SUPREME COURT OF THE UNITED STATES
DEPARTMENT OF COMMERCE, et al., PETITIONERS v.
NEW YORK, et al.
on writ of certiorari before judgment to the united states court of appeals for the second circuit
[June 27, 2019]
Justice Breyer, with whom Justice Ginsburg, Justice Sotomayor, and Justice Kagan join, concurring in part and dissenting in part.
I join Parts I, II, IV–A, and V of the Court’s opinion (except as otherwise indicated in this opinion). I dissent, however, from the conclusion the Court reaches in Part IV–B. To be more specific, I agree with the Court that the Secretary of Commerce provided a pretextual reason for placing a question about citizenship on the short-form census questionnaire and that a remand to the agency is appropriate on that ground. But I write separately because I also believe that the Secretary’s decision to add the citizenship question was arbitrary and capricious and therefore violated the Administrative Procedure Act (APA).
There is no serious dispute that adding a citizenship question would diminish the accuracy of the enumeration of the population—the sole constitutional function of the census and a task of great practical importance. The record demonstrates that the question would likely cause a disproportionate number of noncitizens and Hispanics to go uncounted in the upcoming census. That, in turn, would create a risk that some States would wrongfully lose a congressional representative and funding for a host of federal programs. And, the Secretary was told, the adverse consequences would fall most heavily on minority communities. The Secretary decided to ask the question anyway, citing a need for more accurate citizenship data. But the evidence indicated that asking the question would produce citizenship data that is less
accurate, not more. And the reason the Secretary gave for needing better citizenship data in the first place—to help enforce the Voting Rights Act of 1965—was not convincing.
In short, the Secretary’s decision to add a citizenship question created a severe risk of harmful consequences, yet he did not adequately consider whether the question was necessary or whether it was an appropriate means of achieving his stated goal. The Secretary thus failed to “articulate a satisfactory explanation” for his decision, “failed to consider . . . important aspect[s] of the problem,” and “offered an explanation for [his] decision that runs counter to the evidence,” all in violation of the APA. Motor Vehicle Mfrs. Assn. of United States
v. State Farm Mut. Automobile Ins. Co.
463 U.S. 29
, 43 (1983). These failures, in my view, risked undermining public confidence in the integrity of our democratic system itself. I would therefore hold that the Secretary’s decision—whether pretextual or not—was arbitrary, capricious, and an abuse of discretion.
Three sets of laws determine the legal outcome of this case. First, the Constitution requires an “actual Enumeration” of the “whole number of persons in each State” every 10 years. Art. I, §2, cl. 3; Amdt. 14, §2. It does so in order to “provide a basis for apportioning representatives among the states in the Congress.” Baldrige
455 U.S. 345
, 353 (1982); see also Art. I, §2, cl. 3. The inclusion of this provision in the Constitution itself underscores the importance of conducting an accurate census. See Utah
536 U.S. 452
, 478 (2002) (recognizing “a strong constitutional interest in [the] accuracy” of the enumeration).
Second, the Census Act contains two directives that constrain the Secretary’s ability to add questions to the census. Section 195 says that the Secretary “shall, if he considers it feasible,” authorize the use of statistical “sampling” in collecting demographic information. That means the Secretary must, if feasible, obtain demographic information through a survey sent to a sample
of households, rather than through the short-form census questionnaire to which every
household must respond. The other relevant provision, §6(c), says that “[t
]o the maximum extent possible
and consistent with the kind, timeliness, quality and scope of the statistics required, the Secretary shall acquire and use information available” from administrative sources “instead of conducting direct inquiries.” (Emphasis added.) These provisions, taken together, reflect a congressional preference for keeping the short form short, so that it does not burden recipients and thereby discourage them from responding.
Third, the APA prohibits administrative agencies from making choices that are “arbitrary, capricious, an abuse of discretion, or otherwise not in accordance with law.”
5 U. S. C. §706(2)(A). We have said that courts, in applying this provision, must decide “whether the decision was based on a consideration of the relevant factors and whether there has been a clear error of judgment.” Citizens to Preserve Overton Park
401 U.S. 402
, 416 (1971). The agency must have “examine[d] the relevant data and articulate[d] a satisfactory explanation for its action[,] including a ‘rational connection between the facts found and the choice made.’ ” State Farm
, 463 U. S., at 43. An agency ordinarily fails to meet this standard if it has “failed to consider an important aspect of the problem, offered an explanation for its decision that runs counter to the evidence before the agency, or is so implausible that it could not be ascribed to a difference in view or the product of agency expertise.” Ibid.
Courts do not apply these principles of administrative law mechanically. Rather, they take into account, for example, the nature and importance of the particular decision, the relevance and importance of missing information, and the inadequacies of a particular explanation in light of their importance. The Federal Government makes tens of thousands, perhaps millions, of administrative decisions each year. And courts would be wrong to expect or insist upon administrative perfection. But here, the Enumeration Clause, the Census Act, and the nature of the risks created by the agency’s decision all make clear that the decision before us is highly important to the proper functioning of our democratic system. It is therefore particularly important that courts here not overlook an agency’s (1) failure to consider serious risks of harm, (2) failure to explain its refusal to minimize those risks, or (3) failure to link its conclusion to available evidence. My view, like that of the District Court, is that the agency here failed on all three counts.
A brief history of how the census has worked over the years will help the reader understand some of the shortcomings of the Secretary’s decisionmaking process. The Framers wrote into the Constitution a mandate to conduct an “actual Enumeration” of the population every 10 years. Art. I, §2, cl. 3. They did so for good reason. The purpose of the census is to “provide a basis for apportioning representatives among the states in the Congress,” Baldrige
, 455 U. S., at 353, ensuring that “comparative state political power in the House . . . reflect[s] comparative population,” Evans
, 536 U. S., at 477. The Framers required an actual count of every resident to “limit political chicanery” and to prevent the census count from being “skewed for political . . . purposes.” Id.
, at 500 (Thomas, J., concurring in part and dissenting in part).
Throughout most of the Nation’s history, the Federal Government used enumerators, often trained census takers, to conduct the census by going door to door. The enumerators would ask a host of questions, including place of birth, citizenship, and others. But after the 1950 census, the Bureau began to change its approach. Post-census studies revealed that the census had failed to count more than 5 million people and that the undercount disproportionately affected members of minority groups. See M. Anderson, The American Census: A Social History 201−202 (1988); Brief for Historians and Social Scientists as Amici Curiae
15. Studies showed that statistical sampling would produce higher quality data. Anderson, American Census,
Beginning with the 1960 census, the Bureau consequently divided its questioning into a short form and a long form. The short form contained a list of questions—a short list—that the census would ask of every household. That list included basic demographic questions like sex, age, race, and marital status. The short form did not include, and has never included, a question about citizenship. See ibid.
; Dept. of Commerce, U. S. Census Bureau, Measuring America: The Decennial Censuses From 1790 to 2000, p. 128 (2002). By way of contrast, the long form set forth a host of questions that would be asked of only a sample of households. In 1960, the long form was sent to one in every four households; in subsequent years, it was sent to approximately one in every six. See 351 F. Supp. 3d 502, 520 (SDNY 2019). And it was more recently replaced by the American Community Survey (ACS), which is sent to approximately 1 in 38 households each year. The long form (and now the ACS) has often included a question about citizenship.
In 1970, the Census Bureau made another important change to the census. It significantly reduced its reliance upon in-person enumerators. See Anderson, supra
, at 206.
Instead, it sent nearly all households a questionnaire by mail. Most households received the short form, and a small sample received the long form. Instructions on the form told each household to fill out the questionnaire and return it to the Census Bureau by mail. Enumerators would follow up with households that did not return the questionnaire.
To maximize accuracy and minimize cost, the Bureau tried to bring about the highest possible “self-response” rate, i.e.
, to encourage as many households as possible to respond by mail. For that reason, it tried to keep the short form as short as possible. And it consistently opposed placing a citizenship question on that form. It feared that adding a question about citizenship would “inevitably jeopardize the overall accuracy of the population count,” partly because of added response burden but also because, as it explained, noncitizens faced with a citizenship question would be less likely to respond due to fears of “the information being used against them.” Federation for Am. Immigration Reform
, 486 F. Supp. 564
, 568 (DC 1980).
Likely for similar reasons, Congress amended the Census Act in 1976, enacting the two statutory provisions to which I previously referred. These two provisions,
13 U. S. C. §6(c) and §195, together encourage the Secretary not to ask demographic questions on the short form if the information can be obtained either through the long form or through administrative records.
With this statutory and historical background, we can more easily consider the agency decision directly under review. That decision “reinstate[s] [a] citizenship question on the 2020 decennial census.” App. to Pet. for Cert. 549a−550a (Memorandum from Wilbur L. Ross, Jr., Secretary of Commerce, to Karen Dunn Kelley, Under Secretary for Economic Affairs (Mar. 26, 2018)). The agency’s decision memorandum provided one and only one reason for making that decision—namely, that the question was “necessary to provide complete and accurate data in response to” a request from the Department of Justice (DOJ). Id.
, at 562a. The DOJ had requested the citizenship question for “use [in] . . . determining violations of Section 2 of the Voting Rights Act.” Id.
, at 548a.
The decision memorandum adds that the agency had not been able to “determine definitively how inclusion of a citizenship question on the decennial census will impact responsiveness. However, even if there is some impact on responses, the value of more complete and accurate data derived from surveying the entire population outweighs such concerns.” Id.
, at 562a. The Secretary’s decision thus rests upon a weighing of potentially adverse consequences (diminished responses and a less accurate census count) against potentially offsetting advantages (better citizenship data). In my view, however, the Secretary did not make reasonable decisions about these potential costs and benefits in light of the administrative record.
Consider first the Secretary’s conclusion that he was “not able to determine definitively how inclusion of a citizenship question on the decennial census will impact responsiveness.” Ibid.
Insofar as this statement implies that adding the citizenship question is unlikely to affect “responsiveness” very much (or perhaps at all), the evidence in the record indicates the contrary.
The administrative record includes repeated Census Bureau statements that adding the question would produce a less accurate count because noncitizens and Hispanics would be less likely to respond to the questionnaire. See App. 105, 109–112, 158. The Census Bureau’s chief scientist said specifically that adding the question would have “an adverse impact on self-response and, as a result, on the accuracy and quality of the 2020 Census.” Id.
, at 109. And the chief scientist backed this statement up by pointing to “[t]hree distinct analyses.” Ibid.
The first analysis compared nonresponse rates for the short-form census questionnaire (which did not include a citizenship question) to nonresponse rates for the ACS (which did). Obviously, more people fail to respond to the ACS than to the short form. Yet taking into account the fact that the nonresponse rate will be greater for the ACS than for the short form, the Bureau found that the difference between the two is yet greater for noncitizen households than for citizen households (by 5.1%, according to the Bureau). Id.
, at 111. This led the Bureau to say that it was a “reasonable inference” that the presence of the citizenship question accounted for the difference. Ibid.
The Bureau conducted two additional studies, both analyzing data from the ACS. One study looked at response rates for particular questions on the ACS. It showed that the “no answer” rate for the citizenship question was “much greater than the comparable rates” for other census questions (for example, questions about age, sex, race, and ethnicity). Id.
, at 110. And it showed that the “no answer” rate for the citizenship question was significantly higher among Hispanics. Id.
, at 109−110. The last study examined “break-off” rates, i.e.
, the rate at which respondents stopped answering the questionnaire upon reaching a particular question. It found that Hispanics were significantly more likely than were non-Hispanics to stop answering at the point they reached the citizenship question. Id.
, at 112. Together, these two studies provided additional support for the Census Bureau’s determination that the citizenship question is likely to mean disproportionately fewer responses from noncitizens and Hispanics than from others. Ibid.
Putting numbers upon these study results, the Census Bureau estimated that adding the question to the short form would lead to 630,000 additional nonresponding households. Id.
, at 114. That is to say, the question would cause households covering more than 1 million additional people to decline to respond to the census. When the Bureau does not receive a response, it follows up with in-person interviews in an effort to obtain the missing information. The Bureau often interviews what it calls “proxies,” such as family members and neighbors. But this followup process is subject to error; and the error rate is much greater than the error rate for self-responses. Ibid.
The Bureau thus explained that lower self-response rates “degrade data quality” by increasing the risk of error and leading to hundreds of thousands of fewer correct enumerations. Id.
, at 113−115. The Bureau added that its estimate was “conservative.” Id.
, at 115. It expected “differences between citizen and noncitizen response rates and data quality” to be “amplified” in the 2020 census “compared to historical levels.” Ibid.
Thus, it explained, “the decrease in self-response for citizen households in 2020 could be much greater than the 5.1 percentage points [it] observed during the 2010 Census.” Id.
, at 115−116. Its conclusion in light of this evidence was clear. Adding the citizenship question to the short form was “very likely to reduce the self-response rate” and thereby “har[m] the quality of the census count.” Id.
, at 105, 158.
The Census Bureau’s analysis received support from other submissions. Several States pointed out that noncitizens and racial minorities had been undercounted in every prior census. Administrative Record 1091−1092. They also drew attention to recent surveys indicating that noncitizens had significant concerns about the confidentiality of census responses. Ibid.
Former directors of the Census Bureau wrote that adding the citizenship question so late in the process “would put the accuracy of the enumeration and success of the census in all communities at grave risk.” Id.
1057. The American Sociological Association and Census Scientific Advisory Committee echoed these warnings. See id.
, at 787, 794−795. On the other hand, the Secretary received submissions by other groups that supported adding the question. See, e.g.
, at 1178−1179, 1206, 1276. But as far as I can tell (or as far as the arguments made here and in the District Court inform the matter), none of these latter submissions significantly added to, or detracted from, the Census Bureau’s submissions in respect to the question’s likely impact on response rates.
The Secretary’s decision memorandum reached a quite different conclusion from the Census Bureau. The memorandum conceded that “a lower response rate would lead to . . . less accurate responses.” App. to Pet. for Cert. 556a. But it concluded that neither the Census Bureau nor any stakeholders had provided “definitive, empirical support” for the proposition that the citizenship question would reduce response rates. Id.
, at 554a. The memorandum relied for that conclusion upon a number of considerations, but each is contradicted by the record.
The memorandum first pointed to perceived shortcomings in the Census Bureau’s analysis of nonresponse rates. It noted that response rates are generally lower overall for the long form and ACS than they are for the short form. Id.
, at 552a−554a. But the Bureau explained that its analysis accounted for this consideration, see App. 111, and no one has given us reason to think the contrary.
The Secretary also noted that the Bureau “was not able to isolate what percentage of [the] decline was caused by the inclusion of a citizenship question rather than some other aspect of the long form survey.” App. to Pet. for Cert. 554a. But the Bureau said attributing the decline to the citizenship question was a “reasonable inference,” App. 111, and again, nothing in the record contradicted the Bureau’s judgment. And later analyses have borne out the Bureau’s judgment that the citizenship question contributes to the decline in self-response. See, e.g.
, at 1002−1006, 1008 (August 2018 Census Bureau study).
The memorandum next cast doubt on the Census Bureau’s analysis of the rate at which people responded to particular questions on the ACS. It noted that the “no answer” rate to the citizenship question was comparable to the “no answer” rate for other questions on the ACS, including educational attainment, income, and property insurance. App. to Pet. for Cert. 553a. But as discussed above, the Bureau found it significant that the “no answer” rate for the citizenship question was “much greater” than the “no answer” rate for the other questions that appear on the short form
—that is, the form on which the citizenship question would appear. App. 110, 124. The Secretary offered no reason why the demographic variables to which he pointed provided a better point of comparison.
Finally, the memorandum relied on information provided by two outside stakeholders. The first was a study conducted by the private survey company Nielsen, in which questions about place of birth and time of arrival had not led to any appreciable decrease in the response rate. App. to Pet. for Cert. 552a. But Nielsen, which in fact urged the Secretary not
to add the question, stated that its respondents (unlike census respondents) were paid
to respond, and it is consequently not surprising that they did so. Administrative Record 1276. The memo- randum also cited statements by former Census Bureau officials suggesting that empirical evidence about the question’s potential impact on response rates was “limited.” App. to Pet. for Cert. 558a−559a; see also id.
, at 552a. But there was no reason to expect the former officials to provide more extensive empirical evidence as to a citizenship question when they were not privy to the internal Bureau analyses on this question. And, like Nielsen, the former officials strongly urged the Secretary not
to ask the question. See Administrative Record 1057.
The upshot is that the Secretary received evidence of a likely drop in census accuracy by a number somewhere in the hundreds of thousands, and he received nothing significant to the contrary. The Secretary pointed out that the Census Bureau’s information was uncertain, i.e.
not “definitive.” But that is not a satisfactory answer. Few public-policy-related statistical studies of risks (say, of many health or safety matters) are definitive. As the Court explained in State Farm
, “[i]t is not infrequent that the available data do not settle a regulatory issue, and the agency must then exercise its judgment in moving from the facts and probabilities on the record to a policy conclusion.” 463 U. S., at 52. But an agency confronted with this situation cannot “merely recite the terms ‘substantial uncertainty’ as a justification for its actions.” Ibid.
Instead, it “must explain the evidence which is available” and typically must offer a reasoned explanation for taking action without “engaging in a search for further evidence.” Ibid.
The Secretary did not do so here. He did not explain why he made the decision to add the question without following the Bureau’s ordinary practice of extensively testing proposed changes to the census questionnaire. See App. 624−630, 641 (discussing testing process); see also, e.g.
, Brief for Former Census Bureau Directors as Amici Curiae
17−21 (discussing prior examples of questions that the Bureau decided not to add after many years of pretesting). Without that testing, the Secretary could not treat the Bureau’s expert opinions and its experience with the relevant surveys as worthless merely because its conclusions were not precise. The Bureau’s opinions were properly considered as evidence of likelihoods, probabilities, or risks.
As noted above, the consequences of mistakes in the census count, of even a few hundred thousand, are grave. Differences of a few thousand people, as between one State and another, can mean a loss or gain of a congressional seat—a matter of great consequence to a State. See 351 F. Supp. 3d, at 594. And similar small differences can make a large difference to the allocation of federal funds among competing state programs. Id.
, at 596−597; see also Baldrige
, 455 U. S., at 353−354, n. 9. If near-absolute certainty is what the Secretary meant by “definitive,” that insistence would itself be arbitrary in light of the constitutional and statutory consequences at stake. And if the Secretary instead meant that the evidence does not indicate a serious risk of a less accurate count, that conclusion does not find support in the record.
Now consider the Secretary’s conclusion that, even if adding a citizenship question diminishes the accuracy of the enumeration, “the value of more complete and accurate data derived from surveying the entire population outweighs
. . . concerns” about diminished accuracy. App. to Pet. for Cert. 562a (emphasis added). That conclusion was also arbitrary. The administrative record indicates that adding a citizenship question to the short form would produce less “complete and accurate data,” not more.
The Census Bureau informed the Secretary that, for about 90% of the population, accurate citizenship data is available from administrative records maintained by the Social Security Administration and Internal Revenue Service. App. 146. The Bureau further informed the Secretary that it had “high confidence” that it could develop a statistical model that would accurately impute citizenship status for the remaining 10% of the population. Ibid.
The Bureau stated that these methods alone—using existing administrative records for 90% of the population and statistical modeling for the remaining 10%—would yield more accurate citizenship data than also asking a citizenship question. Id.
, at 159. How could that be so? The answer is somewhat technical but readily understandable.
, consider the 90% of the population (about 295 million people) as to whom administrative records are available. The Government agrees that using these administrative records would provide highly reliable information about citizenship, because the records “require proof of citizenship.” Id.
, at 117. By contrast, if responses to a citizenship question were used for this group, the Census Bureau predicted without contradiction that about one-third of the noncitizens in this group who respond would answer the question untruthfully, claiming to be citizens when they are not. Id.
, at 147. Those incorrect answers—about 9.5 million in total—would conflict with the administrative records on file for those noncitizens. And what would the Census Bureau do with the conflicting data? If it accepts the answer to the citizenship question as determinative, it will have less accurate data. If it accepts the citizenship data from administrative records as determinative, asking the question will have served no purpose.
Thus, as to 295 million people—the overwhelming majority of the population—asking the citizenship question would at best add nothing at all. I say “at best” because, for one thing, the Census Bureau informed the Secretary that asking the question would produce 1 million more people who could not be linked to administrative records, which in turn would require the Census Bureau to resort to a less accurate source of citizenship data for these people. See id.
, at 147−149; see also 351 F. Supp. 3d, at 538−539. For another, the policy of the Census Bureau has always been to use census responses rather than administrative records in cases where the two conflict. App. 147. In this case, that practice would mean accepting 9.5 million inaccurate responses even though accurate administrative records are available. See ibid.
The Census Bureau could perhaps change that practice, but the Secretary’s decision memorandum said nothing about the matter. It did not address the problem.
, consider the remaining 10% of the population (about 35 million people) for whom the Government lacks administrative records. The question here is which approach would yield the most “complete and accurate” citizenship data for this group—adding a citizenship question or using statistical modeling alone? To answer this question, we must further divide this group into two categories—those who would respond to the citizenship question if it were asked and those who would not.
Start with the category of about 22 million people who would answer a citizenship question if it were asked. Would their answers regarding citizenship be more accurate than citizenship data produced by statistical modeling? The Census Bureau said no. That is because many of the noncitizens in this group would answer the question falsely, resulting in an estimated 500,000 inaccurate answers. See id.
, at 148. And those who answer the question falsely would be commingled, perhaps randomly, with those who answer it correctly, thereby casting doubt on the answers of all 22 million, with no way of knowing which answers are correct and which are false. By contrast, the Bureau believed that it could develop a statistical model that would produce more accurate citizenship data than these census responses. The Bureau therefore informed the Secretary that it could do better. As the Bureau’s chief scientist explained, although “[o]ne might think” that asking the question “could help fill the . . . gaps” in the administrative records, the data did not support that assumption. Id.
, at 157. Instead, he explained, responses to the citizenship question “may not be reliable,” which “calls into question their ability to improve upon” the Bureau’s statistical modeling process. Ibid.
Next, turn to the more than 13 million remaining people who would not answer the citizenship question even if it were asked. As to this category, the Census Bureau would still need to use statistical modeling
to obtain citizenship data, because there would be no census response to use instead. Hence, asking the citizenship question would add nothing at all as to this group. To the contrary, as the Government concedes, asking the question would reduce
the accuracy of the citizenship data for this group, because the relatively inaccurate answers to the citizenship question would diminish the overall accuracy of the Census Bureau’s statistical model. See Brief for Petitioners 34 (conceding that the Census Bureau model will be “highe[r] quality” without the question than with it); 351 F. Supp. 3d, at 640 (explaining that asking the question would “corrup[t] . . . the data generated by extrapolating from self-responses through imputation”).
In sum, in respect to the 295 million persons for whom administrative records exist, asking the question on the short form would, at best, be no improvement over using administrative records alone. And in respect to the remaining 35 million people for whom no administrative records exist, asking the question would be no better, and in some respects would be worse, than using statistical modeling. The Census Bureau therefore told the Secretary that asking the citizenship question, even in addition to using administrative records, “would result in poorer quality citizenship data” than using administrative records alone, and would “still have all the negative cost and quality implications” of asking the citizenship question. App. 159. I could find no evidence contradicting that prediction.
If my description of the record is correct, it raises a serious legal problem. How can an agency support the decision to add a question to the short form, thereby risking a significant undercount of the population, on the ground that it will improve
the accuracy of citizenship data, when in fact the evidence indicates that adding the question will harm
the accuracy of citizenship data? Of course it cannot. But, as I have just said, I have not been able to find evidence to suggest that adding the question would result in more accurate citizenship data. Neither could the District Court. After reviewing the record in detail, the District Court found that “all of the relevant evidence before Secretary Ross—all
of it—demonstrated that using administrative records . . . would actually produce more accurate [citizenship] data than adding a citizenship question to the census.” 351 F. Supp. 3d, at 650.
What consideration did the Secretary give to this problem? He stated simply that “[a]sking the citizenship question of 100 percent of the population gives each respondent the opportunity to provide an answer,” which “may eliminate the need for the Census Bureau to have to impute an answer for millions of people.” App. to Pet. for Cert. 556a. He therefore must have assumed, sub silentio
, exactly what the Census Bureau experts urged him not to assume—that answers to the citizenship question would be more accurate than statistical modeling. And he ignored the undisputed respects in which asking the question would make the existing data less accurate. Other than his assumption, the Secretary said nothing, absolutely nothing, to suggest a reasoned basis for disagreeing with the Bureau’s expert statistical judgment.
The Government now maintains that the Secretary reasonably discounted the Census Bureau’s recommendation because it was based on an untested prediction about the accuracy of its model. But this is not a case in which the Secretary was presented with a policy choice between two reasonable but uncertain options. For one thing, the record is much less uncertain than the Government acknowledges. Although it is true that the Census Bureau at one point told the Secretary that it could not “quantify the relative magnitude of the errors across the alternatives at this time,” App. 148, it unequivocally stated that asking the question “would result in poorer quality citizenship data
” than omitting it, id.
, at 159 (emphasis added). Thus, even if the Bureau could not “quantify” the relative accuracy of the options, it could and did conclude that one option was likely more accurate than the other. Even in the face of some uncertainty, where all available evidence indicates that one option is better than the other, it is unreasonable to choose the worse option without explanation.
For another thing, to the extent the record reflects some uncertainty regarding the accuracy of the Census Bureau’s statistical model, that is because the model needed to be “developed and tested” before it could be employed. Id.
, at 146. But the Secretary made his decision before any such development or testing could be completed. Having decided to make an immediate decision rather than wait for testing, the Secretary could not dismiss the Bureau’s prediction about the inadvisability of that decision on the ground that the prediction reflected likelihoods, probabilities, and risks rather than certainties.
Finally, recall that the Census Act requires
the Secretary to use administrative records rather than direct inquiries to “the maximum extent possible.”
13 U. S. C. §6(c). That statutory requirement highlights what should be obvious: Whether adding a citizenship question to the short form would produce more accurate citizenship data is a relevant factor—indeed, a critically important factor—that the Secretary was required to consider. Here, the Secretary did not adequately explain why he rejected the evidence that adding the question would yield less accurate data. He did not even acknowledge that the Census Act obliged him to use administrative records rather than asking a question to the extent possible. And he did not explain how obtaining citizenship data that is no better or worse than the data otherwise available could justify jeopardizing the accuracy of the census count.
In these respects, the Secretary failed to consider “important aspect[s] of the problem” and “offered an explanation for [his] decision that runs counter to the evidence before the agency.” State Farm
, 463 U. S., at 43.
The Secretary’s failure to consider this evidence—that adding the question would harm the census count in the interest of obtaining less accurate
citizenship data—provides a sufficient basis for setting the decision aside. But there is more. The reason that the Secretary provided for needing more accurate citizenship information in the first place—to help the DOJ enforce the Voting Rights Act—is unconvincing.
The Secretary stated that adding the citizenship question was “necessary to provide complete and accurate data in response to the DOJ request.” App. to Pet. for Cert. 562a. The DOJ’s request in turn asserted that the citizenship data currently available from the ACS was not “ideal” for enforcing the Voting Rights Act. Id.
, at 567a. One of the DOJ’s principal complaints was that ACS data is reported for groups
of census blocks rather than for each census block itself. The DOJ letter stated that adding a citizenship question could provide it with individual block-by-block data which, the DOJ maintained, would allow it to better enforce the Voting Rights Act’s protections for minority voters. Id.
, at 568a.
This rationale is difficult to accept. One obvious problem is that the DOJ provided no basis to believe that more precise data would in fact help with Voting Rights Act enforcement. Congress enacted the Voting Rights Act in 1965—15 years after the census last asked every household about citizenship. Actions to enforce the Act have therefore always
used citizenship data derived from sampling. Yet I am aware of no one—not in the Department of Commerce proceeding, in the District Court, or in this Court—who has provided a single example in which enforcement of the Act has suffered due to lack of more precise citizenship data. Organizations with expertise in this area tell us that asking the citizenship question will not help enforce the Act. See, e.g.
Brief for NAACP Legal Defense & Educational Fund, Inc., as Amicus Curiae
30−36. Rather, the question will, by depressing the count of minority groups, hurt those whom the Act seeks to help. See, e.g.
Brief for Leadership Conference on Civil and Human Rights et al. as Amici Curiae
Another problem with the Secretary’s rationale is that, even assuming the DOJ needed more detailed citizenship data, there were better ways of obtaining the needed data. The Census Bureau offered to provide the DOJ with data using administrative records, which, as I have pointed out, are likely just as accurate, if not more accurate, than responses to a citizenship question. The Census Bureau offered to provide this data at the census block level, which would resolve each of the DOJ’s complaints about the existing ACS data. See Administrative Record 3289. But the Secretary rejected this alternative without explaining why it would not fully respond to the DOJ’s request. That failure was particularly problematic given that the Census Act requires the Secretary to use other methods of obtaining demographic information if at all possible. See §§6(c), 195.
Normally, the Secretary would be entitled to place considerable weight upon the DOJ’s expertise in matters involving the Voting Rights Act, but there are strong reasons for discounting that expertise here. The administrative record shows that DOJ’s request to add a citizenship question originated not with the DOJ, but with the Secretary himself. See Administrative Record 3710. The Voting Rights Act rationale was in fact first proposed by Commerce Department officials. See ibid.
DOJ officials, for their part, were initially uninterested in obtaining more detailed citizenship data, App. 414, and they agreed to request the data only after the Secretary personally spoke to the Attorney General about the matter, see Administrative Record 2651. And when the acting director of the Census Bureau proposed alternative means of obtaining better citizenship data, DOJ officials declined to meet to discuss the proposal. See id.
, at 3460.
Taken as a whole, the evidence in the administrative record indicates that the Voting Rights Act rationale offered by the Secretary was not just unconvincing, but pretextual. And, as the Court concludes, further evidence outside the administrative record but present in the trial record supports the finding of pretext. See Part V, ante.
Among other things, that evidence reveals that the DOJ official who wrote the letter agreed that adding the question “is not necessary for DOJ’s VRA enforcement efforts.” App. 1113. And that official further acknowledged that he did not “know whether or not [citizenship] data produced from responses to the citizenship question . . . will, in fact, be more precise than the [citizenship] data on which the DOJ is currently relying for purposes of VRA enforcement.” Id.
, at 1102.
The Court explains, and I agree, that a court normally should not “reject an agency’s stated reasons for acting simply because the agency might also have had other unstated reasons.” Ante
, at 24.
But in this case, “the evidence tells a story that does not match the explanation the Secretary gave for his decision.” Ante
at 27. This evidence strongly suggests that the Secretary’s stated rationale was pretextual. I consequently join Part V of the Court’s opinion (except insofar as it concludes that the Secretary’s decision was reasonable apart from the question of pretext). And I agree that the pretextual nature of the Secretary’s decision provides a sufficient basis to affirm the District Court’s decision to send the matter back to the agency.
* * *
I agree with the Court that the APA gives agencies broad leeway to carry out their legislatively delegated duties. And I recognize that Congress has specifically delegated to the Secretary of Commerce the authority to conduct a census of the population “in such form and content as he may determine.” §141(a). But although this delegation is broad, it is not without limits. The APA supplies one such limit. In an effort to ensure rational decisionmaking, the APA prohibits an agency from making decisions that are “arbitrary, capricious, [or] an abuse of discretion.”
5 U. S. C. §706(2)(A).
This provision, of course, does not insist that decisionmakers think through every minor aspect of every problem that they face. But here, the Secretary’s decision was a major one, potentially affecting the proper workings of our democratic government and the proper allocation of hundreds of billions of dollars in federal funds. Cf. ante
at 10. Yet the decision was ill considered in a number of critically important respects. The Secretary did not give adequate consideration to issues that should have been central to his judgment, such as the high likelihood of an undercount, the low likelihood that a question would yield more accurate citizenship data, and the apparent lack of any need for more accurate citizenship data to begin with. The Secretary’s failures in considering those critical issues make his decision unreasonable. They are the kinds of failures for which, in my view, the APA’s arbitrary and capricious provision was written.
As I have said, I agree with the Court’s conclusion as to pretext and with the decision to send the matter back to the agency. I do not agree, however, with several of the Court’s conclusions concerning application of the arbitrary and capricious standard. In my view, the Secretary’s decision—whether pretextual or not—was arbitrary, capricious, and an abuse of his lawfully delegated discretion. I consequently concur in the Court’s judgment to the extent that it affirms the judgment of the District Court.