Improving Estimates of Mean Welfare and Uncertainty in Developing Countries

Reliable small-area estimates of economic welfare significantly inform the design and evaluation of development policies. This paper compares the accuracy of wealth estimates obtained from the empirical best predictor (EBP) of a linear nested error model, Cubist regression, extreme gradient boosting, and boosted regression forests. The evaluation draws two-stage samples from unit-level household census data in seven developing countries, combines them with publicly available geospatial indicators to generate small area estimates of assets for all seven countries and poverty for two, and evaluates these estimates against census-derived benchmarks. Extreme gradient boosting and Cubist regression generally produce more accurate predictions than traditional EBP models. A proposed two-stage residual bootstrap procedure slightly underestimates confidence intervals, but leads to higher coverage rates than the parametric bootstrap approach used for EBP predictions. These results demonstrate that, given a sufficiently large sample of enumeration areas, predictions from extreme gradient boosting or Cubist regression with a two-stage residual block bootstrap generally provide more accurate point and uncertainty estimates for generating small-area welfare estimates.