Quarterly Small Area Estimates of Extreme Poverty in Brazil Using Transformed Fay-Herriot Models

The Brazilian National Statistical Office (IBGE) uses its main household survey, PNADC, to produce annual estimates of the extreme poverty rate at the country level, as well as for each of its 27 Federative Units (26 states and the Federal District). Although poverty monitoring could benefit from more frequent and spatially disaggregated estimates, those obtained via direct estimation from the PNADC are considered too unreliable for informing public policy due to their large uncertainty. In this article, we illustrate the use of transformed Fay-Herriot models to combine PNADC data with covariate information obtained from an administrative register of beneficiaries of social programs in Brazil (CadUnico). Our results show that, while the CadUnico alone cannot be a substitute for official estimates of extreme poverty, it can be used as a covariate source to produce more frequent and disaggregated estimates of extreme poverty. More precisely, we use CadUnico data to produce quarterly model-based estimates for each of the Geographic Strata, an IBGE geography that partitions Brazil into 146 domains. Use of small area estimation leads to model-based estimates that are more precise, as quantified by the estimated MSEs, than their direct counterparts. Due to their increased frequency, the model-based estimates also provide a more detailed description of the evolution of extreme poverty rates. The article focuses on those elements that ensure an appropriate implementation of Fay-Herriot models, including the choice of transformations, use of administrative data, and variance smoothing. The article contributes by proposing improvements to the estimation strategy used by IBGE to derive spatially disaggregated and timely small area estimates.