The new American Community Survey (ACS), which contacts approximately 250,000 households across the United States each month, collects an unprecedented amount of information about a large number of households distributed spatially across the United States. Even with this large national sample size, however, the sample sizes in many geographic regions such as census tracts are too small for estimates to have acceptable variances. This project will result in new small area estimation methods that take advantage of the ACS's spatial structure and ongoing data collection to give more accurate estimates of characteristics such as poverty rate for geographic areas with insufficient ACS sample size. A new multivariate approach with continuous and binary variables will combine information from different regions, time periods, and data sources to yield more accurate small area estimates without additional data collection cost. New multivariate Bayesian spatial models, allowing nonstationarity in the estimation, will exploit the spatial information in the ACS to model spatial and temporal patterns in the data, improve small area estimates, and enable detection of changes over time. The investigators will study properties of computer-intensive methods for estimating mean squared errors of estimators, and develop numerically stable and computationally efficient methods for calculating mean squared errors.
Estimates from the ACS are used for income and poverty assessments, funding allocation, transportation planning, allocation of resources for the disabled, studying population patterns and migration, and many other purposes. The new statistical methods are expected to give more precise small area estimates from the ACS and other surveys that have spatial information, thereby improving the quality of the information available for making resource allocation decisions. The methods also will allow researchers in many subject areas to take advantage of the spatial and temporal aspects of survey data to study phenomena such as spatial distribution and local discontinuities in poverty, distributional changes following events such as hurricane Katrina, relationships between variables measured in other data sets (for example, environmental contaminants or criminal victimization) and data from the ACS, and changes in transportation patterns. The statistical methods may be used to model the spatial distribution of pollutants, detect environmental or introduced contaminants, and model effects of interventions in education, among many other applications. Large-scale surveys such as the ACS are expensive; the statistical methods developed in this project will help researchers extract more information from them without additional budgetary costs. The research is supported by the Methodology, Measurement, and Statistics Program, the Statistics and Probability Program, and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.
