How early is early enough? Temporal sensitivity analysis of dropout prediction across course timelines using clickstream data
NexusTech · 2026
DOI: 10.31893/tech.2026019 ↗Kurz gefasst
Fachartikel in der Zeitschrift NexusTech, 2026. Verfasst von Rani Santika. Die Arbeit nennt 35 Literaturangaben.
Abstract (Originalsprache)
Early warning systems for student dropout face a core tension: predictions become accurate only late in a course, after the intervention window has largely closed, whereas earlier predictions are feasible but error prone. Dropout is among the most widely modelled outcomes in learning analytics, yet the temporal trade-off between accuracy and lead time is poorly characterised, as most studies report performance at a single course point. Using the Open University Learning Analytics Dataset (32,593 students; over 10 million clickstream records), we built cumulative behavioural features at six checkpoints (weeks 2, 4, 8, 12, 16, 20) and trained logistic regression, random forest, and XGBoost at each under a presentation-forward protocol that excludes already-withdrawn students. Discrimination rose only from AUC 0.58 to 0.61 at week 2 to 0.68 to 0.75 at week 20, far below the 0.90 or higher usually reported for full-course designs. Precision for the dropout class fell as the residual base rate dropped from 20.8% to 7.8%. No checkpoint met the pre-specified acceptability criteria (recall ≥ 0.70, precision ≥ 0.50); under a relaxed precision tolerance (≥ 0.20), logistic regression was actionable at weeks 2 and 4, and at a tolerance of ≥ 0.15 through week 12, so early detection carried little recall cost, the binding constraint being alert precision. Demographic features helped only in sparse early windows (change in AUC of 0.033, gone after week 2), and SHAP attribution shifted from early engagement to assessment trajectories. We propose a temporal utility frontier linking predictive performance to remaining intervention time, guiding when institutions schedule warnings.
Zitieren
APA
Santika, R. (2026). How early is early enough? Temporal sensitivity analysis of dropout prediction across course timelines using clickstream data. NexusTech. https://doi.org/10.31893/tech.2026019
Vancouver
Santika R. How early is early enough? Temporal sensitivity analysis of dropout prediction across course timelines using clickstream data. NexusTech. 2026. doi:10.31893/tech.2026019
BibTeX
@article{Santika2026,
author = {Santika, Rani},
title = {How early is early enough? Temporal sensitivity analysis of dropout prediction across course timelines using clickstream data},
journal = {NexusTech},
year = {2026},
doi = {10.31893/tech.2026019}
}
EUVIMED