Abstract

In 2022, the United Nations World Food Programme (WFP) provided food assistance to 160 million people across 120 countries and territories affected by conflicts, extreme weather, and epidemics. WFP conducts surveys designed to be representative at the admin 1 (provincial) level, but mobile phone surveys are often demographically unrepresentative and are not powered for admin 2 (district-level) targeting.

This study demonstrates how Bayesian computational approaches can transform these survey programs into fine-grained, population-representative maps of food security. We combine Multilevel Regression and Poststratification (MRP), raking, and Small Area Estimation (SAE) with efficient Bayesian methods for real-time, scalable inference. Using data from Zimbabwe—where WFP collects monthly mobile phone survey data (n = 1,200) and the Zimbabwe Food and Nutrition Council conducts the annual district-representative Rural ZimVAC Household Survey (n = 14,000)—we explore various modeling approaches and validation techniques to convert mobile surveys into fine-grained maps.

Our method reduces sampling bias, especially when combined with face-to-face surveys, and provides real-time estimates to better target food aid. We further enhance predictive accuracy by incorporating satellite imagery to extract environmental and climate information, such as weather patterns, vegetation health, drought, and land degradation.