Most countries measure poverty with absolute income thresholds (i.e. a fixed standard values considered as the minimum income to meet their basic needs). Among the multiple challenges faced by poverty metrics, obtaining reliable income information is probably the biggest.

This project uses survey data from the Brazilian Pesquisa Nacional por Amostra de Domicilios (PNAD) 2012, the largest national household survey in Brazil. Households in the database have been classified as poor using a very simple comparison of their income with the respective regional poverty line.

The models used predict household poverty using PNAD's rich data on individual and household characteristics--broadly classified into demographics, education, housing, assets, employment, and access to services.

The analysis evaluates AUC performance of six different algorithms: Logistic Regression, Random Forest Classifier, Gradient Boosting Classifier, AdaBoost Classifier, and two ensemble models: Logistic Regression + Random Forest Classifier, and Gradient Boosting Classifier.

Moreover, policy implications in terms of targeting leakage and undercoverage are taken into account through a baseline cost-benefit analysis.

This repo builds on work for Advanced Quantitative Methods (API-209) of the MPA/ID Curriculum at Harvard Kennedy School.