|
Abstract:
|
The American Community Survey (ACS) multiyear estimates provide detailed economic and demographic information at a census tract level. However, many users of the ACS require customized regional estimates, amounting to a change of support problem. The assumption of stationarity throughout the country is dubious, and so we present a formulation of local stationarity that can take into account variations of the covariance structure across census tracts. In addition, we adopt a nonparametric approach to modeling that remains agnostic about specific distributional features; for instance, we do not make a Gaussian process assumption. Covariances of larger areal units, such as counties, can be estimated from the local covariance structure of census tracts. These in turn facilitate the calculation of residuals, whereby the model adequacy can be verified. Finally, these covariance estimates can be applied via kriging formulas to generate customized estimates.
|
