Geotargeting in Urban Areas Informed by
Community-Driven Insights

Abstract

Recent advances in poverty prediction at a national scale employ new data sources and machine learning. However, the performance of these models for households experiencing poverty, food insecurity, and nutritional deficiency, in urban areas is unknown. This research explores how geospatial indicators, particularly those informed by community insights, improve poverty prediction and household targeting for social transfers in urban settings. We conducted a survey of 300households in Lusaka, Zambia. We combine this information with geo-features, such as neighborhood structure, access to infrastructure and services (based on Point of Interest data from crowd-sourced maps), satellite imagery features, connectivity data, and Twitter. Using machine learning techniques, these features are used to predict welfare based on our four different indicators: assets, food security, iron deficiency, and a combination of asset and food dimensions. We find that using community information in the selection of geographic predictors provides a parsimonious way to identify households in need. Geographic features combined with household-level information can efficiently predict asset-based deprivation, food insecurity and iron deficiency, and households with multi-deprivation. Our best models also reduce targeting errors by 20-22 percentage points compared to baseline models that only use socio demographic characteristics to identify asset, food, and nutrition- deficient households. Our results show that community-informed spatial features can be used for geographic targeting for asset and food-related dimensions of poverty to effectively reach vulnerable populations while improving the legitimacy of targeting methods.