For the first time in modern history, the Census Bureau may be prevented from incorporating the best available statistical methods in the national census and the result will be a "failed census." Instead, 1990 census methods will run a second time with strategic enhancements. For the third time in American history, the Congress is demonstrating that the national census is ultimately about politics. Come 2001, Census Bureau data may lose its luster as "best available." Local governments will need to develop legally defendable decisions and policies regarding use of census data with known undercounts and/or adjusted data-and then use the data more carefully. However, preliminary results of a national survey of municipal planners who use census data show that skills in evaluating census data quality and making adjustments are little known and not even perceived as a high-priority training need.
Note About The Abstract, Paper, And Presentation
The paper and presentation ended up somewhat different from the abstract and more focused on actual data that had become available from the Dress Rehearsal. I thought actually working with data would be better than talking about working with data. I do want to clarify some points raised in the abstract.
The "failed census" comment refers to, at the time I wrote the abstract, the possibility of the Supreme Court prohibiting sampling. It did in January 1999. After the 1990 Census, the Census Bureau had two primary goals for planning Census 2000: 1) reduce costs and 2) increase accuracy (Report to Congress: The Plan for Census 2000, August 1997, Bureau of the Census). If Census 2000 meets neither goal, it will have 'failed' notwithstanding the considerable efforts of the Bureau's professional dedicated staff.
Webster's definition of 'fail' includes 'to be unsuccessful' and 'to disappoint the expectation of trust' and 'to leave undone' (Mirriam Webster's Collegiate Dictionary, 10th Edition). The General Accounting Office (GAO) March 1998 Report to the U.S. Senate used the 'F' word several times in its assessment of key 2000 Census of Population and Housing (Census 2000) plans and activities.
The key activities that GAO examined for this report, including address list development, local outreach and promotion, staffing, and statistical sampling, are all still facing the developmental and/or implementation challenges that led GAO in 1997 to raise concerns about the high risk of a failed census in 2000 (author's emphasis) (GAO, Report to the Committee on Governmental Affairs, U.S. Senate. "2000 Census Preparations for Dress Rehearsal Leave Many Unanswered Questions, pg 3, March 1998).
The "third time in American politics" phrase refers to two previous periods in American history when the Congress choose political expediency. The first was the counting of slaves as 2/3rds of a person for purposes of apportioning Congress. The second was the failure of a rural-dominated Congress to adopt a reapportionment bill after the 1920 Census because urban areas with high immigrant populations made up over 50 percent of the 1920 population.
The national survey refers to 350 responses to a questionnaire mailed to 1800 subscribers to the Planning Advisory Service of the American Planning Association. The final results are in the possession of the APA Chicago Office.
1999 Esri User Conference Paper
The title of this presentation asks the question, "Will Census 2000 be a repeat of the 1990 Census?" The answer is, "Basically, yes but with several important exceptions." Census 2000 is already underway and the first detailed data will be released by April, 2001. After the U.S. Supreme Court decision last January upholding Title 13 prohibition on using 'statistical methods' for the reapportionment of the U.S. House of Representative, Census 2000 was reconfigured to essentially repeat the 1990 Census by eliminating 'sampling' during field follow-up operations.
This paper is a 'work in progress' intended to investigate issues related to Census 2000 that are significantly different from the 1990 Census. Additional analyses, tables and maps will be added later. This paper and presentation presume a working knowledge of 1990 census geography and census data and is designed for urban planners, GIS professionals, and others primarily in local government and small organizations that rely on non-expert census data users but that desire more than pre-prepared tabulations.
In two years time, we may look back nostalgically at the 1990 Census as the last census that was rather 'simple' to use. Here's why:
Topics being researched are:
My research uses 1990 tract level adjusted and unadjusted and 1998 adjusted and unadjusted data for the city of Sacramento. The 1998 data come from the Dress Rehearsal census that was a test of Census 2000 procedures. This dataset, then, is the only large urban dataset that mimics what most of you will be working with after Census 2000 data are made available on the Internet.
As in the 1990 Census, there will be two types of questionnaires, commonly called the short (or "100 percent") form and the long (or "sample") form. The Census 2000 short form includes seven simple questions of each person listed in the household roster: age, sex, relationship (to the first person in roster), race, Hispanic origin, and tenure. Marital status, number of rooms, units in structure, value, and monthly rent are all moved to the sample questionnaire.
The long form asks 36 questions: the seven that appear on the short form; the five questions transferred from the short form; plus repeated 1990 Census questions about education, income, ancestry, housing costs, housing quality, and so on with the following three exceptions: grandparents as caregivers (new), children ever born (dropped), year last worked (dropped). As in previous censuses, the long form will be sent to roughly one out of six addresses and sample data will be available at the Block Group level and higher.
Census 2000 questions and were determined by the needs of the Federal Government. Each question is either required by law, needed to implement a program (and no comparable data are available), and/or needed to continue key time-series data sets. Above all these reasons is the Constitutional mission to provide data for the reapportionment of the U.S. House of Representatives and then for redistricting of Congressional districts within states (with more than one Representative).
You will not have housing data at the block level as in 1990. These data will be at the block group level as the housing questions (except for tenure) are moved to the long form.
The single most important content change is in the revised race question. For the first time next year, the question is, "What is this person's race�mark one or more races." Also, the Asian Pacific Islander (API) category is divided into two categories which then include major subgroupings.
The challenge is tabulating the 'more than one race' data, not only for Census 2000 but also in relating Census 2000 data to local data still using 1990 or some other categories.
In the 1990 Census there were five possible response choices:
For Census 2000 there will be seven response choices:
'No Response' may also be an additional category in the tabulations.
Depending on the diversity of your community, how the multi-race data are presented could be very 'political.' Overall, about two percent of the population is likely to indicate more than one race, rising to about five percent in urban areas.
You may have to explain the changes in your work and/or footnotes. More importantly, state and local governments and businesses will likely have to migrate to the seven Census 2000 categories over time, which is already the tentative policy of the Federal Government.
The Hispanic Origin question is moved for Census 2000 to precede the race question, hoping that the confusion between the two would decrease. The Dress Rehearsal data indicate the confusion remains, with many Hispanics marking 'Other' for race. Hispanic Origin will need to be cross-tabbed with the seven race responses. That's 14 tabulations.
The Accuracy in Coverage Evaluation (ACE, called PES in 1990) that coincides with the regular census will produce detailed undercount estimates. Even if the Census Bureau does not 'adjust' Census 2000 data, the undercount estimates will surely be released, giving opportunity to make adjustments anyway. In short, there will be several sets of Census 2000 data, and none likely deemed 'the official set.'
You may need different counts for different applications within your organization or as required by law. Your neighboring communities may use different counts.
In all, there will be (at least) five sets of Census 2000 population counts:
Since anyone with Internet access and a printer will be able to produce Census 2000 thematic GIS maps, you could end up explaining why population maps differ because of which and how the data are summarized and which counts are used.
Methodology refers to how Census 2000 data is collected and processed in the coming year. Census 2000 will essentially repeat the procedures used for the 1990 Census, only with greater emphasis on local civic participation and a paid publicity campaign. There are several changes in methodology that are important to GIS professionals:
All Census 2000 block numbers will be four digits with no alpha-suffixes. Data users will need to obtain the block number equivalency/correlation lists from the Census Bureau and modify local records, etc. for the revised coding.
There are no Block Numbering Areas (BNAs) for Census 2000. Instead, all 1990 BNAs become Tracts in Census 2000.
The Census Bureau's mission is to count people where they live. There are six ways people may be missed in the census:
It is the last two error types that triggered the Census Bureau's change in housing unit definition for purposes of taking the census. In order to maximize the effort to find and locate all possible housing units, the Census Bureau's definition is changed from the 1990 Census to not require any eating or kitchen facilities. The new definition is
"live separately from other residents, and be able to enter their living quarters without going through someone else's house or apartment."
Note there is no requirement for a kitchen or bathroom.
The new definition could result in Census 2000 housing unit counts that include converted garages, SROs, and other types of units not listed as housing in the 1990 Census and probably not considered housing units in local zoning and building codes.
Census 2000 statistics on housing quality such as number of rooms, presence of kitchen and/or bathrooms, and persons per unit could reflect the change in definition. When linked to 1990 Census data, misleading trends could emerge.
One of the principal ways used to create population projections is the 'housing unit method' that relies on local records of the number of housing units completed, demolished, or converted. How the new Census Bureau definition and local code-defined definitions of housing unit will work together is untested.
There are still no plans for a Post-Census Local Review that, in 1990, gave cities and counties several weeks to review block-level housing unit counts while census operations were still in the field - late summer 2000. While local governments will be greatly encouraged to reduce the undercount from 1990, without the Post-Census Local Review, they may have little feedback on how well they did.
Census 2000 originally included a new procedure--Integrated Coverage Measurement (ICM). Under this plan, similar to the 1990 Post Enumeration Survey (PES), the bureau would have conducted an intensive second census on 750,000 blocks selected to represent the national population. Names would be matched with the "official" census, various categories of people would be estimated, and counts would be corrected at all geographic levels. This process was expected to reduce the net error in Census 2000 to about one percent at the tract level.
The Census Bureau revised the Plan for Census 2000. It now calls for a PES/ICM reduced in scope from 750,000 to 300,000 households and renamed the Accuracy and Coverage Evaluation (ACE). Results of ACE could be used to 'adjust' Census 2000 (this was the subject of the numerous 1990 Census lawsuits). Even if the Census Bureau does not adjust, the estimated undercount will be public record and could be used by anyone to adjust Census 2000 data.
If the 1998 dress rehearsal censuses are an indicator, the urban undercount this time around is likely to be about 6.5 percent, going as high as 10 percent for minorities.
Any Census 2000 adjustment runs into the trouble because the undercount estimating process is very sensitive to small errors. The Census Bureau estimated that corrections for identified measurement errors in the 1990 Census PES would have reduced the net undercount from 2.1 to 0.9 percent (Mulry, 1992; P-16 Report: Total Error in PES Estimates). This does not mean the undercount was that low in 1990. It happened that the national 1.8 percent PES estimate was close to the independent 1.6 percent estimate made through demographic analysis - that the various PES errors offset themselves at the national level. This was one of the reasons for the 1997 decision not to adjust the 1990 Census as the PES-based estimates would be unreliable at smaller geographic levels. For Census 2000, optical character recognition (OCR) software will read and match millions of names and the likelihood of matching error is still present. Each mismatch could translate to 1000's of people in the undercount estimates.
Unlike previous censuses that relied on public service announcements, Census 2000 will have a $100 million paid 'Madison Avenue' publicity campaign. The multi-media campaign will need to raise public awareness, recruit applicants for the 800,000 estimated jobs, and motivate the public to complete and return the questionnaires.
C. Access/products
Census 2000 data will be available primarily via Internet, probably at the Census Bureau's server at www.census.gov as follows:
[April 1, 2000 Census Day]
December 31, 2000 State and national
population totals
April 1, 2001 'Raw' unadjusted basic population
only (no housing) to
the block level in
PL94-171 geography
Summer thru Fall 2001 Releases of complete data
[may be 'adjusted' as well?]
The most frequently sought data and tables will be downloaded as PDF files for most all geographic levels. Data users will be able to submit custom requests from a series of 'pull-down' menus that will produce a PDF and/or DBF file. Users will be able to aggregate geography through a GIS-like process.
While all this public access sounds great, it raises some issues that may end up involving the GIS staff because of your past and current use of census data and GIS.
Current Census Bureau plans call for several layers of PDF reports and maps that users would 'pull' from the Census 2000 server. If you want a 'book', print it yourself!
Data users will be able to select custom geography and data with cross-tabulations from through a GIS-like interface. Esri is working on that contract.
2 Matching Census 2000 to 1990 Census Statistical Geography
All Census 2000 block numbers will be different from the 1990 Census. Correlation files should be available from the Census Bureau, but local datasets using 1990 blocks will need to be eventually updated.
For some data users this will be a significant issue and task: cities that newly incorporated after the 1990 Census; annexed or de-annexed territory; changed Tract boundaries; changed names; adjusted blocks, and so on.
Casual census data users are not aware of some of the nuances of census terms and data such as 'reference person', 'median vs. mean' vs. 'household vs. family' income, the 'universe' for education is different from labor force - and so on. Casual users would either not read or not fully understand metadata.
With every Internet-linked computer able to quickly access census data but likely to do so in different ways, someone will need to establish official Census 2000 data, even just for general public reference, given the confusion that may incur with race data and adjustment. This task may fall to the GIS professionals, in whole or in part.
For the first time, two complete censuses will be readily available to casual users, free, on the Internet. This enables the scenario of many 1990 to 2000 Census trends and projections to 2010, or later. Both censuses have some level of undercount and error and there are the unknown implications of using adjusted and/or unadjusted 1990 and 2000 census data in trends and projections.
As I said at the beginning, the 1990 Census - with its one set of numbers that were somewhat difficult to access initially - will seem like the 'good 'ole data days.'
The next part of the paper turns to using real data that mimic what GIS professionals will face with Census 2000.
III 1990 and 1998 Sacramento Census Data sets
The City of Sacramento, state capital of California and home to 350,000 to 400,000 people (depending on what count you use), was the site of the 1998 Dress Rehearsal Census that incorporated adjusting the data for an estimated undercount. When combined with 1990 Census data, it is the only real example of what working with Census 2000 (and 1990 Census) data will be like.
Sacramento was selected because is a self-contained media market and has a diverse urban population. Dress Rehearsal (DR) censuses seldom achieve the public response rate of the actual census. The DR included a post-census survey to estimate the undercount and 'corrected' data were calculated and released. The Sacramento data are the only data available that mimic the situation GIS professionals will face when Census 2000 data become available. A basic reference map is attached.
The following research is based on the following four datasets (counts):
1990 Census
1990 Census adjusted (PL94-171 only)
1998 Census (PL94-171 only)
1998 Census adjusted (PL94-171 only)
The following tabulations and analyses are based only PL94-171 apportionment data for the City of Sacramento. As only the apportionment counts were adjusted for the 1990 Census, subject content is limited to basic population and housing counts for 1990. No housing counts were released in PL94-171 1998 data.
The city had 100 whole or partial census tracts in 1990. As part of the Dress Rehearsal the city was able to split several large tracts and account for an annexation, resulting in 108 tracts in 1998.
The first step, then, is relating the 1998 geography back to the 1990 Census. This was done simply by recombining the split tracts back into the parent 1990 Census tract (resulting in some large populations for tracts). The annexation, luckily, involved no population so that geography could be ignored.
Table B-2 is the new Census 2000 population tabulation format showing the revised Asian and Pacific Islander categories and the new 'More than One Race' category.
Again, adjusted and unadjusted data are shown and the effect on population groups is apparent. There are 63 potential single and multiple race categories, 6 for 'one race only' and 57 for 'two or more combinations.' Combined with the Hispanic question, there are 126 categories (63 X 2).The first major task is deciding how to link 1998 categories to the 1990 categories.
OMB outlines several ways of linking Census 2000 to 1990 data. They are:
Step 1: 'One Race Only' are simply matched to their corresponding category(ies)
1990 Census Census 2000
White White
Black Black
American Indian Alaska Native American Indian , Eskimo, or Aleut
Asian and Pacific Islander Asian + Native Hawaiian and
Other Pacific Islander
Other Other
[not reported] [not reported]
Step 2: The 'More than One Race' category could be treated several ways:
Deterministic Whole - assign entire count to one of the single categories:
'White and another race' to the other group
'two groups (but not White)' assign to largest single count
'two or more (including White)' to largest single count assign based on NHIS data
Deterministic Fractional - rules for weighting of multiple response
Assign in equal fractions to single race counts (50-50, 33-33-33, etc)
Assign based on NHIS historic data (60%-40%, for example)
There is no guarantee that any of the above or other OMB methods will ever be adopted as Federal policy. Even if one was, is would be subject to criticism and possibly legal action by some interest group or another. Nor is it likely the Census Bureau would be able to create the tabulations before Census 2000 data are expected for release. The 'regular' tabulations will still be released.
OMB is, in effect, creating a third set of Census 2000 tabulations that may eventually be required by certain Federal agencies, but would not be required for other census data users. My advice is, do not wait for OMB's solution, if it ever comes.
The data I have do not allow for any of these options as they necessitate working with raw data. I choose, instead, to look for easy and practical ways to tabulate and connect 1998 to 1990 data and tested three methods.
This option links the 1998 'more than one race' to the 1990 'Other' category. This is an easy way to deal with what will typically be about two percent of the total population, higher in California and New York, perhaps some other localities.
Hispanic Origin as Exclusive Category
The second option uses the Non-Hispanic race tabulations and again adds 'two or more' to the 'Other' category. This method treats Hispanic as essentially equivalent to a race category and more accurately reflects how most Hispanics actually answered. It also allocates any 'Two or More' Hispanic respondents to the Hispanic tabulation so that 'Two or More/Other' effectively has no Hispanic component.
Table D-1 illustrates this option for only one dataset (unadjusted 1990 to unadjusted 1998).
Proportioned to Total: Hispanic Origin as Exclusive Category
The third option allocates any response, single or multi-racial, to all the component race responses, resulting in an overall overcount. Then, the totals are proportioned back to the 'official' total.
Tables E-1 to E-3 illustrate this option.
Page 1 Page 2 Page 3Summary Comparison of Options
Option 1: 1998 More than One Race Category Grouped with Other Race, Hispanic Origin Tabulated Separately (Tables C-1 to C-5)
Good for quick and easy linking to 1990, but problem for any trends/projections as
'Other/More than one race' will show very high growth 19990-2000 change rate.
Comparison of 4 permutations of 2 data sets (adjusted and unadjusted) shows that 1990 to 2000 change and trends/projections vary widely and sub-population groups even more variation than the totals, especially linking 1990 adjusted to 2000 unadjusted.
This option continues to treat Hispanic Origin data as non-exclusive resulting in sum of population groups totaling more than 100%. I
Could try taking Hispanics into 'Other' and then separately listing Hispanic.
Option 2: 1998 More than One Race Category Grouped with Other Race,
Hispanic Origin as Exclusive Category (Table D-1)
This table totals to 100%. It mixes data in the sense that Hispanic Origin is treated as equivalent to race, yet Hispanics may include any race.
Combing 'Two or More' with 'NH-Other' still gives a very high 1990-2000 change rate that would be problematic for trends/projections.
Option 3: 1998 More than One Race Responses Allocated to All Categories
Proportioned to Total: Hispanic Origin as Exclusive Category
This option creates a new data set but may be best suited for redistricting analysis unless other data sets are prepared from raw data that 'divide up' the 'more than one race' responses to their respective race categories.
Option 3 captures the maximum count by race and Hispanic Origin but effectively double (or more) counts persons who respond as 'more than one race.'
The remainder of the paper was not presented at the UC and is not yet completed.
How are projections affected by using adjusted vs. unadjusted.
Using best option for re-tabulating 1998 data.
Table H-1 Adjustments by Tract and Population Group: 1990
Table H-2 Adjustments by Tract and Population Group: 1998
Table H-3 Impacts of Adjustment: 1990
Table H-4 Impacts of Adjustment: 1998
Map H-1 Characteristics of Adjusted Data: 1990 and 1998
Adjustments are not made to the housing count, what happens?
Table G-1 1990 Housing and 1990 and 1998 Population Datasets
Map G-1 Persons per Unit based on 1990 Raw Census Data
Map G-2 Persons per Unit based on 1990 Adjusted Census Data
Waiting for 1998 housing data to analyze and compare to 1990.
Look for 'new' small units with poor quality indicators
How will 'new' units affect CHAS , Redevelopment, and other planning and housing programs?
Work with lawyers on standing, probable 'causes of action' related to undercounts (even with methodology errors); data use injunctions, remedies;
Measures to head off legal action.
I Public Safety Issues Related to Two Censuses with Undercounts
Use results of housing and population research to determine how Census 2000 data will impact projections to 2010,
Use unadjusted and adjusted datasets.
See if public safety is affected by a 'structural' undercount now fixed in two censuses?
How will the change in housing unit definition affect the housing unit method for projections.
Goals and Objectives, why set up a local server?
Extracting Census 2000 (and 1990 Census) data for a local area,
Planning an analysis program,
Controlling access to 'official' data,
Linking to other databases,
Technical training for local data users,
Fee for census data related services,
Liability issues,
Preparing localized data products and maps,
Setting up a server for local users, and
Routinely integrating administrative and other data to census data.
ACS will provide small area 'sample' data on a rolling survey,
Planning to integrate ACS with local server and data
Planning for Census 2010
Create grid (surfaces) for vector data: income, race, etc.
How to related census data to MapObjects?
other
This is a work in progress. No conclusions, just issues and ideas for now. Comments are welcome by contacting the author at 'cbwillia@usc.edu'
REFERENCES (not in any order)
Report to Congress: The Plan for Census 2000, Bureau of the Census, August 1997
Census 2000 Operational Plan, Bureau of the Census, July 1997, Revised January 1999
Census Monitoring Board, April 1, 1999 Report #2
Census Monitoring Board, Feb 99, 1999 Report #1
Sampling and the Census: A Case Against the Proposed Adjustments for Undercount, Kenneth Darga (The AEI Press, Washington, DC), 1999
Understanding the Census: A Guide for Marketers, Planners, Grant Writers, and Other Data Users
Michael R. Lavin (Epoch Books, 1996)
2000 Census Preparations for Dress Rehearsal Leave Many Unanswered Questions: Report to the Committee on Governmental Affairs, US Senate, US General Accounting Office, March 1998
2000 Census Advisory Committee Meetings: Proceedings, Bureau of the Census (authors notes)
Maps and More: Your Guide to Census Bureau Geography, Bureau of the Census, April 1994
Creating a Census for the 21st Century: the Plan for Census 2000, Bureau of the Census, February 1997
Twenty Censuses: Population and Housing Questions 1790-1980, Bureau of the Census, October 1979
Wire services, e-mail newslists (www.census2000.org), and major newspapers
Counting People: the Census in History, Hyman Alterman (Harcourt, Brace & World; New York), 1969
The American Census: A Social History, Margo Anderson (Yale University Press), 1988
Census Bureau, PL94-171 Data (Adjusted), Technical Documentation, July 1991
Census 2000 Advisory Committee: Questions and Answers about Census 2000, 1998
Census Bureau Questionnaires, and Technical Documentation
Final Audit Reports on the 1998 Dress Rehearsals, Office of the Inspector General, U.S. Department of Commerce, Sept 1998, ATL-11050-8-0001, 11052-8-0001, IPE-10753, IPE-10756, ESD-10784-8-0001.
Los Angeles Times, various articles
Washington Times, various articles
AUTHOR INFORMATION
Dr. Christopher Williamson
Research Associate Professor of Geography and Planning
Census 2000 Research Fellow, American Planning Association
Director, Graduate Certificate Distance Learning GIS Program
Geography Department, KAP 416
University of Southern California
Los Angeles, CA 90089-0255
213-740-1767
213-740-0056 FAX