Probability of poverty: PPI analysis by machine learning

Abstract

Recently, poverty has been recognized as a global problem. Poverty Probability Index (PPI) is one of the tools to measure it. Based on the survey results on household characteristics and asset ownership, the PPI calculates the likelihood that a household lives below the poverty line. PPI is currently used by more than 400 organizations around the world – international NGOs, social services, donors, investors, multinational corporations, government and other organizations in various sectors including agriculture, health, education, energy and finance. The most famous PPI-based projects include the “Hunger” and “Electronic Warehouse” projects, Starbucks' strategy for Colombian farmers. However, the basic model with two classes (poor-rich), which underlies the index, does not classify the majority of the population with an average level of income, which has a chance of both getting rich and falling into the poor class over time and under the influence of various exogenous factors. Therefore, the work suggests a clustering model, which allows to identify 3 categories of the population: in addition to the poor and the rich, it also considers people with average earnings. 1) The class of the poor includes people of middle and old age living in villages. In most cases, these are married women with low literacy rates, who do not have their own business, bank account, and often a telephone. 2) An average earner is often a young married man with a good education. In most cases, he is neither an investor nor a business owner, he does not have a home to rent. At the same time, he usually owns at least 2 phones. 3) The class of the rich includes people of both sexes, both single and with a family. These are highly educated people who most likely have a business, investments, apartments for rent. The proposed model will help to develop more accurate tools for both poverty alleviation and prevention.