Dan Iancu, Sarah Billington, Antonio Skillicorn | Interpretable Machine Learning and Mixed Datasets for Predicting Child Labor in Ghana’s Cocoa Sector

This exploratory work examines how two surveys that measure child labor in Ghana (NORC and GLSS7), but differ in quality and scale, can be jointly leveraged for less biased prediction and to identify key predictors of child labor risk. We further investigate whether district-level satellite indicators, including yield-weighted cocoa-driven deforestation, newly lit area, and newly urban area, enhance predictive performance and play important roles in shaping model predictions. Using non-parametric machine learning models (XGBoost, Random Forest) paired with cross-validation and a hyperparameter grid search, we find that the best-performing model in classifying child laborers achieves an out of sample AUC of 0.95 and F1 of 0.84. Model interpretability tools (SHAP values, partial dependence plots) highlight influential predictors such as child age, cocoa-driven deforestation, school commute time, newly lit area, and household herbicide expenditures. In addition to emerging as the second most explanatory feature, cocoa-driven deforestation also shows a clear nonlinear association with predicted child labor risk. Our approach demonstrates new ways of grappling with data scarcity and bias in child labor measurement, while our findings provide actionable risk profiles to support monitoring efforts and underscore the complex interconnections between child labor and environmental practices.