This paper used the MIMIC-interaction model to detect the uniform and non-uniform DIF at the same time, and compared the performance with MIMIC method, and IRT-LR-DIF method. Although results showed the MIMIC-interaction had great power in detecting non-uniform DIF, type I error for this method was inflated. MIMIC-interaction could not show better performance when compared with IRT method.
Q & F:
1. Data were simulated based on IRT model, then the IRT-LR-DIF used true model to do data analysis, so the results should be expected to favor the IRT-LR-DIF method. What would happen if data were generated based on the assumption and framework under MIMIC? And what would be IRT method’s performance to analyze this kind of data?
2. it is strange why LMS was used. Stated by the author, LMS was recommended by Mplus’s user guide, but they also stated that, LMS is not appropriate when one variable is categorical. Then when the type I error inflated for MIMIC-interaction model, the author said it was the reasonable results for LMS and future research needs to select a good alternative method. In other words, before the data analysis, the disadvantage for LMS was obvious, then after using this method, the disadvantage became the excuse for the bad results. This logic sounds very strange….
3. In order to specify correct anchor items, purification procedure is recommended for future study.
4. what kinds of data are recommended to use MIMIC-interaction method, instead of IRT-LR-DIF method?