Statistical and machine learning models for predicting university dropout and scholarship impact.

Although student dropout is an inevitable aspect of university enrollment, when analyzed, universities can gather information which enables them to take preventative actions that mitigate dropout risk. We study a data set consisting of 4,424 records from a Portuguese higher institution. In this study, dropout is defined from a micro-perspective, where field and institution changes are considered as dropouts independently of the timing these occur. The purpose of this analysis is twofold. First, we aim to build predictive models to learn of the significant socioeconomic and academic features associated with students' dropout risk. Another goal is to understand the relationship between financial status and dropout, especially the causal effect of being a scholarship holder. Propensity score matching is conducted first with the training set to better estimate the causal effect of being a scholarship holder on dropout status while controlling confounding variables. The predictive classifiers evaluated are Lasso regression, generalized additive model, random forest, XGBoost and single-layer neural network. The XGBoost model has the highest F1-score 0.904. According to this model, the most important features predicting dropout status are the student's second semester grades and the number of units they are credited. Whether a student's tuition fee is up to date, whether they owe money to a debtor, whether they are scholarship holders, and students' age at enrollment are also found to be important features. The Generalized Additive Model (GAM) performs competitively and offers clear interpretability, revealing how changes in actionable variables influence dropout risk. It shows that receiving a scholarship leads to the reduction in the odds of dropping out by nearly 40%, or by 22.2% in terms of probability when holding other factors fixed. As the study is based on data from a single institution and time period, and unobserved confounders cannot be fully ruled out, results should be interpreted with caution.