– 4 –

Redistricting and Related Legal Uses

Joe Salvo (New York City Department of City Planning) moderated a session on the use of decennial census data for legislative redistricting and related legal issues. He noted that redistricting of the U.S. House of Representatives is next in line after reapportionment of seats in the House in terms of priorities for the census. The P.L. 94-171 file, which provides data for redistricting down to the block level, is used for redistricting of state and local legislative bodies in addition to the U.S. House and is to be delivered to the states by March 31, 2021, is a critically important census data product.¹

The session included three speakers: Justin Levitt (Loyola Law School), who addressed differential privacy in relation to redistricting and other provisions of the Voting Rights Act (VRA); Andrew Beveridge (Queens College and

___________________

¹ Pursuant to 13 U.S.C. § 141(b) and (c), the Census Bureau is required to deliver apportionment counts within nine months of Census Day and the P.L. 94-171 redistricting files within 12 months—for the 2020 Census, by December 31, 2020, and March 31, 2021, respectively. Due to the novel coronavirus pandemic, the Census Bureau suspended 2020 Census field operations in March 2020 and the administration requested that Congress extend the deadlines by four months, continuing nonresponse follow-up operations through October 31, 2020, and changing the deadline for apportionment totals to April 30, 2021, and redistricting files to July 31, 2021. However, action was not taken in both houses of Congress on this request and, on August 3, 2020, the director of the Census Bureau announced the Census Bureau’s intent to end field operations by September 30 and meet the legislatively mandated deadlines. A federal district court enjoined the September 30 end of data collection and the December 31 delivery of apportionment population counts. Ultimately, on appeal, a U.S. Supreme Court order cleared the way for the Census Bureau to end data collection for the 2020 Census on October 15. At the writing of this summary, the timing of data product delivery remains unresolved, undergoing legal challenge in several federal courts.

Graduate Center, City University of New York), who used New Rochelle, New York, as a case study for the use of differential privacy for redistricting; and Michael McDonald (University of Florida), who addressed differential privacy and redistricting with the state of Georgia as a case study. Floor discussion followed their presentations.

4.1 REDISTRICTING AND THE VOTING RIGHTS ACT

Justin Levitt (Loyola Law School) began by noting that he is a lawyer talking about statistics, and his remarks should be taken accordingly. He had not had much time to analyze the 2010 Demonstration Data Products (DDP) but had made some observations from the work of others. Mostly, he wanted to set the legal framework for the other two speakers.

He noted that legislative districts are “off-spine,” that is, not in the hierarchy of geographic areas for which the TopDown Algorithm (TDA) maximizes accuracy (as shown in Figure 2.1), and so are built up from the lower-level block groups and blocks. Levitt discussed census data use cases in terms of accuracy versus privacy pertaining to redistricting, provisions of the VRA outside of redistricting, and other electoral impacts of census data and differential privacy.

4.1.1 Issues in Redistricting

Levitt referred to work by David Van Riper (Minnesota Population Center), comparing the now-public returns of the 1940 Census with the Census Bureau’s release of 1940 Census data subjected to an early version of the 2020 Disclosure Avoidance System (DAS). This analysis determined the degree to which the numbers got more accurate that as ϵ increased. This was less true in smaller jurisdictions and for smaller populations within smaller jurisdictions. Levitt said that this feature could have a big impact on redistricting, although the concerns would be greater with respect to the utility of the data for smaller areas—cities and counties—than for congressional districts. Also, for redistricting purposes, the race numbers would matter more than the total population numbers. Smaller race group populations would mean bigger equity concerns. While understanding the funding implications of wanting to make sure that the total population numbers were as accurate as possible, accuracy of the race numbers would be far more important for this use case.

Levitt raised questions about the value of ϵ that might be used for 2020, ϵ having been set to 6 in the 2010 DDP (4 for the person-based data tables and 2 for the housing unit-based tables). He wondered whether the DDP revealed the expected utility in the event that some of the privacy budget was eaten up by block-level citizenship data. A similar concern is how much of the privacy budget might be eaten up by the release of further data products along the way—-

such as data comparable to the 2010 Summary File 2 (SF2) or the American Community Survey (ACS).

Levitt turned to the issue of equal population legislative districts. He observed that there was a notion in the 1990, 2000, and 2010 cycles that congressional districts must be exactly equally populated; it is deeply embedded in the minds and hearts of redistricters. It was true, and for many people it is still muscle memory. But it is no longer true. State laws and legal opinions may vary, but Levitt argued that the truth is that congressional districts should be about the same size, as clarified by the U.S. Supreme Court in 2012 in upholding the Court’s 1983 conclusion that deviations in district size could be justified if “necessary to achieve some legitimate [government] objective.”² By comparison, Levitt said that the courts have permitted deviations up to 10 percent in state legislative district sizes. Levitt believed it likely that small differences in population size would not be disqualifying for redistricting purposes after the 2020 Census, for either the congressional or state legislative district standards, at fairly large ranges of ϵ in the 2020 Census system.

Levitt referred to analyses of the 2010 DDP data by the Caliper Corporation³ that showed that differences from the original 2010 Summary File 1 (SF1) were small for congressional districts. Yet even small deviations, such as 600, 1,000, or 2,000 people, had a systematic bias. The Caliper maps suggest that the 2010 DDP increased the rural population relative to the urban population. This would represent a systematic shift in political power allocated to rural populations, not by very much in any individual district but quite a bit overall. Levitt noted that, as expressed throughout the workshop, the problem appears to stem from the post-processing part of the TDA to produce census tables without negative values. Correcting negative numbers introduced by the differential privacy injection of statistical noise meant that smaller populations only got bigger while larger populations got both bigger and smaller.

The upshot, depending on the value of ϵ, seemed to be that adding noise would be unlikely to create a problem for larger congressional districts, but would likely represent a systemic shift towards rural populations. Adding substantial levels of noise, depending on the final ϵ, could create a sizable problem in smaller districts, such as city council districts and county commission districts where there were very few people in a county.

___________________

² Levitt cited Tennant v. Jefferson County Commission (567 U.S. 758, 2012), a challenge to West Virginia’s 2011 congressional redistricting plan, which itself derives the “legitimate state objective” test from Karcher v. Daggett (462 U.S. 725, 1983). He also cited Brown v. Thomson (462 U.S. 835, 1983), which centered on the apportionment of state House seats (with at least one per county) in Wyoming.

³ See https://www.caliper.com/census-differential-privacy-maps/index/default.htm.

4.1.2 Voting Rights Act

Levitt moved to discussing relevant provisions of the VRA. Under specified conditions, certain jurisdictions have the responsibility under the VRA to draw districts to facilitate minority political power. Levitt briefly reviewed these so-called “Gingles conditions” (named after their specification by the Supreme Court in Thornburg v. Gingles, 478 U.S. 30, 1986) that apply: a minority group of sufficient size and geographic concentration, in a context in which voting is polarized along those demographic lines and in which the preferred candidate of that minority group typically loses elections. Levitt said that these conditions, or threshold liability determinations, are critically important because they deal directly with winning elections unlike most of the other use cases, which deal with relative power over a broader area.

Levitt identified two primary ways in which differential privacy methods might affect enforcement of the VRA in redistricting. The first was whether a minority group would be of sufficient size to be essentially more than 50 percent of the electorate in a district-sized population, citing the Supreme Court’s ruling in Bartlett v. Strickland (556 U.S. 1, 2009). Would the minority population be big enough?

Using the 2010 DDP, Levitt looked at relatively small state legislative districts, specifically, the Delaware State House, which has districts of about 15,000 people. Obviously, all of this would be amplified for smaller city council districts or county commission districts.

Looking specifically at the African American population, Levitt found that, at the delivered levels of ϵ for the 2010 DDP, most of the black voting-age population estimates in Delaware State House districts lined up with the reported numbers. However, the threshold (50 percent or more for VRA enforcement) really matters, and Delaware has a district that is right on the edge, a district that would get VRA enforcement with the numbers as delivered in 2010 but would not get VRA enforcement with the numbers delivered in the DDP. The percentage difference was relatively small but enough to put the district under the threshold.

4.1.3 Polarization of Voting by Race

Levitt noted that precinct-level data are needed to analyze voting preferences of race groups, which factors into determination of whether voting is sufficiently polarized to create potential liability. Discussions about whether minority communities favor different candidates than majority communities most of the time are based on ecological inference models. Those depend on the data in each precinct and seeing whether they form a pattern and if noise were to be introduced into the precinct composition, noise would affect the correlation of whether voting was polarized based on race. The result

might be to reduce the opportunity to discover polarization where it actually existed. This phenomenon would have more of an impact in homogenous communities, because of the non-zero editing to produce the DDP. If there were a homogenous black precinct that voted for one particular candidate and noise were introduced that made that precinct appear less homogenous, then it would appear as if there was more agreement across racial lines than there might actually be, limiting VRA enforcement possibilities.

Levitt said he did not know how the noise introduced by differential privacy would affect polarization analysis. Similarly, he did not know whether the effects would be exacerbated in municipal elections in which district populations were smaller than in, say, state legislative districts. More study is needed.

4.1.4 Language Access Determinations

Section 203 of the VRA (codified at 52 U.S.C. § 10503) stipulates that jurisdictions must make special accommodations for voting for language groups that make up 5 percent or more of the citizen voting-age population (CVAP) and are not proficient in English. As the act has been amended, the law explicitly directs the use of ACS data for this purpose. Levitt indicated that a real question was how much differential privacy would affect the ACS. Even if differential privacy was not directly applied to the ACS, its use in the census would affect the ACS. There are two reasons why the decennial would affect the ACS: (1) the census is the foundation for the ACS sampling frame and census data are used in the process of weighting up the ACS data; and (2) census data have been used to help refine the precision of estimates from the ACS for Section 203 purposes by reducing the sampling error.

What Levitt did not yet know was which decennial data would be used for the ACS, pre-DAS or post-DAS. Even for internal use, Levitt understood that using the predisclosure avoidance decennial would consume some privacy budget, which would mean that less of the privacy budget would be available for the publicly released tabulations. A postdisclosure avoidance decennial would not eat up any more privacy budget, but then the noise would already be baked in for developing ACS weights or for refining ACS results.

In addition to the information needed for Section 203 of the VRA, SF2 tabulations of detailed race and ethnicity in the 2010 Census (and analogues in previous censuses) have proven very helpful to precincts in determining their obligations to provide voting assistance. Levitt said it was not at all clear whether the privacy budget for 2020 would allow for a full release of SF2-type tabulations for the 2020 Census.

Levitt concluded by expressing the hope that he had provided a framework for the next two presenters.

4.2 IMPACTS ON REDISTRICTING: THE CASE OF NEW ROCHELLE, NEW YORK

Andrew Beveridge (Queens College and Graduate Center, City University of New York) opened his presentation by noting that he wears several hats. He is a litigation consultant and a professor, so he has used census data for research and in a variety of court settings. By examining a particular redistricting case in which he was involved in 2011, for New Rochelle, New York, he hoped to answer or at least raise three questions. In his presentation, he labeled the 2010 DDP as “synthetic data” to emphasize that the data—in his view—are massively changed and materially different from the original 2010 redistricting data.

Beveridge’s first question was whether synthetic data could be used to assess accurately equal populations of districts among redistricting plans. His second question was whether synthetic data could give an accurate measure of the presence of minority groups in a district in order to construct a majority-minority district. His third question had to do with using blocks to develop alternative district boundaries for evaluation. For this purpose, the block data must be accurate, otherwise the resulting districts could be affected. Beveridge cautioned that his analysis was based on one case.

4.2.1 Redistricting the City of New Rochelle

Beveridge helped draw a redistricting plan for the city of New Rochelle, New York, in 2007 based on 2000 Census data and again in 2011 based on 2010 Census data. New Rochelle, which has about 77,000 people and abuts New York City, is on a four-year cycle for its council districts. Beveridge said that the 2007 redistricting plan arose from the city’s settlement of a VRA case filed against the existing plan, which had argued that the then-current redistricting plan diluted a majority African American district. Beveridge said that the solution in the 2007 redraw involved joining two majority-white areas by a very narrow (as little as 15 feet) strip. With elections scheduled for 2011, the city faced the decision to redistrict in 2011 or postpone until the 2015 election round. A variety of factors, including desire to remedy the narrow-strip workaround in the 2007 redistricting, led them to choose to redistrict in 2011. In fact, Beveridge said, the change evident in 2010 Census numbers (relative to 2000) largely forced the city’s determination: the city’s six council districts now had a sizable 23 percent deviation in total population using the new 2010 Census results. Beveridge briefly displayed a set of criteria for redistricting that were widely publicized during the 2011 effort, noting that New Rochelle made intensive effort to seek out community input to identify neighborhoods that should be kept together.

Beveridge compared deviations from equality of the 2011-drawn districts under the original 2010 Census redistricting data and the 2010 DDP (see

Figure 4.1). The latter did not preserve population equality but rather, through particular differences in Districts 2 and 6, pushed the total population deviation over the 10 percent threshold Levitt cited in his presentation. It did not exceed the threshold by very much, Beveridge said, but any excess deviation is very consequential.

He also examined changes by race and ethnicity within the six districts in New Rochelle and found that, in percentage terms, populations were moved around a lot (see Figure 4.2). In District 6, the difference between the original and synthetic data was more than 100 percent for the Latino population and more than 80 percent for the African American population. In the majority Hispanic District 1, there was a large percentage change for African Americans. Moreover, the 2010 DDP data reduces the number of African Americans in District 3 while increasing the number of whites, a problem for the majority-minority district that would likely put it below the 50 percent threshold in CVAP data.

4.2.2 Block-Level Data

Beveridge next turned to examining blocks, which not only are used for building districts but also impact assessments of voting assistance needs for precincts. As a mayor, opponent of a mayor, or constituent generally, one would like to have all the blocks look “right.” Beveridge looked at the percentage change between the 2010 redistricting data and the 2010 DDP data, which was averaged over the 969 blocks in New Rochelle, for total population, occupied housing units, and population by race and ethnicity. It turned out that 14 percent of the total population moved around, the corresponding figures for non-Hispanic whites, African Americans, and Hispanics being 18 percent, 31 percent, and 30 percent, respectively.

Beveridge believed this outcome was a feature of the algorithm and the question was whether it generally pertained. He did the same analysis for blocks in Westchester County and in New York State and found the same kinds of changes going on between the 2010 redistricting data and the 2010 DDP data.

4.2.3 Concluding Comments

Beveridge said that there had been a lot of talk about the effects of data swapping on the original 2010 data, but he doubted that it affected 14 percent of the population as had occurred with the 2010 DDP. Moreover, population counts at the block level and higher geographic levels were kept invariant under the swapping used in 2010. He asked what remedy was available when it was clear that, at least for New Rochelle, the use of differential privacy undermined the determination of equal districts and diluted minority population shares. A related question is how the proposed administrative records-based CVAP would be affected by differential privacy.

The changes due to differential privacy in various population groups would make district drawing difficult. It would also be harder to get input from local stakeholders that reflected reality on the ground.

Beveridge noted how discussion had focused on ϵ because it is an ingredient of the differential privacy approach. In terms of accuracy, though, data utility is key for redistricting, and still appeared to be an open question for the redistricting use case.

4.3 REDISTRICTING AND DIFFERENTIAL PRIVACY

Michael McDonald (University of Florida) said his presentation would address some of the issues that the previous two speakers already covered: equal population and VRA compliance. His take was somewhat different, using the state of Georgia as a case study. McDonald was involved in challenges to congressional and state legislative plans submitted by Georgia to the U.S.

Department of Justice for VRA review. He worked with U.S. Representative John Lewis and his office and other members of the Congressional Black Caucus and with representatives from the state legislature and the congressional delegation from Georgia. He had also conducted other analyses of redistricting.

4.3.1 Equal Populations

McDonald analyzed the deviations of the precincts or voting tabulation districts (VTDs) in the 2010 DDP from the original 2010 P.L. 94-171 file. He noted that the VTDs were off-spine geography. He had current boundaries and was matching up election results with them right now in preparation for the 2020 round of redistricting.

As he was collecting VTD boundaries from localities across the country, it become apparent for smaller jurisdictions that precincts aligned very closely with other governmental units such as supervisory districts and city council districts. Precincts are not only important for redistricting purposes and election purposes, but also meaningfully describe local government boundaries.

His analysis for Georgia showed imbalances in population among districts due to the introduction of noise with the DDP data. While Levitt correctly stated that districts do not have to be exactly equal in population, there is a question of what it means when a district is close to the boundary of a 1 percent total population deviation that is generally deemed permissible for congressional districts.⁴ McDonald concluded that confidence intervals of some sort were needed.

The effort to produce multiple data sets as a way of generating confidence intervals could be helpful, but for legal purposes, the Census Bureau would need to publish measures of uncertainty for different levels of geography. In a courtroom, it would be important to be able to reference an official government document.

4.3.2 Voting Rights

It struck McDonald that while total counts would be constrained to equal the original census counts at the state level for reapportionment purposes, that would not necessarily be true for any other cell at the state level. For example, in Georgia, the application of differential privacy to the 2010 data created 876 African Americans statewide and subtracted 270 (looking at VTDs), a small number out of a state population of almost 10 million. However, there was

___________________

⁴ In positing this 1 percent deviation standard, McDonald cited Tennant v. Jefferson County Commission (567 U.S. 758, 2012), just as Levitt had. In that case, the Supreme Court concluded that new advances in redistricting software and data failed to make a 0.79 percent deviation in a congressional redistricting plan—upheld by the Karcher v. Daggett (462 U.S. 725, 1983) Court—any more problematic or consequential in 2012 than in 1983. Consequently, the “legitimate state objective” served by the deviation was judged to prevail.

a consistent bias for the black voting-age population. Not surprisingly, there were deviations below 0 for VTDs with higher VTD population size and fewer deviations below 0 for smaller VTDs. Indeed, from a simple linear regression line, African Americans were subtracted from precincts with large African American populations and added back into white precincts, confirming what Beveridge showed for New York State. This could be consequential, affecting the ability to provide effective representation for minority communities who might otherwise deserve it.

McDonald said he would prefer that noise be shifted to other data so that the data were perfect for his use case, but of course, every user wants that outcome. He wondered if it might be computationally feasible, say, if one were going to subtract an African American from a census block, to give a higher probability of adding one back into a neighboring census block. Such a method would at least constrain where the noise was being added spatially, which would be helpful for majority-minority districting.

With regard to racially polarized voting, ecological inference must be used because the secret ballot means that only aggregate (and not individual-level) election results are available. By merging the aggregate election results with race and ethnicity data from the census, a statistical analysis can be carried out to infer individual voting behavior from the aggregate data. McDonald presented a simple linear regression analysis for the 2010 governor’s election in Georgia using the published 2010 data and the 2010 DDP data. The two data sets were quite comparable in terms of the percentage of black voting-age population that participated in the gubernatorial vote. If one were to look at other groups besides non-Hispanic whites and blacks in Georgia, then the analysis would get trickier, as such groups as Hispanics, Asian Americans, Native Americans, and others comprise small populations in Georgia. So one would be adding noise to estimates that are already small, which would lead to instability in regression coefficients.

Looking at the 12th Congressional District of Georgia, which was challenged in 2011, the differences between the 2010 published and DDP data were an order of magnitude larger than for the state as a whole. The standard errors also increased, but, still, both the original and DDP data gave the correct conclusion regarding racially polarized voting.

McDonald acknowledged that he had been involved in other litigation involving analysis that was not so cut and dried. He expressed the need to find time to conduct analyses down to the state legislative level and local level, as Beveridge did. He thought it quite possible that there would be consequential differences in those instances.

4.3.3 Concluding Comments

McDonald asserted that it would be helpful to have estimates of uncertainty around point estimates from the census products. He acknowledged that people did not often use confidence intervals, and he was as guilty as anyone. However, when it came to adjudicating these issues in a courtroom, McDonald preferred to have them. He noted that, in a legal realm, one would want as much information as possible to defend one’s analyses and conclusions.

4.4 FLOOR DISCUSSION

Floor discussion began with a comment from the speaker’s dais. Levitt wanted to address a point made by McDonald, namely that even noisy precinct information, aggregated statewide or for a congressional district of 700,000–800,000 people, could produce a clear pattern. As McDonald showed, at least for Congressional District 12 in Georgia, the use of the DDP data did not materially change the statistical significance or overrule a finding of polarization. Levitt encouraged more experimentation with local elections in much smaller areas, Given all the other results presented thus far, the noise in a few precincts in a 3,000-, 5,000-, or 7,000-person jurisdiction might well preclude the finding of polarization where it in fact existed.

Alexis Santos-Lozada (Pennsylvania State University) asked what confidence intervals would represent for a census. Where would the variation come from, given that the census produces total counts? McDonald suggested that the simulation-based approach used by previous speakers, seemingly referring to the simulation of multiple datasets done to calculate the spatial statistics Moran’s I (Section 3.1.3), might be brought to bear. The distribution of the counts over those multiple simulated datasets could give some idea of the potential confidence intervals for a census data product. McDonald said that he did not know enough about the differential privacy methodology to assess if there could be a way to analytically derive confidence intervals. Yet a simulation or running of multiple datasets and an observation of differences could be a way to describe these intervals in some manner. This appeared feasible, but the question was whether there was time and resources to actually produce these kinds of simulations.

Levitt added that there could be differences of opinion on the matter—particularly whether having confidence intervals would in fact be advantageous in a court. Also, if running multiple versions of disclosure avoidance to generate a sort of confidence interval might eat up the privacy budget, then people might prefer to get more accurate individual figures. Beveridge suggested that it could be possible and useful to produce additional DDP-type runs with the 2010 Census data using a bunch of different values of ϵ, and that these might inform the use of simulated confidence intervals. However, he noted that the Bureau

would only be releasing one single run in 2020, in the 2020 Census data products themselves. If there were an analytical approach to constructing confidence intervals, vendors could perhaps produce such for use in redistricting. However, there would be sensitivity to releasing data that were “estimates,” litigation-wise.

James Tucker (Native American Rights Fund) thanked the workshop organizers for including voting rights in the discussion. His organization litigates redistricting issues for American Indian reservations, which often have fewer than 3,000 people. He cited the example of San Juan County, Utah, which shows the importance of having accurate racial data at the census block level, as Levitt stressed. This county, which has about 14,500 people, received a decision from the U.S. Court of Appeals for the Tenth Circuit⁵ that involved a five-member school district with 2,850 people in each part of the district and a three-member county commission with just under 5,000 people in each district. The Navajo population was just over the 50 percent threshold in the county so, if the census block data were not accurate, it would mean the difference between having two Navajo members on the three-member county commission and potentially having one or even just an “influence district” and not a majority-minority district. Tucker hoped the Census Bureau would take this issue to heart, as one of critical importance for this redistricting cycle.

Levitt followed by noting that, in some other applications outside of the electoral realm, gains and losses due to differential privacy more or less netted out. In the voting rights arena, when a population did not qualify for enforcement when it should have because of noise added by differential privacy, this would inhibit that population’s ability to elect candidates and would do harm to the community. On the other hand, artificially increasing a population that was actually too small for enforcement would not help that community because the actual votes would not be there. There would be no netting out: if the size of one population was inflated and another deflated, this would not cancel out, particularly in the electoral threshold use case.

Salvo asked what might happen in the redistricting arena, given the fact that there was going to be noise around what were considered to be absolute counts in the previous era, even with confidence intervals. Would a massive reeducation effort be needed? Were there going to be challenges in the courts? In response, Levitt cited the joke involving three men alone on a desert island with a single can of food—a physicist, a mechanical engineer, and an economist. The physicist said to put the can in the sun and wait for the air to expand and blow it open. The mechanical engineer said, nonsense, drop a boulder on the can and open it that way. The economist said, “Assume a can opener”—adding the punchline that lawyers also like to assume can openers. So, yes, Levitt replied, there would be litigation—guaranteed—but the courts would be likely to take whatever the Census Bureau delivered. He said it would likely that

___________________

⁵ Reference is to Navajo Nation v. San Juan County, No. 18-4005 (10th Circuit, 2019).

the courts would ignore things like confidence intervals as extra noise, meaning legal noise, not just statistical noise. Courts do not like to hear about noise.

McDonald expressed the opinion that the use of census data with noise introduced by differential privacy for redistricting would ultimately wind up before the U.S. Supreme Court. A litigant would go to the Supreme Court to ask the Census Bureau to release the original, unaltered counts. Beveridge concurred and expressed the view that it was likely that a judge would order the Census Edited File (CEF) (prior to the use of differential privacy) to be delivered under seal, since judges have absolute power in their courtrooms. Beveridge’s favorite example was when a judge was contemplating ordering an election to be conducted in 2003, not in 2007. The board of elections representative said she did not know if her organization could meet this deadline. The judge said that, if the board were ordered to be ready for elections in 2003 and did not carry out the order, then people would be eating federal food in prison.

Sierra Watt (National Congress of American Indians) asked if the presenters could speak to examples of elections where the minority population was smaller. The example that came to mind was North Dakota where a number of margins were quite close in majority-minority precincts. Levitt and Beveridge said such an analysis could be done if someone were to develop block equivalency files for small districts on Indian reservations.

Simson Garfinkel (U.S. Census Bureau) stated that Title 13, Section 9, exempted census-collected data from legal process. So, he said, it was a “fantasy” to think that a judge would be able to compel the Census Bureau to provide tabulations from the CEF. Levitt replied that the lawyers of today know never to call any legal outcome a “fantasy.” He agreed that Garfinkel had correctly stated that the statute appears to flat out block access to the raw census data, but Levitt also agreed with Beveridge’s points that judges and courts interpret statutes differently, as they see fit. McDonald noted that the Constitution trumped any statute, with Beveridge adding that the Constitution’s mandate for an “enumeration” would be the central question.