Applying an Unsupervised Machine Learning Approach to Analyze the Non-Income Poverty Indicators Used in the Listahanan 2

Poverty targeting has been used in developing countries as a means to provide social protection programs and services directly to the poor. As information on households' welfare is unavailable or difficult and costly to acquire in the developing world, the proxy means test (PMT), which uses proxy variables to estimate an unobservable welfare variable such as household income or consumption, has become a commonly used method for targeting. This study uses an unsupervised machine learning approach on the set of household- and individual-specific non- income poverty indicators used to estimate household income in the PMT models for the Philippines' National Household Targeting System for Poverty Reduction or Listahanan, in order to examine whether differences between households across these indicators reflect differences in their income. Applying the Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) algorithm onto the Listahanan 2 indicators shows that households naturally cluster into three to four groups. However, these clusters seem to be unrelated to income and expenditure. The richest and poorest households appear to be alike and cannot be differentiated on the basis of the non-income poverty indicators considered. This suggests that these indicators alone may not be sufficient for the PMT models to accurately target the poor. However, this study is a preliminary analysis on the limited data available. A more comprehensive analysis is required to produce conclusive results.