The use of outliers for the evaluation of public policy activities: the first year college drop out rate in Florence.

The use of robust procedures in regression model estimation produces, as a “sub-product”, outlier data that could be usefully utilized to obtain information on specific characteristics of subsets of the population. The aim of this study is to analyze the problem of first year college drop outs at the University of Florence (Italy). A dataset formed by administrative data, collected at enrollment time, combined with information collected through a specific survey of the students of the University of Florence which were enrolled in the academic year 2001-2002, was used for this purpose. In order to identify the most important variables (covariates) affecting the response variable (Y = 1, if the student dropped out; Y = 0, otherwise), the data was first fitted with generalized linear models estimated with classical methods. The same models were then estimated with robust methods that allowed the detection of groups of outliers. These in turn were analyzed by descriptive methods to determine the possible common individual and/or contextual characteristics. These characteristics should represent useful pieces of information for the implementation of academic policy changes that could affect the drop out rate of college students.