Important though information from the decennial census has become for applications ranging from public health surveillance to genealogy, the census is inherently political. Its constitutional mandate is to inform the redistribution of power through the apportionment of seats in the U.S. House of Representatives and, arguably, its most important secondary purpose is providing the raw material for the redrawing of political districts at the federal, state, and local levels. Inheriting as it does from the census and its previous long-form samples, the American Community Survey (ACS) likewise plays an important political role. ACS demographic data—and their detailed information on variables like primary language spoken at home, educational attainment, and physical or mental disability—play important roles in assessing the fairness of new districting plans and securing appropriate access to voting materials and public resources.
Section 7–A summarizes work in using the ACS to implement the Voting Rights Act’s provisions for language services for the diverse population of New York City, while Section 7–B describes the ways in which ACS data have been used in legal cases challenging redistricting plans. Expanding the scope to include the ACS’s role in monitoring social equity generally, Section 7–C summarizes work on studying disparate impacts in housing. To distill the themes from these presentations, the workshop steering committee asked Terri Ann Lowenthal—former staff director of the U.S. House of Representatives subcom-
In 2006—anticipating that the 2010 census would not include a long-form sample—the federal Voting Rights Act was amended to explicitly direct that ACS data be used to identify localities that must provide bilingual voting materials and assistance in areas with significant numbers of linguistically isolated households. Details of these requirements and examples from the first-time application of ACS data to these “Section 203” language-assistance requirements are described in Box 7-1. The mandate is etched in federal law but its implementation is—like the conduct of elections generally—inherently a function of state and local governments. As mentioned in the box, the New York City borough of Queens is obliged by the Census Bureau’s ACS-based calculations to provide foreign language election services in four language categories. In his presentation, Joseph Salvo (director of the Population Division of the New York City Department of City Planning [DCP], and presenting joint work with colleague Peter Lobo) described the challenges of implementing these services for Queens’ Asian Indian population.
Starting to operationalize the problem, Salvo said that DCP focused on what he called the CVLEP population—U.S. citizens of voting age (18 and over) who are limited English proficient (LEP). With that, the basic task may be restated as deploying language services to election poll sites with high concentrations of Asian Indian members of the CVLEP population. However simple that task might seem, stated in a few words, it is extremely complex on at least three conceptual lines:
- First, the “Asian Indian” category is far from homogeneous linguistically—India has nearly two dozen languages listed in its constitution. Queens’ population of 2.3 million would rank it as the sixth largest U.S. city if treated on its own; Salvo added that almost half of the borough’s population was foreign-born and that Queens was home to roughly half of the Asian population of New York City. The combination of the multiple languages under the “Asian Indian” category and Queens’ size means that there are likely to be sizable pockets of several Asian Indian languages. Indeed, Salvo’s analysis of 2010 ACS Public Use Microdata Sample (PUMS) data found many Asian Indian languages among Queens’ CVLEP population; the top five such languages are Bengali, Panjabi, Hindi, Gujarati, and Urdu.
1One speaker on the agenda, Kimball Brace of Election Data Services, was to speak on the use of ACS data in the construction of districting plans, but was unable to attend.
In 1975, P.L. 94-73 amended the Voting Rights Act to prohibit discrimination of access to voting by providing English-only voting materials in areas with concentrations of non-English language minorities; these provisions were included in a new Section 203 to the act, so the bilingual assistance requirements have come to be known as “Section 203 requirements.” The new law instructed the director of the Census Bureau to determine political subdivisions with such concentrations of language minorities or illiterate persons. Further amendment in 1992 (P.L. 109-246) provided additional detail on jurisdictions to be covered by these bilingual requirements and publish the same in the Federal Register, and 2006’s P.L. 109-246 changed the generic reference to “census data” as the source of the determination to explicitly direct that the American Community Survey (ACS) data be used. At the same time, the legal change provided that these determinations be updated every 5 years, rather than the implicit once-a-decade update based on the census long-form sample.
The current law mandates that states or subdivisions thereof are “covered” (and hence must provide bilingual voting materials) if (excerpting from 42 USC § 1973aa-1a(b)(2)(A)):
the Director of the Census determines, based on the 2010 American Community Survey census data and subsequent American Community Survey data in 5-year increments, or comparable census data, that—
(i) (I) more than 5 percent of the citizens of voting age of such State or political subdivision are members of a single language minority and are limited-English proficient;
(II) more than 10,000 of the citizens of voting age of such political subdivision are members of a single language minority and are limited-English proficient; or
(III) in the case of a political subdivision that contains all or any part of an Indian reservation, more than 5 percent of the American Indian or Alaska Native citizens of voting age within the Indian reservation are members of a single language minority and are limited-English proficient; and
(ii) the illiteracy rate of the citizens in the language minority as a group is higher than the national illiteracy rate.
(An entire state may be subject to Section 203 requirements under the 5 percent linguistic isolation threshold, but subdivisions under the state like a county are exempt from those requirements if their local share of that particular language minority group falls under 5 percent.) Under the same section of law, the director’s determinations are held to be “effective upon publication in the Federal Register and shall not be subject to review in any court.”
In October 2011, the Census Bureau published the first set of Section 203 determinations using 2005–2009 ACS data, listing covered states, county-level equivalents, and cities (76 FR 63602–63607); eligibility for Section 203 coverage was considered for 7,892 possible jurisdictions and 64 possible language minority groups (including Hispanic [Spanish-language assistance] and numerous American Indian and Alaska Native dialects). In addition to the new data source, the new determinations made use of a new methodological approach. As detailed by Joyce et al. (2012), the Bureau constructed hierarchical models after partitioning the data into a set of mutually exclusive “minority estimation groups,” generating estimates using empirical Bayes methods.
- Three states were deemed to be covered under Section 203—California, Florida, and Texas, all for Hispanic/Spanish-language assistance.
- As further examined in Section 7–A, Queens County, New York, was deemed covered with respect to four groups (Hispanic, Asian–Asian Indian, Asian–Chinese, and Asian–Korean).
- As discussed in Section 5–D, the Navajo Nation covers a land area roughly the size of West Virginia, spreading across parts of Arizona, New Mexico, and Utah. In all, 3 Arizona counties, 7 New Mexico counties, and 1 Utah county are obliged by the Section 203 determinations to provide Navajo language assistance.
- The nation’s most populous county—Los Angeles, California—qualified under 8 separate categories (Hispanic and 7 Asian languages, including “Other Asian—Not specified”).
- Relative to the prior set of determinations (from 2002; 67 FR 48871–48877), Vietnamese became a covered group in three additional large counties. It remained covered in Los Angeles, Orange, and Santa Clara Counties, California, and Harris County, Texas, and became covered as of 2011 in Alameda and San Diego Counties, California, and King County, Washington. Similarly, coverage of Filipino language groups grew from 5 counties (Los Angeles, San Diego, and Santa Clara, California, and Honolulu and Maui, Hawaii) to 9 (adding two Alaskan boroughs [Aleutians East and West]; Alameda, California; and Clark, Nevada); one previously covered county, Kodiak Island Borough, Alaska, dropped out of coverage.
- One jurisdiction qualified for Section 203 bilingual coverage for Bangladeshi speakers—the city of Hamtramck, Michigan.
- One way to simplify the problem is to focus on the segment of the population that identifies “Asian Indian” as their race, because many who are not Asian Indian also speak Indian languages. Indeed, Salvo noted that his ACS PUMS tabulation found a roughly even split within the single largest Asian Indian language group in the CVLEP population: Roughly half of CVLEP Bengali speakers self-identified as Asian Indian race, but roughly half did not. Three of the top five language groups—Panjabi, Hindi, and Gujarati—hadmuch higher levels of self-identification as Asian Indian race (80 percent or higher), but were not complete. Urdu, the national language of Pakistan, is also classified as an Asian Indian language, but only about 13 percent of Queens’ Urdu-speaking CVLEP population checked the Asian Indian race category.2
- Because the basic objective is to provide election language services, it is natural to focus on the finest-grained operational unit of electoral geography: the poll site (the actual location at which the language services must
2Pakistani is one of the explicitly mentioned write-in examples under the “Other Asian” race category, two items below the “Asian Indian” category, in the ACS race question: Person Question 6 on the 2012 questionnaire.
be deployed) and its covered jurisdiction. However, this electoral geography needs not—and does not—neatly correspond with statistical geography. Salvo noted that there are 669 census tracts and 316 poll sites in the borough of Queens; clearly, the arithmetic does not work to evenly match tracts to poll sites. In fact, Queens’ poll sites and their coverages are defined so that—on average—a poll site is built from parts of about five census tracts.
Salvo described the entire Voting Rights Act compliance project as a function of three basic components: ACS data, administrative data, and input from the Asian community. Having outlined some of the conceptual complexity of the problem, Salvo said that he wanted to reveal the bottom line at the outset and then describe how DCP arrived there, based on findings from (and limitations of) each of the components. They made the simplifying assumption to base calculations and analysis on the population that self-reported Asian Indian as their race, and the ultimate decision was to provide—in designated poll sites—written assistance in Bengali and oral language assistance to all other Asian Indian groups in the Hindi language.
The first major task in this project was to pull tract-level estimates of the CVLEP population for Asian Indian language groups, using 5-year ACS data (2006–2010)—and this work experienced an immediate, curious snag. Census Bureau estimates for the tracts in Queens yielded tract-level CVLEP counts for Gujarati, Hindi, and Urdu speakers but not the two major language categories of Bengali and Panjabi; those languages were lumped into an “Other Indic” language category. Salvo said that DCP had to bump up to a higher level of geography—Public Use Microdata Areas (PUMAs)—and look at the percentages within the “Other Indic” category who speak Bengali or Panjabi. They then applied those percentages to the “Other Indic” counts in each of the tracts in a particular PUMA to approximate the share of Bengali and Panjabi speakers in each tract.
As a second major task—to corroborate these estimates—DCP requested a special tabulation from the Census Bureau: tract-level tabulation of CVLEP estimates for the Bengali, Panjabi, Hindi, Gujarati, and Urdu language group. They initiated the request cognizant of the “massive disclosure issues” associated with it, but felt the need to seek the tabulation to determine if their work was “grounded in reality.” The agreement was far from perfect but it did suggest that DCP’s approximations were in the ballpark. For instance, the special tabulation identified 18 Queens tracts with 50 or more CVLEP Bengali speakers while the department’s approximate calculations from ACS data indicated 28 such tracts; the two sources agreed on 14 tracts and the 14 tracts flagged only by DCP’s calculations were all geographically proximate to the 14 jointly agreed-upon tracts. This provided some reassurance that the ACS-based calculations could be useful in identifying small areas of interest.
The third major task was to start the process of identifying needs at the poll site level, and was the point at which administrative data and input from the local community entered the mix. DCP sought lists of Asian Indian surnames compiled by various immigrant advocate groups, for each of the major Asian Indian language groups, and matched them to Queens voter rolls to generate counts of each language group by poll site. This provided auxiliary information on the concentration of CVLEP persons for each language group that could be contrasted with the ACS-based estimates—effectively, borrowing strength in finalizing estimates.
The fourth task was to confront the geographic mismatch mentioned at the outset: converting the ACS-based tract-level estimates to poll site levels. To do this, the Asian Indian CVLEP population by census block had to be obtained, which could then be aggregated perfectly into poll site areas. To get the CVLEP population by census block, a key assumption had to be made that the census block-to-census tract distribution of Asian Indian voters would also hold for the ACS CVLEP tract-level population.
The result of this task was, for each of the major Asian Indian language groups, poll-site-level estimates of the number of self-identified Asian Indians speaking that language who are also CVLEP. A map of those estimates could then be compared directly with the surname-and-voter-roll-based map for the final step: selecting the poll sites at which to supply language assistance. DCP settled on some uniform rules, to recommend that a poll site be identified for language assistance provision if it ranked high on both the ACS- and administrative data-based maps (say, for Bengali, a poll site with 50 CVLEP speakers in the ACS and 35 Bengali surnamed voters in the administrative data). Again, similar geographic clustering in both sets of data provided some reassurance—some “modicum of reliability”—that appropriate sites were being flagged.
The final details of the decision came about from input from the community. Bengali was the most common of the Asian Indian languages found in the data for Queens so it was a fairly natural selection. DCP learned from the Asian Indian community that Panjabi speakers commonly understand Hindi as well, Hindi being the national language of India. Because that permitted two of the major language groups to be combined, DCP reasoned that Bengali and Hindi would provide the most effective combination of languages to offer.
Salvo commented that “the great thing about this project is this—we are going to know whether this worked or not, because there’s going to be hell to pay” if it does not. Time—and flak from reaction as voters go to the polls in November—will ultimately determine whether this analysis steered language assistance resources appropriately. Of the experience, Salvo said that this work is indicative of the challenges federal mandates place on localities. But, more positively, he views it as a good illustration of how political (anecdotal and qualitative), statistical, and administrative data can all be brought together to craft solutions that work. Moreover, he argued, it serves as an argument for more
Criteria for the construction of political districts rank among the central protections in the federal Voting Rights Act. To preserve the right of protected, demographic minority groups to elect representatives of their choice, the act requires that so-called majority minority or supermajority districts be drawn in areas where there are sufficient concentrations of protected groups (rather than diluting their voice over multiple majority-driven districts). The act further specifies that general political redistricting lines cannot be drawn in ways that weaken the political power of the members of these protected groups. With that in mind, Jeanne Gobalet (Lapkoff and Gobalet Demographic Research, Inc.) began her remarks at the workshop by preemptively answering a question raised in several of the discussion sessions. She argued that, if the ACS were suddenly no longer available, “we will be back in the Dark Ages of the 1980s,” when important legal and voting equity determinations had to be made more through guesswork than solid quantitative evidence.
In her applied demography work, Gobalet said that she has used ACS data for several years in a variety of political and legal settings: in housing discrimination cases, studying different kinds of housing tenure and occupancy rates; in review of jury selection systems, ensuring that juries chosen in trials truly represent their communities; in assessing fair access to education by examining private school enrollment rates within public school districts; and in challenges of racially polarized voting. However, with the time constraints of the workshop, Gobalet said that she would focus on only one area where ACS data have usefully been brought to bear: comparing and assessing different scenarios for political redistricting.
Specifically, Gobalet said that the three case studies described in her presentation share three basic characteristics:
- Each case involved a California county governed by a Board of Supervisors, with those supervisors elected by district; the districting scenarios all concerned the supervisor districts.
- Each county in question has a large Hispanic population, many of whom are immigrants.
- Each case depended critically on the special Citizen Voting Age Population (CVAP) tabulation prepared by the Census Bureau from 5-year ACS
data; two different vintages of CVAP tables have been produced (for 2005– 2009 and 2006–2010).3
Procedurally, the work is complicated by the fact that the target districts are sufficiently small that census blocks are the necessary building blocks for supervisor districts—not higher-level geographic aggregates. So, Gobalet said, she goes through a process similar to what Salvo described for inferring finer detail from the available ACS data; specifically, she develops citizenship rates (by different racial and ethnic categories) at the census tract level from the CVAP tabulation, then applies those rates to voting-age-population counts—by census block—from the 2010 census. She added that she uses the CVAP tract-level data even though the CVAP releases estimates down to the block group level in order to reduce the problem of large margins of error.
The motivating question behind Gobalet’s first case study is one of basic feasibility: Can districts with Hispanic majorities be drawn to properly comply with the Voting Rights Act? The county she used to illustrate the fundamental challenges of drawing such lines is Monterey County, including the city of Salinas. To answer the basic question, she laid out a two-step approach—map the Hispanic CVAP shares for various subareas of the county and then compare those rates with the corresponding share of the total voting age population that is Hispanic (without regard to citizenship). The first map (making use of the ACS’s information on citizenship) might be thought of as the true eligible-to-vote map for Hispanics in the county, while the other—limited only to decennial census data provided in redistricting data files—reflects the information actually used to draw the lines. What Gobalet observed through this process is a negative correlation between Hispanic citizenship rates and Hispanic population shares—the higher the concentration of Hispanics in a census tract, the lower the citizenship rate (and thus the effective eligible voter population). She said that this demonstrated that—to preserve Hispanic voting rights—simply clearing the 50 percent threshold of Hispanic population in a district in this county would not ensure that the district would include a majority of Hispanics eligible to vote. An effective “majority minority” district in this case would have to have an extremely large Hispanic population share to meet the Voting Rights Act requirements—raising the possibility that such a district would look (on the surface) like an extreme case of “packing” the district.
For her second case study, Gobalet described a case in which her firm was engaged to examine four different redistricting scenarios that would partition a county (unnamed) into five supervisor districts; the task was to determine which of the scenarios comply with Voting Rights Act requirements, with the slight complication that Hispanic advocate groups had gravitated toward one particular plan where they sensed the plan would give them the ability to elect Hispanic
3The CVAP files and support documentation are available at https://www.census.gov/rdo/data/voting_age_population_by_citizenship_and_race_cvap.html.
supervisors in two of the five districts. In this project, Gobalet said that she examined numerous data sources, but arguably the most compelling work came down to the type of analysis outlined before: work with tract-level information from the ACS CVAP tabulation to estimate the Hispanic CVAP population in each district, for each of the four plans. Gobalet said that the real difficulty in this work comes in calculating the margins of error for these districts, formed by piecing together Hispanic CVAP shares from smaller-level geographies. Ultimately, though, she said that she was able to derive the margins of error and thus compute 90 percent confidence intervals for the Hispanic CVAP share by district.
Each plan—arbitrarily labeled A, B, C, and D—shared some basic traits. Hispanics made up the majority of the total population in four of the five supervisor districts in all of the plans; indeed, all of the plans included at least one district that is at least 70 percent Hispanic. The same holds true restricting attention to the voting-age population—Hispanic majorities in four of five districts. But working with the ACS to add in consideration of citizenship (and thus baseline eligibility to vote), Gobalet found that only two of the plans—B and C—included an effective Hispanic majority district, in which the 90 percent confidence interval for the Hispanic CVAP share of the population lies completely above 50 percent. Both Plan B and C included a second district where the confidence interval overlapped 50 percent, suggesting that there might be a second majority Hispanic district (but not necessarily); only in Plan C did the point estimate for the Hispanic CVAP share in this second possible district exceed 50 percent. Neither Plan A nor Plan D could be said with confidence to have an effective majority Hispanic district; A had two districts with the Hispanic CVAP confidence interval overlapping 50 percent and D only had one.
Gobalet said that the “punchline” of the second case study underscores another reality in the application of the ACS to legal and political matters: namely, that data do not always dictate the decisions ultimately made. In this particular case, the Board of Supervisors ultimately chose Plan D, the one with arguably the weakest claim to any effective majority Hispanic district. Gobalet briefly mentioned follow-up research in this case that paralleled parts of Salvo’s work—examination of surname lists from the county’s voter rolls. Even with the large Hispanic shares of the total population and the voting age population for most of the districts in all of the plans, none of the districts—in any plan—has a majority of voters (registered and actual) with Hispanic surnames. Gobalet said that this work further reinforces the notion that proposed districts with very high Hispanic concentrations may not really have effective Hispanic majorities; it determined that none of the districting scenarios really have the effective Hispanic majorities that advocates had hoped would obtain for two of the five districts. Gobalet added that the county in question faces a strong possibility of Voting Rights Act litigation.
The third case study follows up from the methodology briefly mentioned
at the end of the second case study, in the interest of answering the question of whether local administrative data on voters can be used instead of or in addition to the ACS data in Voting Rights Act analyses. Returning the focus to Monterey County, California—a jurisdiction for which she has done a lot of redistricting work—Gobalet said that she has had the opportunity to do detailed geocoding and surname analysis of the voter rolls. She emphasized that surname analysis might be feasible for Spanish-origin names but is not believed to be reliable for identifying blocs of other minority groups; African American surnames are not distinctive enough to be singled out, and she said that the methodology has been found to have limited utility for identifying Asian American voters. But—if workable to identify Hispanic voters—the surname-based voter roll data has the attractive feature of being linked to precise addresses (and thus exact geographic locations) that can be linked to arbitrarily small geographic areas.
Gobalet said that her work with the Monterey County registered voter data has suggested fairly strong concordance between the 2006–2010 ACS CVAP tabulation and the surname-based Hispanic voter shares, with the latter computed separately from registered voter rolls as of November 2008 and the rolls as of November 2010. Monterey County is subject to U.S. Justice Department pre-clearance of districting plans under the Voting Rights Act, so she conducted this work in the context of evaluating two alternative districting plans partitioning the county into five districts—one developed by county authorities (and ultimately adopted) and a last-minute alternative proposal backed by the county’s largest city, Salinas. Across both proposals and both vintages of the surname-based Hispanic shares, the Hispanic CVAP percentage and Spanish-surnamed registered voter share differed by no more than 5 percent for any district. She concluded that it is reassuring that both data sources seem to be measuring the same basic thing so, in some cases, it may not be necessary to rely solely on the Hispanic CVAP data from the ACS. That said, she reiterated that the CVAP data have more utility for examining concentration of minority groups other than Hispanics.
She closed by repeating her bottom-line conclusion from all three case studies—that the ACS-based CVAP special tabulation has become essential to ensure compliance with the letter and intent of the Voting Rights Act. Absent the ACS, things would necessarily return to the days of basing important determinations on guesswork and hunches.
Andrew Beveridge (Queens College and Graduate Center, City University of New York, and Social Explorer, Inc.) began his remarks by describing himself as “sort of an accidental demographer; I stumbled into this field without knowing too much about it, after graduate school.” Getting involved in civil
rights issues and then serving as president of the Yonkers, New York, school board changed the focus of his work and research, and he has since been actively involved in using demographic data in a variety of areas under the broad heading of civil rights: redistricting challenges, jury selection systems, review of New York Police Department policies, and the topic that he elected to focus his workshop presentation upon, the study of housing disparity. He started his comments by observing that the standard tabulations from the ACS provide much information about the differences (and potential disparities) between various racial and ethnic groups, and analysis is even more powerful using the ACS PUMS data to tailor areas to the question of interest and choose how the tabulation is done; the frequency of ACS releases also make it preferable to the once-a-decade long form. Hence, as Gobalet did before him, Beveridge stated up front his view that—were the ACS to go away—the types and the richness of disparate impact analysis that are now possible for understanding housing conditions would be impossible.
Reviewing the established standards for demonstrating disparate impacts—not just in housing but also in other areas like employment discrimination—Beveridge emphasized that the key feature of such cases is that one need not prove intent to discriminate. The policy or action in question may be facially neutral yet still be found to have a disparate impact on minority groups if all three of the following are proven:
- The policy or action affects some members of some minority group more than the rest of the population.
- The housing referenced in the policy or action is more likely occupied—or more likely to be occupied—by minority groups.
- The building, failure to build, or destruction of housing like that covered in the policy or action is likely to have a disparate impact on such minority groups in general.
Beveridge said that the ACS is emerging as a particularly valuable tool in analyses on the second and third of these points.
Beveridge used as an example a recent filing in a case to block the construction of subsidized housing in Kenosha, Wisconsin. A developer proposed building low-income housing through the use of Low Income Housing Tax Credits (LIHTC) and the project was readily approved by the municipal government, but soon experienced serious public opposition. The city responded by blocking the project; the developer sued to force development, and so the question occurred: Are there disparate impacts associated with the city’s blocking the new housing development? As additional context, defining the parameters of the study and the affected population, Beveridge noted that the LIHTC program is targeted toward construction of units serving people who make less than 60 percent of the median income in a particular area. The proposed new housing was also to include some units open to people with Section 8 public housing choice
Beveridge’s analysis in this case made use of ACS data for the PUMAs in the city of Kenosha, for the surrounding Kenosha County, and the tri-county region comprised of Kenosha County and its immediate neighbors to the north and south (Racine County, Wisconsin, and Lake County, Illinois, respectively).4 For exploratory purposes, he began by plotting a choropleth map shaded based on the percent of the population coded as non-Hispanic black; he overlaid on this map a layer of circles, centered on the location of existing LIHTC-funded projects in the area and drawn with the radius proportional to the number of units in the project. Zooming into the city of Kenosha, Beveridge observed that the proposed development site was in an area with moderate to strong minority populations but that is not already being served by similar LIHTC projects of any substantial size.
A next step in the analysis was to compare the composition of the Section 8 housing choice voucher waiting list for Kenosha County with the underlying demographics of the county as reflected in the 2005–2009 ACS PUMS data. Beveridge concluded that the disparate impact of getting into new Section 8 housing is fairly evident: the number of non-Hispanic blacks on the voucher wait list represents almost 75 percent of the non-Hispanic black population of Kenosha County as a whole, while wait-listed non-Hispanic whites make up just under 3 percent of the county’s total non-Hispanic white population.
Those initial steps suggested the possibility of a disparate impact case, but fuller analysis requires examination of income data. The Census Bureau now regularly prepares a special tabulation from ACS data for the U.S. Department of Housing and Urban Development (HUD) known as the Comprehensive Housing Affordability Strategy (CHAS) data; earlier vintages of CHAS were prepared using 1990 and 2000 census long-form-sample data.5 The CHAS data provide household counts (as fine grained as the place/city level) of households in need of housing assistance and with incomes that qualify for HUD programs like LIHTC—including shares of households that fall below various percentages of the prevailing area median income defined by HUD. For the Kenosha example, Beveridge looked at data for households meeting the two basic criteria to be eligible to move into the proposed housing development: household income not exceeding 60 percent of area median income and median rent not exceeding 30 percent of the area median income. Again, Beveridge concluded, evidence
4Kenosha County also shares shorter borders with Walworth County, Wisconsin, to the west and McHenry County, Illinois, to the southwest, and both those counties show up in some of Beveridge’s analytical maps. Kenosha County and city is bordered on the east by Lake Michigan.
5To date, the Census Bureau has produced two iterations of CHAS data using 3-year ACS data (for 2005–2007 and 2006–2008); per HUD’s website, HUD requested and the Census Bureau delivered in early 2012 a 5-year CHAS data file for 2005–2009. See http://www.huduser.org/portal/datasets/cp/CHAS/data_doc_chas.html.
of disparate impacts are fairly clear; only a small percentage of Kenosha’s non-Hispanic white population meets the eligibility criteria for the proposed development (roughly 8 percent), but just over 20 percent of the city’s non-Hispanic black and Hispanic populations qualify. Similar disparity is evident when looking at the entire tri-county area.
A final piece of evidence in the case concerned the eligibility of disabled persons for entry into the proposed development; in addition to the income requirement (not exceeding 60 percent of median area income), households must have one member classified as disabled to be eligible for Section 8 vouchers. Running the eligibility criteria again, this time for disabled (technically, households qualified as disabled) versus nondisabled, the divide was again stark; Beveridge found that just over 25 percent of the nondisabled population in Kenosha met the eligibility requirements compared to nearly 50 percent of the disabled population.
Ultimately, Beveridge said, the disparate impact among African American and Hispanic households of not building the LIHTC housing became evident to all parties. The city of Kenosha and the developer agreed to settle the breach-of-contract case out of court, with the developer winning a multimillion-dollar settlement for damages and expenses from the city.
Beveridge concluded his remarks by briefly discussing his other work with census and ACS data, through Social Explorer—directing workshop users to the company’s website at http://socialexplorer.com—and offering a few general comments on census and ACS data. First, having worked with The New York Times on several projects, he commented that he agreed with Ford Fessenden’s approach to mapping (see Section 4–C)—to consider the meaning of small-count areas or gaps but not to let them interfere with telling a story with data, and to drop them if necessary. On a related note, and concurring with Fessenden’s and Salvo’s approaches (see Section 7–A), he recalled a favorite example in making the point that it is possible to become overly paralyzed by margins of error—even the “perfect,” “real” data of the decennial census has curious anomalies. He said that he distinctly recalls getting a phone call shortly after the release of detailed data from the 2000 census, from a Times reporter with a curious question: “What’s going on with the 59 people that live in Central Park?” The reporter had noticed, amidst the census block counts, that population total corresponding to the park; the reporter called the Census Bureau for clarification, and said that the Bureau staff did not have a good answer to the question. As it turns out, Beveridge said, there has been a small population credited to Central Park since the 1980 census, but there is not a very clear answer to the question of why they are counted there. A large part of it might be homeless people encountered during operations like 1990’s S-Night (Shelter and Street Night) and 2000 and 2010’s Service Based Enumeration program, but some might be from data swapping to preserve confidentiality.
Moderating the discussion session, Terri Ann Lowenthal (Funders Census Initiative and the Census Project, and former staff director of the U.S. House subcommittee with oversight of the census) began by commending the presenters for what she called a strong set of presentations on issues that all, in some form, tie back to the original purpose of the census itself. The constitutional mandate for the decennial census is to fairly allocate representation through apportionment and, later, redistricting; through the long series of Supreme Court cases on redistricting and the enactment of related law like the Voting Rights Act, issues of ensuring social equity—fairness in representation writ large—have emerged as major uses of census data, so it is good to see the ACS data satisfying the same mandates.
That said, by way of summarizing both the session and what she had heard in the workshop as a whole, Lowenthal said that she had noted three “big disconnects” from the political and policy perspective that she said need to be considered going forward. First, Lowenthal observed a disconnect between ACS data users and the current political climate. She noted that several presentations and discussion sessions had ended in some “wish lists,” some for more questions or topics, some for more frequent products, and many for a larger sample size. Though understandable and sensible within the community of data users, Lowenthal argued that users need to recognize that such wishes are “complete anathema to what is going on in the political environment right now.” She argued that it is a major, looming communication challenge for all sides; ACS data users “have to hear” and listen to what ACS critics are saying (even though, Lowenthal senses, the critics are, “frankly, vastly in the minority”) and understand their arguments. At the same time, data users have to recognize that the language of some of their wishes—“we really need a larger sample” mapping roughly to “we want more people to get that burdensome form that you don’t want anyone to get”—are not arguments that will carry weight with critics.
The second “disconnect” Lowenthal observed is between the Census Bureau and the respondents to the ACS (“otherwise known as the American public”). She said that the Census Bureau has made efforts to explain why specific questions are asked in the ACS—and, to a lesser degree, how those questions and the answers to them benefit the public. But she said she is concerned that the Census Bureau is not answering the concern in quite the right way. She recalled that, during her time as staff director for the House census subcommittee in the mid-to late 1980s, the Census Bureau went about addressing the problem of justifying the questions on the census long form by compiling “The Notebook”—a binder that listed each long-form question and attempted to outline the legal and regulatory requirements behind the question. “The Notebook” has endured, and continued into the ACS era, albeit now in a digital form: roughly, “we collect
data, or publish data, on journey work because Title X or Section X of this act says that we need these data to implement this program.” This information is all well and good, but Lowenthal argued that it is simply “too esoteric for the average citizen, and dare I say, for the average member of Congress.” She said that specific examples of ACS use like the work presented at the workshop are likely to be “much more powerful” and persuasive to decision makers. In short, she said, she would encourage the Census Bureau and ACS stakeholders “to really talk in more basic terms” and concrete examples in discussing the need for the ACS.
Continuing, Lowenthal argued that ACS users—and the specific examples of ACS uses for important purposes—need to take care in how they justify some of the ACS questions that draw the most confusion and ire:
- Lowenthal observed that the questions “that are being ridiculed the most right now are those on disability.” Read cold, absent any context, the language of the core disability question—about having “serious difficulty concentrating, remembering, or making decisions” or having “difficulty dressing or bathing”—comes across as “a really bizarre personal question,” and it is unfortunately easy to lampoon the question as such.6 What has been missing from the discussion, she argued, is a clear explanation that the question is not asked to pry into any individual person’s life situation but rather to get information about various dimensions of disability—pieces that can be constructed to derive information on the disabled population generally—to make important policy decisions. Beveridge’s example of the disability question being essential to judge fairness in housing policy is one of many others that could be then used to make a solid case for the question.
- Another frequently lampooned or criticized question is the one at the heart of journey-to-work data: “What time did this person usually leave home to go to work LAST WEEK?”7 Absent context and good examples, the question is easy to criticize as trying to dupe respondents into revealing the best times to ransack their homes. Again, Lowenthal suggested, a missing step in the logic is that “they [the Census Bureau] really don’t care when you leave the house and when you get home”; they’re asking the question to get a sense of when roads in the vicinity might experience peak traffic and when transportation routes are being most (and least) utilized.
6See Person Question 18 on the 2012 questionnaire; the text excerpts from parts a and c of the question, and part b asks “Does this person have serious difficulty walking or climbing stairs?” Question 19 adds: “Because of a physical, mental, or emotional condition, does this person have difficulty doing errands alone such as visiting a doctor’s office or shopping?”
7Person Question 33 on the 2012 ACS questionnaire.
- Finally, some of the housing stock questions—the number of rooms and the number of bathrooms—are commonly criticized; Lowenthal said that what is missing is a rationale (with examples) for asking a series of questions rather than asking the overly blunt and likely misleading question “Is this house overcrowded?”—“because if you’re mad at your wife that day, you just might say it is.”
The third disconnect Lowenthal noted—and, she admitted, a hard one to resolve—is between Congress and the work that it does. She said that the Congress that asks “Why are we gathering these data?” and hammering the ACS as intrusive is the same Congress that periodically has to reauthorize programs in major surface transportation bills—seemingly missing the connection that the allocation of funds under those acts, and all of transportation plans that local and state governments have to submit, relies heavily on the journey-to-work data that are currently being collected systematically only by the ACS. The same phenomenon holds true in other policy domains as well—all the more reason, Lowenthal concluded, that clear examples of how ACS data (at fine levels of geography) that are essential to responding to congressional mandates are going to be “more powerful arguments for lawmakers than many other things that we can say.”
Turning to specific questions, Lowenthal asked the panel of presenters the same question that was raised in other discussion sessions—what would you do, and how would your analyses change, if the ACS were no longer available. Gobalet’s presentation had discussed the possibility of Hispanic surnames from voter registration rolls as one possible alternative data source, but—though those data were workable in the specific Monterey County example she examined—she concluded that the ACS data were more flexible and ultimately more useful. Accordingly, Gobalet answered Lowenthal that she would see no alternative but to go back to the 2000 census long-form data and to the “last,” most recent iteration of ACS data. Lowenthal probed on that point and asked how long that would last; Gobalet answered that, at some point, it would be clear that those antique data are effectively useless, but that the basic argument that she would have to make is “these data are useless but they are all we have.” Beveridge said that his sense of what would happen is that some of the commercial data vendors—“some of whom may even be in this room”—would derive products that are more model-based—necessarily involving more conjecture, being based in part on old data—and that there would be considerable uncertainty in exactly how the estimates are derived. Salvo said that his work with Lobo pushed ACS data to their limits, but that having those data gives him a distinct power: the ACS data are far from perfect, and no data are completely neutral, but the ACS data impart power by forming “a basis for discussion” and policy debate. He said that his concern is that, were the ACS to go away, that basis for discussion
On specific alternatives—the use of surname lists having been invoked—Salvo added that it was challenging to come up with a comprehensive list for Asian Indian languages, but the advocacy groups were up to the task. It is, he said, a task that “borders on insanity” and was tremendously difficult. But—just as Gobalet found for the Hispanic surname list in Monterey County—Salvo conceded that it was interesting that the process yielded good results, when “combined with other items and combined with reality checks on the ground.” (Or at least, he hastened to add, the process yielded good results—with the “goodness” yet to be determined with the performance of poll sites in November.)
With the floor opened up to general questions and comment, Steve Murdock added a comment concurring with Gobalet about what would happen without the ACS. He recalled a time in the mid-1980s, working in Texas and being delivered a bunch of data from the Texas Department of Human Resources—“just incredible breakdowns,” for all manner of demographic subgroups, for all 254 counties in the state. When he called the department and asked how they had arrived at what was presented as current and detailed data, they replied—essentially—that they had “made up population estimates for the counties and we used the same rates, or whatever, that were in the 1980 census.” His concern is that the same form of crude extrapolation, from increasingly old and unreliable data, would be the alternative if the ACS were to be discontinued.
Lester Tsosie (Navajo Nation) asked the presenters to comment on the social equity uses of the ACS in light of the changing demographics of the nation, with longstanding racial and ethnic minorities becoming more significant segments of the total population. The state of New Mexico is already at or near “majority minority” status, with about half the population in the 2010 census being of Hispanic origin; do the presenters have a sense of implications—more injustice or less injustice—from changing demographics, first in states like Arizona and New Mexico and later in other parts of the nation, particularly if the ACS were weakened? Gobalet answered bluntly that her sense is that neither the letter nor the intent of the federal Voting Rights Act—as currently written and interpreted by the courts—can be enforced without ACS data. Beveridge replied to Gobalet’s point that, unfortunately and ironically, “there are people who would be fine with that.” That said, he agreed that—without the ACS data—it is simply impossible to accommodate the demographic changes Tsosie referred to with respect to voting.
Alan Zaslavsky (Harvard University) said that he was struck in this session, and in some earlier sessions, that the point is made about small counties in rural areas suffering from lack of ACS samples—even though, operationally, their rates are higher in some cases because of concern about producing estimates for local civil divisions. Given that much of the work described here involves very small areas (and very small areas within urban areas as well), are there impli-
cations about sampling these different kinds of areas in being able to meet the needs for ACS data? Salvo agreed, recalling that he listened to Andrew Conrad’s presentation about economic development uses in Iowa (see Section 6–A)—and its recitation about how few cities and counties in Iowa have populations over 20,000—with jaw agape; “we have neighborhoods in New York with 30,000 people, 40,000 people.” So he conceded being a bit embarrassed to make things sound like an argument for an increased sample in a borough like Queens—but “the law is the law, and Section 203 implementation requires detail,” and that is going to push the limits of the ACS sample even in the densest of areas. He agreed that more effective ways to sample in rural areas need to be considered—indeed, he said it would be a real death knell for the ACS if the sample in New York City were increased at the expense of rural areas, because the ACS would lose value as purely “an urban area/New York City thing.” Lowenthal said that she wanted to raise and put on the table an idea that Ken Hodges (Nielsen) has raised in the past—instead of trying to increase the ACS sample “in fits and starts,” try to find a way to convince Congress to link the sample size to some automatic measure (like the estimated number of housing units in the nation). She said that such incremental, automatic changes to sample size might be more palatable than getting approval for larger chunks of sample.
Struck by Lowenthal’s description of “the most ridiculed question” on the ACS, Stephen Tordella (Decision Demographics, Inc.) said that he had to wonder if putting legislators on the spot might be an effective argument—countering ridicule of the disability question by asking, in reply, “Are you ridiculing the disabled?” Beveridge answered that his own reaction to Lowenthal’s discussion of the disability question was that it would be a good idea to get disability advocates involved and engaged in discussions of the ACS. Likewise, with some other questions, he wondered whether it would be a good idea to get civil rights advocates more engaged in ACS-specific efforts. Lowenthal agreed with Beveridge, adding that the Leadership Conference on Civil and Human Rights is currently one of the most active organizations in support of the ACS and the census; she said that getting disability advocates more involved could bolster the rationale for the question, and is certainly consistent with resolving the disconnects she described in her opening remarks.