Multilevel analysis of student ratings with missing level-two covariates: a comparison of imputation techniques

We analyse the relationship between student ratings of university courses and several characteristics of the student, the course and the teacher. In particular, we exploit data from a survey collecting information about teacher beliefs and practices at the University of Padua in a.y. 2012/13. Student ratings are nested into classes, calling for multilevel modelling. However, due to survey non-response, the information about beliefs and practices is missing for about half of the teachers, posing a serious issue of missing data at level 2. To avoid listwise deletion, we make multiple imputation via fully conditional specification, exploiting information at all hierarchical levels. The proposed approach turns out to be effective. We found that some of the teacher beliefs and practices are significantly related to student ratings.