Latent Impact and Differential Item Functioning Analysis for Asymmetric IRT Models
Gabriel Wallin, Qi Huang · arXiv (2026) · doi:10.48550/arxiv.2605.05684 · gemini-3.1-flash-lite
TL;DR
Educational assessments often assume test items function symmetrically, but this frequently conflicts with real student response processes. The authors provide a statistical method to detect unfair test items (DIF) and student group differences even when we do not know the group labels or which items are fair. This helps educators and test developers distinguish between actual test bias and legitimate differences in student ability levels.
Method and sample: The study utilized a simulation design (N=500, 1000, 3000) and two empirical datasets: TIMSS mathematics items (N=605, Singapore/Morocco) and a university mathematics test (N=3000). The model uses a complementary log-log (CLL) link and Gumbel-distributed latent traits.
#IRT#psychometrics#differential item functioning#asymmetric models#latent class analysis#measurement invariance