基于GRM模型的CAT分层方法在校准误差中的应用研究

在计算机化自适应测验(CAT)中,0-1评分模型下b组块a分层的方法(BASTR)可以提高测量准确性的同时平衡项目的曝光率,但在多级评分模型中项目难度/步骤参数有多个,无法直接使用该方法; 又因为信息函数可以较好地综合被试能力和项目参数,但最大信息量选题策略的测验安全性太低.因此,将多级评分模型中的多个参数综合成一个指标作为b分块的依据,模仿BASTR方法,提出5种新的B分块a分层方法,并且采用“影子题库”下最大信息量的选题方法.在等级反应模型(GRM)下蒙特卡洛实验结果表明,新方法在测验精度、题库利用率和机会红利等评价指标中总体表现良好,B_max-min分块方法表现最优.

The Several Stratified Methods of CAT in the Presence of Calibration Error on GRM

For dichotomous scoring,the a-stratified method with b blocking(BASTR)is an effective and safe method for computerized adaptive testing(CAT).But it could not be applied to the polytomous scoring CAT,because there are too many parameters in the polytomous item response model.It is well known that the Fisher information function is a good comprehension of all item parameters as well as the ability parameter,but the maximum Fisher information(MFI)method derogates the security of CAT.Five new stratified methods are proposed in this paper.The new methods are comprehension of all information of the item parameters for polytomous items and play the role of BASTR.Because "shadow pool" can improve the uniformity of item bank, so the item select strategy is MFI under "shadow pool".The results of Monte Carlo study of graded response model(GRM)show that the new methods has better effect, and B_max-min method is the best one.