To reduce the magnitude of the unexplained location specific component, we estimate a separate model to explain the cluster specific error terms. As regressors, cluster means of the household specific variables are obtained from the census at enumeration area level and merged into the survey data set. This is a common procedure in poverty mapping. It amounts to explaining spatial autocorrelation between factors common to a household in a given PSU. To the extent that households attend the same school, make use of the same source of fuel wood, or water and have similar access to markets, this procedure is likely to go a long way in explaining spatial autocorrelation. Yet, various rather obvious determinants of spatial autocorrelation cannot be obtained from the census. Population and tree density, soil type and quality, access to infrastructure are examples of such information. By building an integrated dataset with census and biomass information, we are able to include such bio-physical information in explaining spatial autocorrelation. We estimate equation 1 taking into consideration the location and heteroscedasticity component of the disturbance term. Survey weights are included in some of the regressions depending on the Hausman test (see Deaton, 1997) results for whether the regressions should be weighted or unweighted.