Statistical models for detection of differential item functioning

0570033 - ÚI 2023 RIV CZ eng D - Dizertace
Hladká, Adéla
Statistical models for detection of differential item functioning.
Matematicko-fyzikální fakulta, Univerzita Karlova v Praze. Obhájeno: Praha. 17. 3. 2021. - Praha: Univerzita Karlova, Matematicko-fyzikální fakulta, 2021. 166 s.
Grant CEP: GA ČR(CZ) GA21-03658S
Institucionální podpora: RVO:67985807
Klíčová slova: differential item functioning * generalized logistic regression * nonparametric methods * differential distractor functioning
Obor OECD: Statistics and probability
https://dspace.cuni.cz/handle/20.500.11956/125037

ZÁKLADNÍ ÚDAJE: Disertační práce. Matematicko-fyzikální fakulta, Univerzita Karlova v Praze. Obhájeno: Praha. 17. 3. 2021. ABSTRAKT: This thesis focuses on topic of Differential Item Functioning (DIF), a phenomenon that can arise in various contexts of educational, psychological, or health-related multiitem measurements. We discuss several statistical methods and models to detect DIF among dichotomous, ordinal, and nominal items. In the first part, generalized logistic regression models for DIF detection among dichotomous items are introduced, which account for possibility of guessing and/or inattention. Techniques for estimation of item parameters are presented, including a newly proposed algorithm based on a parametric link function. Two simulation studies are presented. The first compares the generalized logistic regression models to other widely used DIF detection methods. The second illustrates differences between the techniques to estimate item parameters. Implementation of the models into the R software and its difNLR package is illustrated. In the second part, generalized logistic regression models for DIF detection among polytomous items are discussed. Cumulative logit, adjacent category logit, and nominal models are introduced together with the maximum likelihood method to estimate itemparameters and with examples of implementation in the difNLR package. The third part deals with a nonparametric comparison of regression curves for DIF detection based on kernel smoothing. We discuss several settings and we newly propose an estimate of an optimal weight function for a test statistic to identify DIF. Nonparametric approaches are compared to the logistic regression method in a simulation study. In the fourth and last part, further topics of DIF detection are discussed, including item purification, multiple comparison corrections, and DIF effect sizes. Different approaches are compared in a complex simulation study on three of the most used DIF detection methods.
Trvalý link: https://hdl.handle.net/11104/0341403