PhD in PsychMethods & ClinicalPsych with @sverreuj @SachaEpskamp | Prev @UvAmsterdam @UWaterloo | (Network) Psychometrics; (Intensive) Longitudinal Data; Natural Language Processing; Applied Statistics
Xinkai Du
Loading...
Beyond network comparison, we believe IVPP has broader insights for SEM/psychometrics and meta-analysis in:
▶︎ Determining measurement/parameter invariance
▶︎ Parameter meta-analysis, which reveals parameters that are shared across samples and parameters that are unique to each sample.
Unfortunately, the captions of Figure 2-5's subfigures now have wrong fonts and font sizes. This is an error that they failed to correct during the proof examination, and I have contacted them to fix it. The error is expected to be fixed in three weeks. The fonts in the preprint are fine.
2) Previous approaches to compare dynamic networks are only viable when intensive measurements are available. IVPP allows comparing networks when only a few data points (t >= 3) are available per person, a situation common in large-scale longitudinal surveys.
Thrilled to share that our work (with Sverre Urnes Johnson and @sachaepskamp.bsky.social) on invariance partial pruning (IVPP), a novel approach to comparing networks in time-series and panel data is now online at Psychological Methods 10.1037/met0000824 (preprint link: 10.31234/osf.io/vb8dz_v2).
IVPP fills in two essential gaps:
1) Previous approaches to comparing dynamical networks only report the presence/absence of heterogeneity. In contrast, IVPP uncovers edge-level differences, providing a more meaningful network-difference test that reveals the mechanisms of heterogeneity.
In simulations with both synthetic and empirical networks, DFI showed transparent and often superior type I & II error rates vs. Hu/Bentler cutoffs (especially in empirical networks). We argue that the transparency and consistency of DFI provide more reliable model evaluation of network models.
The link to the initial study on fit indices and Hu/Bentler cutoffs mentioned earlier: psycnet.apa.org/record/2026-31171-001
Our recent study on PsychMethod showed that SEM fit indices had desirable sensitivity to the misspecification in (dynamic) networks, yet were also sensitive to sample and model characteristics (e.g., N and network size), so we created DFI for networks to accommodate design-specific characteristics.
My latest work with @sachaepskamp.bsky.social on creating dynamic fit index cutoffs for Gaussian graphical models is now out as a preprint:
osf.io/preprints/ps..., accompanied by an R package, netDFI: github.com/xinkaidupsy/....
