This article compared five item selection procedures in the context of multidimensional computer adaptive testing with the conditions of contents control or not. The five procedures are list as the following: (a)minimum error variance on optimized weight which proposed by the author (V_2), (b) minimum error variance on equal weight (V_1), (c) maximum information in the direction (Ag), (d) maximum determinant of information (Vm), and (e) maximum posterior KL information (KL). To evaluate the adaption of these five procedures, the absolute bias, correlation, test reliability, time used and item usage. The results shown that (a) in the condition of without contents control, the Ag and Vm performed best and the V_2 performed worst with serious bias in theta estimation. (b）While in the condition with contents control, V_2 performed better with slightly Superior than other four procedures. At the end of the paper, the author proposed the complex structure multidimensionality should be investigated in the future, as well as the procedures with item exposure control.
Questions and Comments:
1. The calculation of reliability of multidimensional test might be problematic. The formula offered in this article is identical with those provided in Segall (2001) and they are follow the definition of reliability. However, the reliability of the whole test should take into account the violation of unidmensionality (Widhiarso, 2007).
2. As Prof. Wang questioned, if the items wasn't administered in order with respect to dimension, will the item selection procedure still appropriate to be used? Is there any burden for select the next item from different dimension with the last item?
3. If involving the stopping rule, how will the performance of the five item selection procedures?