College student dropout prediction at different time points and for different populations

What is it about?

The information a university administration has about its students can be useful in creating an early warning system, that indicates which students are at risk of not finishing their studies and may drop out. Our work investigates which information is the most useful for this prediction at different times of the student trajectory, for example at the time of enrollment, after one term, and so on. We find that the most helpful information drastically changes during this period. We also analyze if early warning systems may work better for specific subgroups, for example, better for females than others, and if the most helpful information is different among these groups. We find that the differences are not very pronounced.

Featured Image

Photo by Vasily Koloda on Unsplash

Photo by Vasily Koloda on Unsplash

Why is it important?

Students who do not finish their college degrees suffer from financial and emotional consequences, and institutions are also motivated to have high completion rates. If we can successfully predict students at risk of dropping out, we may intervene before they happen. Modern prediction techniques, such as Machine Learning models are powerful tools but complex to understand. Our work offers novel insights into these models to make college dropout prediction cheaper for administrations by focusing on the most important variables. Moreover, we find that early warning systems may be ethical regarding their fairness between different student groups.

Perspectives