areas except for a few large states, it could play an important role in small-area model-based estimates from the ACS by providing controls so that the ACS estimates reflect the best available national and regional estimates.
A substantial amount of work needs to be carried out to make indirect estimation a reality for the ACS. The Census Bureau has already taken some important initial steps (see Bell, 2006; Chand and Alexander, 1997; Huang and Bell, 2005). In addition, the Census Bureau’s Small Area Income and Poverty Estimates (SAIPE) models and its Small Area Health Insurance Estimates (SAHIE) models are closely related to the models that might be worthwhile to develop for the ACS, as are the models used for the Bureau of Labor Statistics’ Local Area Unemployment Statistics (LAUS) program.5
The SAIPE models of poverty and median income for states, counties, and school districts currently use data from the CPS Annual Social and Economic Supplement (CPS ASEC), federal income tax records, food stamp records, the 2000 census long-form sample, and census-based population estimates. Census Bureau researchers have conducted work on the potential for school lunch program records, earned income tax credit records, and Medicaid records to improve the SAIPE models. The SAHIE models of health insurance coverage for states and counties currently use data from the CPS ASEC, federal income tax records, food stamp records, Medicaid records, and census-based population estimates.
The LAUS models of employment and unemployment for states and a few other large areas currently use data from the monthly CPS (current and historical estimates); the monthly Current Employment Statistics (CES) program, which surveys a large number of nonfarm business establishments; and state unemployment insurance (UI) records. The LAUS estimates for smaller areas, such as counties and cities, are constructed through a building-block approach that uses data from the CPS, the CES program, state UI systems, and the 2000 census long-form sample.
Presumably, the inclusion of the ACS in all of these models, which are designed to improve the CPS estimates, could result in small-area estimates that are more precise than the current model-based estimates. As noted above, models could also be developed to improve the ACS direct estimates by producing more precise small-area estimates that represent a current (or recent) time period instead of averages over a longer time period.
Three caveats are in order. First, it is not clear how strong a predictive model can be developed that would improve on the ACS period estimates for many of the characteristics of interest. Second, the effort required to