On stability and robustness of students' dropout prediction

High dropout rate is a serious problem at universities all over the world. To study factors of students' dropout we analyzed data about students who enrolled for bachelor study at one Czech university in five subsequent academic years 2013/14–2017/18. Using decision trees and logistic regression we created several classification models to predict who will finish the study successfully and who will not. Analyzing data collected on a semester basis, we found that the most important variable for all five years is the percentage of lost credits in the most recent semester. We also assess the stability and robustness of the classification models by testing them on data about students who enrolled in the subsequent years. Here, we found that the decision trees are more stable and robust than logistic regression models.