Nodewise Parameter Aggregation for Psychometric Networks.

Area of Science:

• Psychometrics
• Network Analysis
• Statistical Modeling

Background:

• Psychometric networks are valuable for understanding complex relationships between variables.
• Nodewise regression is a method for estimating these networks, especially when direct computation is challenging.
• Edge weights in these networks represent conditional associations between variables.

Purpose of the Study:

• To investigate the accuracy of nodewise regression in estimating edge weights for psychometric networks with continuous variables.
• To identify potential biases in current aggregation methods for regression coefficients.
• To propose and validate improved methods for obtaining true partial correlations from nodewise regression.

Main Methods:

• Utilized nodewise regression, fitting generalized linear models with each node as the outcome.
• Examined the aggregation of two regression coefficients per link to derive edge weights.
• Introduced and evaluated two novel aggregation techniques: multiplying coefficients and taking the square root, and rescaling by residual variances.

Main Results:

• Standard averaging of regression coefficients can lead to asymptotically biased partial correlation estimates, particularly when predictor correlations with control variables differ.
• This bias is pronounced in networks where variables have heterogeneous correlations with other nodes.
• The proposed methods (multiplying coefficients/square root and rescaling by residual variances) successfully recovered true network structures and edge weights.

Conclusions:

• The aggregation method for nodewise regression coefficients is critical for accurate psychometric network estimation with continuous variables.
• Simple averaging is insufficient and can introduce bias.
• Multiplying coefficients (with square root) or rescaling by residual variances are robust alternatives for obtaining accurate partial correlations and network structures.