An Empirical Evaluation of Poverty Mapping Methodology: Explicitly Spatial versus Implicitly Spatial Approach

Poverty maps provide information on the spatial distribution of welfare and can predict poverty levels for small geographic units like counties and townships. Typically regression methods are used to estimate coefficients from the detailed information in household surveys, which are then applied to the more extensive coverage of a census. One problem with standard regression techniques is that they do not take into account the ‗spatial dependencies‘ that often exist in the data. Ignoring spatial autocorrelation in the regression providing the coefficient estimates could lead to misleading predictions of poverty, and estimates of standard errors. Household survey data usually lack exact measures of location so it is not possible to fully account for this spatial autocorrelation. In this paper, we use data from Shaanxi, China with exact measures of distance between each household to explicitly model this spatial autocorrelation. We also investigate which set of augmenting variables (i) census means or (ii) environmental variables mainly from satellite imagery have the most impact in soaking up unwanted spatial autocorrelation.