Calibration of Response Data using MIRT Models with Simple and Mixed Structures
Items may involve more than one construct and they belong to multidimensional model; otherwise, they belong to unidimensional model. In the past, parameters would be misestimated if multidimensional items with high correlation between latent traits were treated unidimensional items; however, it seemed to not much difference if correlation between latent traits is low. The results support that the second study conducted by the current paper. When mixed structures (an item involves more than one construct) are treated as simple structure (an item belong to only one construct), correlation between latent traits would be overestimated and item parameters would be incorrectly estimated. Actually, it is apparent to understand that it would result misestimated output if data is analyzed by wrong model. Moreover, the author investigated performance on calibration for multidimensional and unidimensional models for simple structure. The results illustrated that performance on calibration for two different models did not affected by JMLE estimation method. For MMLE estimation method, multidimensional model has better calibration when test length is short; however, unidimensional model shows advantage once test length is large. Complex models can be considered in the further studies.