|
Abstract:
|
Official statistics agencies often publicly release microdata with a rich set of measured variables, but geocoded at a coarse scale of spatial resolution to protect the identity of observed households. On the other hand, many statistical agencies also release tabulated estimates of selected variables at finer spatial scales --- for example the number of households that fall into various race, education, and other demographic categories. This trade-off is ever present in official statistics: data users can have a high degree of spatial resolution, or a rich set of measured variables, but never both.
Fortunately data users can use both data sources to downscale the microdata to finer spatial scales. We show that one popular method for downscaling, penalized maximum entropy dasymetric modeling (PMEDM), can be viewed as the posterior mode of a particular model with a particular prior distribution. Additionally, we explore the Bayesian framework for PMEDM, including alternate priors and alternate methods for obtaining PMEDM estimates. We apply our methods to Public-use Microdata Samples and tabulated estimates released by the Census Bureau from the American Community Survey.
|
