From Prediction to Prevention: A Causal AI Framework for Student Dropout Risk and Targeted Interventions

Abstract

This study proposes a Causal AI framework that advances student dropout analytics from mere prediction to actionable prevention. Using the UCI ‘Predict Students’ Dropout and Academic Success’ dataset (N = 3,630), we implement a seven-stage pipeline integrating XGBoost-based dropout prediction (ROC-AUC = 0.959), SHAP-driven explainability, domain-grounded causal Directed Acyclic Graph construction, and heterogeneous treatment effect estimation via Inverse Probability Weighting, Doubly Robust estimation, and Causal Forest models. Results demonstrate that scholarship provision causally reduces dropout probability by 14.3 percentage points (95% CI: [-0.198, -0.088]) and tuition fee support by 21.7 percentage points (95% CI: [-0.268, -0.166]). Conditional Average Treatment Effect analysis reveals substantial heterogeneity across student subpopulations, enabling individualized, CATE-ranked intervention recommendations. This framework bridges the gap between correlational prediction and causal, targeted educational interventions.