Title: On the Relation between Item Response Theory and Network Models
Authors: Charlotte Tanis
Affiliation: University of Amsterdam
Abstract:
Item response theory (IRT) models are a cornerstone of psychometrics.
Fundamentally, they assume that the observed correlations between items are
due to an underlying common cause - a latent variable. Recently, the
so-called "network approach" has become popular in psychometrics,
eschewing the assumption of latent variables and instead modelling the
associations between items directly (e.g., Epskamp, Maris, Waldorp & Borsboom,
2018). One such model is the Curie-Weiss model, originating from the analysis
of magnetic spins of atoms in statistical physics. Although IRT and network
models differ in their assumptions and ontology, there exists a surprising
statistical equivalence between the extended-Rasch model, the marginal Rasch
model, and the Curie-Weiss model (Marsman et al. 2018). In practice, empirical
examples in educational measurement exist where we cannot distinguish a
Curie-Weiss model from an extended-Rasch model. In this talk, I will dive
deeper into this equivalence, discuss the statistical properties of the
Curie-Weiss model, and show it can be estimated.