Robust inference in poverty mapping

Small area estimation (SAE) methods are widely used for estimating poverty indicators at finer levels of a country’s geography. Three unit-level SAE techniques – the ELL method (Elbers, Lanjouw, and Lanjouw, 2003), also known as the World Bank method, the Empirical Best Prediction (EBP) method (Molina and Rao, 2010) and the M-Quantile (MQ) method (Tzavidis et al., 2008) have all been used to estimate micro-level FGT poverty indicators (Foster, Greer, and Thorbecke, 1984). These methods vary in terms of their underlying model assumptions particularly differences in consideration of random effects. This thesis provides results from a numerical comparison of the statistical performance of these three methodologies in the context of a realistic simulation scenario based on a recent Bangladesh poverty study. This comparison study shows that the ELL method is the better performer in terms of relative bias but also significantly underestimates the MSEs of its small area poverty estimates when its underlying area homogeneity assumption is violated. A modified MSE estimation method for ELL-type poverty estimates is therefore developed in this thesis. This method is robust to the presence of significant unexplained between-area variability in the income distribution. This ELL-based MSE estimation methodology is based on a separate bootstrap procedure for MSE estimation, where a correction factor is v used to generate cluster-specific random errors that capture the potential between-area variability unaccounted for by the explanatory variables in the ELL regression model.