Determinants of the Armenian household poverty : an econometric and machine learning approach

Abstract

The aim of the paper is to use combine econometric analysis and machine learning modeling to explain the multidimensional nature of the Armenian household poverty. The multinomial logistic regression results show that there are monetary and socioeconomic variables affecting poverty. Food and Non-Food related purchases in dram, members of the household, settlement, which includes Yerevan, other urban and rural towns, income received from abroad, educational level of the head of the household and a few other variables have a significant effect on the poverty status. After measuring the direct variable impact, Neural Network and Decision Tree models are constructed. All three models are fit on the same training data and later evaluated on the same testing data to find out how well they perform the task of classifying Poor and Very Poor Households. From the original data, less than 2 percent of the observations fall under Very Poor Category, so correct results for this class are the most prioritized. Neural Networks provide the best results in terms of correctly classifying the Poor and Very Poor Households from the testing data, followed by Decision Tree and Logistic Regression. As a main classification metric F1 score is taken.