Multilevel models for evaluating efficacy of educational programmes: some notes on model stability

We discuss the limits in applying mixed effects models in a routine evaluation of the efficacy of public policy actions. In particular we discuss the sensitivity of these models to changes in the underlying assumptions with reference to a dataset on the evaluation of public-funded programmes for professional education in Northern Italy. The limits are evaluated with respect to the variable selection process, comparing differences in the selected model if the underlying assumptions on the random effects, on the observation correlation and on the variance function form are changed.