glossaryPage.heroH1

Significant. Per NEPC research, AACE peer-reviewed work, and SoLAR ethics framework: (1) Historical bias — models trained on historical institutional data encode historical patterns of bias against under-represented student groups (which historically had lower retention because of institutional and societal factors, not because of ability). The model then perpetuates the pattern. (2) Self-fulfilling prophecy — students flagged at-risk who are treated as at-risk may internalise the label, with longitudinal effects. (3) Resource-allocation bias — if intervention resources are limited and the model concentrates resources on flagged students, students just-below-the-flag-threshold may receive less support. (4) Demographic correlation — risk scores often correlate with race, socioeconomic status, and first-generation-college status not because those factors cause failure but because they correlate with institutional under-support. Best practice: regularly audit per-demographic-group flagging rates, retain human-in-the-loop intervention decisions, and treat the model as advisory.

Per Civitas Learning aggregate data and similar research, well-tuned models identify roughly 60-80% of eventually-failing or eventually-withdrawing students 6-12 weeks before midterm grades make the problem visible. False-positive rates of 10-20% are typical (students flagged who would not actually have failed without intervention — though the design goal is to prevent failure, so all flagged students get intervention). Per EAB Student Success Collaborative research, model accuracy matters less than institutional response capacity — institutions with strong advisor-follow-up workflow see retention lift even with mediocre models; institutions with strong models but weak advisor response see minimal lift. Institutional readiness to act on the prediction is the determining factor.

Student performance prediction is a subset of learning analytics focused specifically on per-student outcome forecasting (will this student pass / fail / drop out). Learning analytics covers a broader scope: per-student outcome forecasting, but also curriculum effectiveness analysis (which courses correlate with degree completion vs act as bottlenecks), pedagogical pattern analysis (which teaching methods correlate with learning outcomes), and program-level effectiveness review. Per SoLAR research, learning analytics is the umbrella category; performance prediction is one application within the category.

Approaches vary. Some institutions disclose the predictive model to students at orientation and via institutional policy publishing; some treat it as internal administrative tooling and disclose only on student request; some publish aggregate model behaviour (per-demographic-group flagging rates, signal-level attribution methodology) without per-student score disclosure. Per UNESCO 2024 AI in Education transparency guidance and EU AI Act 2024 high-risk AI requirements, the EU regulatory framework requires disclosure to data subjects of high-risk AI processing affecting them; institutions in EU-affecting deployments handle disclosure formally. Best practice per SoLAR ethics framework: publish the institutional policy, provide opt-out where institutionally feasible, and retain human-in-the-loop intervention review.

Civitas Learning publishes practitioner-focused research on early-alert system implementation across member institutions. EAB Student Success Collaborative publishes peer-comparison and benchmark research for member institutions. SoLAR (Society for Learning Analytics Research) hosts peer-reviewed research at the annual LAK conference. NEPC publishes critical-perspective research on bias and equity in predictive education models. Educause Learning Analytics Initiative tracks US higher-ed adoption. The 2014 SoLAR-published "Ethics of Learning Analytics" framework remains a primary reference. Per OECD AI Principles and EU AI Act 2024, the regulatory framework for high-risk education AI provides governance baseline.