A comparison of item response models for a test consisting of testlets

Abstract

A set of items that have common stimulus is called testlet. Responses to the items in a testlet often violate the assumption of local independence. In this article, we compared three types of IRT models that take local dependence into account and standard IRT models in terms of accuracy of latent trait's estimation. Simulation results showed that the values of bias, root mean square error and correlations between latent trait and its estimates for four types of models were similar, not depending on the sample size, the proportion of locally dependent items and the degree of local dependence affected the accuracy of latent trait's estimation.