long-form-type data on a continuous basis: point-in-time reference periods; close agreement for small areas between the sample estimates and the complete counts by age, sex, race, ethnicity, and housing type; and—under some assumptions—cost savings from taking advantage of the census infrastructure. On the other hand, the paired strategy had at least three disadvantages: decreasing relevance of the data the longer the period between Census Day and the time for which estimates are desired; impaired data quality because of the priority given to completing the head count; and infrequent opportunities to revise the questionnaire. In addition, the paired strategy introduced inefficiencies into the census operations and—perhaps—impaired the completeness of the census head count.
The long-form-sample data, like all census data, referred to a single point in time, which in 2000 was Census Day, April 1 (even though questionnaires were mailed in mid-March and follow-up operations to complete the enumeration spread out over several months). For some economic characteristics, such as income, the data referred to a single reference period, which was the preceding calendar year. Accompanying this point-in-time reference period was a concept for assigning people to a specific “usual residence,” which was defined as the place where the person lived or stayed most of the time. These concepts were easy for users to understand and work with. In contrast, a continuous survey requires users to become accustomed to somewhat different and more complex concepts of reference periods and residence. In the case of the ACS, the estimates for a calendar year are based on aggregating data over the 12 months of data collection; the reference period is either the time when a household fills out the questionnaire or the preceding 12 months (for income, weeks and hours worked, and some housing costs); and residence is defined using a 2-month residence concept.
The paired strategy had the advantage that estimates for head counts and basic demographic characteristics from the long-form sample could be controlled to conform to the full census figures for small areas, using a statistical procedure called a raking ratio adjustment. This procedure reduced sampling and nonsampling error in the long-form-sample data products and produced a high level of consistency between estimates for demographic groups and small areas from the complete enumeration and the sample. Postcensal estimates are used to control the ACS, but they contain more error than the census counts and are not available for very small areas.
The “piggy-backing” of the long-form sample on the existing infrastructure for the short-form census may have had the advantage for the Census Bureau of reducing the costs of administering the long form. “Infra-