The framers of the Constitution created the census as a mechanism to apportion political power and tax dollars among the various constituencies of the population. Every 10 years the federal government would count the population and reapportion Congress on the basis of the results. Like elections, the census and reapportionment process shifts political power in the United States. And as with elections, the losers have to concede to the winners. Over the past two centuries, the census has basically accomplished its goal. Congress has managed to reapportion successfully despite profound population growth and migration, ethnic and racial and political, social and economic change.

“Beginning in the 1940s, statisticians and demographers perfected methods of measuring census accuracy, using the techniques of demographic analysis and dual systems estimation.”

The concept of “census undercount” is a function of the development of modern probability sampling methods and their use in statistical methodology. Census officials from George Washington and Thomas Jefferson to the present have been aware of the extraordinary difficulties of counting a dynamic, diverse, and mobile population. Until the 1940s, census officials had hunches but few tools to measure accuracy, and certainly no mechanism to “correct” for undercounts, overcounts, or miscounts beyond doing the enumeration over again. Beginning in the 1940s, statisticians and demographers perfected methods of measuring census accuracy, using the techniques of demographic analysis and dual systems estimation. Sample post-enumeration survey (PES), statistical inference from sample to national rates and comparison to the vital registration system through demographic analysis replaced the hunches and laments about census error with increasingly reliable point estimates of error and intervals of confidence about those estimates.

Since 1940, the Census Bureau has conducted evaluation studies based on probability samples of subsets of the population to measure the level of accuracy of census results, in terms of both coverage errors—that is, under- and over-enumeration—and content error—that is, incorrect information on the characteristics of the population. The often reported statistic of the net undercount, 1.8% in 1990, represents, at the national level, the net undercount once undercounts are subtracted from overcounts and erroneous enumerations. It does not mean that 98.2% of the population was counted accurately. The evaluation studies of census coverage have themselves improved over the past 20 years, incorporating new measures of accuracy, including measures of gross error in the census—that is, the sum of overcounts, undercounts, and erroneous enumeration.

The purpose of the 1950 PES was coverage evaluation, with the goal of identifying areas in which to increase accuracy. At about the same time the Census Bureau expended considerable resources to develop a statistical model for census error. The measure of correlated response error due to interviews led, for example, to the trial in 1960 of the mail-out, mail-back approach to enumeration which worked to eliminate the role of the interviewers. With the recognition during the 1980s that most coverage improvement efforts had failed to address the differential undercount, attention focused anew on using the PES to correct to raw enumeration results. In the end, the 1990 design was a compromise, a traditional census complete with coverage improvement programs and a PES for possible use in correcting the count.

The sample size of the 1990 PES, of approximately 165,000 households, was a compromise between the original design of 300,000 households and the Republican administration efforts to eliminate the PES. Despite the generally acknowledged success of the 1990 adjustment as an improvement over the raw enumeration counts, the small sample size had two major impacts on accuracy. First, the level of sampling error was a serious component of the overall census error; and second, the data at low levels of geography were sufficiently sparse as to require smoothing of the adjustment results, even across state lines. The public discourse on the 1990 census and adjustment has conflated “sample size” problems with a more general condemnation of sampling in the context of census-taking, and has led to the myth that “sampling” itself is a suspect enterprise.

“The 2000 census plan proposes to employ some time-honored methods of counting and some innovations.”

The 2000 census plan proposes to employ some time-honored methods of counting and some innovations. The time-honored methods include the use of a mail census as the primary means of contacting households. The mail census was first used in 1970. In that year about 60% of American households received their census form in the mail and were instructed to fill it out and mail it in. In 1980 and 1990 over 90% of households were contacted by mail. For the parts of the country that cannot be reached by mail, the Census Bureau uses enumerators to canvass a particular geographic area, the fundamental enumeration procedure from 1790 to 1960. The Census Bureau will also use enumerators to contact households that do not return the mail census form in a timely manner, in the counting phase known as the nonresponse follow-up. The 2000 plan proposes collecting additional detailed information on the population through the use of a long form sample (which began in 1940) in conjunction with the short form complete count. And the Census Bureau will evaluate the quality of coverage of the count with a post-enumeration sample survey, a procedure used in one form or another since 1950.

The new methods envisioned for 2000 originally included sampling for nonresponse follow-up. It is this new procedure that generated some of the most heated objection in Congress. Census officials know from past experience that residents at about one-third of the addresses will forget to fill out and mail back a census form, will ignore the form or perhaps not receive it in the first place. The Census Bureau follows up by sending an enumerator to the address. This phase of the count starts in late April, and is designed to retrieve information from the households that have not responded to that point. In 1970, 1980, and 1990, nonresponse follow-up was conducted for all households that did not mail back their census forms. The evaluation results of the last two censuses indicted that the quality of data collected by enumerators from nonresponding households got much poorer the longer it took the enumerators to collect it. That is, responses gathered from households in June or later were significantly more error-filled than those collected in April and May. Thus the Bureau concluded that a higher quality sampling process for nonresponse follow-up would produce better data than 100% follow-up because the process could use better-trained employees and be done much more quickly. The census plan guarantees that 90% of the households in each census tract will be counted; they will make inference for the residual nonresponders derived from the sampled nonresponse follow-up households. The opponents of sampling for nonreponse follow-up, as noted above, claim that the Census Bureau has given up the effort to contact everyone, and will make up people, a process that could be manipulated to the benefit of Democrats.

The other significant and controversial innovation of the 2000 Census Plan was the integration of the post-enumeration survey process into the traditional enumeration. In 1990 the Census Bureau took a post-enumeration survey and produced adjusted census counts on the basis of the survey results. But the 1990 census did not fully integrate the PES and the traditional enumeration to produce adjusted census counts on the basis of dual systems estimation. Rather the bureau released the results of the April, or traditional, enumeration in December 1990 and then released adjusted results in June 1991. There were eight years of litigation on the quality and legality of the two sets of figures. This decade, the Census Bureau proposed a one-number census—that is, procedures that would produce a final census count which could not easily be disaggregated into the traditional enumeration and the adjustments made on the basis of the results of dual systems estimation. Again, critics charge, the adjustment process is subject to political manipulation, to making up people.

“There remain major differences of opinion regarding this use of sampling in the census context between the Congressional Republicans and the Clinton administration and some state and local governments.”

In its January 25, 1999 ruling in Department of Commerce v. House of Representatives, Nos. 98-404 (1999), the Supreme Court held that the Census Act prohibits the proposed use of statistical sampling to determine the population for congressional apportionment purposes. The ruling invalidated the Census Bureau’s plans to use statistical sampling for the nonresponse follow-up phase of the enumeration, but it did not explicitly ban sampling in the census for uses other than apportionment. The major use of sampling in the Census 2000 plan would be via a post-enumeration survey, the result of which could be used to adjust those from the “traditional” enumeration for purposes other than apportionment. There remain major differences of opinion regarding this use of sampling in the census context between the Congressional Republicans and the Clinton administration and some state and local governments. Hence it is reasonable to expect the constitutional questions to be raised again in litigation in the future, and for Congress to debate legislation to resolve the controversy.

The bureau’s plans for the PES discussed above were intended to produce direct state estimates for apportionment and thus called for a very large sample of 750,000 households. Without sampling for nonresponse follow-up, the bureau will need to spend more time in the field doing traditional follow-up, which will delay the start of the PES. As a consequence, it plans to reduce the size of the PES sample to about 300,000 households and to relax requirements for direct state estimates.

We agree with the leaders in Congress who point out that there are political implications of using one or another counting technique. Deciding to use integrated coverage measurement can move a seat in Congress from one state to another, as can a “computer mistake.” But we would like to direct attention to the discussion of how best to avoid mistakes, and to enhance the accuracy of the census. Many other decisions about census or apportionment methodology also move seats among states. Deciding to count the overseas military in the apportionment totals may shift a seat. So may changing the apportionment formula or the size of the House. We expect that Congress will continue to focus on the details of census taking in the year ahead. Changing census methods may move seats in Congress between states, not the two dozen seats some claim, but one or two.

It is not a coincidence that the undercount controversies emerged on the heels of the Supreme Court “one person, one vote” decisions of the 1960s. Once the courts required congressional districts to adhere to that rule, not only did urban representation increase dramatically, but the accuracy of state populations with reference to one another, and for within state districting, also became very important to the political life of the nation. The differential undercount of racial minorities is also a differential undercount by state, and a differential undercount that impacts on districting within states. And at these lower levels of geography the national net differential undercount of a few percent gets magnified many times over. As a consequence, the differential undercount is a partisan issue because of the political fact that minority groups in the United States vote heavily Democratic. And thus the litigation that has emerged, and the political splits within Congress, pit Democrats against Republicans, and apportionment losers against apportionment winners.

Thus, in an apparent paradox, as the technical capabilities of the Census Bureau to measure and adjust for the differential undercount have improved over the past three censuses, the political controversies of the legality and constitutionality of doing so have heightened. To resolve the controversies, one must separate the technical questions from the political questions and determine who has responsibility and authority to address each.

Excerpted from Who Counts? The Politics of Census-Taking in Contemporary America by Margo Anderson and Stephen E. Fienberg (Russell Sage Foundation, August 1999).

Margo Anderson is Distinguished Professor of History & Urban Studies at the University of Wisconson–Milwaukee. She specializes in American social, urban, and women’s history and has research interests in both urban history and the history of the social sciences and the development of statistical data systems, particularly the census. Stephen E. Fienberg (1942-2016) was University Professor of Statistics and Social Science at Carnegie Mellon University. He was a founder and editor-in-chief of the Journal of Privacy and Confidentiality, and he received multiple awards and accolades throughout his career. Anderson and Feinberg coauthored Who Counts? The Politics of Census-Taking in Contemporary America (Russell Sage Foundation, August 1999).

This essay originally appeared in Items Vol. 53, No. 1 in March 1999. Visit our archives to view the original as it first appeared in the print editions of Items.