The Use And Comparison Of Three IRT Models For Analyzing Differential Item Functioning

Currently,in the field of psychological and educational measurement,research on fairness of tests are mainly focused on differential item functioning(DIF).As the detection and estimation of DIF are conducted based on the comparison of different groups,the division of examinees is a very fundamental and important step of relevant research.Traditionally,methods of DIF analysis,such as IRT models,divide examinees according to manifest grouping variables;however,more recently,a number of research advocate grouping by latent,rather than manifest,categorical variables,and have proposed some corresponding methods like mixture IRT models.Considering both of them,a study combined the two orientations of grouping and developed a new model,named mixture IRT-MIMIC model,to analyze DIF on the basis of manifest grouping and latent classification simultaneously.Moreover,the study compared this model with IRT model and mixture IRT model.Nonetheless,previous studies did not examine mixture IRT-MIMIC model under the conditions in which ability difference between groups or/and ability difference between classes exists,and did not compare it with IRT model or mixture IRT model in such conditions.In response to this situation,the author conducted a simulation study.Aiming at exploring the performance of Rasch model,mixture Rasch model,and mixture Rasch-MIMIC model on DIF analysis in different conditions,the simulation study included "distribution of ability","overlap ratio.of manifest group and latent class",as well as "proportion of DIF items exhibiting both kinds of DIF" as variable factors,manipulated these factors and simulated 4×3×2=24 conditions,generated response data in accordance with each condition,and then estimated parameters of each model via MCMC algorithm.Finally,the main results are as follows:(1)If an item exhibits DIF between latent classes,mixture Rasch-MIMIC model would perform better than Rasch model on the estimation of DIF between manifest groups as well as difficulty of the item;(2)If an item exhibits DIF between manifest groups,mixture Rasch-MIMIC model would perform better than mixture Rasch model on the estimation of DIF between latent classes as well as difficulty of the item;(3)In conditions with high overlap ratio of manifest group and latent class as well as 50%DIF items exhibiting both kinds of DIF,mixture Rasch-MIMIC model performs better than Rasch model and mixture Rasch model on the detection of DIF items.The results suggested that the advantage of mixture Rasch-MIMIC model to Rasch model and mixture Rasch model is not absolute,but conditional.The construction of the dissertation is as below:Firstly,it was an overview of relative research at home and abroad.Based on it,the research problem as well as research purpose were fixed,and research significance as well as research idea were illuminated;Secondly,according to research problem and research purpose,the theoretical basis of the three models and their application to DIF analysis was elaborated pointedly;Thirdly,the paper set relevant elements in the simulation study,including fixed factors,variable factors,the method to estimate parameters,priors of models,indicators to evaluate models,data generation,data analysis,and so on;Finally,based on convergence of estimation,estimation and detection of DIF,as well as estimation of difficulty,the results were organized,discussed,and concluded.