Uruguay's Machine-Learning Early-Warning Model for Primary-School Repetition Risk (Panel-EIT / ANEP)

Uruguay's early-childhood impact assessment (Panel-EIT) tracked 15,529 children from initial education in 2016 through primary grades 1–3 (2017–2019). Researchers from Universidad de la República, working with ANEP-CODICEN data, built supervised machine-learning models (logistic regression, Bayesian networks, classification trees) to predict which students would repeat a grade, intended to inform early-warning and intervention policy.

The best models reached 80% accuracy and an AUC of 0.81, with sensitivity of 62–64% and specificity of 83%; 16.2% of the cohort repeated at least one grade. Removing the early-childhood assessment score from the model dropped accuracy to 70–72% and sensitivity as low as 42%, confirming it as the single strongest predictor. UNICEF context data shows the stakes: only 44% of Uruguayan adolescents complete secondary education by ages 21–23, and completion for the poorest quintile (14%) lags nearly five-fold behind the richest quintile (79%).

This is strong predictive research but weak deployment evidence: the Inter-American Development Bank's own fAIrLAC pilot page on this initiative leaves its implementation status blank, and the peer-reviewed authors explicitly caution that the models establish statistical association, not causal mechanisms, over a narrow three-year observation window. It should be read as a validated research prototype rather than a proven, actively operating national early-warning system.

Read the full analysis: https://www.redalyc.org/journal/270/27069733005/html/