Maximin criterion for item selection in computerized adaptive testing.

Abstract

In computerized adaptive testing (CAT), information-based item selection rules (ISRs), such as maximum Fisher information (MFI), often excessively rely on discriminating items, leading to unbalanced utilization of the item pool. To address this challenge, the present study introduced the MaxiMin Information (MMI) criterion, which is grounded in decision theory. MMI calculates each item's minimum information (I_min) within the current confidence interval (CI) of the trait level, selecting the item with the maximum I_min to be administered. For examinees with broader CIs (less precise trait estimates), MMI leans toward administering less discriminating items, which tend to yield larger I_min. Conversely, for narrower CIs, MMI aligns more closely with MFI by favoring items with higher discrimination. This indicates that MMI's item selection is tailored to each examinee based on his or her provisional trait estimate and its estimation precision. Five simulation studies were conducted to assess MMI's performance in CAT under various conditions. Results demonstrate that although MMI is comparable with other ISRs in terms of trait estimation precision, it excels in balancing item pool utilization. By fine-tuning confidence levels, MMI not only efficiently schedules the use of discriminating items toward the test's later stages to enhance test efficiency but also effectively adapts to different testing scenarios. From these findings, we generally recommend applying MMI with a confidence level of 95% to optimize item pool utilization without compromising trait estimation accuracy. With its evident advantages, MMI holds promise for practical applications, especially for high-stakes tests requiring utmost test efficiency and security.