Micro-estimates of Multidimensional Child Poverty in sub-Saharan Africa

Child poverty maps allow governments and other organizations to design policies to track and evaluate their impact in the fight against child poverty. However, reliable data on the geographic distribution of child poverty is scarce, sparse in coverage and expensive to collect. For some countries, the only available measurements are at the country level. In this thesis, we propose to train Machine Learning models to obtain finely grained predictions of child poverty using heterogeneous and publicly available data sources as geographical, demographic and economic georeferenced inputs. Benchmarks of child poverty, computed from nationally representative household survey data, are used as targets to train and calibrate our proposed prediction models. The multidimensional child poverty index has six dimensions: sanitation, water, education, housing, health and nutrition, and is defined such that the predictions can be compared across countries. Using the techniques that are introduced in this thesis, we compute and release a complete and publicly available set of micro-estimates of prevalence, depth and specific poverty dimensions at a 5.2 km2 resolution for sub-Saharan African countries. Prediction intervals are included to facilitate responsible downstream use. The resulting micro-estimates have the potential of being used to deepen the understanding of the causes of child poverty in sub-Saharan Africa and to gain insights on the impact of future actions.