Fair, Explainable Early Warning for First-Year Student Dropout: Accuracy, Group Equity and Support Allocation

Working paper (v1.0, October 2026). Not peer reviewed; results are preliminary and may contain errors. Colleges use early-warning systems to decide which students receive outreach, tutoring and advising, yet published audits show these tools can be inaccurate and unequal across student groups. This working paper asks three questions an institution should answer before relying on such a tool: Is a transparent model accurate enough? Is it equally accurate for every student group? And who should receive support when capacity is limited? Using public records for 4,424 students (UCI Predict Students’ Dropout and Academic Success), we compare transparent logistic and multinomial-logit models, built at enrollment and after the first semester, against a grades-only baseline. We audit their errors across four student groups, and we formulate the allocation of limited support as an optimization problem with fairness quotas, proving that a simple greedy rule solves it optimally. Key findings (means across five data splits): A first-semester model that never sees demographic attributes reached AUC 0.897 and caught 66.8% of eventual dropouts at a 5.5% false-alarm rate. Entry grades alone reached AUC 0.564. The model under-detected dropouts among scholarship holders by 27 percentage points and among younger students by 25 points, consistently across splits. Group-specific thresholds narrowed these gaps but sharply increased false alarms. With support capacity for 10% of students, model-based targeting reached three times as many eventual dropouts as random selection. Fairness quotas cost about 2 of 132 support slots. Files: the working paper (PDF), and fair-early-warning.zip, which contains R code that reproduces every table and figure (base R only). The data are available from the UCI Machine Learning Repository (CC BY 4.0, DOI 10.24432/C5MC89) and are not redistributed here.