Comparing Item Selection Criteria in Multidimensional Computerized Adaptive Testing for Two Item Response Theory Models

Abstract

Multidimensional computerized adaptive testing is one of the most popular research issues in statistical and psychological measurement. The purpose of this study is to compare several commonly concerned item selection criteria in different typical testing conditions for dichotoumous and polytomous testing data. Two simulation studies were conducted to explore ability parameter estimation accuracy and item exposure rate for these criteria with the assumption of multidimensional two parameter logistic model and multidimensional graded response model could fit the testing data well, individually. Results showed that the criterion of Bayesian A-Optimality generally performs best both for the two item response theory models from the perspective of the above evaluation indices. As for the three-dimensional case based on the two models, A-Optimality was a relatively bad criterion in terms of ability parameter estimation accuracy.