The local item independence and person independence are the essential assumption in IRT. We know that the violation of those assumptions was investigated separately. The proposed model for dealing with local item dependence was testlets model. On the other hand, the sampled students clustered within school represented the person dependence, and this situation was modeled as multilevel IRT. This study proposed the four level IRT model for dealing with the item dependence (testlets) and person dependence effects at the same time. The research methods included a simulation study and real data analysis. The proposed four level model, the multilevel model (for person dependence), the testlets model, and Rasch model were compared in both methods. In simulation part, the author checked the parameter recovery by bias, SE, and RMSE. The result showed the bias of ability had no difference on four models but the SE on proposed model was larger. In real data analysis part, the model data fit index (DIC) was checked, and the proposed model was better fitting than others.
1) I thought the multilevel IRT for dealing with the clustering person group just wants to obtain the group ability parameter. I mean the person dependence effect can be explained by the multilevel situation was not in my thinking. I thought the person dependence situation is more like the structure of the high covariance between person abilities rather than higher level ability structure.
2) It is not friendly to read the conditions as S1, S2, S3, and S4 in Table. It will be better if he list the operated variance value in the Table.
3) This model can be extended to the multilevel higher order model, this model will fit many practical situations, such as the scale with many subscale was used to assess the teacher’s performance by student’s rating.