Back to the basics: Rethinking partial correlation network methodology.

Area of Science:

• Psychology
• Statistics
• Network Analysis

Background:

• Gaussian graphical models (GGMs) are popular in psychology for analyzing variable relationships via precision matrices.
• The graphical lasso (glasso) is a common estimation method, but it's optimized for high-dimensional data (p > n), which is rare in psychological research.
• Existing methods like glasso may not be ideal for typical psychological research settings.

Purpose of the Study:

• To propose a more suitable method for estimating GGMs in psychology, particularly in low-dimensional settings (p << n).
• To demonstrate the effectiveness of a non-regularized maximum likelihood approach combined with Fisher Z transformed confidence intervals.
• To compare this proposed method against the graphical lasso for accuracy and error rate control.

Main Methods:

• Estimating the precision matrix using non-regularized maximum likelihood.
• Utilizing Fisher Z transformed confidence intervals to identify significant partial correlations (non-zero relationships).
• Conducting simulations in low-dimensional settings (p << n) for performance comparison.

Main Results:

• The proposed method demonstrates superior performance in detecting non-zero effects compared to the graphical lasso in low-dimensional settings.
• The graphical lasso was found to be inconsistent for model selection and did not control the false discovery rate.
• The proposed method converges on the true model and effectively controls error rates.

Conclusions:

• The proposed basic maximum likelihood method with Fisher Z confidence intervals is a more appropriate and accurate approach for estimating GGMs in psychology than the graphical lasso, especially in low-dimensional contexts.
• This method offers better control over statistical errors, ensuring more reliable identification of variable relationships.
• Researchers in psychology should consider this alternative for more robust network analyses.