Evaluating Data-Minimized Early-Warning Systems for Educator Decision Support: Evidence Across Review Capacities and Educational Stages

Educational early-warning systems rank students for review when support capacity is limited. We examined how removing background information changes retrospective review-list performance in the Open University Learning Analytics Dataset (OULAD) and two separately analyzed UCI education datasets. Full-information models were compared with reduced-input models across several educational stages and four review capacities (5%, 10%, 20%, and 30%). The primary outcome was the difference in review-list precision, with paired bootstrap intervals adjusted across stages and capacities within each policy family. Sex/gender removal showed no statistically detectable difference after this adjustment; this does not establish equivalence. Removing socioeconomic/family information reduced whole-cohort predictive performance in OULAD and UCI 697 and reduced precision in selected larger OULAD lists and at UCI 697 enrollment. UCI 320 estimates were inconclusive. Some adjusted precision findings varied across calibration methods. Diagnostic models showed that excluded attributes remained partly predictable from retained records. These post-development analyses evaluate predictive performance, not causal effects on educator workload, interventions, or student outcomes. The findings support evaluating data minimization at the intended educational stage and review capacity, while separately assessing whole-cohort performance and list membership. Prospective evaluation with educators and students is needed before drawing conclusions about educational benefit.