The purpose of this study is to investigate the most appropriate alternative IRT parameter estimation models among bi-factor model, testlet based model, and second-order IRT model when the IRT assumptions are not met. A simulation study was conducted to compare the alternative IRT parameter estimation models when the assumptions are unsatisfied. First, the comparison of the IRT models using the simulation data set without the satisfaction of the unidimensionality assumption indicated that bi-factor IRT model appeared the best fitting model. Second, when the local independency assumption was violated, the testlet based model appeared the best fitting model. The results of this study indicated it is necessary to estimate alternative IRT models by going through the process of anticipating the possibility of IRT assumptions violation due to the test forms and domains of their contents. This study also suggested that such a process will provide the basis for applying the IRT more precisely in order to estimate the capability of item and person characteristics
Jiyoung et al. (Tue,) studied this question.