Files
Abstract
Response styles may be found in self-report data, particularly when using Likert-type scales, and can result in response biases. This is because these biases can contaminate observed scores and the factor structure of the data due to spurious correlations between items and scores. When this occurs, the validity and reliability of the data are threatened. Item response tree (IRTree) models have been used for studying response styles although current studies using the IRTree models have focused only on measuring the preference of response styles. As a result, little research has been reported fully reflecting the reality of the impact of these styles. The purpose of this study is to propose a new psychometric model, a mixture item response tree model (the MixIRTree) that is a combination of a mixture item response theory (MixIRT) model and an IRTree model.
In this dissertation, we present two studies examining this general model. In the first, the relative performance is examined of the MixIRT and IRTree models in analyzing Likert-type response data. First, it appeared that the best candidate model was different depending on the model selection methods. Second, the target latent traits of both models differed because the effect of the response category "Not sure" was accounted for in the IRTree model.
In the second study, the proposed MixIRTree model is investigated further, including an extension of the model to a random component (the RC-MixIRTree model) for incorporating an item covariate information such as reverse-coded items. The MixIRTree and the RC-MixIRTree models were applied for detecting differential item functioning (DIF). In empirical study, DIF items were shown to be detected based on the content of item (i.e., the direction process) and response style (i.e., the intensity process). In simulation study, the results suggest that the MixIRTree and RC-MixIRTree produced similar performance of recovery item and person parameters regardless of the true generating models. Further, the MH method would be the recommended DIF detection method for the suggested models.