Model selection methods for unidimensional and multidimensional IRT models.

摘要

Item response theory (IRT) consists of a family of mathematical models designed to describe the performance of examinees on test items. Efficient fit of the model to the data is important if the benefits of IRT are to be obtained. Although there is now an extensive research literature on IRT, relatively little has been done to help practitioners evaluate the suitability of specific models for item response data.;Several model selection methods, including the likelihood ratio test, information theoretic methods, Bayesian methods, and new cross-validation approach, were investigated in choosing the best model among many available unidimensional or multidimensional IRT models. Some indices appeared to function better under particular conditions than others, and for some generating models than for others. Through the simulation studies, in general, two Bayesian model selection methods (DIC and CVLL) appeared to be more stable and accurate in model selection than the other indices in finding the correct IRT model.;Deciding which of these models is most appropriate for a particular test or set of data is generally difficult at best, since the true model is not known for real data. It is hoped that this study may inform decisions as to which indices provide the most consistent and accurate results under certain conditions.