A Multilevel Testlet Model for Dual Local Dependence
This paper proposed a new model that considers dual dependence of item and person to deal with much complexity of testlet format and multilevel structure. In terms of item dependence, we may think a situation that the probability of specific item would affect another item which belongs to the same testlet. Some researchers developed testlet related models to extract the confounding trait from testlet. As a result, local item independence can be built and the latent trait we concern can thus be estimated accurately. The authors considered person clusters as well. However, it seems like just multilevel structure, I don’t really understand the true meaning of the problem of person dependence here. Simulations were conducted and a real data analysis was reported by using Winbugs. Results illustrated that the proposed model performed quite well and had the best fit rather than another three models. According to analysis of real data, it was found that testlet effects were slight, which is that assumption of item independence existed.