Estimação de indicadores de pobreza em pequenas áreas para o Brasil com modelos beta

Abstract

The fight against poverty is of utmost importance, being the primary objective of the Sustainable Development Goals established by the United Nations and a fundamental goal of the Brazilian Constitution. However, to effectively tackle poverty, it is necessary to understand its various features and its distribution over the country. Poverty can be monitored using three main measures: poverty incidence, poverty gap, and poverty severity. Therefore, the aim of this dissertation is to produce reliable estimates of these three poverty measures for local areas defined as groups of Brazilian municipalities (146 geographic strata), using data from the Continuous National Household Sample Survey (PNADC) from IBGE. Although the survey has allowed the estimation poverty measures by geographic strata since 2012, they can be imprecise at this level of disaggregation. One approach to overcome this challenge is the use of small area estimation (SAE) methods. Therefore, to improve the precision of poverty estimates, SAE models were implemented. Bayesian small area models were developed, with a Beta distribution and logistic link function, to estimate poverty incidence (head count ratio), gap and severity for geographic strata from 2012 to 2022. The Beta regression models incorporate area level socioeconomic auxiliary variables, as well as area and time random effects. These models yielded satisfactory results, improving overall precision of estimates, without evidence of bias. Building on the success of the models based solely on 2021 survey data, the method was extended to a repeated survey context, incorporating survey data from 2012 to 2022. The improvement in the precision of estimates obtained from temporal models were even more pronounced, in a way that all estimates of the three poverty measures became suitable for publication. Furthermore, in the temporal model, the use of a historical series spanning at least four years of data notably enhanced the precision of estimates when compared to direct estimates.