When Administrative Networks Fail: Curriculum Structure, Early Performance, and the Limits of Co-enrolment Social Synchrony for Dropout Prediction in Engineering Education

Social integration theories suggest that students embedded in supportive peer networks are less likely to drop out. In learning analytics, this has motivated the use of social network analysis (SNA) from institutional co-enrolment data to predict attrition. This study tests whether such administrative network features add predictive value beyond a leakage-aware, curriculum-graph-informed model in a long-cycle Civil Engineering programme at a public university in Argentina. Using a three-semester observation window and a 16-fold leave-cohort-out design on 1,343 students across 15 cohorts, we compare four configurations: a baseline model (M0), baseline plus network features (M1), baseline plus curriculum-graph features (M2), and a full model (M3). After a leakage audit removed two post-outcome variables that had produced implausibly perfect performance, retrained models show that M0 and M2 achieve F1 = 0.9411 and ROC-AUC = 0.9776, while adding network features systematically degrades performance (M1 and M3: F1 = 0.9367; ROC-AUC = 0.9768). We conclude that in curriculum-constrained programmes, administrative co-enrolment SNA does not provide additional risk information beyond curriculum topology and early academic performance.