from the smaller overall initial sample size of the ACS and the use of CAPI subsampling.
The above discussion of sampling errors did not have the advantage of actual data from the full ACS. Now that the first year of data collection for the full ACS has been completed, the Census Bureau can begin to estimate the expected 5-year sampling errors for small governmental units and census tracts in larger jurisdictions, investigate disparities in sample allocation among states that differ in governmental organization, and determine the extent of other anomalous situations, such as jurisdictions with similar populations that fall into disparate sampling rate categories. Using that information, the Census Bureau should review the sample design decisions that led to the initial sample sizes and effective sample sizes after CAPI subsampling and consider alternatives that might reduce anomalies and make the allocation of the sample as equitable as possible. A review should be conducted of such alternatives as making the CAPI subsampling rates a smoother function of mail and CATI response rates and informing the choice of subsampling rates by the theoretical results on optimum subsampling rates for initial nonrespondents developed by Hansen and Hurwitz (1946).
Yet whatever the particulars of the sample design, given the available budget, the bottom line is that the sampling error of ACS estimates for small governmental jurisdictions will be larger, often substantially so, than the corresponding long-form-sample estimates. The same conclusion applies to ACS estimates for census tracts in larger jurisdictions, although these estimates can much more readily be combined into larger areas for analytical purposes.
The panel thinks that it is critically important to maintain and, if possible, increase the overall size of the ACS sample. A goal could be to increase the final ACS 5-year sample size (after subsampling for CAPI follow-up) to at least the number of housing units in the 2000 long-form sample, which was 16.4 million. This increase would provide a sample about 55 percent larger than the current ACS. To attain this larger final sample size would require an initial 5-year ACS sample size of about 23.5 million housing unit addresses instead of the current 15 million.6
Even with an increase in the ACS sample size of the magnitude just outlined, many small-area estimates, particularly for small population groups,