A Bayesian approach to poverty mapping

• Main Author: Migotto, Mauro
• Published: University of Reading 2012
• Subjects: 339.460151954
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.577980

Poverty mapping can be defined as the disaggregated spatial representation (estimation) of poverty, inequality and other well-being indicators. Substantial amounts of money are spent and allocated on the basis of such estimates, and in some cases poverty maps have become part of the regular statistical service. Assessing the reliability of different estimators in different circumstances becomes therefore essential. However, the empirical evidence on the performance of econometric poverty mapping methods is still limited, given the data constraints. The World Bank method, which is the most used, has been found to be less robust and reliable than claimed by its proponents. The alternative methods proposed to date have some disadvantages, especially with respect to out-of-sample prediction. The aim of this thesis is to contribute to this literature by comparing Hierarchical Bayes estimators with the World Bank method on simulated and real data. Hierarchical Bayes estimators provide an appealing and robust alternative, especially when there is substantial geographical model variation. They also provide more accurate measures of precision. This is important given that the World Bank method tends to overestimate the precision of its estimates. Although Bayesian estimators are more robust when the restrictive assumptions of the World Bank method do not hold, there remain two areas that pose difficult problems and that need further research. The first is out-of-sample prediction in presence of substantial model heterogeneity across space. A Bayesian spatial model is proposed which, although it does not perform satisfactorily with the data used, needs further consideration. The second is the selection of the predictors. More research is also needed on the relative performance of different methods. Census-like, better quality data is needed to test different estimators.