Measuring the Unmeasurable through Machine Learning Regressions and Classifications: Multidimensional Poverty Predictions in the Poorest Region of Luzon, Philippines

Abstract

Abstract Poverty is notoriously difficult to quantify, it is multidimensional and is deemed unmeasurable in the field of social science. Current poverty measurements are time-consuming, labor-intensive, and cost-expensive. Hence, policy targeting becomes challenging for policymakers to implement poverty alleviation programs. Thus, this work proposes new measures of poverty in the poorest region of Luzon, Philippines by training and testing the community-based system datasets. We have utilized machine learning regression and classification algorithms matched with advanced econometrics models. For regression, we applied 7 algorithms, for 273 ensemble runs, and for classification, we employed 12 algorithms, for 468 ensemble runs to analyze 34 locals, 4 sectors at disaggregation system, and then combined. Random forest regression outperforms all models with MSE(0.0792), RMSE(0.3298), and R-square(0.92075), while random forest classifier outperforms all models with the highest accuracy(91.08% at random and 95.95% at pipeline). It also validates the existing correlation and causation between multidimensional attributes (27 variables) and poverty outcomes (Incidence, gap, severity, and watts). This work highlights the feasibility of machine learning for poverty prediction that can minimize cost, reduce labor, and maximize time, particularly in the poorest regions of the Philippines. Finally, the output has provided policy targeting tools for poverty reduction for various locals at different poverty configurations.