Comparison Study Of Two Multiple Group DIF Methods

Detection of Differential item functioning (DIF) items is the first and a crucial step for ensuring the fairness of a test, which therefore has drawn a lot attention around the world. But currently, the studies regarding DIF is mostly focused on the procedures that are used to detect DIF between two groups. While with the development and popularity of international assessment programs, such as PISA, TIMSS, and the need for developing and studying DIF procedures for multiple groups increases. But there is only a little studies about multiple DIF procedures. And through a literature research, I found there is no simulation comparison study about genralized Logistic Regression and generalized Lord X2 method, thus this study chose these two methods.Firstly, this paper briefly introduces the concept of DIF and DIF procedures for detecting between two groups, and then reviews the literatures of multiple DIF detecting methods; secondly, introduces the model and formulae of these twwo multiple DIF methods:GLR and GLord. At last, there are three studies in this paper:Study 1, as a pre-experiment, simply compares these two mutiple DIF method with their corresponding two-group methods(LR,LR with a bonferroni adjusted a level, Lord X2 and LordX2 with a bonferroni adjusted a level). The results show that the type I error rate of LR and Lord X2 are quite inflated with the number of focal groups increases. And the BLR and BLord X2 has low power rates with type I error rate is controled within nominal error level. And GLR and GLord X2 can control the type I error rate well within nominal error level. So these twwo methods are recommended when there are multiple groups.Study 2 emphasizes on the comparion between GLR and GLordX2. Rasults show that the type I error rates of both methods decreases with sample size increases but increases with the number of DIF groups increases. When the ability distributon of groups differ, both methods’type I error rate are inflated. About power,no matter whether the ability distribution differs or not, it always increases as sample size increases,while decreases and then decreases as the number of DIF items and DIF groups increases.Study 3 uses part of PISA 2009 reading iteracy data to compare the results of two methods. Results show that the items and the number of items both method identified as DIF items are almost the same. And the percentage of items identified as DIF is quite large.