Investigating Differential Severity Across Linguistic Subgroups in Automated Scoring of Student Argumentation

This study uses Many-Facet Rasch Measurement (MFRM) to examine the extent to which computer scoring models for assessing students’ argumentation in science might be more or less severe when scoring students who have been designated as English Learner (EL) students than humans scoring the same data. We found that while no one machine scoring approach produced significant bias, performance on certain items demonstrated that one machine model had significant potential to widen performance gaps.