From prediction to early policy action: machine learning early warning for primary school dropout in Burkina Faso's EMIS

---

Résumé

School dropout remains a major obstacle to achieving Sustainable Development Goal 4 in low-income and fragile contexts, where it reinforces intergenerational poverty and social inequalities. In Burkina Faso, rural poverty, insecurity, and gendered social norms combine to produce persistent dropout, particularly in hard-to-reach areas. This article examines how machine learning models can be integrated into routine education data systems to support early detection and prevention of primary school dropout. Using administrative data on 12,217 pupils from 55 schools taking part in the KIX-SEEDS ‘Une classe dans la main’ initiative (2024-2025), six supervised algorithms were evaluated on a modelling cohort of 11,365 pupils, including 200 cases of school dropout, using training, validation and test sets comprising separate schools. Despite the rarity of the events, administrative data enabled the identification of useful predictive signals, but performance declined as the number of available cases decreased. In the first term, the probit model achieved an F1 score of 0.50, whilst Random Forest had an AUC-PR of 0.49, substantially higher than the observed prevalence. In the second term, logistic regression achieved an F1 score of 0.26 and an AUC-PR of 0.22, detecting 11 of the 34 dropouts amongst 52 pupils flagged. The average of the Random Forest and probit probabilities produced a more limited list of alerts, with six dropouts detected amongst 20 pupils flagged. In the third term, only eight dropouts were present in the test set and no model exceeded an F1 score of 0.03. The main predictors varied by term and included, in particular, absences, parents' educational attainment, academic performance, year group, the operation of the school canteen and teachers' training. At the school level, term-by-term risk scores could help teachers and headteachers identify at-risk pupils as early as Term 1 and prioritise home visits, remedial support, or referrals to social services. At local government level, the prominence of absenteeism, repetition, and parental education supports investments in teacher professional development, school feeding programs, and community-based family literacy initiatives. At national level, embedding early warning algorithms into the Education Management Information System (EMIS/SIGE) would enable systematic, cost-effective monitoring of dropout risk, strengthening the implementation of the sector plan and the SDG 4.1 agenda on out-of-school children. The study thus contributes to debates on data-driven governance in education and the role of EMIS-based early warning systems in making progress towards more equitable and inclusive education in sub-Saharan Africa.

Mots-clés

School dropout, Machine learning, Early warning system, Interpretability, Education policy, Burkina Faso, Sub-Saharan Africa