Computerized adaptive tests (CATs) were originally developed to obtain an efficient estimate of an examinee's ability. For classification problems, applications of the Sequential Probability Ratio Test (Wald, 1947) have been shown to be a promising alternative for testing algorithms which are based on statistical estimation. However, the method of item selection currently being used in these algorithms, which use statistical testing to infer on the examinees, is either random or based on a criterion which is related to optimizing estimates of examinees (maximum (Fisher) information). In this study, an item selection method based on Kullback-Leibler information is presented, which is theoretically more suitable for statistical testing problems and which can improve the testing algorithm for classification problems. Simulation studies were conducted for two- and three-way classification problems, in which item selection based on Fisher information and Kullback-Leibler information were compared. The results of these studies showed that the performance of the testing algorithms with Kullback-Leibler information-based item selection are sometimes better and never worse then algorithms with Fisher information-based item selection. (A)
Abstract