Integrating the generalized graded unfolding model into a latent state-trait theory framework using the Bayesian Hamiltonian Monte Carlo Algorithm

Latent state-trait (LST) theory has been widely applied to psychological measures to account for person (trait), situation (state), and person by situation interaction effects on scores. However, previously defined LST models were developed based on linear assumptions, thereby ignoring potential nonlinear relationships between traits and responses. The goal of this dissertation was to integrate the most popular nonlinear model, the generalized graded unfolding model (GGUM; Robert, Donoghue, & Laughlin, 2000), into LST frameworks and compare its performance with linear models. Specifically, the GGUM was integrated into three types of LST models: latent trait (LT), correlated states (CS), and orthogonal methods (OM).

An overall sample of 1165 respondents completed the Conscientiousness subscale from the Big Five Inventory (BFI; John, Donahue, & Kentle, 1991) on two occasions separated by one week. Results revealed that LST GGUMs fit better than corresponding LST linear models in relation to conventional structural equation modeling (SEM) fit indices, even though the scale examined was intentionally developed under linear assumptions. Effects of using LST GGUMs were mainly in producing less-extreme responses for negatively phrased items. Among all calibrations, linear models best explained linearly constructed data, whereas the OM-GGUM with method factor loadings allowed to vary across occasions had the best predictive performance in terms of the posterior means. Recommendations for future investigations of LST GGUMs and uses of LST GGUMs for construction of inventories are discussed.