The UN Millennium Development Goals have recognized poverty reduction as the main goal of global development policy. A comprehensive framework to evaluate the effectiveness of single policy measures and policy packages with respect to poverty reduction is still lacking, though. Policy evaluation is exposed to manifold uncertainties given the dependency of the preferred outcomes on a chosen policy, available information, and policy makers' preferences. We show that Bayesian Model Averaging (BMA) is most valuable in this context as it addresses the parameter and model uncertainty inherent in development policies. Using data for the 61 Vietnamese provinces we are able to ascertain the most important determinants of poverty from a large number of potential explanatory variables.