Advancing test security: a new detection model for examinees with item preknowledge and compromised items

Test security is a common concern in both academic and professional assessments. When examinees are exposed to live test items before administration, their scores may no longer accurately reflect their true knowledge, abilities, or skills. A common form of test fraud involves examinees having prior knowledge of compromised items (CIs). However, most existing detection methods rely on the prior identification of these CIs. When the status of CIs is misidentified or only partially identified, the predictive performance of such models is significantly reduced.

To address this limitation, the present study proposes a new model, the Deterministic Gated Rasch-Lognormal Response Time (DG-RLn) model, designed to simultaneously detect examinees with item preknowledge (EWPs) and CIs by integrating both response accuracy and response time data.

The model s predictive performance was first evaluated using simulated data under varying conditions, including different proportions of EWPs, CIs, and aberrance levels. It was then applied to real testing data to assess its practical effectiveness.

Results indicated that when the proportion of CIs reached 40%, the proportion of EWPs reached 20%, and the aberrance level exceeded 1.5, the model demonstrated robust and reliable detection performance. The DG-RLn model presents a promising approach to improving test security by identifying EWPs and CIs without requiring prior knowledge of compromised content. However, its performance was more limited in real-data applications, potentially due to the increased complexity of examinee behavior and the presence of other types of test fraud beyond preknowledge of CIs.