Name: Na Liu

School of Psychology – Ph.D. Dissertation Defense Meeting

Date: Friday, May 1st, 2026

Time: 2:00 PM - 3:30 PM

Location: Na Liu's Dissertation Defense | Meeting-Join | Microsoft Teams

Dissertation Committee Chair/Advisor:
James Roberts, Ph.D. (Georgia Tech)

Dissertation Committee Members:
Audrey Leroux, Ph.D. (Georgia Tech)
Dingjing Shi, Ph.D. (Georgia Tech)
Mark Himmelstein, Ph.D. (Georgia Tech)
Hongli Li, Ph.D. (Georgia State University)

Title: Enhancing Precision in the Generalized Graded Unfolding Model (GGUM) Using Successive Interval Judgements Indicative of Item Location

Abstract: The Generalized Graded Unfolding Model (GGUM) is an ideal-point item response theory model suited to measuring non-cognitive constructs such as attitudes, emotions, and personality. A practical limitation of the GGUM is its requirement for large samples (N > 750) to achieve stable parameter recovery. This dissertation evaluated whether integrating collateral item location information from the Method of Successive Intervals (MSI) into the traditional GGUM could improve parameter recovery and reduce sample size requirements. A Monte Carlo simulation study crossed two sample sizes (N = 300, 750), two response formats (2- and 6-category), and two MSI–trait correlations (r = .50, .80) across 10 replications per cell. Parameter recovery for item discriminations (α), item locations (δ), person locations (θ), and response category thresholds (τ) was evaluated using root mean square deviation (RMSD), posterior standard deviation (PSD), and relative parameter bias (RPB). An empirical comparison was also conducted using attitude-toward-abortion data from 40 items administered to undergraduate participants at the Georgia Institute of Technology. Simulation results showed that estimation quality was driven primarily by sample size and response format, whereas model-related effects were uniformly small and did not favor the new model. In the empirical comparison, posterior summaries for α, θ, and τ were highly similar across models, and δ differences were broadly homogeneous across replications. Overall, adding MSI-based collateral information to the GGUM did not produce consistent or practically meaningful improvements in parameter recovery under the conditions examined.