Files and links (1)
Thesis-XZ2.38 MB
Embargoed Access, Embargo ends: 06/29/2028
Dissertation
Interpretable explanatory item response models with machine learning
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
DOI: 10.25820/etd.008346
Abstract
Educational assessment increasingly requires models that are not only accurate, but also interpretable and fair across diverse groups of learners. Traditional psychometric models, including item response theory (IRT) and explanatory item response models (EIRMs), provide a principled framework for modeling person- and item-level variation, yet they often rely on restrictive assumptions that may fail to capture complex nonlinearities and interactions in modern educational data. In contrast, machine learning (ML) methods offer substantial predictive flexibility, but they typically ignore the hierarchical person-item structure central to psychometric modeling and are often criticized for limited transparency. This dissertation addresses this gap by proposing a hybrid framework, Explanatory Item Response Models with Machine Learning Integration (EIRM-ML), which augments the explanatory structure of EIRMS with ML-based predicted probabilities while preserving random person and item effects.
The dissertation is organized around three central questions: whether EIRM-ML improves predictive performance and calibration relative to conventional EIRMS and standalone ML models; whether interpretable ML tools can deepen substantive understanding of person- and item-level effects; and whether the hybrid approach supports more equitable prediction across demographic subgroups. To answer these questions, the study develops the conceptual and methodological foundations of EIRM-ML, presents a two-stage estimation strategy based on cross-validated ML predictions embedded in a mixed-effects psychometric model, and evaluates the framework using both simulation and empirical evidence.
A comprehensive simulation study examines the performance of EIRM-ML under linear, nonlinear, interaction, and differential item functioning scenarios. The results are designed to show that the proposed framework remains competitive when the linear model is correctly specified and provides clear gains in discrimination, calibration, and subgroup robustness when nonlinearities or interaction effects are present. An empirical application using the Open University Learning Analytics Dataset (OULAD) illustrates how the framework can be implemented in practice to model student-item response behavior using both learner characteristics and item features. The application also demonstrates how feature importance measures, partial dependence plots, accumulated local effects, and subgroup calibration diagnostics can be used to interpret predictive mechanisms and assess fairness.
This dissertation contributes to educational measurement by offering a unified framework that bridges psychometric explanation and algorithmic prediction. By combining hierarchical modeling, predictive flexibility, and interpretability, EIRM-ML provides a practical and theoretically grounded approach for next-generation assessment analytics.
Details
- Title: Subtitle
- Interpretable explanatory item response models with machine learning
- Creators
- Xiaoting Zhong
- Contributors
- Stephen Dunbar (Advisor)Catherine Welch (Committee Member)Deborah Harris (Committee Member)Kai Wang (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Psychological and Quantitative Foundations (Educational Measurement and Statistics)
- Date degree season
- Spring 2026
- DOI
- 10.25820/etd.008346
- Publisher
- University of Iowa
- Number of pages
- xiii, 129 pages
- Copyright
- Copyright 2026 Xiaoting Zhong
- Language
- English
- Date submitted
- 04/28/2026
- Description illustrations
- graphs, tables
- Description bibliographic
- Includes bibliographical references (pages 112-118).
- Public Abstract (ETD)
Educational tests are often used to make important decisions about students, programs, and learning support. Because of this, prediction models used in assessment should be accurate, easy to understand, and fair across different groups of learners. This thesis studies how to improve those models by combining ideas from educational measurement and machine learning.
Traditional psychometric models are useful because they explain how student characteristics and item features relate to performance. However, they may miss complex patterns in real data. Machine learning can detect those patterns, but it often works like a “black box,” making its results harder to interpret. This thesis proposes a hybrid framework, called Explanatory Item Response Models with Machine Learning Integration (EIRM-ML), that brings these two approaches together. The model keeps the clear structure of explanatory item response modeling while adding information learned from machine learning predictions.
The study asks three main questions: whether this hybrid method improves prediction, whether it gives meaningful explanations for why students respond as they do, and whether it supports fairer results across subgroups. To answer these questions, the thesis uses simulation studies and an empirical application based on the Open University Learning Analytics Dataset.
This research contributes a practical way to build assessment models that are more flexible than traditional methods, but still transparent enough for educators, researchers, and decision-makers to trust and use.
- Academic Unit
- Psychological and Quantitative Foundations
- Record Identifier
- 9985176870302771
Metrics
1 Record Views