Answers are grounded in this paper and cite their sources.
Evaluating measurement invariance is essential for ensuring score comparability across individuals and groups. While unidimensional item response theory (IRT) models are commonly used, their assumption of local independence is often violated. Bifactor models provide an alternative by incorporating both a general factor and specific factors, making them suitable for assessments with locally dependent item clusters, such as reading comprehension tests based on shared text passages. This study examines measurement invariance in bifactor IRT models using marginal maximum likelihood (MML) estimation and likelihood ratio testing, with simulations motivated by the Norwegian National Reading Test (2022–2023). We first investigate the implications of using a misspecified unidimensional model to evaluate differential item functioning when data follow a bifactor structure. We then evaluate how alternative item parameter constraints in multiple-group bifactor models, constraining general slopes alone versus both general and specific slopes, impact invariance assessment. We further compare MML estimators for likelihood ratio testing, including adaptive Gauss–Hermite quadrature, Laplace approximation, and the expectation-maximization algorithm with dimension reduction. The results show that all methods yield accurate size when specific-factor slopes are the same across groups. However, when these slopes differ across groups, models that fail to account for such differences can produce inflated size.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: