A comparison of multidimensional IRT models for assessing test dimensionality

In educational and psychological measurement, the study of test dimensionality is a fundamental issue. The aim of the paper is to compare item response theory (IRT) models with different structures in the latent abilites, in the case of a test consisting of different subscales. In fact, a unidimensional model may not properly fit the data while multidimensional models, also assuming the existence of general and specific traits, may be more accurate. IRT models with a multidimensional ability structure are compared through simulation studies in order to understand their effective capability of recovering different test structures. Model estimation is conducted via Markov chain Monte Carlo (MCMC) methods, adopting a fully Bayesian approach. An application is also conducted on real data on Italian standardized student assessments.