Network accuracy across levels of analysis using stochastic block models

Area of Science:

• Psychometrics
• Network Science
• Statistical Modeling

Background:

• Network psychometric methods are widely used for analyzing complex relationships.
• Existing simulation studies often lack community structures and focus solely on edge recovery.
• The accuracy of network estimation methods is crucial for valid inferences across different analytical levels.

Purpose of the Study:

• To evaluate the performance of four network estimation methods across four levels of analysis.
• To assess the impact of sample size and data type on network recovery accuracy.
• To provide empirical benchmarks for selecting appropriate network analysis methods and measures.

Main Methods:

• A large-scale simulation study using stochastic block models.
• Empirically informed edge weight distributions from 293 psychological networks.
• Evaluation of EBICglasso variants, ggmModSelect, GGMncv, and GGMnonreg.

Main Results:

• Large sample sizes (≥1,000) are generally needed for adequate recovery across most network levels.
• EBICglasso variants showed better performance in smaller samples and dichotomous data, though overall recovery was inadequate.
• Continuous data and larger samples (≥2,500) yielded adequate accuracy for most methods.
• Centrality accuracy varied, with network loadings reliably recovered but bridge strength being unreliable.
• Community detection was robust except under small sample sizes and dichotomous data.

Conclusions:

• Sample size and data type significantly impact the accuracy of network estimation methods.
• Recommendations are provided for reporting centrality measures (raw values over rankings).
• Established interpretation benchmarks for network similarity metrics and community detection robustness.