis.

Area of Science:

• Psychological network analysis
• Statistical modeling
• Missing data imputation

Background:

• Handling missing values in network analysis often relies on regularization techniques.
• Nonregularized methods like neighborhood selection using Bayesian Information Criterion (BIC) are effective for psychological networks, especially with uneven missingness.
• Existing methods require evaluation for missing data handling in this context.

Purpose of the Study:

• To evaluate multiple imputation (MI) and expectation-maximization (EM) procedures for handling missing values in neighborhood selection via BIC.
• To assess different sample size calculation methods for log-likelihood and BIC computation.
• To identify optimal methods for missing data handling and sample size definition in psychological network analysis.

Main Methods:

• Comparison of stacked MI, grouped MI, direct EM, and two-step EM procedures for missing data.
• Evaluation of various sample size calculation approaches (e.g., local number of observations per node).
• Simulation study assessing edge recovery, partial correlation bias, and network statistics.

Main Results:

• Stacked MI demonstrated superior overall performance in handling missing values.
• The two-step EM approach is computationally efficient and performs adequately under specific conditions (large N, low missingness, small network).
• Using the local number of observations per node for sample size calculation minimized bias, especially with high missingness.

Conclusions:

• Stacked MI is recommended for robust missing data handling in BIC-based neighborhood selection for psychological networks.
• The R package 'mantar' implements the most effective methods for this analysis.
• Careful consideration of sample size calculation is crucial, particularly with substantial missing data.