Surveys such as the American Community Survey provide samples that cover the country’s geography more evenly than highly clustered personal interview surveys. As a result, small areas such as census tracts will usually contain sample both within a tract and also within nearby tracts. The potential for using nearby data to increase the efficiency of tract level estimates, with spatial models, is investigated for a group of census tracts in the Delmarva Peninsula. Using Census 2000 counts of housing unit vacancy and occupancy rates and corresponding 1990 data as covariates, small area estimates of vacancy rate and person per housing unit are made incorporating conditional auto-regressive (CAR) spatial models.

The following issues related to making model-based estimates are addressed: fitting a parametric small area model to the housing unit data, comparison of CAR spatial models with a traditional hierarchal model without spatial components, the practical reduction in variance achieved by using a CAR spatial model and, lastly, the sampling properties of the small area estimates drawn from samples of 2000 census counts using both the CAR spatial model and the hierarchical model.