When machine learning meets econometrics: Can it build a better measure to predict multidimensional poverty and examine unmeasurable economic conditions?

Abstract

Poverty is as old as human civilization, hard to eradicate, multidimensional, and difficult to measure. The methods used to measure poverty today are costly, labor-intensive, and time-consuming. Therefore, policymakers find it difficult to target policies when putting poverty reduction initiatives into action. Indigenous communities are among the most disadvantaged and vulnerable populations in society. Their socioeconomic situations are complex and multifaceted. While research on poverty is usually generic, prone to large sampling errors, and intended to guide national policy, research on indigenous people is qualitative. Thus, to measure multidimensional poverty with disaggregated techniques, this work blends machine learning and econometrics. Researchers who have been studying poverty worldwide can replicate all of the approaches, strategies, and resources used in this study. With the best R-square and accuracy, random forest models perform better than all regressors and classifiers combined. It also confirms the causal relationship and current econometric association between multidimensional characteristics and poverty consequences. This study demonstrates the viability of using machine learning to predict poverty in a way that can save costs, cut labor, and maximize time to empower indigenous communities and alleviate the poverty of impoverished societies in the poorest region of Luzon, Philippines.