Latent Dirichlet Allocation-Augmented Structural Equation Modelling for Psychometric Calibration of Adaptive Formative Assessment Instruments in Higher Education Contexts

Abstract

Contemporary adaptive formative assessment frameworks increasingly demand psychometrically robust calibration pipelines capable of reconciling construct validity with algorithmic personalisation. This study proposes a novel hybrid methodology integrating Latent Dirichlet Allocation (LDA) topic-modelling with Structural Equation Modelling (SEM) to optimise item-response calibration within computer-adaptive testing environments across undergraduate STEM cohorts. Employing a multi-stage validation design across four institutional sites (N = 2,847), the framework demonstrated superior discriminant validity (RMSEA = 0.041; CFI = 0.963) relative to conventional Item Response Theory baselines. Differential Item Functioning analysis further confirmed cross-demographic measurement invariance. Findings indicate that LDA-augmented latent trait estimation substantially reduces construct-irrelevant variance in adaptive item sequencing, establishing a replicable, scalable optimisation protocol for formative assessment instrument development in post-secondary educational settings.