Interpretable machine learning for predicting school absenteeism at-scale in Brazil: evidence from 22,000 schools

Area of Science:

• Education
• Machine Learning
• Public Health

Background:

• Persistent high rates of school absenteeism negatively impact student development.
• Effective interventions require identifying at-risk schools and understanding contributing factors.

Purpose of the Study:

• To predict and identify secondary schools in Brazil at risk of elevated student absenteeism.
• To analyze the influence of school-level factors on absenteeism rates.

Main Methods:

• Utilized data from 22,476 Brazilian secondary schools, including school infrastructure, human resources, regional characteristics, attendance, and demographics.
• Compared five machine learning models, selecting Random Forest for its predictive performance and interpretability.
• Applied SHapley Additive exPlanations (SHAP) for detailed predictor importance analysis across school complexity levels and conducted "what-if" counterfactual scenarios.

Main Results:

• Identified key school-level predictors (infrastructure, human resources, regional characteristics) for student absenteeism.
• Demonstrated how predictor importance varies with school complexity, offering nuanced insights.
• Counterfactual analyses suggest policy-relevant adjustments can mitigate absenteeism risk.

Conclusions:

• Machine learning models effectively identify schools at risk of high absenteeism.
• Understanding the interplay of structural, staff, and contextual features, moderated by school complexity, is vital for targeted interventions.
• Findings provide actionable insights for developing more equitable educational policies and interventions to reduce student absenteeism.