Development Of a Poverty Prediction Model Using Geospatial Data in The Oshikoto Region, Namibia

Abstract

This study addresses the challenge of eradicating poverty in developing nations by exploring machine learning (ML) as a tool for efficient poverty prediction. Traditional poverty assessments rely on decennial household surveys, which are resource-intensive and infrequent, especially in African countries. The research leveraged census data from Namibia's Oshikoto Region, training three ML models - Logistic Regression, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LGBM) to classify households as poor or non-poor. The models were trained with 10-fold cross-validation using two feature selection methods: Filter and SHAP.

With the full feature set (350 variables), the study achieved a maximum prediction accuracy of 88.21%. Using the SHAP method, the top 50 features achieved 85.23% accuracy, while the top 20 and 10 features yielded accuracies of 83.67% and 75.79%, respectively. In contrast, the Filter method significantly underperformed, achieving 63.99% accuracy with the top 50 features. Additional metrics, including Area under the curve (AUC), receiver operating characteristic curve (ROC), precision, recall, and Cohen’s kappa, confirmed the models' reliability.

The findings demonstrate the feasibility of using ML for accurate poverty prediction with reduced feature sets, highlighting the SHAP method’s effectiveness in preserving accuracy. This enables shorter, cost-effective surveys to be conducted more frequently, empowering policymakers and aid organizations to target resources effectively. By focusing on explanatory features, the study provides a scalable framework for addressing poverty in Oshikoto and similar regions, ensuring timely interventions and better resource allocation.