Termination Criteria in Computerized Adaptive Tests: Do Variable - Length CATs Provide Efficient and Effective Measurement?

Abstract

This simulation study examined a number of computerized adaptive testing (CAT) termination rules using the item response theory framework. Results showed that longer CATs yielded more accurate trait estimation, but there were diminishing returns with a very large number of items. Standard error termination performed quite well in terms of both administering a small number of items and having high accuracy of trait estimation if the standard error level used was low enough, but it was sensitive to the item bank information structure. Change in estimated  performed comparably to standard error termination but was less sensitive to the bank information structure. Fixed-length CATs performed either slightly worse or comparable to their variable-length termination counterparts; previous findings stating that variable-length CATs are biased were the result of artifacts, which are discussed. Recommendations for CAT termination are provided.