A Generalized Model with Internal Restrictions on Item Difficulty for Polytomous Items
The standard item response model with internal restrictions on item difficulty (MIRID) usually considers component items and composite items into the model. A component item is to measure the specific component and the item difficulty of a composite item is a combination of several component items. The test developer is focus on measuring the latent trait that composite is designed to measure. Existed model with MIRID is only for dichotomous items, thus the authors proposed the extend model to be used in polytomous items. The present paper proposed two approaches of polytomous items: cumulative logits (i.e. CL-MIRID) and adjacent-category logits (i.e. ACL-MIRID). Also, the authors considers above models into multilevel framework (i.e. latent trait is regress of gender in the study). A series of simulations and an empirical data were conducted and analyzed here. As regards the simulations study, results shown that the 2P-ML-CL-MIRID and the 2P-ML-ACL-MIRID models have well parameters recovery. In terms of empirical data, results illustrated that the 2P-ML-CL-MIRID has the best data fit, followed by the 2P-ML-ACL-MIRID. Overall, cumulative logits approach is better than adjacent-category logits one.
Comments, Questions and Future Study
1) A model itself is not attractive to me, it seems like we just used mathematic symbols to create another complex model. However, the authors used proposed models to fit the guilt data, models and data are meaningful at the same time. Data follows the assumption of models, so that models can be effectively explained, thus we can obtain meaningful information.
2) The authors suggested that the future studies can release some constrains on regression weight of thresholds for component items and error for composite items. This kind of adjustment allows model having more flexibility. On my opinion, it is reasonable and easily to be implemented in the presented study due to the fact that the corresponding data is existed. However, if the corresponding data does not exist on the world, we have to think about the necessary on creating complex models.
3) Maybe we can apply DIF and Person-fit procedures on the proposed models.