Name: Zhaoyu Wang

Dissertation Proposal Meeting

Date: Monday, August 18th, 2025

Time: 4:30 PM

Location: Virtual, click here

Dissertation Committee Chair/Advisor:

James Roberts, Ph.D. (Georgia Tech)

Dissertation Committee Members:

Susan Embretson, Ph.D. (Georgia Tech)

Brian Habing, Ph.D. (University of South Carolina)

Audrey Leroux, Ph.D. (Georgia Tech)

Dingjing Shi, Ph.D. (Georgia Tech)

Title: Integrating the Multidimensional Generalized Graded Unfolding Model (MGGUM) and Multidimensional Scaling (MDS) to Improve Estimation Accuracy and Precision 

Abstract: This dissertation proposes a hybrid model to enhance the Multidimensional Generalized Graded Unfolding Model (MGGUM) by integrating Multidimensional Scaling (MDS) to refine item location parameter estimates and reduce sample size requirements. The MGGUM, an extension of the unidimensional GGUM, has been applied across various psychological domains. It often requires large samples, especially with multidimensional data, which leads to a difficulty in the practical application of the model in social science research where smaller samples prevail. To address this limitation, our approach incorporates MDS scale values derived from similarity judgments between stimulus pairs into the MGGUM framework. This strategy is expected to improve the precision of parameter estimates within the MGGUM. Specifically, the proposed model framework simultaneously estimates the parameters of MGGUM and item coordinates of MDS by connecting them through a multivariate normal distribution. The variance covariance matrix of the multivariate normal distribution specifies the relationship between the perceptual item locations (MDS scale values derived from dissimilarities data) and preferential item locations (MGGUM item locations derived from preference or [graded] disagree-agree response data), which allows the information from these two sources to simultaneously contribute to parameter estimation across models during an MCMC procedure. We conduct a parameter recovery simulation study and perform a real data analysis to compare the performance of the hybrid model to the original MGGUM. We anticipate that the hybrid model will improve the precision of MGGUM item location estimates and perhaps reduce the need for large sample sizes to some degree. The improvement will extend the application universe of the MGGUM.