Title: A simulation-study on pseudo likelihood estimation for item response models of binary data under uncertainty
Authors: Mia Johanna Katharina Mueller-Platz - RWTH Aachen University (Germany) [presenting]
Maria Kateri - RWTH Aachen University (Germany)
Irini Moustaki - London School of Economics (United Kingdom)
Abstract: Latent variables are widely used to measure constructs that are not directly observed such as attitudes, intelligence, abilities, etc. The latent variables are measured using multiple observed indicators often of categorical nature (ordinal, binary, nominal). A full information maximum likelihood approach for model fitting is based on the marginal distribution of the response vector, integrating out the latent variables. In this way, the distributional assumptions made for the latent variables introduce a further source of uncertainty and possible misspecification. Moreover, multidimensional latent variable models require the approximation of multidimensional integrals and often become intractable. An attractive alternative is the pseudo likelihood estimation procedure, which tackles the model fitting with a flexible and feasible approximation of the marginal distribution. This procedure is adopted in the context of IRT models for binary responses, targeting at developing a fast model fitting technique, appropriate for settings with many items and many latent variables. A simulation-study is used to study the robustness of the proposed procedure against model misspecification and non-normal latent variable distributions.