Multilevel Modelling of Educational Data

Abstract

Educational science was the first social science to develop fully multilevel modelling, although it had already been long used in statistics under the form of “random-effect models” or “mixture models” (Eisenhart et al., 1947). In fact, for many years, the literature on education research had been the forum for substantive discussions on the most relevant analytical unit for measuring scholastic attainment: should statistical analysis focus on the class or on the student? As Harvey Goldstein explains in this first chapter, Aitkin et al. (1981) came up with a genuine multilevel model inspired by an earlier study (Bennett, 1976), which recognised only students and teaching styles as the units of analysis, ignoring teacher groupings and class groupings: Ait-kin and his colleagues introduced the effect of these groupings while at the same time admitting a specific student effect. They thus managed to show the important role played by the teacher: this erased the sizeable differences found between students subjected to a “formal style of teaching” and to an “informal” style in the earlier study, which also referred to a “mixed style of teaching”. The choice between the class and the student ceased to be relevant. In fact, Aitkin and his co-authors argued, the effects of both aggregation levels should be examined together, so that their respective actions can be properly separated. Since that date, education researchers have accepted the need to take aggregation levels into account in order to understand the differences observed between students.