Assessing the quality of institutions' rankings obtained through multilevel linear regression models

The aim of this paper is to assess the quality of the ranking of institutions obtained with multilevel techniques in presence of different model misspecifications and data structures. Through a Monte Carlo simulation study, we find that it is quite hard to obtain a reliable ranking of the whole effectiveness distribution while, under various experimental conditions, it is possible to identify institutions with extreme performances. Ranking quality increases with increasing intra class correlation coefficient and/or overall sample size. Furthermore, multilevel models where the between and within cluster components of first-level covariates are distinguished perform significantly better than both multilevel models where the two effects are set to be equal and the fixed effect models.