DİF analysis with a multidimensional multilevel mixture model with covariates

The aim of this study is to compare the manifest groups using the differential item functioning (DIF) and the latent classes created with mixture models in terms of the number of items with DIF and effect size levels, and to investigate the nature of the DIF. In this direction, a data set of 20 items (10 mathematics and 10 science items) from eTIMSS 2019 booklet 1, which includes 22 countries, was created. The structure of the data set was analyzed and it was determined that it displayed a multidimensional and multilevel structure. DIF analyses were conducted for mixture models with the variables of gender and the number of books at home, which were taken as manifest group variables. The number of books at home was selected as a covariate for the mixture models. According to the results of the study, for the manifest group variables, there were negligible DIF items only at effect level A, while many items were found at effect levels B and C according to the mixture models. In the latent country classes created for each model, the country distributions of the high and low scoring classes were examined. As a result of the model comparisons, it was found that the 2 PL model fit the data better. In addition, in the context of the data set, the subject and cognitive domain of mathematics and science items and the structure of latent classes and DIF sources emerged more clearly.