Predicting Student Retention via Expectancy Value Theory Using Data Gathered before the Semester Begins

Abstract

This full-length research paper presents results from a machine-learning analysis of engineering student persistence at a large southeastern university. Students leave engineering school for many reasons, ranging from low math preparation to a low sense of belonging in engineering, which can be viewed through the Situated Expectancy Value Theory (EVT) framework of academic decision-making. Prior work has found many strong predictors of persistence from first-semester data, including EVT variables, but when it comes to identifying interventions, it might be better to identify predictors from earlier in the first semester. In this study, we attempted to predict student persistence using two machine learning techniques, neural networks and decision trees, and only using early data from the beginning of the first semester (EVT survey data, standardized test scores, demographic data, and Pell eligibility). We found that decision trees were better able to predict retention rates from the beginning of the semester than neural networks, as neural networks struggled to find clear signals that indicated if a student was likely to drop out of engineering school. Grouping students together who are at-risk of leaving engineering school from the beginning of the semester will allow instructors and advisors to focus their attention on those groups, and therefore improve retention rates.