A new perspective on detecting performance decline: A change-point analysis based on Jensen-Shannon divergence

• Published in: Behavior research methods Vol. 55; no. 3; pp. 963 - 980
• Main Authors: Tu, Dongbo, Li, Yaling, Cai, Yan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2023; Springer Nature B.V
• Subjects: Behavioral Science and Psychology; Bias; Cognitive Psychology; Computer Simulation; Humans; Probability; Psychology; Statistical analysis; Person-fit statistics; Performance decline; Jensen-Shannon divergence; Change-point analysis
• ISSN: 1554-3528; 1554-3528
• DOI: 10.3758/s13428-021-01779-z

A common observation in ability assessment is that the probability of an examinee giving a correct response drops for end-of-test items due to low motivation, time limits or other factors. On the test-takers’ side, this change can be considered performance decline (PD), which can strongly affect test validity and bias respondents’ ability estimators. Currently, there is an increasing interest in the detection of PD among researchers and practitioners. Researchers and practitioners found that PD detection fails to achieve acceptable power, which is typically below 0.55. Change-point analysis (CPA), a well-developed statistical method, can be applied to item response sequences to identify whether an abrupt change exists. Existing CPA methods cannot be directly used to detect PD because they are appropriate for two-sided alternative hypotheses. To address these issues, this research firstly develops a CPA method based on Jensen-Shannon divergence to detect PD. Additionally, existing CPA statistics were converted into one-sided statistics to accommodate PD detection. Then, a simulation study was conducted to investigate the performance of the proposed method and compare it with modified CPA statistics. Results show that the proposed CPA method can detect PD with higher power while generating a well-controlled Type‐I error rate. Compared against modified CPA statistics, the proposed method exhibits an augmentation in power from 1.0% to 8.2%, with average of 5.7% and higher accuracy in locating the change point. Finally, the proposed method was applied to two real datasets to demonstrate its utility.