In the literature, computerized adaptive testing is used in conjunction with compensatory Multidimensional Item Response Theory (MIRT) models (denoted as MCAT-C). Since items may measure multiple latent traits that cannot compensate each other, noncompensatory MIRT models have been developed to account for such items. This study aimed to develop computerized adaptive testing algorithms that were based on noncompensatory MIRT models (denoted as MCAT-N). Four item selection methods, namely, the Fisher information method, the Kullback-Leibler information method, the Shannon entropy method, and the mutual information method, were investigated under two major factors, namely, the correlation between latent traits and the level of termination criterion. Results of a series of simulations showed that all the four item selection algorithms were successfully implemented; a high correlation between latent traits or/and a strict level of termination improved measurement precision and test reliability; test reliabilities for all dimensions were similar regardless of test administration stage because of the noncompensatory nature of MCAT-N. Moreover, among the four methods, the Fisher information method was the superior and the Kullback-Leibler information method the inferior. Copyright © 2017 International Meeting of the Psychometric Society.
|Publication status||Published - Jul 2017|