Modeling Nonlinear Ability Trajectories and Learner Heterogeneity in Online Learning: A Bayesian Nonparametric Dynamic IRT Framework
Zhihua Ma, Alice Xu, Icy Zhang, Guanyu Hu · arXiv (2026) · doi:10.48550/arxiv.2606.15525 · gemini-3.1-flash-lite
TL;DR
The researchers created a new statistical model to track how student knowledge changes over time during an online statistics course. They found that students generally fit into four distinct groups, but their ability levels were remarkably stable throughout the semester. Surprisingly, simple measures of student activity, such as total time spent or the number of sessions, did not help predict improvements in student performance.
Method and sample: The sample consists of 198 undergraduate students (out of 213 enrolled) in an introductory statistics course using the CourseKata platform during Winter 2023. The methodology employs a Bayesian nonparametric dynamic IRT model incorporating B-spline basis expansions for engagement effects and a Mixture-of-Finite-Mixtures (MFM) prior. The model is validated via a Monte Carlo simulation study with 100 independent replications mimicking the empirical setting (N=198, T=9, J=230).
#Dynamic IRT#Bayesian Statistics#Learning Analytics#Nonparametric Modeling#Educational Measurement#Psychometrics