Local Multilevel Modeling for Comparison of Institutional Performance

We propose a general methodology for evaluating the quality of public sector activities such as education, health and social services. The traditional instrument used in comparisons of institutional performance is Multilevel Modeling (Goldstein, H., Multilevel statistical models, Arnold, London, 1995). However, rankings based on confidence intervals of the organization-level random effects often prevent to discriminate between institutions, because uncertainty intervals may be large and overlapped. This means that, in some situations, a single global model is not sufficient to explain all the variability, and methods able to capture local behaviour are necessary. The proposal, which is entitled Local Multilevel Modeling, consists of a two-step approach which combines Cluster-Weighted Modeling (Gershenfeld, N., The nature of mathematical modeling, Cambridge University Press, Cambridge, 1999) with traditional Multilevel Modeling. An example regarding the evaluation of the “relative effectiveness” of healthcare institutions in Lombardy region is discussed.