Kernel density estimation on grouped data: the case of poverty assessment

Grouped data have been widely used to analyze the global income distribution because individual records from nationally representative household surveys are often unavailable. In this paper we evaluate the performance of nonparametric density smoothing techniques, in particular kernel density estimation, in estimating poverty from grouped data. Using Monte Carlo simulations, we show that kernel density estimation gives rise to nontrivial biases in estimated poverty levels that depend on the bandwidth, kernel, poverty indicator, size of the dataset, and data generating process. Furthermore, the empirical bias in the poverty headcount ratio critically depends on the poverty line. We also undertake a sensitivity analysis of global poverty estimates to changes in the bandwidth and show that they vary widely with it. A comparison of kernel density estimation with parametric estimation of the Lorenz curve, also applied to grouped data, suggests that the latter fares better and should be the preferred approach.