可解释性
I类和II类错误
差异项目功能
统计能力
正规化(语言学)
样本量测定
统计
项目反应理论
计算机科学
统计假设检验
数学
统计分析
计量经济学
人工智能
样品(材料)
选择偏差
机器学习
特征选择
实证研究
选择(遗传算法)
观测误差
选型
统计模型
字错误率
大样本
标识
DOI:10.1177/01466216261506348
摘要
Test bias is a critical concern in educational and psychological measurement. Differential item functioning (DIF) is widely used to identify potential test bias. Although regularization methods have shown promise in DIF detection, interpreting the resulting DIF parameters remains challenging. This study proposes a two-stage approach, termed Reg-MH, that combines regularization with the Mantel–Haenszel (MH) method. Specifically, regularization is first used to select anchor items, followed by MH testing to identify DIF items, combining the anchor selection advantages of regularization with the interpretability of MH. A pilot simulation study examined alternative regularization-based anchor selection strategies and the number of anchor items. A main simulation study evaluated the performance of Reg-MH against traditional MH and regularized DIF (Reg-DIF) methods across varying sample sizes, DIF magnitudes, DIF proportions, and ability distributions. An empirical illustration using PISA 2012 data further demonstrated its practical application. The simulation results showed that Reg-MH generally achieved a more favorable balance between Type I error control and statistical power than the competing methods. Compared with traditional MH, Reg-MH reduced Type I error rates while maintaining higher statistical power than Reg-DIF. Type I error rates generally increased with sample size, DIF magnitude, and DIF proportion, whereas statistical power increased with sample size and DIF magnitude but decreased as DIF proportion increased. Overall, Reg-MH provides a promising approach to DIF detection by integrating regularization-based anchor selection with conventional statistical testing.
科研通智能强力驱动
Strongly Powered by AbleSci AI