Extraordinary or Ordinary at Best? - An Empirical Study on the Application of Machine Learning Tools for Proxy Means Tests in Poverty Targeting

Sammanfattning: Proxy means tests are a widely used approach in development programs where the beneficiaries need to be determined through targeting. These tests apply a standard econometric method, ordinary least squares, to predict consumption levels using household characteristics as input variables. Yet, they still exhibit substantial misclassification rates when it comes to determining whether a household is poor or not. In this thesis, we investigate whether three alternative statistical approaches, penalized regressions, random forests or neural networks, could be applied to decrease these misclassification rates. For this purpose, we use two multi-topic household panel surveys from India and Indonesia and apply an out-of-sample validation procedure. Additionally, we evaluate how good the different methods predict poverty over time. While neural networks yield the lowest misclassification rates for most of our analyses, overall, we conclude that the precision of the methods does not differ from each other both from a statistical and economic perspective. These results are robust for important subgroups, a different set of input variables and a lower poverty line. Additionally, we find that the targeting accuracy of all methods is very stable over time.