Evaluating Alternative Approaches to Small Area Estimation of Poverty with Survey and Census Data

This paper uses five rounds of Mexican and Brazilian census extracts to evaluate the accuracy of different model specifications and estimation methods that use survey and census data to generate small area estimates of poverty. Models that utilize more granular data for prediction (household- and/or village-level predictors) tend to produce more accurate estimates of poverty than models estimated only using area-level predictors. Differences in accuracy across models and methods that utilize household or village level predictors are minor. Models that omit household-level predictors tend to be more robust than unit-level models to the use of old census data and classical measurement error in survey predictors. The performance of the Fay-Herriot area-level model falls in the presence of sample selection bias and small sample sizes. Rescaling sample weights is important in Mexico, where the sample is informative within areas. Applying raw sample weights without rescaling in this case greatly reduces the accuracy of estimates from linear models and distorts methodological comparisons. Overall, no one approach dominates across all contexts, but when sample weights are rescaled there is no downside to using more granular data for prediction.