Evaluating measurement invariance is essential for ensuring comparable test scores across individuals and groups. Although unidimensional IRT models are widely used, their local independence assumption is often violated in assessments with item clusters, such as reading comprehension tests, where bifactor models offer a practical alternative. Using data from the Norwegian National Reading Test (2022–2023) and simulations, we examine the impact of applying misspecified unidimensional models and alternative parameter restrictions in multiple group bifactor models (whether specific slopes are constrained across groups). Results show that unidimensional models yield biased inferences when specific slopes differ, highlighting important implications for applied measurement invariance practice.
CREATE Open Seminar - Assessing Measurement Invariance with Bifactor IRT Models
This seminar will be held by Lu Zhang. In this talk, she will explore how different statistical modeling choices can affect whether test scores are truly comparable across groups. Using examples from large-scale reading assessments and simulation studies, the seminar highlights why commonly used models may sometimes lead to misleading conclusions, and how alternative approaches can improve measurement practice.
Lu Zhang, Doctoral Research Fellow, CREATE, CEMO, University of Oslo.
Published Mar. 5, 2026 2:32 PM
- Last modified Mar. 5, 2026 2:32 PM