ContentMIRT: Context-Augmented Multidimensional Item Response Theory with Neural Latent Regression and Sparse Moderation
Abstract In large-scale educational assessments (e.g., TIMSS, PISA), item calibration and background conditioning are conventionally conducted via multi-stage pipelines, where linear latent regression fails to capture non-linear saturation thresholds while ignoring item-level Differential Item Functioning (DIF). To address these challenges, we propose ContentMIRT (C-MIRT), an end-to-end variational deep generative measurement framework that unifies multidimensional item response modeling, non-linear neural latent regression, and sparse moderation. C-MIRT incorporates a Tabular Neural Context Prior to map heterogeneous background variables and digital process timing paradata directly into individualized latent trait prior distributions. Simultaneously, an $L_1$-regularized Sparse Moderation Decoder isolates item-specific DIF shifts, decoupling construct proficiency from measurement bias under formal majority-invariance conditions. To ensure metric determinacy without ad-hoc rotational restrictions, we employ positive softplus discrimination constraints alongside an exact Orthogonal Procrustes alignment protocol. Across Monte Carlo simulations spanning 30\% to 85\% response sparsity, C-MIRT consistently surpasses standard and linear latent regression MIRT in out-of-item predictive accuracy and discrimination recovery, while two-stage screening secures high DIF detection precision. Across six empirical assessment cohorts from TIMSS 2023 and PISA 2022 spanning three performance tiers (Singapore, USA, Chile), C-MIRT exhibits strong linear alignment with operational Plausible Values ($r_\theta > 0.81$), yields a 26.0\% (PISA Singapore) to 39.2\% (TIMSS Singapore) reduction in posterior trait uncertainty, and reveals via gradient attributions that behavioral timing paradata account for over 55\% of prior predictive influence. These findings demonstrate that integrating neural latent regression and sparse moderation provides an expressive, interpretable, and scalable framework for context-augmented cognitive measurement.
Authors
- QIN Hong
- Yong Zhang
- Bo Li
Institutions
- Zhongnan University of Economics and Law (CN)
- Central China Normal University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-10972910/v1
- Primary Topic
- Psychometric Methodologies and Testing
- Type
- preprint