OPINION

The Department of Commerce (“DOC”) appeals the district court’s grant of summary judgment ordering it to disclose statistically adjusted data generated as part of Census 2000 in response to plaintiffs’ Freedom of Information Act (“FOIA”) request. DOC challenges the district court’s finding that the adjusted data are neither predecisional nor deliberative as required to permit nondisclosure under the “deliberative process” privilege in Exemption 5 to FOIA. We affirm.

I. FACTUAL AND PROCEDURAL BACKGROUND

It is generally accepted that the decennial census results in a net under-count of the population, particularly with respect to minority and disadvantaged groups. The federal government recognizes that the undercount has significant consequences for affected communities. Thus, the Bureau of Census (“Bureau”), an agency of DOC, has endeavored to obtain more accurate population estimates for purposes of intra-state redistricting and allocation of federal funds by generating adjusted census figures through statistical extrapolation based on targeted surveys of representative population blocks.

During Census 1990, the Bureau employed a large-scale statistical adjustment for the first time in a capture-recapture methodology to compensate for the presumed undercount in the unadjusted data. Compilation of the initial enumeration for the entire population was followed by post-enumeration surveys (“PES”) of representative households. The Bureau compared the results of the PES to the unadjusted data in order to determine a rate of error. That rate was then extrapolated to the entire population to adjust the results of the initial census count in an attempt to estimate more accurately population counts at the national, state, and local levels. However, experts differed as to whether the adjusted figures were more reliable than the unadjusted data. The Secretary of Commerce (“Secretary”) ultimately decided not to release the adjusted figures as the official data for intra-state redistricting or allocation of federal funds.

For Census 2000, the Bureau further refined the processes used to generate both the unadjusted and adjusted population counts. First, it made intensified efforts to obtain an accurate initial enumeration by involving local governments in the process of building address lists; designing a simplified questionnaire; developing a multi-step mailing strategy; implementing a paid advertising campaign; and restructuring the pay scale for temporary workers. Then, to generate adjusted data, the Bureau employed a capture-recapture sampling methodology called Accuracy and Coverage Evaluation (“ACE”), which was designed to improve on the adjustment process used in Census 1990. ACE involved an intensive survey of 314,000 representative households. Demographers and mathematical statisticians categorized individuals in both the initial enumeration and the ACE sample into 448 groupings, or “post-strata,” based on variables such as age, race, gender, and homeowner/renter status. The Bureau then compared the results from the ACE samples to those of the corresponding portions of the initial enumeration to create “coverage correction factors.” Applying those factors to the initial enumeration for each post-stratum within each representational block, the Bureau created adjusted block-level estimates. The Bureau then aggregated the block-level data to generate adjusted population estimates at the local, state, and national levels.

Next, the Bureau formed a committee called Executive Steering Committee for Accuracy and Coverage Evaluation Policy (“ESCAP”) to evaluate the accuracy of the adjusted data in order to advise the Secretary in deciding which figures to release as the official numbers for intra-state redistricting and allocation of federal funds. ESCAP compared the ACE totals to those produced by the Demographic Analysis (“DA”), ah independent measuring tool that looks to records of births, deaths, documented immigration, Medicare enrollment, and estimates of emigration and undocumented immigration. ESCAP initially advised the Secretary to release the unadjusted data for intra-state redistricting because it could not resolve discrepancies between the ACE totals and those produced by the DA, as well as other technical concerns, before the statutorily prescribed deadline to release that data to the states. After reviewing ESCAP’s recommendations, the Seéretary decided to release only the unadjusted data as the official Census 2000 numbers for intra-state redistricting.

ESCAP initially withheld judgment as to whether the adjusted figures should be used as the official data for allocation of federal funds, and the Bureau continued to consider whether to use the adjusted data for that purpose. But, after further consideration, ESCAP recommended in a second report issued in October 2001 that only the unadjusted data should be released as the official figures for allocation of federal funds and other non-redistricting purposes. The Secretary again reviewed ESCAP’s recommendations about the accuracy of the data adjusted using the ACE, and accordingly decided to release only the unadjusted data as the 2000 Census numbers for all official purposes.

Through its website, the Bureau has released ESCAP’s reports with recommendations and attachments. That information includes, inter alia, detailed and technical information about the adjusted data, including specifications for the ACE methodology; quality indicators for both the initial enumeration and ACE; equations used to measure error in the initial enumeration and the ACE; the criteria on which the decision to use unadjusted data was based; undercount percentages as measured by the ACE on the national level for each race category, males and females, renters and homeowners, and various age groups; and minutes of the ESCAP meetings. The Bureau has not released any adjusted figures below those aggregated at the national level.

Plaintiffs, two Oregon state legislators, issued a FOIA request to the Bureau for all data adjusted using the ACE for each of the jurisdictions covered by the entire 2000 census. In response, the Bureau invoked the “deliberative process” privilege under Exemption 5 to FOIA and denied their request. Plaintiffs then' appealed to DOC’s Assistant General Counsel, who upheld the Bureau’s decision. They thén filed suit against DOC in federal district court seeking declaratory judgment and injunctive relief. Relying on Assembly of California v. United States Department of Commerce (“Assembly ”), 968 F.2d 916 (9th Cir.1992), the district court found, that the deliberative process privilege did not permit nondisclosure of the adjusted numbers because, they were neither predeci-sional nor deliberative. It, therefore, granted summary judgment for plaintiffs and ordered DOC to release the adjusted data.

II.JURISDICTION

The district court had jurisdiction under FOIA, 5 U.S.C. § 552 (a)(4)(B). We have jurisdiction under 28 U.S.C. § 1291 .

III.STANDARD OF REVIEW

Although this case is on appeal from a grant of summary judgment, the standard of review is not simply de novo. See Fiduccia v. United States Dep’t of Justice, 185 F.3d 1035, 1040 (9th Cir.1999); Assembly, 968 F.2d at 919 . In a FOIA case, we will overturn the district court’s factual findings underlying its decision only for clear error. Fiduccia, 185 F.3d at 1040 ; Assembly, 968 F.2d at 919 . After giving deference to such factual findings, we then review de novo whether a particular FOIA exemption applies. Fiduccia, 185 F.3d at 1040 . As in Assembly, this case essentially turns on “whether disclosure of the requested information would reveal anything about the agency’s deci-sional process. This is a fact-based inquiry where deference to the district court’s findings is appropriate.” 968 F.2d at 919 .

IV.DISCUSSION

A FOIA request similar to the one at issue in this case was made for adjusted data generated as part of Census 1990. See id. at 918-19 . DOC advanced the same arguments in Assembly that it presents here. There, we affirmed the district court’s finding that the adjusted data from Census 1990 did not fall within the deliberative privilege exception in Exemption 5 because they were neither predecisional nor deliberative. The district court in this case evaluated the purported factual differences surrounding the 1990 and the 2000 adjusted data and determined that any distinctions were not legally significant. We agree.

FOIA requires government agencies to disclose to the public any requested documents. 5 U.S.C. § 552 (a); Assembly, 968 F.2d at 920 . An agency may avoid disclosure only if it proves that the documents fall within one of nine exceptions. 5 U.S.C. § 552 (b)(1)-(9); Assembly, 968 F.2d at 920 . Because FOIA’s purpose is to encourage disclosure, its exemptions are to be narrowly construed. Dep’t of Justice v. Julian, 486 U.S. 1, 8 , 108 S.Ct. 1606 , 100 L.Ed.2d 1 (1988); Assembly, 968 F.2d at 920 . The government bears the burden of proving that a requested document is exempted. 5 U.S.C. § 552 (a)(4)(B); Assembly, 968 F.2d at 920 .

Exemption 5 permits nondisclosure of “inter-agency or intra-agency memorandums [sic] or letters which would not be available by law to a party other than an agency in litigation with the agency.” 5 U.S.C. § 552 (b)(5). This provision shields “those documents, and only those documents, normally privileged in the civil discovery context.” NLRB v. Sears, Roebuck & Co., 421 U.S. 132, 149 , 95 S.Ct. 1504 , 44 L.Ed.2d 29 (1975). Accordingly, it includes a “deliberative process” privilege. Dep’t of Interior v. Klamath Water Users Protective Assoc., 532 U.S. 1, 8 , 121 S.Ct. 1060 , 149 L.Ed.2d 87 (2001). The purpose of this privilege is “to allow agencies freely to explore possibilities, engage in internal debates, or play devil’s advocate without fear of public scrutiny.” Assembly, 968 F.2d at 920 (citing Sears, 421 U.S. at 150-54 , 95 S.Ct. 1504 ; EPA v. Mink, 410 U.S. 73, 85-94 , 93 S.Ct. 827 , 35 L.Ed.2d 119 (1973)). Thus, Exemption 5 covers “documents reflecting advisory opinions, recommendations and deliberations comprising part of a process by which governmental decisions and policies are formulated.” Klamath, 532 U.S. at 8 , 121 S.Ct. 1060 (citation and internal quotation marks omitted).

To fall within the deliberative process privilege, a document must be both “predecisional” and “deliberative.” As sembly, 968 F.2d at 920 . We elaborated on these terms in Assembly:

968 F.2d at 920 (quoting Formaldehyde Inst. v. Dep’t of Health and Human Servs., 889 F.2d 1118, 1122 (D.C.Cir.1989)) (internal citations omitted) (alteration in original). Although these inquiries are to be analyzed separately, they tend to overlap in practice. Assembly, 968 F.2d at 920 .

A. Predecisional

“A document may be considered predecisional if it was ‘prepared in order to assist an agency decisionmaker in arriving at his decision.’ ” Id. at 921 (quoting Renegotiation Bd. v. Grumman Aircraft Eng’g Corp., 421 U.S. 168, 184 , 95 S.Ct. 1491 , 44 L.Ed.2d 57 (1975)). “Material which predates a decision chronologically, but did not contribute to that decision, is not pre-decisional in any meaningful sense.” As sembly, 968 F.2d at 921 . In Assembly, as here, the adjusted data were “prepared solely for the purpose of post-decision dissemination.” Id. Giving deference to the district court’s factual findings, we held in Assembly that the adjusted data generated as part of Census 1990 were not predeci-sional because those data, rather than the Bureau’s evaluation of the processes used, did not “contribute” to the choice of which data to release. Id.

DOC argues that, as opposed to the adjusted data at issue in Assembly, the figures adjusted using the ACE in 2000 did in fact contribute to the decisions regarding which data to release for intra-state redistricting and allocation of federal funds. DOC maintains that the primary reason for ESCAP’s recommendation to release unadjusted data for intra-state redistricting “was the unexplained difference” between estimates produced by the ACE and DA. This discrepancy purportedly proves that both estimates themselves contributed to the decision to use unadjusted data for redistricting. In contrast, DOC asserts that “[t]he 1991 decision rested primarily on methodological concerns about the underlying statistical model and on policy concerns about disrupting ongoing redistricting and creating perverse incentives with respect to conducting the headcount.” The 2001 decision, DOC argues, is therefore “substantially different” from that in 1991.

This argument is unpersuasive. First, just as in 1991, the Secretary’s decision in 2001 not to release the adjusted data was due to questions regarding the reliability of the data as determined by an evaluation of the methods used to adjust the original count. The Principal Associate Director for Programs at the Bureau stated in his declaration that “[t]he inconsistency between DA and the ACE raised the substantial possibility of an as yet undiscovered problem in the ACE methodology,” and that “[o]ther, more technical concerns are discussed in the ESCAP Report, including the possibility that balancing error or synthetic error had adversely affected the ACE data.” Further, the 1991 decision was also based, at least in part, on discrepancies between the adjusted data and DA.

DOC also argues that the extent of agency deliberations was markedly different in 1991 and 2001. Assembly rejected the argument that the adjusted data generated during Census 1990 were predeci-sional based on the possibility that they might be considered for purposes other than the redistricting decision at some undisclosed time in the future. Id. Here, DOC asserts that the 2000 adjusted data were the subject of actual deliberations after the initial decision with respect to intra-state redistricting, and “did in fact contribute to the October 2001 funding decision.” Again, this purported “difference” is irrelevant. The same reasons supporting the district court’s finding that the adjusted figures did not “contribute” to the Secretary’s decision regarding intrastate redistricting also support its finding that they did not “contribute” to his decision as to funding. And Assembly forecloses DOC’s argument on this point to the extent it asserts that the adjusted figures are predecisional because they are under consideration for other possible future uses. See id. (“Characterizing these documents as predecisional simply because they play into an ongoing audit process would be a serious warping of the meaning of the word.”) (citation and internal quotation marks omitted).

B. Deliberative

“The underlying purpose of the deliberative process privilege is to ensure that agencies are not forced ‘to operate in a fishbowl.’ ” Id. We must, therefore, “focus on the effect of the materials’ release: the key question in Exemption 5 cases [is] whether the disclosure of materials would expose an agency’s decision-making process in such a way as to discourage candid discussion within the agency and thereby undermine the agency’s ability to perform its functions.” Id. (citation and internal quotation marks omitted) (alteration in original). Thus, “[p]redecisional materials are privileged to the extent that they reveal the mental processes of decision-makers.” Id. (citation and internal quotation marks omitted).

Based on several factors, Assembly affirmed the district court’s conclusion that the adjusted data generated in Census 1990 were not deliberative. Id. at 922-23 . Significantly, release of the data would not have enabled the public to reconstruct any protected deliberative process, because it would have been impossible “to derive the formulas, or the process that created the formulas, from the adjusted data.” Id. at 922-23 . Assembly also noted that DOC already had disclosed most of its decisional process by revealing the method used to generate the adjusted data, as well as the adjusted census figures for the national, state, and city levels, and one-half of the adjusted figures at the block level. Id.

The district court in this case found that, just as in Assembly, release of the bare numbers by themselves would not reveal any protected decision-making process. DOC argues, however, that factual distinctions take this case outside the result in Assembly. DOC had attempted to withhold in Assembly only one-half of the block-level adjusted data, whereas here it attempts to withhold all adjusted figures below those aggregated at the national level. We agree with the district court that this distinction is not legally significant. Although we noted in Assembly that the extent the adjusted numbers had been disclosed weighed in favor of the district court’s conclusion that the remaining block-level adjusted data would not reveal any deliberative process, that fact was not crucial to our holding. Significantly, our decision in Assembly did not turn on waiver because, as we explained, “[ajgencies should not be penalized for openness.” See id. at 922-23 n. 5. Thus, as in Assembly, “[w]e consider prior disclosures only to determine whether the disclosure of [the adjusted data ] would expose the decision-making process any more than it has already been disclosed.” Id.

DOC further maintains that the 2000 adjusted figures are deliberative because they are the product of a complex and elastic statistical model which constitutes a protected opinion. We recognize that numbers can constitute deliberative material where they represent opinions and subjective judgments created to assist in the decision-making process, or where they would otherwise so expose that process. However, we rejected DOC’s argument on this ground in Assembly:

Id. at 922 .

Here, as in Assembly, the district court found that the adjustment process cannot be determined from the adjusted figures alone. Additionally, the Bureau has already disclosed the method and procedures used to generate adjusted data in 2000, as well as the factors it considered important in deciding which data to release as the official census numbers. Though derived through a refined process designed to improve upon that used in 1990, the adjusted data generated as part of Census 2000 retain the same character as the adjusted data at issue in Assembly. Thus, we agree with the district court that disclosure of the adjusted numbers would not expose any protected deliberative process.

Finally, we reject DOC’s argument that Exemption 5 applies because disclosure will chill future adjustment decisions. The thrust of its “chilling effect” argument is that DOC will be less likely to adjust census data in the future if forced to disclose the adjusted data generated during Census 2000 because it will not want to be forced again to release unreliable data to the public. But this argument did not permit non-disclosure in Assembly. See id. at 923 (“[Ijnaccuracy is not a basis for FOIA exemption.”). See also Petroleum, 976 F.2d at 1436-37 (noting that any concerns with public confusion caused by release of erroneous information could be allayed by warning FOIA requesters that the information is unofficial and disclaiming responsibility for “any errors or gaps”). Accordingly, DOC’s “chilling effect” argument does not permit nondisclosure under FOIA here.

V. CONCLUSION

The district court’s finding that the adjusted data were neither predecisional nor deliberative is supported by the record and in accordance with binding Ninth Circuit law. We, therefore, AFFIRM the judgment of the district court.

. The Census Clause of the U.S. Constitution, art. I, § 2, cl. 3, requires Congress to conduct a census every ten years. . Congress delegated its responsibility under the Census Clause to the Secretary of Commerce through the Census Act. See 13 U.S.C. § 4 . The Secretary of Commerce in turn delegated part of his responsibility to the Bureau of Census. . The Census Act authorizes the use of sampling methodology in conducting the census. 13 U.S.C. § 195 . But only unadjusted data may be used for apportioning seats in the United States House of Representatives. Dep’t of Commerce v. United States House of Representatives, 525 U.S. 316, 339-40 , 119 S.Ct. 765 , 142 L.Ed.2d 797 (1999). ."Unadjusted data” refers to the "initial enumeration,” “initial count,” or "initial headcount”; it is compiled primarily based on information obtained through mail-in surveys and personal visits. However, unadjusted data embody some statistical adjustment through imputation of demographic characteristics of neighboring households when census takers are unable to obtain information about a household. . See 56 Fed.Reg. 33,582 (July 22, 1991) (setting out the factors underlying the Secretary’s decision in 1991, which included comparison of estimates of the size of the population from the original enumeration, DA, and PES). . See Nat’l Wildlife Fed’n v. U.S. Forest Serv., 861 F.2d 1114, 1121 (9th Cir.1988) (holding that estimates of cost and usage associated with land use plans were deliberative because they were "opinions that figure[d] heavily in the formulation of Forest Service policies”); Quarles v. Dep’t of the Navy, 893 F.2d 390, 392-93 (D.C.Cir.1990) (holding that cost estimates prepared to assist in selection of home ports for ships were deliberative because they derived "from a complex set of judgments” and were a part of "that elasticity that has persuaded courts to provide shelter for opinions generally”). Cf. Petroleum Information Corp. v. U.S. Dep’t of Interior, 976 F.2d 1429, 1435 (D.C.Cir.1992) (holding that objective acreage estimates were not protected under the deliberative process privilege because they revealed nothing about "preliminary positions or ruminations about how to exercise discretion on some policy matter”).