The DIF-Free-Then-DIF Strategy for the Assessment of Differential Item Functioning
This paper first introduced three methods to build common metrics for DIF assessment, which are the EMD, the AOI, and the CI. There are assumptions of these three methods, and they were not easy to meet the assumptions for the EMD and the AOI once DIF rate is high. Thus, only CI is recommended for such situation. Actually, Woods (2009) proposed the RB method to assess DIF under large sample size and polytomous model. Results shown that accuracy was quiet high, however, it would decrease with increasing of DIF percentage. As aforementioned, the RB method itself does not consider purification and the performance of the RB under dichotomous item and small sample size is unknown. Therefore, the authors here conducted two simulations to assess DIF. For the first simulation, the RB and the RB-S (RB with purification) were implemented to evaluate DIF; and for the second simulation, different conventional DIF methods and the RB-S were used to assess DIF. The results illustrated that the RB-S was outperform than the RB, particularly it could well controlled type one error when ASA is constant for the first simulation; results of the second simulation, the RB-S method also could well controlled on type one error than all the other DIF methods.
Sharing, Question and Future study:
1) DFTD strategy is a quite interesting idea. However, as the authors mentioned in the conclusion, once many DIF items exist, the risk would be increase, that means the anchor items may easily include DIF items and the performance of DIF would be affected by unclean anchor set.
2) The RB-S method can be used to assess uniform DIF and compared with the MIMIC model.