SHAP-Explained Machine-Learning Models Predict Child Dropout Risk from Malawi's National Household Panel Survey

School dropout in Malawi, as in many low-income countries, results from an accumulation of household and school-level factors that are hard to disentangle without granular data — and data-collection and management systems in such settings are often financially and technically constrained.
Researchers Hazal Colak Oz, Çiçek Güven and Gonzalo Nápoles (Tilburg University) used Malawi's nationally representative Integrated Household Panel Survey (IHPS) to build and compare four machine-learning approaches — random forest, LASSO, ridge regression and a multilayer neural network — for predicting which children were at risk of dropping out of school. Applying sample weights to correct for the survey's design improved recall substantially: LASSO reached 80.6% recall and the neural network 78.8%, roughly 11-12 percentage points higher than unweighted versions. The authors used SHAP (SHapley Additive exPlanations) to interpret feature importance, finding that a child's involvement in household farm labour and the father's education level were among the strongest predictors of dropout risk.
The work is a methodological, data-science study rather than a deployed intervention: it was not tested as an operational early-warning tool inside Malawian schools or the Ministry of Education, and the authors' own framing is about improving the science of targeting rather than proven policy impact. Using household socioeconomic markers such as child labour to flag risk also carries a real risk of stigmatising the same poor families it aims to help, if such indicators are ever used punitively rather than to trigger supportive interventions.

Read the full analysis: https://research.tilburguniversity.edu/en/publications/school-dropout-prediction-and-feature-importance-exploration-in-m/