Lucia Guastadisegni, Dipartimento di Scienze Statistiche, Università di Bologna
Date: 11 MARCH 2021 from 12:04 to 12:06
Event location: Modalità telematica, mediante sistema di videoconferenza su piattaforma Microsoft Teams
Type: Statistics Seminars
In the context of goodness-of-fit of Item Response Theory (IRT) models, the Lagrange Multiplier
(LM) test has been widely used to detect local dependence, measurement non-invariance and
deviation from the parametric model. However, in this framework, the LM test and its generalizations
have not been extensively studied under model misspecification.
In this work, we first evaluate the performance of different versions of the LM test under model
misspecification to detect measurement non-invariance in IRT models for binary data by means of a
simulation study. Two kinds of model misspecification are considered, local dependence among items
and non-normal distribution of the latent variable. The results highlight that, under mild model
misspecification, in general all tests have a good behaviour. Under strong model misspecification,
none of the tests performs well. Motivated by these findings, we focus on the scenario of nonnormality
of the latent variable and, to improve the tests performance, we assume a seminonparametric
specification (SNP) for the distribution of the latent variable that is flexible enough to
allow for an asymmetric, multi-modal, heavy or light tailed smooth density. We show some
preliminary results of the performance of the LM tests in the SNP-IRT framework.
Gli Organizzatori Il Direttore
Prof. Silvia Cagnone Prof. Angela Montanari
Prof. Christian Hennig