Hello everyone,
I’m interested in the feasibility of applying Graph Neural Networks (GNNs) to analyze psychometric scale data, for example, datasets consisting of around 40 items and 500 participants.
Traditionally, methods like factor analysis and Structural Equation Modeling (SEM) are used to uncover latent structures in such data. I’m curious whether GNNs could serve as a viable alternative — especially for capturing nonlinear relationships and learning latent structures automatically.
Specifically, I’d like to ask:
- Has anyone seen or tried applying GNNs (or related graph-based models) to questionnaire or psychometric data?
- How would the outcomes or interpretability compare with SEM?
- Is a sample size of around 500 participants sufficient for training meaningful graph representations?
I’d appreciate any thoughts, shared experiences, or recommended readings on this topic. Thank you!
An interesting question. Perhaps it is best to send this one to the SEMNET email listserv, which has a lot of SEM experts on it:
https://listserv.ua.edu/cgi-bin/wa?A0=SEMNET
(SEMNET requires registration but it is free.)
Dear Hong Du,
My student and I used EGA ( here a test 10.3390/psych3030032 and here the R package: Package ‘EGAnet’).
I know that it is a different proposition, but it may be useful.
In his work testing a scale, he compared different EFA alternatives and between them, EGA solution works better in the CFA test.
I hope it is useful for you.
There is a Dutch group of researchers which are working with network models, like you describe. Some names are Dany Borsboom, Eiko Fried, Sasha Epskamp. There is a lot of literature from this group, and also some software in R. One early paper concerning general network theory in psychology is Borsboom Cramer: Network Analysis DOI: 10.1146/annurev-clinpsy-050212-185608 · Source: PubMed. A paper on psycometric analysis:
Epskamp: Network Psycometric.
DOI: 10.48550/arXiv.1609.02818
- Badges
- Science topic