Using Mixed-Effects Modeling to Compare Different Grain-Sized Skill Models

Abstract

Most assessments, like the math subtest of the SAT or the GRE, are unidimensional, in that they treat all questions on the test as sampling a single underlying “skill”. Can we predict state tests scores better if we tag the questions with fine-grained models the skills needed? Psychometricians don’t do this presumably because they don’t get better fitting model, for a variety of reasons. We are investigating if we can do better prediction with finer-grained skill models. Our result gave a confirmative answer to this question.