UCLA

In the field of psychometrics, there has been an increase in interest concerning the evaluation of fairness in standardized tests for all groups of participants. One possible feature of standardized tests is a group of testlets that may or may not contain differential item functioning (DIF) favorable to one group of participants over another. A testlet is a cluster of items that share a common stimulus. In this dissertation, a DIF detection method useful for testlet based data was developed and tested for accuracy and efficiency. The proposed model is an extension of the generalized full-information item bifactor analysis model. Unlike other IRT-based DIF detection models, the proposed model is capable of evaluating locally dependent test items and their potential impact on the DIF estimates. This assures the new capability of the bifactor DIF detection method that was not evident in previous methods. Item parameters were estimated using a maximum likelihood estimation (MLE) method producing expected a posteriori (EAP) scores. Using the restrictions of a bifactor model, the dimensionality of integration can be analytically reduced and the efficiency can be increased. Following prior research regarding DIF on a PISA dataset, the proposed DIF model was applied to mathematics items of the Program for International Student Assessment (PISA) 2009 dataset to confirm the utility of the model. After the meaning of results to the PISA research community is conveyed, a simulation study was conducted to provide concrete evidence of the model's utility. Finally, limitations of this study from computational and practical standpoints were discussed, as well as directions for further research.