FracksonM/malawi-dhs-2024-geoai

# Malawi Child Stunting Prediction: Survey-Weighted Machine Learning (2024 MDHS) **Author:** Frackson Makwangwala **Contact:** fracksonmakwangwala@gmail.com ## Overview Survey-weighted multi-algorithm machine learning for nationally representative prediction and district-level decomposition of child stunting risk in Malawi, using the 2024 Malawi Demographic and Health Survey (n = 5,122 children, 767 enumeration area clusters). ## Key results - Best model: Random Forest, ROC-AUC = 0.6856 (95% CI: 0.666, 0.705) - Survey-weighted stunting prevalence: 37.5% (95% CI: 35.9%, 39.1%) - Top predictor: maternal height (mean absolute SHAP = 0.0558) - District SHAP decomposition: 3 risk archetypes across 32 districts ## Notebooks | Notebook | Description | |---|---| | 10 | Survey-weighted model training: RF, XGBoost, LightGBM, LR | | 10b | Model enhancement: calibration, DeLong test, subgroup analysis | | 11 | SHAP global analysis and district decomposition | | 12 | Wasting and anemia models | | 13 | Spatial GeoAI: GPS integration, spatial lag, environmental covariates | ## Application `app.py` — Streamlit prediction tool with HSA, DHO, and Researcher dashboards. ```bash streamlit run app.py ``` ## Citation Makwangwala F. (2025). Survey-Weighted Machine Learning for Nationally Representative Prediction and District-Level Decomposition of Child Stunting Risk in Malawi: A Multi-Algorithm Analysis of the 2024 Demographic and Health Survey. ## Data access Raw DHS microdata are not included in this repository (licence restricted). Access the 2024 MDHS at dhsprogram.com