dissenting.
Section 195 of the Census Act requires that, for purposes other than apportionment, the Secretary of Commerce “shall” authorize the use of statistical sampling methods to compensate for errors in the census “if he considers it feasible.” 13 U.S.C. § 195. The majority approves the Secretary’s decision not to authorize statistical sampling, although the record shows that the Secretary did not consider whether such sampling was feasible, or at least did not base his decision on that consideration, as the statute requires. Furthermore, even if the Secretary had considered whether sampling was feasible, he could not have reached the conclusion that he did under a reasonable definition of the word “feasible.”
No' census has fever succeeded in counting every person residing in the United States. The 1990 census, according to the Census Bureau’s estimate, resulted in an undercount of 1.8%, or 4.7 million people. BUREAU OF THE CENSUS, U.S. DEP’T OF COMMERCE, REPORT TO CONGRESS — THE PLAN FOR CENSUS 2000, at 2 (revised Aug. 1997)[herein-after PLAN FOR CENSUS 2000], Studies by the Bureau since 1940 have also shown that, decade after - decade, some groups of people are much more severely undercounted than others. Groups subject to this “differential undercount” include racial and ethnic minorities, children under *878the age of 18, and people who rent rather than own their homes.1
The differential undereount has a profound and unjust effect on polities and communities in which members of under-counted groups disproportionately reside. Billions of dollars of state .and federal funds marked for distribution on the basis of population are, in fact, distributed on the basis of these inaccurate census data. See, e.g., P-RICE WATERHOUSE COOPERS, EFFECT OF CENSUS 2000 UN-DERCOUNT ON FEDERAL FUNDING TO STATES AND LOCAL AREAS, 2002-2012, at i (2000) (study prepared for U.S. Census Monitoring Board). States also use the census figures as their measure of population when redrawing election districts. PLAN FOR CENSUS 2000, at 4. The result is that residents of some places — particularly places with large numbers of racial and ethnic minorities, children, and non-homeowners — receive less than their fair share of public funds and services and wield less than their proportionate share of electoral power.
Statistical sampling is a means of correcting the undercount and the inequities it causes. It was for this purpose that Congress in 1976 amended § 195 of the Census Act to its current form, requiring the Secretary to use statistical sampling “if he considers it feasible.” Pub.L. No. 94-521, § 10, 90 Stat. 2459, 2464 (1976) (codified as amended at 13 U.S.C. § 195). In 1991, in response to the inaccuracy of the 1990 census, Congress went still further, directing the Commerce Department to contract with the National Academy of Sciences (NAS) to “study ... the means by which the Government could achieve the most accurate population count possible.” Decennial Census Improvement Act of 1991, Pub.L. No. 102-135, § 2(a), 105 Stat. 635. The 1991 Act specifically required the NAS to “consider ... the appropriateness of using sampling methods, in combination with basic data-collection techniques or otherwise, in the acquisition or refinement of population data.” Id. § 2(b).
It is important to observe how these two pieces of legislation work together. The 1976 amendment to the Census Act instructs the Secretary that if he “considers” sampling to be feasible, he must authorize its use; the 1991 Act establishes a specific method for “considerfing]” the use of sampling, namely a scientific evaluation to be conducted by the NAS.
The NAS did conduct a study and concluded not only that sampling was appropriate, but furthermore that “improved coverage or data quality” could not be achieved by “traditional Census methods of physical enumeration,” even with additional funding. See PLAN FOR CENSUS 2000, at 7(quoting NAS panel). “[T]o continue trying to count every last person” physically, the NAS reported, “is fruitless.” Id. Following the various NAS analyses, the Census Bureau prepared a report to Congress in 1997 that included a plan for Census 2000 providing for a “limited use of sampling” in order to ensure an “accurate and cost-effective census.” Id. at 11. The Bureau’s Accuracy and Coverage Evaluation (A.C.E.) Report, issued in June 2000, provided a detailed evaluation of the feasibility of the proposed method of statistical sampling and concluded that it *879was both “operationally and technically feasible,” and that barring unforeseen circumstances the adjusted data “will be more accurate than the uncorrected data.” Accuracy and Coverage Evaluation, 65 Fed.Reg. 38,374, 38,375 (2000). Then. Secretary of Commerce William Daley adopted these conclusions, although a final decision about whether to issue adjusted numbers was deferred until the data were collected and analyzed for reliability.
After a change in administration following the disputed presidential election of 2000, however, the Executive Steering Committee on A.C.E. policy (ESCAP), the committee of the Census Bureau charged with evaluating the effectiveness of the A.C.E. sampling program, issued a report recommending the release of the unadjusted figures, rather than the corrected figures, as the official Census 2000 data. ES-CAP made this recommendation despite its own evaluation that the “A.C.E. was a design and operational success,” that “the A.C.E. operations appear to have been in control, performed as expected, and produced data as good or better than the data produced by the 1990 [post-enumeration survey],” and that “[u]nder what the Committee considered reasonable assumptions, state, congressional district, and county level analyses showed a marked improvement for adjustment.” Notice of Recommendation and Report, 66 Fed.Reg. 14,004, 14,004-06, 14,012.
ESCAP based its recommendation not to use the A.C.E. figures on its lack of certainty that sampling would improve the accuracy of the census. Id. at 14,005. It cited an unexplained discrepancy between those figures and those that resulted from the Bureau’s demographic analysis (DA), which provides an independent benchmark for evaluating Census data. Id. at 14,005, 14,007. ESCAP concluded that the discrepancy most likely resulted from one of the following three causes: (1) inaccuracies in the 1990 Census, resulting in incorrect assumptions for the DA; (2) inaccuracies originating in the DA itself, or (3) inaccuracies in the Census 2000, as corrected by the A.C.E. Id. at 14,012. In other words, ESCAP recommended against using the A.C.E. figures because one of the possible causes of the discrepancy between those figures and the DA involved a problem with the A.C.E. figures, and despite the fact that “a majority of the evidence indicates both the continued existence of a differential 'undercount of the population and the superior accuracy of the adjusted numbers.” Id. at 14,005. Although it found that the adjusted numbers were probably superior, ESCAP recommended against using them because of “[t]he potential for a reversal of these findings.” Id. at 14,014 (emphasis added). As the ESCAP concluded, the experts could not be certain within the time frame allotted that the adjusted data would improve the accuracy of the census. Id. at 14,005. That, as I explain below, does not constitute a determination as to feasibility, or even a consideration of that question.
The DA and the A.C.E. did agree in one important respect: Both showed that “Census 2000 perpetuated the historic phenomenon of the differential undercount.” Id. at 14,007.
Acting Census Bureau Director William Barron, Jr., who was a member of ESCAP, forwarded to the newly appointed Secretary of Commerce Donald Evans his recommendation that the Secretary not release the adjusted data. On March 6, 2001, Secretary Evans adopted the ES-CAP recommendation and ordered that the unadjusted data be released for redistricting and federal funding purposes.
Because this court is required to review a federal agency’s implementation of its governing statute, the majority is correct that the proper mode of analysis is the *880two-step framework established in Chevron v. Natural Resources Defense Council, 467 U.S. 837, 104 S.Ct. 2778, 81 L.Ed.2d 694 (1984). First, the court must examine the statutory language at issue to determine whether Congress provided a clear directive to the agency. “If the intent of Congress is clear, that is the end of the matter; for the court, as well as the agency, must give effect to the unambiguously expressed intent of Congress.” Id. at 842-43, 104 S.Ct. 2778. If, however, the statute .does not provide a clear answer, the Court must uphold the agency’s interpretation if it is “reasonable.” Id. at 844, 104 S.Ct. 2778.
The statutory language at issue provides that “the Secretary shall, if he considers it feasible, authorize the use of the statistical method known as ‘sampling’ in carrying out the provisions of this title.” 13 U.S.C. § 195. In my view, the Secretary violated the clear intent of Congress by failing to base his decision on a “consideration]” of whether sampling was feasible. I would therefore consider the matter settled under Chevron step one. Furthermore, even if I were persuaded that the Secretary did “consider” feasibility, I believe the record precludes a finding that he considered sampling not to be “feasible” under any reasonable definition of the word. In short, any conclusion that he reached would necessarily be arbitrary and capricious. I would, therefore, hold that his decision must be rejected under Chevron step two as well.
Under the majority’s interpretation, § 195 does not require the Secretary to do anything; rather, Congress’s use of the word “shall” merely “demonstrates Congress’ intent to strongly encourage the use of sampling.” Maj. op., at 869-70 (emphasis added). Courts, however, do not customarily ascribe to the word “shall” the sense of “eneourage[ment].” To the contrary, long established Supreme Court law describes the word “shall” as creating a mandatory obligation. See, e.g., Lopez v. Davis, 531 U.S. 230, 241, 121 S.Ct. 714, 148 L.Ed.2d 635 (2001) (“The statute’s use of the permissive ‘may’ contrasts with Congress’ use of a mandatory ‘shall’ elsewhere in [the statute] to impose discretionless obligations.... ”); Lexecon, Inc. v. Milberg Weiss Bershad Hynes & Lerach, 523 U.S. 26, 31, 118 S.Ct. 956, 140 L.Ed.2d 62 (1998) (“The Panel’s instruction comes in terms of the mandatory ‘shall,’ which normally creates an obligation impervious to judicial discretion.”).
The mandatory effect of the word “shall” in this instance is confirmed by the statutory history. Before the 1976 amendments, § 195 was highly permissive. As adopted in 1957, it read, “[T]he Secretary may, where he deems it appropriate, authorize the statistical method known as sampling.” Pub.L. No. 85-208, 71 Stat. 484 (emphasis added). The 1976 amendments replaced the permissive “may” with the mandatory “shall”: “[T]he Secretary shall, if he considers it feasible, authorize the statistical method known as sampling.” Pub.L. No. 94-521, § 10, 90 Stat. 2459, 2464 (emphasis added). This change in language clearly evinces congressional intent to create a mandatory obligation on the part of the Secretary. Under the majority’s interpretation, however, the new language has precisely the same effect as the old. See United States v. Redman, 35 F.3d 437, 441 (9th Cir.1994) (“The amendment of language ... must properly be understood as evidence of an intent to abandon the standard established by the old language in favor of something different.”).
By changing the language of the statute from “may” to “shall,” Congress demonstrated a clear preference for the use of sampling and limited the scope of the Secretary’s discretion to prevent its use. The *881“may” of the pre-1976 provision allowed the Secretary to base his determination whether to authorize sampling on whatever factors he reasonably found relevant, including whether as a general or abstract matter he considered sampling to be a more desirable process than employing unadjusted data. The decision whether to use sampling might properly have involved a wide range of philosophical and public policy issues. Before 1976, the Secretary was invited to weigh all such considerations in making his determination. By amending the statute, however, Congress effectively answered the indefinite range of policy questions — and answered them in favor of sampling. It concluded that sampling is to be used, when and if feasible. The determination as to feasibility was vested in the Secretary, who was to exercise his reasoned discretion in making that judgment. Under the amended statute, the Secretary’s discretion is limited to one consideration: whether he considers sampling to be feasible in the particular instance. He may not consider whether sampling would, as a general proposition, be fair, what its political consequences would be, or whether it is certain to produce better data. His discretion is limited to the essentially practical determination of whether sampling, in his considered judgment, is feasible, given the particular circumstances.
The import of the change in language is confirmed by the Senate Committee Report, which expresses the desire of Congress to “require” the use of statistical sampling by the Census Bureau:
Section 10 [of the 1976 Census Act] amends section 195 of title 13, U.S.C., to require that the Secretary of Commerce authorize the use of sampling procedures in carrying out the provisions of this title whenever he deems it feasible, except in the apportionment of the U.S. House of Representatives. This differs from present language which grants the Secretary discretion to use sampling when it is considered appropriate. The section as amended strengthens congressional intent that, whenever possible, sampling shall be used.
Report of the Senate Post Office and Civil Service Committee, 94-1256 at 6 (emphasis added), reprinted in 1976 U.S.C.C.A.N. 5463, 5468.
Finally, the Supreme Court has explicitly stated that the change in the language of § 195 created a mandatory obligation. Department of Commerce v. House of Representatives, 525 U.S. 316, 341, 119 S.Ct. 765, 142 L.Ed.2d 797 (“[T]he amendments .... changed a provision that permitted the use of sampling for purposes other than apportionment into one that required that sampling be used for such purposes ‘if feasible’ ”). The majority dismisses the Supreme Court’s statement as dictum, too lightly in my view. See United States v. Montero-Camargo, 208 F.3d 1122, 1132 n. 17 (9th Cir.2000) (“[W]e do not blandly shrug ... off [Supreme Court dicta] because they were not a holding.”) (internal quotation marks omitted).
The majority argues that its unusual interpretation of “shall” is necessitated by the phrase, “if he considers,” which confers discretion on the Secretary. To give the word “shall” its generally accepted meaning, the majority argues, “renders the ‘if he considers it’ portion of the phrase meaningless.” Maj. Op. at 870. The very opposite is the case, however. It is the majority’s interpretation of “if he considers” that renders the phrase “shall” meaningless. Under the majority’s interpretation the Secretary’s discretion is so broad that he may disregard Congress’s preference for sampling, regardless of whether sampling is objectively demonstrated to be manifestly feasible. The majority effectively glosses “if he considers it feasible” as “if he wants to do it.” In its effort to *882give meaning to one key term — “if he considers” — the majority frustrates the basic purpose and intent of the statute.
There is no need to reach this unattractive result. Section 195 requires the Secretary to authorize sampling whenever a certain specified condition is met: when the Secretary “considers” sampling to be feasible. Thus the Secretary can not fulfill his obligation under the statute without “consider[ing]” the question of feasibility, because without considering feasibility, it is impossible for him to know whether he must authorize sampling. Unavoidably, then, the statute requires the Secretary both to consider the feasibility of sampling and to base his decision whether to authorize sampling on that consideration.
“The word ‘consider’ in its ordinary usage means ‘to reflect on’ or ‘think about with a degree of care or caution.’ ” Finley v. Nat’l Endowment for the Arts, 100 F.3d 671, 689 (1981) (quoting Webster’s Third New International Dictionary 483), rev’d on other grounds, 524 U.S. 569, 118 S.Ct. 2168, 141 L.Ed.2d 500 (1998). The meaning of the word “considers” in the context of § 195 is also informed by the language of the Decennial Census Improvement Act of 1991, discussed above, in which Congress ordered the Secretary to contract with the NAS for a scientific study that would “consider ... the appropriateness of using sampling methods, in combination with basic data-collection techniques or otherwise, in the acquisition or refinement of population data.” Decennial Census Improvement Act of 1991, Pub.L. No. 102-135, § 2(b), 105 Stat. 635 (emphasis added). By requiring the Secretary to use a study by the NAS as a means of considering whether sampling is feasible, the 1991 Act shows congressional intent that the Secretary base his consideration of feasibility, at least in part, on objective scientific considerations.
The record shows that the Secretary did not base his decision not to authorize sampling on scientific or on any other considerations as to whether sampling was “feasible” and that he therefore violated the clear intent of Congress. Everything in the record pertaining to the question of feasibility, including the NAS study, the A.C.E. Report, and the ESCAP Report, demonstrates that sampling was feasible-feasible both operationally and technically, as the A.C.E. Report put it — and that it would most likely yield data more accurate than the unadjusted figures. The Secretary disregarded the question of feasibility and instead simply adopted the final recommendation of ESCAP not to authorize sampling because the Committee could not be “certain” that sampling would produce more accurate data. ESCAP did not base its recommendation on whether sampling was feasible, but rather on whether it could be certain that sampling would improve the data. Because the Secretary adopted the recommendation of ESCAP, and therefore, like ESCAP, did not base his decision on a consideration of feasibility, I would hold that he violated the clear intent of Congress and would reverse the decision of the district court.
The majority passes over the Secretary’s failure to base his decision on a consideration of feasibility and focuses instead on the malleability of the word “feasible.” Although I believe that the record shows a failure of the Secretary to consider feasibility, I will also discuss the word “feasible,” in order to show that, if the Secretary did, as the majority believes, consider “feasibility],” only an unreasonable definition of that word could have led him to the conclusion that sampling was not feasible. Because the Secretary could not have employed a reasonable definition of “feasible,” this court would be obligated to reverse the Secretary’s decision under *883step two of Chevron, even if we were not obligated to do so under step one.
The majority asserts that “the plain meaning of ‘feasible’ incorporates whether a particular action is ‘capable of being successfully done or accomplished.’ ” Maj. op. at 872. The term is ambiguous, the majority argues, as to how certain the Secretary must be “that the use of sampling would produce accurate results” in order to consider sampling feasible. Id. at 872. Because a level of uncertainty remained regarding the accuracy of the adjusted numbers, or as the majority says, the experts “could not be certain within the time frame allotted that the adjusted data would improve the accuracy of the census,” the Secretary could, in his discretion, find it not feasible to use them. Id. at 876.
The Secretary was required to consider whether sampling was feasible, not whether it was “certain within the time frame allotted that ... [it would] improve the accuracy of the census.” (Emphasis added.) The majority’s own definition of “feasible” is “capable of being successfully done.” (Emphasis added and emphasis removed.) The majority’s construction of the Secretary’s definition, in contrast, is “certain ” to be successfully done. A reasonable definition of feasible does not extend to certainty, and by allowing the Secretary to define feasibility in those terms, the majority undoes Congress’s limitation of the Secretary’s discretion through its 1976 amendments.
The majority’s argument is incorrect because it fails to take into account the nature of the choice that the Secretary faced. On the one hand, the Secretary could release the unadjusted data alone; on the other, he could release both the adjusted and the unadjusted. If feasibility turns on accuracy, therefore, it must turn on the relative accuracy of the adjusted and the unadjusted data. Although both data sets were certain to be inaccurate in various ways, all the evidence in the record points to the greater accuracy of the adjusted numbers. Most important, the ESCAP report, the recommendation of which the Secretary adopted, found that the sampling procedure was likely to- produce more accurate results. Thus, if accuracy or certainty of success was the Secretary’s measure of feasibility, releasing the adjusted numbers was more feasible than releasing the unadjusted numbers alone. Under the majority’s approach, then, the Secretary permissibly chose the less feasible option, on the ground that he did not consider the more feasible option feasible. Because this conclusion is not possible under a reasonable definition of “feasible,” I would hold that the Secretary’s decision must be rejected under step two of Chevron.
In addition, by choosing the less feasible option, the Secretary arbitrarily reversed the congressional preference for sampling and adopted a preference for using unadjusted data. The record is clear that, if the adjusted and unadjusted systems had been viewed neutrally or without a predisposition in favor of unadjusted data, the Secretary could not have concluded that sampling would not be feasible. Only by presuming that unadjusted data is more accurate, and that such presumption can be overcome only by demonstrating with certainty that sampling is more accurate, could the Secretary have reached the determination he did. Such a presumption is not only contrary to all the evidence in the record, but to the purpose, intent and language of the census statute as well.
Finally, I might note that the majority’s rebanee on the statement that the experts were unable to make a determination with certainty “within the time frame allotted” exceeds all tolerable limits of bureaucratic ineptitude and gives “bureaucracy” an *884even worse name than it already has. The census is conducted every 10 years. That statute setting forth the Secretary’s responsibility had been in effect for well over 20 years. The excuse that the Census Bureau did not have a fair and full opportunity to ascertain feasibility is simply that — a transparent excuse that cannot justify a failure by the Secretary, given all the other facts in the record before him, to conclude that the use of the adjusted data was feasible.
In short, I think it evident that the Secretary did not consider feasibility at all. Moreover, if he did, he would have been required to conclude that the use of the sampling method was feasible. Accordingly, I respectfully dissent.
. In 1990, Asians and Pacific Islanders were undercounted by 2.3%, African-Americans by 4.7%, Hispanics by 5.0%, and American Indians living on reservations by 12.2%. Non-Hispanic Whites, in contrast, were under-counted by only 0.7%. Children under the age of 18 accounted for 52% of the undercount, although they composed only 26% of the population. Renters were undercounted by 4.17% in urban areas and 5.92% in rural areas, whereas homeowners were under-counted by only 0.09% in urban areas and 0.03% in rural areas. PLAN FOR CENSUS 2000, at 3-4.