Kingsley-amg/africa-nutrition-ml

Predicting undernourishment across 49 African countries with 10 ML algorithms + statistical models (OLS, mixed-effects, SHAP). World Bank data, Python. A lesson in honest cross-validation. # Predicting Undernourishment in Africa, 10 ML Algorithms + Statistical Models A machine-learning and statistical study of **food insecurity (undernourishment)** across **49 African countries (2001-2023)**, built in Python from World Bank data. It compares **ten ML algorithms** and backs the findings with classical **inferential statistics** (regression with full inference, VIF, and a mixed-effects panel model). > **Headline insight (a lesson in honest evaluation):** with ordinary random > cross-validation the best model looks excellent (R² ≈ 0.98), but that is > leakage, because the same country appears in both train and test. Under > **country-held-out (GroupKFold)** validation, R² drops to **~0.17**: predicting > undernourishment for a country the model has never seen is genuinely hard, > because levels are persistent and country-specific. --- ## 🔬 What the project does **Machine learning, 10 algorithms compared** Linear Regression, Ridge, Lasso, ElasticNet, K-Nearest Neighbors, SVR (RBF), Decision Tree, Random Forest, Gradient Boosting, and Extra Trees, each evaluated under **two validation designs** (random vs. country-grouped) to expose data leakage, with **SHAP** explaining the best model. **Statistics to support the findings** - **Correlation analysis** of every determinant with undernourishment (with significance). - **OLS multiple regression** with standardised coefficients, 95% CIs, p-values and a **VIF** multicollinearity check. - **Mixed-effects panel model** (random intercept by country), the appropriate model for repeated country observations. ## 🔑 Key findings - **Income is the dominant correlate**: higher GDP per capita is strongly associated with lower undernourishment (r = -0.55). - **Water access, sanitation and health spending** are also strongly protective (r ≈ -0.5). - **Higher fertility and agricultural employment** are associated with *more* undernourishment. - These directions are consistent across the correlation, OLS and mixed-ef …