Envisioning the 2020 Census
~Steve Pierson, ASA Director of Science Policy, email@example.com
Lawrence D. Brown is the Miers Busch Professor in the department of statistics at the Wharton School at the University of Pennsylvania. He chaired the Committee on National Statistics (CNSTAT) panel that prepared the report “Envisioning the 2020 Census.”
The 2010 census is well under way. As of Census Day, April 1, according to the U.S. Census Bureau’s “participation rate” data on the Web, an estimated 52% of households had already mailed back their forms. As this column goes to press, nonresponse follow-up (NRFU) should be in full swing.
Although the census now appears to be running smoothly, serious problems occurred in the planning earlier in the decade that could have jeopardized accuracy of the count. These problems drove estimated census costs up to almost $15 billion, or about $115 per housing unit on average. Back in 1970, when the mailout-mailback paradigm first replaced 100% in-person enumeration, census costs were about $17 per household on average in today’s dollars. It is not at all too early to think seriously about the planning process for 2020—something the Census Bureau, to its credit, has already begun. Indeed, it is necessary now to begin considering fundamental changes to census research and development (R&D) to avoid the pattern of escalating per-housing-unit real-dollar costs that has characterized censuses since 1970.
I recently chaired the Committee on National Statistics Panel to Review the Design of the 2010 Census Program of Experiments and Evaluations (CPEX), which was charged not only to comment on CPEX itself but also to identify lessons learned from the 2010 census that could improve the planning for 2020. The panel’s final report, “Envisioning the 2020 Census,” released February 22, issues a clarion call for a revitalization of the Census Bureau’s R&D infrastructure and a clear set of goals to guide R&D for 2020. (The panel’s previously released interim and letter reports on CPEX are bound with the final report) here.
Since the current approach of a mailout-mailback enumeration based on a master address list with in-person follow-up of nonrespondents was first used, the real dollar-per-housing-unit cost (2009 dollars) of the U.S. census has increased by more than 600%. In contrast, the real dollar-per-housing-unit cost of recent Canadian censuses (in 2008 Canadian dollars) remained constant at $39–$40 for the 1996, 2001, 2006, and 2011 censuses.
The unprecedentedly high per-housing-unit cost expected for 2010 resulted in part from planned long-term investments—improvements to the Census Bureau’s TIGER geographic database (essential for accurate geographic coding and mapping of addresses) and the testing and implementation of the continuous American Community Survey (ACS). High costs also resulted from a failed initiative to automate the nonresponse follow-up operation, which is now using traditional paper-and-pencil questionnaires instead of hand-held computers, as originally planned.
Moreover, contributing to cost increases over the entire 40-year span has been the accretion of operations to address special situations. These operations have not been fully evaluated to determine their benefits and costs. Underlying this accretion and the failure to achieve major steps forward on a timely basis, such as the NRFU automation initiative, is a largely incremental approach to census design and an R&D program that is often unfocused and ineffective.
The overall quality of the census enumeration using the current mailout-mailback methodology has improved between 1970 and 2000, to the point at which major continued improvement is unlikely. The net undercount of the population (measured by the demographic analysis method) decreased from 2.7% in 1970 to 0.1% in 2000. Moreover, the difference in net undercount rates between African-Americans and all others decreased from 4.3 percentage points in 1970 to 3.1 percentage points in 2000.
These figures mask large numbers of offsetting errors of omission: both of people who should have been counted and of duplicate or other types of erroneous enumeration of people who should not have been counted. Given trends that make census-taking more difficult, it is unlikely that the increased costs anticipated for the 2010 census will achieve much, if any, further increase in quality.
Social and Technological Change
Changes in the U.S. population since 1970 that have made census-taking more difficult are well known. They include increased immigration, including illegal immigration, and therefore the existence of communities that are wary of cooperating with the census and in which English is not the primary language. Changing norms in residence and living arrangements have also complicated the concept of a single, usual place of residence for many people, including children in joint custody, people with seasonal homes, and commuter workers or couples who maintain a separate workweek residence.
More broadly, the public’s willingness to respond to surveys has declined significantly over the past three decades. Nevertheless, important dynamics are dramatically changing the environment in which the census is taken, offering opportunities to make the 2020 census markedly more cost effective. It is difficult to paint a reliable picture now of the United States in 2020, but it is easy to conceive of the Internet as being the primary method for communication and conducting household business for the great majority of residents, combined with the likelihood that a large fraction of U.S. households may opt not to receive ordinary mail or may ignore much or all of it. It is also likely that administrative records will provide timely, high-quality, and inexpensive information that would be a useful input to a variety of census operations. Clearly, the design of the 2020 census should be dramatically different to accommodate such changes.