The use of population estimates based on demographic analysis as benchmarks is a common practice. At the national level, demographic analysis is perceived to be relatively error free for counts classified by age, race, sex and ethnicity. By benchmarking to these estimates, coverage errors can be reduced. In addition, benchmarking can also provide some reduction in variance.
If fully implemented, one of the products of the American Community Survey (ACS) will consist of estimates by county. As it is expected that the ACS will have coverage errors comparable to other national surveys, benchmarking to correct county demographic population is appropriate for reducing coverage error. Also, at the county level, benchmarking may provide reduction in variance due to moderate county sample sizes. However, county demographic estimates of population are known to be less accurate than their national level counterparts, primarily due to errors in estimating net county migration. Here, the use of county level benchmarks that are themselves in error becomes an issue in deciding whether or not benchmarking, at this level, results in an overall lowering of mean squared error (MSE).
We focus on identifying the effects of errors in the benchmarks, due to estimating inter-county migration, on the accuracy of ACS estimates. Although the benchmarks are based on migration estimates derived only from tax-filers rates (Sink, 2000), the ACS covers both filers and non-filers allowing comparisons to be made. Since file status is not present in the ACS, we use latent class models to identify likely groups of non-filers and evaluate their migration behavior. However, the use of only tax-filers to estimate inter-county migration may be innocuous if:
To illustrate the relative effect of net-migration errors on the ACS, we consider the mean square error of the benchmarked estimate relative to the expansion estimator, benchmarked only at the national level. The components of the MSE that will be considered to be important are the variance, coverage error bias, bias due to differences in residence rules between ACS and demographic analysis and bias due using an incorrect migration rate in the demographic analysis. Although differences due to residence rules will contribute if the ACS residence population is benchmarked, we will assume, here, that questions can be added to the ACS to produce a resident population comparable to the demographic analysis population. In order to obtain a crude assessment of these components of the MSE, the error of the expansion and benchmarked estimators will be assessed using generalized variance functions. The bias due to undercoverage in the expansion estimator will be assessed using coverage errors estimated from other national surveys. We use the model of inter-county migration error to assess its bias contribution to the MSE.
Future consideration on combining sample-based migration information from the ACS with demographic analysis to, possibly, achieve a hybrid benchmark will be discussed.