Work

Reliability Exploratory Graph Analysis (rEGA): Nodal Variance Decomposition and Coefficient Gamma for Psychometric Networks

Year: 2026
Type: preprint
Abstract: Psychometric network analysis faces a foundational methodological divide: while rejecting the latent common-cause assumption, applied studies evaluate composite score reliability using classical common-cause coefficients (α, ω) or substitute metric precision with bootstrap partition frequencies (bootEGA), incurring indirect reliability induction under REGEMA guidelines and violating AERA/APA/NCME testing standards. We introduce Reliability Exploratory Graph Analysis (rEGA), an analytical framework grounded in support-constrained maximum likelihood refitting (K̂_refit) and the Neumann–Markov resolvent operator (Q = [I − Ω]⁻¹). rEGA establishes the Law of Conservation of the Nodal Triad, partitioning standardized item variance into Within-Community Communality (hᵢ²), Inter-Community Bridging Variance (βᵢ²), and Markov Nodal Uniqueness (ψᵢ²), such that hᵢ² + βᵢ² + ψᵢ² ≡ 1.000. From this decomposition, rEGA defines a purity-weighted adjacency matrix (A^rEGA), Within-Community (Γ_c) and Total Network (Γ_t) reliability coefficients, and the Graph Standard Error of Measurement (SEM^raw). Across a Monte Carlo simulation of 97,200 datasets (972 conditions), rEGA-Leiden outperformed Walktrap, Louvain, and Parallel Analysis in dimensionality recovery (87.65% vs. 64.81%–71.32%), eliminated ℓ₁-shrinkage bias (RMSE_ψ = .035), classified core, bridge, and noise nodes with 86.8%–99.4% accuracy, and yielded unbiased reliability estimates (Bias_Γ = −.007) immune to lexical doublet inflation. Multi-sample validation on the AAQ-II (N = 3,200) and DASS-21 (N = 2,751) demonstrated that the Nodal Triad Leakage Criterion (β̄² > h̄²) rules out spurious over-fragmentation while distinguishing noisy items (A1) from transdiagnostic bridges (D2, S3, S6) with high cross-sample invariance (r > .93). Separating structural validity from score error, rEGA prevents ad hoc scale pruning and reconciles network psychometrics with international measurement standards.
Language: English
Open Access status: green
OpenAlex ID: W7225910094
Cites:
Cited by:
Related to: 10