The major purpose of this study is that to propose a bifactor model to address the construct shift in item response theory vertical scaling. The bifactor model is proposed to estimate a common dimension for all grades and the grade-specific dimension for each grade. Three factors were manipulated to evaluate the performance of the proposed model, they are sample size, number or percentage of common items, and variance of grade-specific factors; resulted in total 27 conditions. The results shown that, first, the parameter estimation and its stability were significantly affected by sample size. Second, the stability of parameter estimation was affected by the variance of grade-specific dimension. Third, the number or percentage of common items has a neglected effect.
Comments & Questions
1. In real data analysis, the constrained bifactor model was the best fitting model according to the information criteria. But the difference between the 2P testlet and Constraint bifactor or 2P UIRT and constraint bifactor is small. Why it is?
2. How to deal with the violation of group factors are orthogonal to one another? For example, extend this model to address the construct shift in multidimensional framework, there is a correlation between group factors or dimensions when multidimensional was used in practice.