A Multinomial and Predictive Analysis of Factors Associated with University Dropout

Abstract

The phenomenon of dropout, by its complexity and educational and social impact, has been extensively studied to understand the specific causes. In this line of research, the purpose of this study was to analyze explanatory and predictive models of student dropout from university studies at the Instituto Tecnológico de Costa Rica (TEC), based on many variables recorded in the institutional system indicators. The first stage of the analysis considered multinomial regression models to identify the influence of these variables on the dropout. In the second analysis, six machine learning algorithms were evaluated in order to find a model that would predict student dropout. Data analysis showed that the probability of dropping out is related to sociodemographic variables, study program, academic history, scholarship and other benefits, and performance after first semester. In addition, the best predictor of dropout algorithm was the “random forest”, a probability of 0.83 to predict the dropout correctly and to capture 34% of the actual student dropout. These results are the first step toward building a more robust predictive model of dropout, which will contribute to preventive decision making in this university.