network-psychometricscivhexacoegauva

Detecting construct-irrelevant variance with network psychometrics (HEXACO-PI)

Exploratory Graph Analysis + Unique Variable Analysis flagged 110 CIV items on the 240-item HEXACO-PI; 89% overlapped with traditional methods.

Citation

Network psychometrics CIV study (Psychol. Int. 2025) (2025). Psychol. Int. — DOI 10.3390/psycholint7040088

5 min · Updated 2026-08-25

Educational context only, not a diagnosis or clinical assessment. Product-authored items are labeled where validated instruments do not apply.

§ 01 — Findings

What the study found

UVA found 22 highly redundant items; EGA/TMFG flagged 59-80 items misaligned with theoretical communities; combined 110 CIV items.
Post-reduction fit improved modestly (RMSEA) but CFI/TLI remained below thresholds — redundancy and misalignment both matter.
Twelve items were flagged only by network methods (instability/redundancy) vs. traditional discrimination/DIF — evidence for complementary use.

§ 02 — Encoding

At 500-scale, redundancy is the failure mode. We will add a semantic dedup gate (embedding cosine >0.82 flagged) before merge.
Daily micro-batches should be facet-dispersed, not clustered, to respect local independence.