【作者】 刘娟； 罗照盛； 郑蝉金；

【Author】 Juan Liu;Zhaosheng Luo;Chanjin Zheng;Jiangxi Normal University;

【机构】 江西师范大学心理学院；

【摘要】 当测验含有题组时,基于单维项目反应理论的模型不再适用,相应的功能性差异检验方法也需更新。此外,在实际测验情境中,能够收集到的被试背景信息不止一种,因此为了更好的检验题组测验在多个背景变量下的公平性,本研究对多背景变量的DBF(differential bundle funcitoning)检验方法进行了分析和探讨,并提出了一种拓展的多组DBF检验方法。由于题组模型的共通性,本研究对基于项目束水平的DBF检验方法和基于项目水平的题组DIF(differential item functioning)检验方法均进行了梳理,在项目束水平上,选取了SIBTEST方法和MIMIC-Bundle-DBF检验方法;在项目水平上,选取了基于Rasch题组反应模型、带有协变量的多维双因素模型两种题组DIF检验方法,对以上方法分别进行了介绍与评鉴,并对方法中涉及的概念、模型进行了汇总比较。然后,对前人关于多组DIF检验方法的研究进行了梳理,并总结了其在多组拓展方式上的异同点。最终本研究提出使用MIMIC-Bundle模型进行多组DBF检验,并通过模拟研究对该拓展模型在多种数据条件下的表现进行了检验。模拟研究发现:(1)当样本比例平衡时,使用MIMIC-Bundle模型进行多组DBF检验的一类错误率和三类错误率在大部分情境下均能控制在可接受的水平(0.1以内),其统计检验力也均在大部分情境下达到了0.9以上;(2)当样本比例不平衡时,一类错误率未出现较大幅度的上升,但三类错误率在部分情境下飙升至0.6以上。其统计检验力在DBF量为0.3的情境下出现了较大幅度的下降。更多还原

【Abstract】 When the test contains testlets, the model based on single-dimension IRT model is no longer applicable to testlet-based data. In addition, the number of individual’s background information which can be selected in a real-world test situation is usually far more the one type(gender), so the current analysis and discusses the detection methods for the differential bundle funcitoning in multiple background variables. In this study, the DBF detection methods based on the bundle level and the DIF detection method based on the item level were first reviewed respectively. At bundle level, the SIBTEST method and the MIMIC-Bundle method are selected; at item level, the Rasch testlet response model(RTRM) and the bi-factor model for testlets with covariates are selected to test the testlet’s DIF. All above methods are introduced and evaluated respectively, and the concepts and models involved in the method are summarized. Then, this study reviews the previous researches on multi-group DIF detection methods, and summarizes similarities and differences in the multi-group expansion methods. Finally, on the basis of literature review, this study proposes using the MIMIC-Bundle model to detect the DBF in multigroup simultaneously, and then tests the performance of the above extension simulation studies. Through analysis of simulation studies, it can be found that:(1) When the proportion of samples is balanced, the Type Ⅰ error and the type three error rate of DBF detection using the MIMIC-Bundle model can be controlled at acceptable levels in most situations(within 0.1), these power rates also reached 0.9 or more in most situations;(2) When the proportion of the sample is unbalanced, there is no significant increase in the Type I error, but the Type Ⅲ error rates are inflated to above 0.6 in some situations. The power rate has a significant decrease under the condition of DBF amount of 0.3.更多还原