The authors explored the use of 4 adaptive Kullback-Leibler (KL) information methods to classify an examinee’s membership under the mixture Rash model-based computerized adaptive testing conditions. Classification accuracy and absolute bias served as outcomes to evaluate how the 4 methods performed. The findings supported the research hypothesis that the use of an ability estimate for individual latent class provides more information regarding the likelihood of a latent class membership, resulting in the highest classification accuracy compared to other item selection methods although the differences were trivial. The findings also suggested that the magnitude of item separation plays a critical role but separation in ability distributions of the 2 latent classes did not, no matter what item selection method was performed. It means that a good item pool is prominent in CAT. (It is applied to each measure, for sure).
Comments and questions:
1) It is not clear why the random item selection yielded significantly the most accurate estimates of person ability, compared to the proposed four methods. This finding is opposite to the expectation where the use of CAT achieved neither higher measurement accuracy nor test delivery efficiency.
2) This paper did not explain why Method 3 yielded the highest classification accuracy, but did poor in the estimation of person ability when the magnitude of item separation was large.
3) I am interested if the mixture model is used in computerized classification, will the findings differ from the present study?