National Ass'n of Clean Water Agencies v. Environmental Protection Agency ( 2013 )


Menu:
  •  United States Court of Appeals
    FOR THE DISTRICT OF COLUMBIA CIRCUIT
    Argued May 3, 2013                 Decided August 20, 2013
    No. 11-1131
    NATIONAL ASSOCIATION OF CLEAN WATER AGENCIES,
    PETITIONER
    v.
    ENVIRONMENTAL PROTECTION AGENCY AND GINA
    MCCARTHY, ADMINISTRATOR, EPA,
    RESPONDENTS
    MAXWEST ENVIRONMENTAL SYSTEMS, INC., ET AL.,
    INTERVENORS
    Consolidated with 11-1167, 11-1185, 12-1236, 12-1237
    On Petitions for Review of a Final Rule of the
    United States Environmental Protection Agency
    Jeffrey A. Knight argued the cause for
    petitioners/respondents-intervenors National Association of
    Clean Water Agencies, et al. With him on the briefs were
    Peter H. Wyckoff and Steven A. Hann.
    2
    James S. Pew argued the cause for petitioner Sierra
    Club. With him on the briefs was Jonathan A. Wiener.
    Lisa Sharp argued the cause for intervenor MaxWest
    Environmental Systems Inc. With her on the briefs was D.
    Cameron Prell.
    Michele L. Walter and Martha C. Mann, Attorneys,
    U.S. Department of Justice, argued the causes and filed the
    brief for respondents.
    Jonathan A. Wiener argued the cause for respondent-
    intervenor Sierra Club. With him on the brief was James S.
    Pew.
    Steven A. Hann, Jeffery A. Knight, and Peter H.
    Wyckoff were on the brief for respondents-intervenors
    National Association of Clean Water Agencies, et al.
    Before: GARLAND, Chief Judge, BROWN, Circuit
    Judge, and SENTELLE, Senior Circuit Judge.
    Opinion for the Court filed by Senior Circuit Judge
    SENTELLE.
    SENTELLE, Senior Circuit Judge: In March 2011, the
    Environmental Protection Agency (“EPA”) issued a final rule
    establishing emission standards for sewage sludge
    incinerators under § 129 of the Clean Air Act, 
    42 U.S.C. § 7429
    . See Standards of Performance for New Stationary
    Sources and Emission Guidelines for Existing Sources:
    Sewage Sludge Incineration Units, 
    76 Fed. Reg. 15,372
     (Mar.
    21, 2011). Determining that sewage sludge incinerators were
    “solid waste incineration unit[s]” as defined in § 129(g)(1),
    3
    the EPA promulgated “maximum achievable control
    technology” (“MACT”) standards for two subcategories of
    sewage sludge incinerators.
    The Clean Air Act cabins EPA’s discretion in setting
    MACT standards, requiring EPA to base the standards on the
    emissions achieved by the best-performing existing
    incinerators. See 
    42 U.S.C. § 7429
    (a)(2). But acting under
    pressure of a court order to establish the MACT standards by
    a set deadline, EPA took a targeted approach to collecting
    emissions data and used several different methods to estimate
    the emissions levels achieved by existing incinerators. See 76
    Fed. Reg. at 15,386.
    The petitioners challenge several different aspects of the
    rulemaking. Petitioners National Association of Clean Water
    Agencies and Hatfield Township Municipal Authority
    (collectively, “NACWA”) challenge EPA’s authority to
    regulate sewage sludge incinerators under § 129, asserting
    that sewage sludge incinerators do not fall within the scope of
    § 129(g)(1)’s definition of “solid waste incineration unit.”
    Petitioners NACWA and Sierra Club seek review of the
    sewage sludge incinerator emission standards, challenging
    several aspects of EPA’s methodology for estimating the
    emission levels achieved by the best performing units. In
    addition to these petitioners, MaxWest Environmental
    Systems, developer of a proprietary biosolids management
    process, intervenes to challenge EPA’s treatment of its
    technology in the sewage sludge incinerator rule.
    For the reasons stated below, we deny NACWA’s
    petition for review as to EPA’s authority to regulate sewage
    sludge incinerators under § 129. As to the petitioners’
    challenges to EPA’s methodology in setting emission
    standards, we agree that in some respects EPA has not
    4
    adequately established that its estimations are reasonable, and
    so remand parts of the sewage sludge incinerator rule to EPA
    for further proceedings without vacating the current standards.
    We otherwise deny the petitions for review, and will not
    consider intervenor MaxWest’s arguments as they are not
    within the scope of issues raised by the petitioners.
    I.      BACKGROUND
    A. Statutory Background
    The Clean Air Act requires EPA to set emission
    standards for polluting sources “to protect and enhance the
    quality of the Nation’s air resources.”          
    42 U.S.C. § 7401
    (b)(1). Section 112 of the Clean Air Act, 
    42 U.S.C. § 7412
    , requires EPA to set emission standards for a list of
    hazardous air pollutants emitted by major sources and area
    sources. 
    Id.
     § 7412(d). Section 129 is more specific,
    directing EPA to establish emission standards for a list of nine
    pollutants emitted by solid waste incineration units. Id. §
    7429(a). Subject to certain exceptions not relevant to this
    case, § 129 defines a “solid waste incineration unit” as “a
    distinct operating unit of any facility which combusts any
    solid waste material from commercial or industrial
    establishments or the general public (including single and
    multiple residences, hotels, and motels).” Id. § 7429(g)(1).
    Under § 129, the standards established by EPA must
    reflect “the maximum degree of reduction” in the emissions
    of a list of pollutants1 that EPA, “taking into consideration the
    1
    Specifically, § 129 requires EPA to set numerical emission
    limitations for: “particulate matter (total and fine), opacity (as
    appropriate), sulfur dioxide, hydrogen chloride, oxides of nitrogen,
    carbon monoxide, lead, cadmium, mercury, and dioxins and
    dibenzofurans.” § 129(a)(4).
    5
    cost of achieving such emission reduction, and any non-air
    quality health and environmental impacts and energy
    requirements, determines is achievable for new or existing
    units” in each category of sources. Id. § 7429(a)(2). EPA
    refers to these standards as the “maximum achievable control
    technology” standards, abbreviated as the MACT standards.
    Northeast Maryland Waste Disposal Authority v. EPA, 
    358 F.3d 936
    , 939–40 (D.C. Cir. 2004).
    Congress set parameters governing EPA’s establishing of
    the MACT standards, which EPA has implemented through a
    two-step process. First, EPA sets a baseline level of
    stringency for emissions controls known as the MACT floor.
    For new units, the MACT floor is the level of emissions
    control “that is achieved in practice by the best controlled
    similar unit,” as determined by EPA. 
    42 U.S.C. § 7429
    (a)(2).
    For existing units, the MACT floor is “the average emissions
    limitation achieved by the best performing 12 percent of
    units” in each category. 
    Id.
     Second, EPA determines whether
    more stringent “beyond-the-floor” MACT standards are
    achievable, taking into consideration the factors listed in §
    129(a)(2).
    The Clean Air Act makes promulgating MACT standards
    under § 112 and § 129 mutually exclusive. Id. § 7429(h)(2).
    Although the statutory directive on setting MACT standards is
    virtually identical under § 112 and § 129, EPA’s decision to
    regulate a source under one section rather than the other has
    practical consequences. For example, the list of pollutants for
    which EPA must set MACT standards differs between the two
    sections. Compare id. § 7412(b) (list of hazardous air
    pollutants), with id. § 7429(a)(4) (list of nine pollutants for
    which EPA must set MACT standards for solid waste
    incinerators). The stringency of regulation for sources
    covered under these sections can also differ, depending on the
    6
    type of source. Under § 129, all solid waste incinerators
    within § 129(g)(1)’s definition of “solid waste incineration
    unit” are subject to the MACT standards that EPA establishes
    for that category of incinerator. See id. § 7429(a). In
    contrast, the MACT standards established under § 112 are
    mandated only for “major sources,” defined as sources that
    have the potential to emit ten tons per year or more of any
    hazardous air pollutants, or twenty-five tons per year or more
    of any combination of hazardous air pollutants. See id. §
    7412(a)(1), (d)(5).
    For sources that are not “major sources”—defined in §
    112(a)(1) as “area sources”—EPA is given the discretion to
    establish standards providing “for the use of generally
    available control technologies or management practices . . . to
    reduce emissions of hazardous air pollutants.”           Id. §
    7412(d)(5). The generally available control technology
    standard is not governed by the same statutory requirements
    as the MACT standard, giving EPA more flexibility in
    regulating area sources. Because EPA determined in 2002
    that no sewage sludge incinerator emitted hazardous air
    pollutants at such a level as to qualify as a major source, the
    generally available control technology standard would apply
    to sewage sludge incinerators if EPA regulated them under §
    112. See National Emission Standards for Hazardous Air
    Pollutants: Revision of Source Category List Under Section
    112 of the Clean Air Act, 
    67 Fed. Reg. 6,521
    , 6,523. (Feb.
    12, 2002). Sewage sludge incinerators also would not be
    subject to monitoring and siting review requirements, which
    are mandated by § 129 but not by § 112. See 
    42 U.S.C. § 7429
    (a)(3), (c).
    7
    B. Regulatory Background
    Publicly-owned      treatment    works,   owned     by
    municipalities or regional authorities, are responsible for
    managing all sewage that enters into the sanitary sewer
    system. Publicly-owned treatment works first treat the
    wastewater, creating sewage sludge in the process, then use
    various methods to dispose of the sewage sludge. Many
    publicly-owned treatment works use sewage sludge
    incinerators to dispose of sewage sludge. EPA’s inventory of
    sewage sludge incinerators stood at 204 at the time of the
    rulemaking. 76 Fed. Reg. at 15,387.
    EPA proposed emission standards for sewage sludge
    incinerators in October 2010, asserting its authority under §
    129 to regulate “other categories of solid waste incineration
    units.” See 75 Fed Reg. at 63,263; 
    42 U.S.C. § 7429
    (a)(1)(E).
    EPA began to develop these standards after the District Court
    for the District of Columbia determined that EPA was failing
    to discharge its non-discretionary duty under provisions of §
    112, and ordered EPA to do so. See 75 Fed. Reg. at 63,264.
    Although the specific § 112 obligations with which EPA had
    failed to comply are not relevant to this petition, the district
    court determined that § 112 required EPA to set emission
    standards for sewage sludge incinerators.2 See Sierra Club v.
    Jackson, 1:01-CV-1537, EFC No. 84 at 23 (D.D.C. filed Aug.
    2, 2006); Sierra Club v. Jackson, 1:01-CV-1537, ECF No.
    150 at 6–8 (D.D.C. filed Jan. 20, 2011). After granting EPA
    multiple motions to extend the deadline for issuing sewage
    2
    The § 112 obligations were to identify and regulate certain area
    sources that account for 90 percent or more of aggregate air
    emissions of 30 hazardous air pollutants identified by EPA under §
    112(k)(3)(B)(i)–(ii). See Sierra Club v. Jackson, No. 1:01-CV-
    1537, ECF No. 80 (D.D.C. filed Mar. 31, 2006); see also 
    42 U.S.C. § 7412
    (c)(3), (k)(3)(B).
    8
    sludge regulations, the district court ultimately required EPA
    to promulgate the final rule by February 21, 2011. 
    Id.,
     ECF
    No. 150, at 25.
    1. Proposed Rule
    On October 14, 2010, EPA issued a proposed rule
    proposing emission standards for sewage sludge incinerators.
    See Standards of Performance for New Stationary Sources &
    Emission Guidelines for Existing Sources: Sewage Sludge
    Incineration Units, 
    75 Fed. Reg. 63,260
    . In the preamble,
    EPA explained that although it had stated in other rules its
    intent to regulate sewage sludge incinerators under § 112, it
    was proposing to regulate sewage sludge incinerators under §
    129 in light of our ruling in Natural Resources Defense
    Council v. EPA, 
    489 F.3d 1250
     (D.C. Cir. 2007). See id. at
    63,263 (citing Standards of Performance for New Stationary
    Sources and Emission Guidelines for Existing Source: Other
    Solid Waste Incineration Units, 
    70 Fed. Reg. 74,870
     (Dec. 16,
    2005)).
    In the proposed rule, EPA established two subcategories
    of sewage sludge incinerators: multiple hearth incinerators
    and fluidized bed incinerators. 
    Id. at 63,268
    . EPA found that
    these were the only two types of incinerators used to combust
    sewage sludge, and determined subcategorization was
    warranted because the combustion design for these two types
    of incinerators varied significantly. 
    Id.
    In proposing the MACT standards for the subcategories
    of incinerators, EPA extensively discussed the methodology it
    used to derive the MACT floors. See 
    id. at 63
    ,269–75. To
    select which units to survey for emissions data, EPA
    identified units equipped with the control technology that it
    believed would achieve the lowest emissions possible for the
    9
    § 129 pollutants. Id. at 63,270. EPA stated that using control
    technologies to select best-performing units was sufficient
    because municipalities were already required to limit the
    concentration of pollutants in sewage sludge under Clean
    Water Act regulations. Id. (citing 40 C.F.R. pt. 503). These
    preexisting regulations, EPA explained, caused incinerators to
    “burn a relatively homogeneous waste,” thus rendering
    control technologies a suitable proxy for targeting the best-
    performing units from which to collect emissions data. Id.
    For its dataset, EPA surveyed 9 municipalities, and
    supplemented the results of that study with data from State
    environmental agencies’ public databases, yielding emissions
    information from 5 fluidized bed incinerators and 20 multiple
    hearth incinerators, although EPA acknowledged that not
    every test contained information on all nine § 129 pollutants.
    Id. Because 12 percent of the existing incinerator population
    based on EPA’s then-current count of incinerators was 7
    fluidized bed incinerators and 20 multiple hearth incinerators,
    EPA acknowledged that it did “not have actual emissions test
    data for the population of units that represent the best-
    performing 12 percent,” for every pollutant. Because EPA
    interpreted § 129 to require a MACT floor dataset
    representative of the best-performing 12 percent of
    incinerators, EPA concluded it needed to determine whether
    its data from fewer than 12 percent of incinerators could
    represent the best-performing 12 percent. Id.
    EPA addressed this issue by explaining that it could use
    “statistical techniques to determine the minimum number of
    observations needed to accurately characterize the distribution
    of the best performing 12 percent of units in each
    subcategory.” Id.; see Memorandum from Eastern Research
    Group, Inc. to Amy Hambrick, U.S. EPA, at 7–9 (Jan. 2011)
    (“Revised MACT Floor Memo”). Based on this statistical
    10
    analysis, EPA concluded that it had collected enough
    observations to conclude that the dataset it used met “the
    minimum size needed to characterize the population of 12
    percent of the best-performing units for all pollutants, when
    late-arriving data are included.” 75 Fed. Reg. at 63,271.
    Nevertheless, EPA requested “that commenters provide
    additional emissions stack test data and supporting
    documentation, as that may enable us to establish a final
    MACT floor based on a more complete data set.” Id. at
    63,270.
    The proposed rule also discussed EPA’s methods for
    addressing variability in the emissions data it collected. EPA
    bases its MACT standards on short-term emissions test data,
    which are not always “representative of the range of operating
    conditions that the best-performing facilities face on a day-to-
    day basis.” Id. at 63,269. Therefore, EPA believed it needed
    to account for variability in emissions performance. Id. EPA
    explained that for two or more tests at a single incinerator
    under what appear to be the same operating conditions,
    “[v]ariations in emissions may be caused by different settings
    for emissions testing equipment, different field teams
    conducting the testing, differences in sample handling or
    different laboratories analyzing the results.” Id. And
    emissions may even vary within a single test, as each test
    comprises at least three separate test runs, and each test run
    captures only a snapshot of an incinerator’s performance. Id.
    To address this variability, EPA proposed using a
    statistical tool it terms the “upper prediction limit.” For future
    observations of emissions from an incinerator, the upper
    prediction limit “is the upper end of a range of values that
    will, with a specified degree of confidence, contain the next
    (or some other pre-specified) randomly selected observation
    from a population.” Id. Thus, a 99 percent confidence-level
    11
    upper prediction limit “represents the value which one can
    expect the mean of future 3-run performance tests from the
    best-performing 12 percent of sources to fall below, with 99
    percent confidence, based upon the results of the independent
    sample of observations from the same best-performing
    sources.” Id. at 63,271.
    EPA’s proposed MACT floor methodology also
    addressed “non-detect data,” which are emission testing data
    too low for the testing equipment to accurately detect. Id. at
    63,272. Rather than estimate that non-detect data was at the
    “method detection level,” i.e., “the minimum concentration of
    a pollutant that can be measured with confidence that the level
    is greater than zero,” EPA Br. at 61 n.20, EPA used a
    different test to determine the MACT floor. Id. at 63,273.
    Under the test, EPA multiplied what it termed the
    “representative method detection level” by three, and
    compared that value to the MACT floor that EPA calculated
    using all data, including non-detect data. Id. If three times the
    representative method detection level was less than the
    calculated MACT floor, EPA would conclude that the MACT
    floor calculation adequately addressed measurement
    variability; if not, EPA would use the three-times value “to
    ensure that the MACT floor emission limit accounts for
    measurement variability and imprecision.” Id.
    For new source MACT floors, EPA explained that it
    would base the floors “on the best-performing single source
    for each regulated pollutant, with an appropriate accounting
    for emissions variability.” Id. at 63,274. Thus, EPA
    identified the lowest emitting incinerator with at least three
    test runs, and applied the 99 percent upper prediction limit.
    Id. While EPA proposed a new source MACT floor for
    fluidized bed incinerators, it did not propose a new source
    MACT floor for multiple hearth incinerators. Id. at 63,272.
    12
    Instead, it proposed that all new incinerators—including
    multiple hearth incinerators—meet the emission limits for the
    best-performing fluidized bed incinerator, explaining that
    industry information suggested that future units constructed
    would likely be fluidized bed incinerators and that industry
    information demonstrated that new fluidized bed incinerators
    “have more efficient combustion characteristics resulting in
    lower emissions.” Id. at 63,272, 63,274.
    In discussing whether to set “beyond-the-floor” MACT
    standards for existing sources, EPA determined that for most
    of the § 129 pollutants, no additional control technologies
    were available that would cost-effectively reduce emissions.
    Id. at 63,275, 63,277. For mercury, EPA concluded that using
    activated carbon injection with some form of particulate
    matter control for multiple hearth incinerators would be a
    cost-effective option for achieving beyond-the-floor emission
    reductions, noting that these combined control technologies
    would also control for dioxins and dibenzofurans. Id. at
    63,276–77. For fluidized bed incinerators, EPA concluded
    that “[i]n light of the technical feasibility, costs, energy, and
    nonair quality health and environmental impacts” discussed in
    the rule, it was not reasonable to establish beyond-the-floor
    MACT standards for new and existing fluidized bed
    incinerators. Id. at 63,277.
    EPA also proposed monitoring requirements for all new
    and existing sewage sludge incinerators. Id. at 63,277–82. In
    relevant part, EPA proposed initial and annual emissions
    performance tests for most pollutants, with continuous
    monitoring as an alternative, and control device parameter
    monitoring for certain control technologies. Id. at 63,277.
    EPA specifically required continuous emissions monitoring
    for carbon monoxide on new sewage sludge incinerators,
    although continuous emissions monitoring for carbon
    13
    monoxide remained optional for existing incinerators, and
    optional for all other pollutants. Id. at 63,278, 63,281.
    2. Final Rule
    EPA promulgated the final rule setting emission limits
    for sewage sludge incinerators on March 21, 2011. 
    76 Fed. Reg. 15,372
    . The final rule remained substantially similar to
    the proposed rule, regulating sewage sludge incinerators
    under § 129 and generally adopting the methodology for
    setting the MACT floors stated in the proposed rule. Id. at
    15,382–92.
    The final rule did contain a few substantive changes.
    While EPA had proposed setting all new incinerator MACT
    floors on the best-performing fluidized bed incinerator, in the
    final rule it decided to set a separate MACT floor for new
    multiple hearth incinerators. Id. at 15,384. EPA explained
    that it had been persuaded by comments pointing out that
    under the proposed regulations, any source that exceeded a
    threshold in modification costs would be considered a new
    unit. See 
    40 C.F.R. § 60.4775
     (defining a new sewage sludge
    incinerator as a unit that “[c]ommenced modification after
    September 21, 2011”); 
    40 C.F.R. § 60.4930
     (defining
    modification as “a change to an existing [sewage sludge
    incinerator] later than September 21, 2011 and that meets one
    of two criteria”). Because it did not want to discourage
    municipalities from modifying multiple hearth incinerators,
    and because there was otherwise no technical reason why
    municipalities could not build new multiple hearth
    incinerators, EPA explained it decided to establish separate
    new incinerator MACT floors. 76 Fed Reg. at 15,384.
    In setting the new multiple hearth incinerator MACT
    floors, EPA’s upper prediction limit analysis on what it
    14
    deemed the best-performing multiple hearth incinerators
    yielded MACT floors for two pollutants that were less
    stringent (i.e., a higher emission limit) than what it had
    calculated for existing multiple hearth incinerators. EPA set
    the new multiple hearth incinerator floors for these two
    pollutants—hydrogen chloride and sulfur dioxide—at the
    same level as existing multiple hearth incinerator floors,
    reasoning that new incinerator MACT floors could not be less
    stringent than existing incinerator MACT floors. 
    Id. at 15
    ,388–89.
    EPA also deviated from the proposed rule by deciding
    not to set beyond-the-floor standards for any pollutants. 
    Id. at 15,380
    . In the final rule, EPA explained that the cost of
    requiring the additional contemplated control technology to
    reduce mercury was $80,000 to $100,000 per pound removed,
    and that, based on this cost and other factors, it determined
    that beyond-the-floor standards were no longer appropriate.
    
    Id. at 15,394
    .
    EPA also made minor changes to its MACT floor dataset,
    such as reducing its inventory of incinerators to 204, and
    consequently, reducing the numbers of incinerators needed to
    represent 12 percent to 18 multiple hearth incinerators and 8
    fluidized bed incinerators. 
    Id. at 15,387
    . Although the
    reduction in inventory decreased the number of incinerators
    necessary to represent 12 percent, EPA still did not have
    emissions data from 12 percent of incinerators for certain
    pollutants. While commenters attempted to supplement that
    dataset by submitting emissions stack test data that EPA
    requested in the proposed rule, EPA rejected that data because
    commenters had not substantiated it with emission test
    reports. 
    Id.
    15
    In the final rule, EPA candidly noted that its MACT floor
    methodology—including the emissions testing dataset from
    less than 12 percent of incinerators—was motivated in part by
    the impending court-ordered deadline to establish emission
    standards. EPA explained that “given the court-ordered
    deadline for EPA to issue the final [sewage sludge
    incinerator] rule, it was not possible to undertake the time-
    consuming process of sending an [information collection
    request] to all the affected [sewage sludge incinerators]
    consistent with the requirements of the [Paperwork Reduction
    Act].” 
    Id. at 15,386
    .
    EPA also responded to comments criticizing EPA for not
    using data available to it to set MACT floors, including data
    about variability in sewage sludge metal concentrations
    collected from the Clean Water Act regulations. EPA
    responded that the upper prediction limit and its survey of
    units from nine different states adequately accounted for
    variability. 
    Id. at 15,391
    . EPA further stated that it “did not
    have sufficient information at proposal to consider if it were
    appropriate to incorporate variability based on sludge
    content,” explaining that the data commenters submitted was
    not adequately supported and therefore insufficient to clarify
    the effect of sewage sludge variability on emissions. 
    Id.
    Sierra Club and NACWA filed petitions for
    reconsideration of EPA’s final rule. EPA denied both
    petitions. See 77 Fed Reg. 25,087 (Apr. 27, 2012). Sierra
    Club filed a petition for review in this court. NACWA, joined
    by the Hatfield Township Municipal Authority, also filed a
    petition for review. We have consolidated all petitions for
    review.
    16
    II.    NACWA’S AND SIERRA CLUB’S PETITIONS
    FOR REVIEW
    A. EPA’S AUTHORITY TO REGULATE SEWAGE SLUDGE
    INCINERATORS
    We first address NACWA’s contention that EPA violated
    the Clean Air Act by setting emission standards for sewage
    sludge incinerators under § 129 rather than § 112.
    Specifically, NACWA asserts that § 129(g)(1)’s definition of
    “solid waste incineration unit” excludes sewage sludge
    incinerators.
    Section 129(g)(1) defines a solid waste incineration unit
    as “a distinct operating unit of any facility which combusts
    any solid waste material from commercial or industrial
    establishments or the general public (including single and
    multiple residences, hotels, and motels).” 
    42 U.S.C. § 7429
    (g)(1). In interpreting the phrase “solid waste material
    from commercial or industrial establishments or the general
    public,” EPA explained in its final rule that “[s]ewage sludge
    clearly originates from the general public, including
    residential and commercial facilities. Simply because the
    waste is treated at a [publicly-owned treatment work] prior to
    combustion does not change the original source of the sewage
    sludge.” 76 Fed. Reg. at 15,383.
    In contrast to EPA’s interpretation, NACWA argues that
    the words “from . . . the general public” “refer only to the
    proximate source of the solid waste material in question,”
    covering, for example, trash a municipality collects from a
    house and transports to a municipal incinerator, but not a
    waste product that the municipality itself creates. NACWA
    Br. at 21–22. Because the sewage sludge incinerated by a
    publicly-owned treatment work is the product of the treatment
    17
    of domestic sewage, NACWA asserts that sewage sludge
    comes from the publicly-owned treatment work, and not
    “from . . . the general public” that produces the domestic
    sewage. Id. at 19
    Because NACWA asks us to review EPA’s construction
    of § 129 as authorizing EPA to regulate sewage sludge
    incinerators under the category of “other . . . solid waste
    incineration units,” see 76 Fed. Reg. at 15,383 (citing 
    42 U.S.C. § 7429
    ), we apply Chevron v. NRDC to determine
    whether EPA is entitled to deference in its interpretation. See
    
    467 U.S. 837
    , 842 (1984). Under Chevron, we first determine
    whether the statute unambiguously forbids EPA’s
    interpretation. 
    Id.
     at 842–43. If the statute is silent or
    ambiguous, we then question whether EPA’s interpretation is
    based on a permissible construction of the statute. 
    Id. at 843
    .
    At first glance, the definition of solid waste incineration
    units in § 129(g)(1) appears ambiguous, a reality even
    NACWA acknowledges. See NACWA Br. at 21 (“Read
    alone, the word ‘from’ does not reveal whether it refers to the
    proximate source of the material or whether it refers instead to
    a distant ‘original’ source of the material.”); see also Oral
    Arg. Tr. at 5:19–23, 7:4–10, 10:5–19 (conceding that, without
    any other context, § 129 “would carry the meaning that EPA
    has ascribed to it”). Among the dictionary definitions of
    “from” is “a function word to indicate the source or original
    or moving force of something: as . . . the place of origin,
    source, or derivation of a material or immaterial thing.”
    WEBSTER’S THIRD NEW INTERNATIONAL DICTIONARY 913
    (1981). Thus, W.B. Yeats may proclaim, “All creation is
    [from] conflict,” and not necessarily mean that creation
    springs directly from conflict rather than through intermediate
    consequences of conflict, while one who states that a man
    “took a dime [from] his pocket” could only be understood to
    18
    mean that the dime originated from a specific location on a
    specific person. Id. Because the word “from” in § 129(g)(1)
    may be susceptible to either sense of the word, we agree with
    both parties that the phrase “from . . . the general public,”
    standing alone, is textually ambiguous. Cf. Environmental
    Defense Fund v. EPA, 
    167 F.3d 641
    , 652 (D.C. Cir. 1999)
    (Williams, J., dissenting) (explaining that the phrase “comes
    from” is ambiguous—for example, “A layabout who says he
    ‘comes from a hardworking family’ can be telling the truth
    even if all his relatives are dead.”).
    But textual ambiguity is not the end of the matter, as we
    have held that “a statute may foreclose an agency’s preferred
    interpretation despite such textual ambiguities,” an analysis
    we undertake by “exhaust[ing] the traditional tools of
    statutory construction to determine whether a congressional
    act admits of plain meaning.” Catawba County v. EPA, 
    571 F.3d 20
    , 35 (D.C. Cir. 2009); Arizona Public Service Co. v.
    EPA, 
    211 F.3d 1280
    , 1287 (D.C. Cir. 2000). Thus, although
    both NACWA and EPA acknowledge § 129(g)(1)’s apparent
    textual ambiguity, both parties also argue that the statute
    unambiguously resolves in their favor. EPA relies on our
    opinion in NRDC v. EPA, 
    489 F.3d 1250
     (D.C. Cir. 2007), in
    which we interpreted the definition of “solid waste
    incineration unit” to be very broad, while NACWA relies
    primarily on the words surrounding “from” and on the overall
    structure of the Clean Air Act and the Clean Water Act.
    We begin with NACWA’s argument that “from . . . the
    general public” requires a proximate cause interpretation, lest
    the words “general public” become superfluous. NACWA
    asserts that because all waste has its origin in the general
    public at some point, Congress could have simply defined a
    solid waste incinerator as “a unit . . . which combusts any
    solid waste material” without having to add “from
    19
    commercial or industrial establishments or the general
    public.” See 
    42 U.S.C. § 7429
    (g)(1). But even if we take as
    true NACWA’s assertion that all solid waste “originates”
    from the general public, and NACWA’s assertion that EPA’s
    interpretation of § 129(g)(1) would therefore cover all units
    that incinerate solid waste (except those specifically excluded
    in § 129(g)(1)), the “‘preference for avoiding surplusage
    constructions is not absolute.’” Amoco Products Co. v.
    Watson, 
    410 F.3d 722
    , 734 (D.C. Cir. 2005) (quoting Lamie v.
    U.S. Trustee, 
    540 U.S. 526
    , 536 (2004)). In some cases,
    redundancy may reflect the broad purpose of a congressional
    statute. See Babbitt v. Sweet Home Chapter of Communities
    for a Greater Oregon, 
    515 U.S. 687
    , 698 n.11 (1995).
    Although defining a covered incinerator as one that
    combusts solid waste “from commercial or industrial
    establishments or the general public” suggests some limitation
    on the coverage of § 129(g)(1) based on the source of the
    waste, the extent of that limitation is unclear. Congress may
    have intended to define solid waste incinerators to exclude
    specific categories of incinerators beyond the express
    exceptions listed in § 129(g)(1); for example, incinerators
    combusting waste directly produced by state or local
    government sources. But it also may have intended to give
    the definition of § 129(g)(1) a broad scope, with “the general
    public” functioning as something akin to a catchall. Thus, the
    fact that the three broad categories of sources of solid waste
    listed in § 129(g)(1)—commercial and industrial
    establishments, and the general public—may be surplusage
    under EPA’s original source interpretation does not
    unambiguously mean that Congress intended for the word
    “from” to have NACWA’s proximate source interpretation.
    In any event, if Congress indeed unambiguously intended to
    exclude municipal sewage sludge incinerators from the
    definition of § 129(g)(1), it chose a strange way to go about it.
    20
    NACWA also asserts as factually incorrect EPA’s
    statement that “[s]ewage sludge clearly originates from the
    general public, including residential and commercial
    facilities,” arguing that sewage sludge does not “originate”
    until a publicly-owned treatment work treats raw sewage.
    NACWA Br. at 22–24. In support of this argument, NACWA
    cites EPA regulations in which EPA recognized that sewage
    sludge results from wastewater treatment and is distinct from
    domestic sewage, and to a Clean Water Act provision from
    which one may infer that the production of sewage sludge is,
    by statutory definition, part of a publicly-owned treatment
    work. See 
    33 U.S.C. § 1292
    (2)(A); 
    40 C.F.R. §§ 60.4930
    ,
    60.5250; Joint Appendix 986–87.
    NACWA’s argument, however, fails to address how
    EPA’s original source interpretation of “from”—i.e., that the
    general public is a “but-for” cause of sewage sludge—renders
    the treatment facility that creates sewage sludge relevant. For
    example, one could say “bread comes from fields of wheat,”
    and be understood, or say “bread comes from the baker,” and
    also be understood. The fact that several intermediate
    processes had to occur to produce the bread—transporting
    wheat from the field, adding different ingredients to produce
    dough, or heating the dough in an oven—does not negate the
    validity of a sentence that uses “from” to link the bread to the
    source of an important ingredient. As we noted above,
    WEBSTER’S, supra 17, at 913, at least one dictionary defines
    “from” as a function word used to indicate, among other
    meanings, “the place of origin, source or derivation of a
    material or immaterial thing.” (emphasis added). Thus, the
    fact that the sewage sludge may not exist in that form until
    treated at a publicly-owned treatment work does not
    unambiguously       invalidate     EPA’s     original    source
    interpretation that sewage sludge is from the general public,
    21
    even if EPA’s interpretation is some steps removed. While it
    is also true that other EPA regulations recognize that sewage
    sludge is distinct from domestic sewage, these regulations are
    not dispositive of EPA’s interpretation of § 129(g)(1). Put
    differently, the fact that EPA determined in other regulations
    that sewage sludge and domestic sewage are distinct does not
    preclude EPA from recognizing that sewage sludge would not
    exist but for domestic sewage, and does not prevent EPA
    from interpreting “from . . . the general public” as meaning
    the original, but-for source of sewage sludge.
    In addition to its textual arguments, NACWA asserts that
    EPA’s interpretation of § 129(g)(1) conflicts with § 112(e)(5),
    which states that EPA “shall promulgate standards pursuant to
    subsection (d) of this section applicable to publicly owned
    treatment works (as defined in title II of the Federal Water
    Pollution Control Act) not later than 5 years after November
    15, 1990.” NACWA argues that when § 112(e)(5) is read in
    conjunction with the Clean Water Act, which NACWA
    claims defines publicly-owned treatment works to include
    sewage sludge incinerators, EPA must establish emission
    standards for the entirety of a publicly-owned treatment work
    pursuant to § 112(d). Because EPA may only establish
    emission standards for a source exclusively under either §
    112(d) or § 129(a), see 
    42 U.S.C. § 7429
    (h)(2), NACWA
    maintains that § 112(e)(5) supports its interpretation that
    Congress did not intend for EPA to regulate sewage sludge
    incinerators under §129(g)(1).
    We agree with EPA, however, that § 112(e)(5) is simply
    a timing provision, and does not prevent EPA from regulating
    aspects of a publicly-owned treatment work to which more
    specific provisions apply. See FTC. v. Manager, Retail Credit
    Co., Miami Branch Office, 
    515 F.2d 988
    , 993 (D.C. Cir.
    1975) (“The principle that a specific statutory provision
    22
    prevails over a more general provision is established beyond
    question.”). If we accept EPA’s interpretation of “from . . .
    the general public,” § 129 would govern sewage sludge
    incinerators, and the exclusivity provision of § 129(h)(2)
    would render § 112 not “applicable to” sewage sludge
    incinerators, leaving no conflict between the texts. Thus, the
    overall structure of the Clean Air Act does not unambiguously
    require NACWA’s interpretation of the word “from.”
    In fact, when EPA issued its rule proposing emission
    standards for publicly-owned treatment works as required by
    § 112(e)(5), it established standards for certain processes at
    publicly-owned treatment works while deciding to regulate
    sewage sludge incinerators under § 129, a decision that went
    unchallenged at the time. See National Emission Standards
    for Hazardous Air Pollutants: Publicly Owned Treatment
    Works, 
    63 Fed. Reg. 66,084
    , 66087 (Dec. 1, 1998) (Proposed
    Rule); 
    64 Fed. Reg. 57,572
     (Oct. 26, 1999) (Final Rule).
    Though EPA’s decision in 1998 to regulate sewage sludge
    incinerators under § 129 does not prove that its interpretation
    is correct in the present, the fact that it established standards
    for other processes within publicly-owned treatment works
    under its § 112 authority demonstrates that sewage sludge
    incinerators are not the only aspect of a publicly-owned
    treatment work to which § 112 may be “applicable.” See 40
    C.F.R. Part 63, Subpart VVV.
    We therefore conclude that the traditional tools of
    statutory construction do not demonstrate that § 129(g)(1)
    unambiguously excludes sewage sludge incinerators. But we
    also disagree with EPA that our opinion in NRDC v. EPA, 
    489 F.3d 1250
     (D.C. Cir. 2007), unambiguously resolves §
    129(g)(1) in EPA’s favor.
    23
    In NRDC, we heard challenges to an EPA rule that
    defined “commercial or industrial waste” to include solid
    waste combusted at incinerators that did not provide for
    energy recovery or operated without energy recovery. Id. at
    1258 (citing 70 Fed. Reg. at 55,572). Defining “commercial
    or industrial waste” in this way effectively created exceptions
    to the definition of “solid waste incineration unit” beyond
    those written in the statute. Id. We vacated and remanded
    EPA’s rule, rejecting its argument that it was resolving an
    ambiguity created by Congress’s failure to define
    “commercial or industrial waste.” Id. As we explained,
    Congress’s use of the word “any” in the definitional phrase
    “any facility which combusts any solid waste from
    commercial or industrial establishments” rendered the phrase
    clear and unambiguous, and EPA had no authority to create
    exceptions not explicitly listed in the statute through its
    definition of “commercial or industrial waste.” See id. at
    1259–60.
    Because the resolution of the present issue depends on
    the role of the word “from” in this statute, however, our
    discussion about the broad scope of § 129(g)(1)’s definition
    of “solid waste incineration unit” based on the word “any” is
    largely irrelevant. As we noted in NRDC, “The word ‘any’ is
    usually understood to be all inclusive,” 
    489 F.3d at 1257
    (internal citation omitted), and EPA presented no compelling
    reason “why ‘any’ should not mean ‘any.’” 
    Id. at 1260
    (internal citation omitted). In contrast, “from” is susceptible
    to different meanings and renders § 129(g)(1) ambiguous—
    even when viewed in light of the traditional tools of statutory
    construction—such that either NACWA’s or EPA’s
    interpretation of § 129(g)(1) is plausible.
    Having determined that the phrase “from . . . the general
    public” is ambiguous under Chevron step one, we now apply
    24
    Chevron step two to determine whether EPA’s interpretation
    “is based on a permissible construction of the statute.”
    Chevron, 
    467 U.S. at 843
    . If EPA’s “choice represents a
    reasonable accommodation of conflicting policies that were
    committed to [EPA]’s care by the statute, we should not
    disturb it unless it appears from the statute or its legislative
    history that the accommodation is not one Congress would
    have sanctioned.” 
    Id. at 845
     (internal quotation marks
    omitted).
    NACWA first asserts that EPA’s interpretation of §
    129(g)(1) is unreasonable under Chevron step two because
    EPA advanced a new rationale for its interpretation of §
    129(g)(1) for the first time in its denial of reconsideration to
    NACWA. Specifically, NACWA cites EPA’s reasoning that
    its conclusion on the coverage of § 129(g)(1) “is based on a
    reasonable interpretation of two provisions of the [Clean Air
    Act], so as to give both meaning,” and that “it is reasonable
    for the EPA to consider both provisions and to conclude that
    . . . section 129(g)’s all-encompassing definition of solid
    waste incineration unit requires regulation” under § 129.
    Joint Appendix 1092 (EPA’s letter denying NACWA’s
    petition for reconsideration). NACWA argues that EPA’s
    explanation is irrational not only because EPA allegedly
    raised it for the first time in the denial of reconsideration, but
    also because it amounts to a conclusory statement and
    because it fails to consider the importance of other
    environmental statutes.           We find these arguments
    unconvincing. EPA explained in its final rule that it viewed §
    112(e)(5) as merely a timing provision, and further stated that
    to interpret § 112(e)(5) more broadly “would conflict with
    section 129(g) and with the DC Circuit’s [sic] interpretation
    of section 129(g).” 76 Fed. Reg. at 15,383. EPA’s reasoning
    in denying NACWA’s petition for reconsideration is not only
    consistent with its reasoning in the final rule, but also a
    25
    reasonable interpretation of the statute for the reasons we
    have just explained.
    Second, NACWA argues EPA’s construction of §
    129(g)(1) is unreasonable because it ignores legislative
    history and the policies underlying the Clean Water Act’s
    provisions on sewage sludge incineration. NACWA contends
    that the discussion of wastewater treatment facilities in the
    legislative history of § 112, and the absence of any reference
    to sewage sludge or sewage sludge incinerators in § 129,
    demonstrate that Congress intended EPA to regulate all
    aspects of publicly-owned treatment works only under § 112.
    But though the legislative history on § 112 mentions
    wastewater treatment plants—an unsurprising fact in the
    context of a section stating the date by when EPA must issue
    § 112(d) emission standards applicable to publicly-owned
    treatment works—NACWA has cited no language in the
    legislative history pertaining to sewage sludge incinerators.
    We need not determine whether legislative history can
    generally suffice to render an agency’s interpretation invalid
    at Chevron step two. Nothing in the legislative history cited
    by NACWA suggests that Congress would not have
    sanctioned EPA’s interpretation of § 129(g)(1) as including
    sewage sludge incinerators or EPA’s interpretation of §
    112(e)(5) as being a timing provision.
    NACWA also contends that one of the congressional
    objectives of the Clean Water Act is to “maintain[] local
    flexibility and control over the means for managing sewage
    sludge,” and that adopting EPA’s interpretation of § 129
    would usurp local control. NACWA Br. at 27–28; see 
    33 U.S.C. § 1345
    (e) (“The determination of the manner of
    disposal or use of sludge is a local determination . . . .”). We
    agree with EPA, however, that this argument is largely
    irrelevant to whether it reasonably interpreted § 129.
    26
    Establishing MACT standards for sewage sludge incinerators
    does not, as a purely legal matter, remove local control over
    which method of sewage sludge disposal to use. Even if the
    presumably increased costs associated with emission
    standards would affect a municipality’s decision on how to
    dispose of sewage sludge, 
    33 U.S.C. § 1345
    (e) is not so
    strongly worded as to completely insulate a municipality’s
    decision-making process from EPA rulemaking.
    To sum, we conclude that the phrase “from . . . the
    general public” is ambiguous. Because EPA’s original source
    interpretation of that phrase is permissible, we give deference
    to its interpretation of the definition of “solid waste
    incineration unit,” and uphold its authority to establish
    emission standards for sewage sludge incinerators under §
    129.
    B. CHALLENGES TO THE MACT FLOOR METHODOLOGY
    Petitioners NACWA and Sierra Club both challenge the
    adequacy of EPA’s methodology in determining the MACT
    floors for existing units. Both challenge EPA’s decision to set
    MACT floors on emissions data from less than 12 percent of
    sewage sludge incinerators, albeit on different legal theories.
    See NACWA Br. at 32–37 (asserting that EPA’s failure to
    base MACT floors on less than 12 percent of incinerators
    violates § 129); Sierra Club Br. at 28–30 (asserting that
    EPA’s failure to base MACT floors on less than 12 percent of
    incinerators is arbitrary and capricious).
    Both petitioners also criticize distinct but related aspects
    of EPA’s rulemaking. Sierra Club contends that EPA’s
    method of selecting the best performers based on control
    technology is unlawful and arbitrary, pointing to other factors
    that may influence emission levels. See Sierra Club Br. at
    27
    18–23. NACWA argues that EPA failed to demonstrate that
    the data it used to set MACT floors represented the
    performance of the best-performing sewage sludge
    incinerators, contending that EPA failed to consider
    variability in sewage sludge contents and its effect on
    emission levels and that the upper prediction limit does not
    account for that variability. See NACWA Br. at 38–42.
    Sierra Club also challenges EPA’s use of the upper prediction
    limit, arguing that EPA does not demonstrate that the upper
    prediction limit represented the “average emissions limitation
    achieved” and was therefore unlawful and arbitrary. Beyond
    these related arguments, Sierra Club argues that EPA’s
    method for accounting for non-detect data is flawed.
    In promulgating the MACT standards for sewage sludge
    incinerators, EPA took a different approach than it has in
    other MACT standard regulations that have come before us
    on petitions for review. First, EPA collected its MACT floor
    dataset—i.e., the emission levels of “the best performing 12
    percent of units” for the existing incinerator MACT floors—
    by targeting the sewage sludge incinerators it believed to
    employ the best air pollution control technology for emissions
    testing. See 75 Fed. Reg. at 63,269–70. Second, after it had
    collected emissions data for the MACT floor dataset, EPA
    applied a statistical analysis, which it termed the “upper
    prediction limit,” to account for variability. See 75 Fed. Reg.
    at 63,269, 63,271 (explaining that EPA “must exercise its
    judgment, based on an evaluation of the relevant factors and
    available data, to determine the level of emissions control that
    has been achieved by the best performing [sewage sludge
    incinerators] under variable conditions.”).
    Both steps in this approach involved several different
    estimations and assumptions. For example, because EPA
    chose to limit its information collection requests to nine
    28
    municipalities, it had to estimate which sewage sludge
    incinerators would have the lowest emissions, which it chose
    to do based on the air pollution control technology the
    incinerators used. See 75 Fed. Reg. at 63,270. And because
    EPA’s limited emission testing did not yield data for 12
    percent of incinerators for every pollutant, it used a statistical
    technique to estimate whether a dataset of fewer than 12
    percent of incinerators could estimate the best-performing 12
    percent of incinerators. Id. Because every test did not
    produce usable data, EPA used an approximation to account
    for emissions test data too low to be accurately measured by
    monitoring equipment. Id. at 63,272. EPA also estimated the
    variability of the sewage sludge incinerators—what they
    would achieve under a range of operating conditions—by
    applying the upper prediction limit. Id. at 63,271.
    We have accorded Chevron deference to EPA’s
    interpretation of § 129 as allowing it to estimate MACT
    floors, noting that the requirement that the existing unit floors
    “not be less stringent than the average emissions limitation
    achieved by the best performing 12 percent of units” does not,
    on its own, dictate “how the performance of the best units is
    to be calculated.” Sierra Club v. EPA, 
    167 F.3d 658
    , 661–62
    (D.C. Cir. 1999). Although EPA would ideally set MACT
    floors by surveying all existing incinerators and identifying
    the best-performing 12 percent of units with hard data, we
    have not required EPA to go that far, recognizing that “EPA
    typically has wide latitude in determining the extent of data-
    gathering necessary to solve a problem.” 
    Id. at 662
    ; see also
    
    id. at 661
     (noting that Sierra Club, in arguing that case, had
    “disavowed any interpretation that would require measuring
    the performance of every last unit”); Oral Arg. Tr. at 53:25–
    54:9 (statement from EPA’s attorney stating, “EPA in a
    perfect world would have data from all 204 units”). Instead,
    we explained that the plain meaning of § 129(a)(2) does not
    29
    “exclude estimation, either by sampling or by some other
    reliable means.” Sierra Club, 
    167 F.3d at 662
    . But we have
    not given EPA free rein in its estimation techniques. EPA
    “must demonstrate with substantial evidence — not mere
    assertions” that its estimation “allows a reasonable inference
    as to the performance of the top 12 percent of units.” 
    Id. at 663
    ; Northeast Maryland Waste Disposal Authority, 
    358 F.3d at 954
     (internal quotation marks omitted).
    Relying on Sierra Club’s holding that EPA may estimate
    “the average emissions limitation achieved by the best
    performing 12 percent” without violating the Clean Air Act,
    we have often held EPA’s attempts to estimate the
    performance of the top 12 percent units to be lawful in theory.
    But we have often held that, in practice, EPA could not
    support the assumptions underlying its estimations with
    substantial evidence. For example, in Sierra Club, EPA based
    existing medical waste incinerator MACT floors on emission
    limits established by state regulations, assuming that “all
    [medical waste incinerators] are . . . achieving their
    [regulatory] limits.” Id. at 663 (second alteration in original).
    Although we held that EPA could, in theory, use regulatory
    data as a proxy, EPA’s use of the data in that case to estimate
    the performance of the top 12 percent was arbitrary and
    capricious because the state emission limits were substantially
    higher than emissions from an uncontrolled incinerator,
    rendering the regulatory data a meaningless proxy for
    emission levels from medical waste incinerators. See id. at
    663–64 (explaining that while the average emission level
    from uncontrolled incinerators was 2,770 parts per million
    volume, the average of the state emission limits appeared to
    be 5,227 parts per million volume); see also Northeast
    Maryland Waste Disposal Authority, 
    358 F.3d at
    953–54
    (rejecting EPA’s use of state emission levels for the same
    reason as in Sierra Club).
    30
    In Sierra Club, we also discussed a method EPA used to
    set new incinerator MACT floors, which are required to be
    “no less stringent than the emissions control that is achieved
    in practice by the best controlled similar unit.” See 167 F.3d
    at 664–65; 
    42 U.S.C. § 7429
    (a)(2). In setting new incinerator
    MACT floors, EPA had chosen what it believed to be the
    most effective control technology used by an incinerator in
    each category, identified the highest level of emissions (i.e.,
    worst) recorded by any incinerator using that technology, and
    then increased that value by 10 percent. Sierra Club, 167
    F.3d at 665. Selecting the control technology used by sources
    with the lowest emission levels and then setting MACT floors
    at the levels achieved by the worst performing source, which
    we termed the “MACT approach” in later cases, was
    supposed to account for the fact that the best-controlled
    similar unit will not consistently achieve the same emission
    level. See Cement Kiln Recycling Coaliation v. EPA, 
    255 F.3d 855
    , 861 (D.C. Cir. 2001). In other words, as we
    explained in Sierra Club, it is reasonable to expect that the
    incinerator on which the MACT floors are based should be
    able to “achieve” the MACT floor “in practice,” which it
    could not do unless “achieved in practice” meant “achieved
    under the worst foreseeable circumstances.” 167 F.3d at 665.
    But though EPA may account for variability and set MACT
    floors at the emission levels achieved by the best-controlled
    source under the worst foreseeable circumstances, we
    concluded in Sierra Club that EPA had not adequately
    explained why adopting the MACT approach would achieve
    that goal, and remanded the medical waste incinerator MACT
    standards to EPA for further clarification. Id.
    In later cases, we addressed EPA’s attempts to expand its
    MACT approach to developing existing source MACT floors.
    See National Lime Ass’n v. EPA, 
    233 F.3d 625
    , 632 (D.C. Cir.
    31
    2000); Cement Kiln, 
    255 F.3d at 859
    ; Sierra Club v. EPA, 
    479 F.3d 875
    , 879–880, 882–83 (D.C. Cir. 2007) (“Brick
    MACT”). For example, in Cement Kiln, EPA identified the
    best-performing 12 percent of sources by emission levels,
    identified the control technology used by sources with
    emission levels equivalent to or lower than the median of that
    12 percent, and then set the MACT floor at the worst emission
    level achieved by any source using that control technology.
    
    255 F.3d at 859
    . As in Sierra Club, EPA’s rationale in setting
    the MACT floors on the worst performer using MACT
    technology was to account for variability in the emission
    levels “achieved” by the best-performing 12 percent. See 
    id. at 862
    , 865–66. While we explained this approach could be
    lawful in theory if, for example, control technology was
    completely or significantly determinative of a source’s
    emission levels, we nevertheless concluded that EPA had not
    adequately demonstrated with substantial evidence that its
    estimation was reasonable. 
    Id.
     at 863–66. Because factors
    apart from air pollution control technology could affect
    emission levels, we concluded that EPA’s assumption that the
    worst-performing unit could represent the best-performing
    units was flawed. 
    Id. at 866
    ; see also Brick MACT, 
    479 F.3d at 882
     (“Given Cement Kiln’s holding that EPA may not use
    emission levels of the worst performers . . . without a
    demonstrated relationship between the two, we conclude that
    the emission floors . . . violate the [Clean Air Act].”); 
    id. at 883
     (“EPA’s decision to base floors exclusively on
    technology even though non-technology factors affect
    emission levels violates the [Clean Air Act].”).
    As these cases demonstrate, establishing MACT floors is
    no simple task. Determining the best performing sources is
    not even as straightforward as simply collecting emission test
    data from all incinerators and ranking them, as incinerators
    that have low emission levels one day may have very high
    32
    emission levels under the worst foreseeable conditions (for
    example, if an incinerator experiences a spike in sludge
    pollutant concentrations during certain times of the year).
    Recognizing that variability in the performance of sources can
    make identifying the best-performing sources based on short-
    term emissions data a nearly impossible task, we have upheld
    EPA’s estimation of MACT floors in at least one case. In
    Mossville Environmental Action Now v. EPA, 
    370 F.3d 1232
    (D.C. Cir. 2004), EPA explained that great variability in
    emissions among sources it sought to regulate made
    comparing sources and selecting the best-performing units
    virtually impossible, and so set the MACT standards for the
    pollutant at issue at the level of preexisting EPA emission
    standards. See 
    id. at 1240
     (“With comparisons between plants
    impossible, and emission variations not related to
    technological performance, the EPA claims it was unable to
    select the best [performing] sources.”). Because EPA pointed
    out that the source with the overall lowest long-term emission
    of the pollutant at issue barely satisfied the preexisting
    emission standards, we upheld EPA’s estimation that its
    preexisting emission standards for that pollutant reasonably
    represented the average emission levels of that pollutant for
    the best-performing units. 
    Id. at 1242
    .
    With this background in mind, we turn to the petitioners’
    challenges to the MACT floor methodology EPA used in
    setting emission standards for sewage sludge incinerators. As
    we explained, EPA’s approach to setting MACT floors had
    essentially two steps: (a) determining the best-performing
    sewage sludge incinerators and gathering data; and (b)
    applying the upper prediction limit to the collected dataset to
    account for variability.
    To determine the best-performing incinerators, EPA,
    mindful of our holding in past MACT floor cases, has devised
    33
    a different approach than it has in other rulemaking we have
    reviewed. First, EPA identified the incinerators it believed
    would have the lowest emissions based on the type of unit and
    installed air pollution controls. 75 Fed. Reg. at 63,270. EPA
    then conducted emission tests from these incinerators to
    develop its MACT floor dataset, which it supplemented with
    data from state environmental agency public databases. Id.
    Because some test runs yielded emissions data at a level that
    EPA’s testing equipment could not accurately measure, EPA
    developed a method for incorporating this non-detect data.
    Id. at 63,273.
    This method of using technology to set MACT floors
    differs from the “MACT approach” discussed in Cement Kiln
    and Brick MACT. In those cases, EPA had first identified the
    sources with the lowest emissions, then identified the primary
    emission control technology used by those sources, and then
    set the MACT floors based on sources that used that
    technology. See Cement Kiln, 
    255 F.3d at 859
    ; Brick MACT,
    
    479 F.3d at 879
    . In contrast, EPA has, in this rulemaking,
    selected which sewage sludge incinerators to survey based on
    their control technology, without first determining their
    emission levels relative to other sources. See Revised MACT
    Floor Memo at 6 (“To select the surveyed owners, EPA
    reviewed the inventory of [sewage sludge incinerators] for the
    control devices being operated, and identified a subset of units
    expected to have the lowest emissions based on the type of
    unit and the installed air pollution controls.” (emphasis
    added)).
    Even after selectively identifying and collecting data
    from incinerators, EPA did not collect data on every § 129
    pollutant from 12 percent of sources. For example, while EPA
    estimated in the final rule that it would need data on eighteen
    multiple hearth incinerators to meet the 12 percent
    34
    requirement, the number of multiple hearth incinerators for
    which it had data (including supplemental reports from state
    environmental agency databases) ranged from nineteen to five
    incinerators depending on the pollutant. See Revised MACT
    Floor Memo at 7. Believing that § 129(a)(2) required it to
    have data representative of at least 12 percent of incinerators,
    EPA devised a method of estimating whether a limited dataset
    could be representative of the best-performing 12 percent of
    units. Specifically, EPA applied a statistical analysis on the
    underrepresented pollutant datasets to estimate whether it had
    enough observations from testing incinerators to represent the
    best-performing 12 percent of incinerators. See 
    75 Fed. Reg. 63,270
    ; Revised MACT Floor Memo at 7–8. Applying this
    statistical analysis to its MACT floor dataset yielded an
    estimate of an estimate; in other words, the limited dataset to
    which EPA was applying this statistical analysis was itself
    already the result of EPA’s estimating the best performers
    based on control technology.
    After it had collected its dataset, EPA applied the upper
    prediction limit to estimate variability in sewage sludge
    incinerator emissions, stating its belief that the MACT floors
    had to be set at such a level that the best-performing
    incinerators “can expect to meet ‘every day and under all
    operating conditions.’” 75 Fed. Reg. at 63,269 (quoting
    Mossville Environmental Action Now, 
    370 F.3d at
    1241–42).
    We will address challenges to EPA’s estimations in the
    following order: (1) whether EPA may use control technology
    as a proxy for best-performing incinerators; (2) whether EPA
    did not adequately account for variability in the characteristics
    of sewage sludge fed into the sewage sludge incinerators, and
    whether it may account for variability with the upper
    prediction limit; (3) whether EPA may apply a statistical
    equation to determine whether EPA had a sufficient dataset to
    be representative of the best-performing 12 percent; and (4)
    35
    whether EPA may incorporate non-detect data by comparing
    calculated MACT floors to a value that is three times the
    representative detection level.
    1. Identifying the best-performing incinerators based on
    control technology
    We first address Sierra Club’s challenge to EPA’s
    selection of the best-performing units based on the type of
    unit and installed air pollution control technology. In
    addressing Sierra Club’s argument that EPA did not establish
    that non-technology factors do not affect emissions, we admit
    some confusion over whether NACWA is also arguing that
    EPA acted arbitrarily in its selection of best performers by
    failing to account for variability, or whether its argument is
    that EPA failed to account for the variability experienced by
    the best-performing units it selected. See NACWA Br. at 39
    (“Commenters argued that EPA’s targeted selection of nine
    [publicly-owned treatment works] based solely on type of
    pollution control makes it impossible for EPA to assume that
    the data are representative of the best-performing [sewage
    sludge incinerators] across the entire category.”) (citing
    NACWA Comments, Joint Appendix 24–25). But see
    NACWA Br. at 41 (“Because these data show the great
    variability of these pollutants, commenters urged EPA to
    determine the emission rates achieved by the best-performing
    sources under the full range of operating conditions.”). The
    closeness of these two arguments is hardly surprising given
    that variability in incinerator operating conditions may make
    the “best performing 12 percent of units” a moving target,
    particularly when EPA uses emission levels as the metric for
    “best performing.” But while both NACWA’s and Sierra
    Club’s arguments on this point share a similar element—
    EPA’s alleged failure to account for sewage sludge variability
    makes its MACT floor methodology arbitrary and
    36
    capricious—Sierra Club focuses on EPA’s assumption about
    control technology installed on the incinerator, while
    NACWA focuses on EPA’s assumptions about the effect of
    sewage sludge characteristics on emission levels. Because
    these contentions are different in kind, we will address them
    separately.
    We begin with Sierra Club’s allegation that EPA’s
    estimate of the best-performing 12 percent of units is
    unlawful and arbitrary. In arguing that this estimate is
    unlawful, Sierra Club relies on Cement Kiln, asserting that
    EPA’s estimation technique can be upheld only if air
    pollution control technologies are “the only factor
    determining emission levels.” 
    Id.
     (quoting 
    255 F.3d at 863
    ).
    Because EPA conceded that some non-technology factors
    affected emission levels, even accounting for Clean Water
    Act Part 503 regulations, see 75 Fed. Reg. at 63,270, Sierra
    Club argues that EPA’s method of estimating is per se
    unlawful.
    Sierra Club reads our holding in Cement Kiln too strictly.
    Later in our opinion in Cement Kiln, we explained that “if
    [EPA] can demonstrate with substantial evidence – not mere
    assertions – that MACT technology significantly controls
    emissions, or that factors other than the control have a
    negligible effect, the MACT approach could be a reasonable
    means of satisfying the statute’s requirements.” 
    255 F.3d at 866
     (emphases added).        Unsurprisingly, EPA cited this
    portion of our opinion when responding to Sierra Club’s
    challenge to using control technology as a proxy for the best-
    performing incinerators. EPA Br. at 55. Sierra Club asserts
    that EPA cannot justify its approach under this softer
    standard, arguing that the significant/negligible standard is
    dicta. See Sierra Club Reply Br. at 3–4.
    37
    While Sierra Club may be correct that this statement is
    dicta, we now elevate it to holding. First, our statement in
    Cement Kiln that the MACT approach would satisfy the
    statute “if pollution control technology were the only factor
    determining emission levels” was a direct quote from our
    opinion in National Lime, where the statement was itself
    dicta. 
    255 F.3d at 863
     (quoting National Lime, 
    233 F.3d at 633
    ); see National Lime, 
    233 F.3d at 633
     (summarizing an
    argument made by Sierra Club that we could not consider
    because it was not properly raised). Moreover, in Cement
    Kiln, we did not rely on the presence of any other factor
    influencing emission levels to hold that EPA’s MACT
    approach failed to satisfy the statutory requirements for
    setting MACT floors, instead resting our holding on the bases
    that several non-control factors could influence emission
    levels, and that EPA’s difficulty in quantifying these factors
    was no excuse for failing to demonstrate its estimate was
    reasonable. Cement Kiln, 
    255 F.3d at
    863–65.
    Second, if we are to give any substance to EPA’s ability
    to estimate the best performing units based on control
    technology, we must allow EPA to use air pollution control
    technology as a proxy for emission levels even if the
    correlation between control technology and emission levels is
    imperfect because non-control factors have a negligible effect
    on emissions. This is so because an estimate, by definition,
    will not accurately account for every variable that may affect
    emissions. WEBSTER’S, supra 17, at 778 (defining “estimate”
    as meaning “to judge the value, worth, or significance of: esp:
    to arrive at (a value judgment that is often valid but
    incomplete, approximate, or tentative)”) (emphasis added).
    Were we to adopt the strict standard Sierra Club argues
    applies, we would effectively be prohibiting EPA from using
    this estimation technique, as we find it impossible to imagine
    any situation where there is a perfect correlation between
    38
    control technology and emissions. See Cement Kiln, 
    255 F.3d at 871
     (“[W]e do not expect the impossible of the [EPA].
    Floors need not be perfect mirrors of the best performers’
    emissions.”). Thus, the fact that air pollution control
    technology is not the only factor affecting emission levels
    does not render EPA’s use of control technologies to identify
    best performers per se unlawful.
    EPA, however, must still demonstrate that its estimate is
    reasonable.    In justifying its approach of “specifically
    [seeking] emissions data from those municipalities that have
    installed and operate more than one of the controls that EPA
    identified as achieving the most reductions possible for the
    Section 129 pollutants,” EPA Br. at 56, EPA points to Clean
    Water Act regulations on sewage sludge disposal as
    accounting for non-control technology factors. See 40 C.F.R.
    Part 503, Subpart E. These regulations include a requirement
    that incinerators comply with Clean Air Act emission
    standards for beryllium and mercury, see 
    40 C.F.R. § 503.43
    (a)–(b), and set limits for the average daily
    concentrations of lead, cadmium, and other metals in the
    sewage sludge that is fed into an incinerator. See 
    40 C.F.R. § 503.43
    (c)–(d). EPA explained in the rulemaking, and now
    before us, that sewage sludge incinerators “receive a more
    homogenous type of waste to burn,” and that because of the
    Part 503 standards sewage sludge incinerators “are already
    incorporating management practices and measures to reduce
    waste and limit the concentration of pollutants in [sewage
    sludge].” See Revised MACT Floor Memo at 6–7; 75 Fed.
    Reg. at 63,270.
    It is true that EPA has pointed to some evidence of
    reduced variability among sewage sludge incinerators, a
    showing EPA failed to make in other MACT floor cases. See,
    e.g., Sierra Club, 
    167 F.3d at
    662–64 (concluding that
    39
    assumption that States would set emission limits at a level
    near what incinerators achieved in practice to be unsupported
    and contradicted by the record); Cement Kiln, 
    255 F.3d at 865
    (determining that EPA did not attempt to support its
    assumption that emission levels were solely dependent on
    control technologies, instead explaining that for some
    pollutants, the “factors other than technology that affect
    emissions . . . are difficult to quantify for the definition of
    MACT”) (internal quotation marks omitted). Nevertheless,
    EPA concedes “that there is some variation of Section 129
    pollutants present in the waste that is burned at individual
    sewage sludge incinerators.” EPA Br. at 58. The question
    thus remains: can EPA demonstrate with substantial evidence
    that some variation in pollutants has a negligible effect on
    sewage sludge incinerators, or that non-technology factors
    aside from sewage sludge content have a negligible effect on
    emissions?
    Sierra Club argues that EPA has not made that
    demonstration, pointing out that EPA’s assumption does not
    account for the fact that incinerator emissions are “affected by
    the fuels they use, the age and design of the individual unit,
    the specific quality and age of control devices at individual
    units, the training and skill of the operators, and the care with
    which they run individual units.” Sierra Club Br. at 18–19
    (internal quotation marks omitted). We found this argument
    persuasive in Cement Kiln, see 
    255 F.3d at
    862–65, and again
    agree with Sierra Club here. The fact that publicly-owned
    treatment works are already required to limit pollutant
    concentrations in sewage sludge before incineration does not
    establish that other variations in the operation of sewage
    sludge incinerators and the air pollution control technology
    they use will have only a negligible effect in emissions.
    Without evidence that air pollution control technology will
    achieve substantially the same performance across
    40
    incinerators without regard to the particular incinerator on
    which it is installed, EPA has made only a “mere assertion”
    that its regulations account for non-technology factors and
    that the type of air pollution control technology used by an
    incinerator is significantly determinative of emissions. See 76
    Fed. Reg. at 15,392; Cement Kiln, 
    255 F.3d at 866
    .
    Sierra Club also argues that variations in sewage sludge
    pollutant content are a factor affecting emission levels apart
    from the air control technology in use by a sewage sludge
    incinerator. Responding to EPA’s arguments that the Part 503
    regulations already require publicly-owned treatment works to
    “apply[] non-technology measures to reduce emissions” and
    therefore produce a more homogenous waste, Sierra Club
    argues that “[e]ven if correct, these arguments merely show
    that the units EPA selected as best performers have varying
    sludge inputs and that these inputs affect sewage sludge
    incinerators’ emissions less than they affect emissions from
    other categories of incinerators.” Sierra Club Br. at 19.
    Although this is a fair point, it is more relevant to NACWA’s
    argument that EPA failed to account for variability in sewage
    sludge pollutants when collecting its dataset, a point we
    address more comprehensively in the following section.
    Suffice it to say, unless EPA can demonstrate that the
    relatively reduced variations in sewage sludge characteristics
    have a negligible effect on emissions, Sierra Club’s argument
    that non-technology factors affect emissions seems even more
    meritorious. See Cement Kiln, 
    255 F.3d at
    864 (citing, in
    support of its holding, an EPA technical memorandum in
    which EPA observed that the variability in emissions among
    sources using MACT technology “indicate[d] that the air
    pollution control device system type . . . may not be the only
    important consideration affecting [dioxin/furan] control; other
    factors such as combustion quality and waste composition . . .
    41
    may also be of importance.”) (emphasis added) (second
    alteration in original). As we will discuss in the next section,
    EPA has alluded to evidence that may demonstrate a low
    correlation between sewage sludge characteristics and
    emissions.      See Joint Appendix 1095 (Letter Denying
    NACWA’s Petition for Reconsideration) (citing data gathered
    during EPA’s information collection request that showed that
    “the contents of . . . pollutants in the sludge itself has little
    relationship to the emissions of the pollutants, because these
    pollutants are removed by the control devices.”). But as we
    will also discuss in the following section, citing this evidence
    for the first time in denying a petition for reconsideration may
    be an insufficient basis upon which to defend EPA’s position
    that its Part 503 regulations account for non-technology
    factors.
    To bolster its argument that non-technology factors have
    a non-negligible effect on emission levels, Sierra Club points
    out that some incinerators EPA did not survey for data
    collection reported superior performance to those EPA did
    survey, undercutting EPA’s assumption that control
    technology is the only factor controlling emission levels. For
    example, Sierra Club cites comments submitted by Palo Alto,
    which reported that its program of reducing mercury inputs
    was effective in reducing mercury emissions to a level far
    below the existing unit MACT floor, and which was not
    included in the best-performing 12 percent. See Joint
    Appendix 628–29.         Sierra Club also notes that the
    supplemental data EPA took from state environmental agency
    public databases showed emission levels lower than those
    EPA decided to survey—a fact EPA candidly acknowledges.
    See 
    76 Fed. Reg. 15,387
     (“For some pollutants, the emissions
    from these supplemental test reports were lower than those
    from the nine [surveyed] sources.”). This is evidence, Sierra
    42
    Club argues, that EPA’s estimate of the best performers based
    on control technology is arbitrary and capricious.
    EPA responds that it did not have to consider the Palo
    Alto incinerator’s reported emission level, because Palo Alto
    failed to comply with EPA’s instruction in the proposed rule
    to provide “supporting documentation” when submitting
    additional emissions stack test data in its comments. See 75
    Fed. Reg. at 63,270. Because EPA could not verify the
    accuracy of Palo Alto’s asserted emission levels, EPA
    explains that it was unable to draw any correlation between
    Palo Alto’s sewage sludge management practices and a
    reduction in mercury emissions. In the past, we have upheld
    EPA’s rejection of data it determined deficient, explaining
    that we give substantial deference to an agency’s expert
    scientific judgment. Similarly, when EPA requests that
    commenters substantiate their reported emission test data, and
    they do not comply, EPA “[is] not obligated under its policy”
    to compare its collected results with the unsubstantiated data
    that commenters submit. See Edison Electric Institute v. EPA,
    
    2 F.3d 438
    , 448 (D.C. Cir. 1993) 
    Id. at 449
    . Thus, we agree
    with EPA that it deserves deference for its decision not to
    draw any correlations between Palo Alto’s change in
    management practice and a reduction in mercury emissions
    based on the data submitted.
    EPA, however, does not respond to Sierra Club’s
    argument that the fact that randomly selected incinerators
    from state environmental agency databases had emission
    levels lower than those from the incinerators EPA chose to
    survey based on technology demonstrates that EPA’s use of
    control technology is unreasonable. Instead, it explains that
    the supplemental information from state environmental
    agency databases “included emissions test data from
    facilities/units that met the same criteria EPA used in issuing
    43
    information collection requests, i.e., units with more than one
    of the controls that EPA identified as achieving the most
    reductions possible for the Section 129 pollutants.” EPA Br.
    at 56–57; compare Joint Appendix 979 (listing incinerators
    from which EPA collected emissions test data and the control
    technologies they used), with Joint Appendix 1012–1040
    (listing incinerator emission test data collected from state
    environmental agency databases along with control
    technologies used). But EPA did not state this rationale in the
    rulemaking, and we cannot “accept appellate counsel’s post
    hoc rationalizations for agency action.” Motor Vehicles
    Manufacturer’s Ass’n v. State Farm Mutual Auto Insurance
    Co., 
    463 U.S. 29
    , 50 (1983). And before this court, EPA does
    not claim that it specifically searched state environmental
    agency databases for data from sewage sludge incinerators
    that use what it considered the best-performing technology,
    explaining only that the supplemental data “included”
    emissions data from facilities that had one or more of the
    controls identified. Even taking as true that the incinerators
    for which EPA had data in supplemental testing used similar
    air pollution controls as the incinerators it selected, the fact
    that similarly-controlled incinerators achieve lower emission
    levels suggests that non-technology factors have a non-
    negligible effect on emission levels.
    Therefore, we agree with Sierra Club that EPA has not
    demonstrated with substantial evidence that non-control
    technology factors apart from sewage sludge content, like
    variations in age, design, or operation of the incinerators
    themselves, would have a negligible effect on incinerator
    emissions. That EPA’s supplemental test reports show
    superior performance to the incinerators EPA chose for
    information collection requests is strong evidence that the
    type of air pollution control technology used itself is not
    significantly determinative of emissions.      Nevertheless,
    44
    because EPA may be able to explain why non-control
    technology factors have a negligible effect on emissions and
    why the data it used from supplemental test reports
    outperform the units it identified as best, we remand this
    portion of the rulemaking to EPA for further explanation
    without vacating the MACT floor regulations.
    2. Accounting for variability
    We now address NACWA’s and Sierra Club’s challenges
    to EPA’s method of accounting for variability. In the
    proposed rule, EPA explained that it was accounting for intra-
    unit variability using a statistical tool it termed the upper
    prediction limit, and was relying on the Part 503 regulations
    to account for variability in sewage sludge pollutant
    concentration between incinerators. See 75 Fed. Reg. at
    63,270–71. Because the upper prediction limit is not a
    straightforward method of accounting for variability, we first
    review and elaborate on this concept. The 99 percent upper
    prediction limit “represents the value which one can expect
    the mean of future 3-run performance tests from the best-
    performing 12 percent of sources to fall below, with 99
    percent confidence, based upon the results of the independent
    sample of observations from the best-performing sources.”
    75 Fed. Reg. at 63,271. EPA calculated the upper prediction
    limit using the following formula:
    1    1
    ̅       0.99,   1
    In this formula, n represents the number of test runs (i.e., the
    sample size), m represents the number of test runs in a
    compliance average, ̅ represents the mean, s represents the
    45
    standard deviation (i.e., a value representing how much
    variation exists from the mean within a dataset), and t(0.99,
    n–1) represents a value called the t-statistic, which is a
    number based on the number of test runs and EPA’s desired
    99 percent significance. Id. Because of the role of the
    standard deviation in this formula, greater variation within a
    dataset will produce a higher upper prediction limit. And
    because EPA incorporated the upper prediction limit into its
    analysis by setting the MACT floor at the level of the upper
    prediction limit (unless its methodology for addressing non-
    detect data, which we will discuss below, yielded a higher
    result), a higher upper prediction limit means a higher MACT
    floor.
    For existing incinerators, EPA did not apply the upper
    prediction limit to each 3-run test for each incinerator, but
    instead applied it to the entire dataset it collected for a
    pollutant. See id.; see, e.g., Revised MACT Floor Memo
    Attachment B-8. For example, when setting the MACT floor
    for sulfur dioxide, for which EPA had 63 test runs, EPA
    calculated the upper prediction limit using a sample size of
    63, a mean and standard deviation based on all 63 test runs, a
    desired 3-run compliance average, and a desired confidence
    level of 99 percent. See 75 Fed. Reg. at 63,271; Revised
    MACT Floor Memo Attachment B-8. The resulting value
    calculated by the upper prediction limit appears to represent,
    with 99 percent confidence, the value EPA predicts a 3-run
    average from an imaginary incinerator—representative of the
    mean and variation among the best-performing 12 percent
    incinerators—could achieve. Predicting the value that a “3-
    run average” will fall below appears confusing at first
    glance—but we note that the focus of the upper prediction
    limit is on the average, and not the values of the three runs
    used to derive that future average. Thus if the upper
    prediction limit (i.e., MACT floor) were 5, it would not matter
    46
    whether the 3-run test was 5-5-5, or 1-5-9, as both 3-run tests
    average to 5. Designing the upper prediction limit this way
    appears to give incinerators testing for compliance with the
    MACT standards some leeway in variations among emissions
    between test runs. See 75 Fed. Reg. at 63,269.
    During the comment period Sierra Club, but not
    NACWA, challenged EPA’s use of the upper prediction limit.
    See Joint Appendix 608. Although NACWA did not criticize
    the upper prediction limit as a method for addressing
    variability in its comments, it did assert that EPA failed to
    account for variability in sewage sludge characteristics
    resulting from regional and seasonal variability. See id. 670–
    72. NACWA also challenged EPA’s use of the Clean Water
    Act Part 503 regulatory data in its rulemaking, asserting that
    EPA’s assumption that sewage sludge was homogenous was
    demonstrably false based on that data.           For example,
    NACWA pointed out that the monthly average lead
    concentration in sewage sludge in January 2009 was 62.23
    mg/dry kg for southerly incinerators and 97.00 mg/dry kg for
    westerly incinerators, while the monthly average lead
    concentration in July 2009 was 123.14 mg/dry kg for
    southerly incinerators and 218.55 mg/dry kg for westerly
    incinerators. Id. 671. NACWA argued that because EPA
    regulations required publicly-owned treatment works to report
    this data to EPA, see 
    40 C.F.R. § 503.48
    , EPA had access to
    the data and should have used it during rulemaking to account
    for variability. 
    Id.
     671–72. Moreover, NACWA noted that
    the Part 503 regulations did not even address all § 129
    pollutants. See id. 672.
    In the final rule, EPA defended its use of the upper
    prediction limit against Sierra Club’s criticisms. See 76 Fed.
    Reg. at 15,389. In response to NACWA’s comments, EPA
    explained that it collected emissions data from nine different
    47
    facilities in nine different states, claiming that these facilities
    were “located in a mix of northern, southern, eastern, and
    western states,” each with its “own unique sludge
    characteristics from different residential and commercial
    populations.” Id. at 15,391. EPA stated that it felt the dataset
    had “sufficient variation in regions, climates and populations”
    to “adequately incorporate[] variability in wastewater
    treatment systems across the U.S.,” and that it had “also
    incorporated variability using the [upper prediction limit].”
    Id.
    Both petitioners now challenge EPA’s use of the upper
    prediction limit to account for variability among sewage
    sludge incinerators. Sierra Club contends that the upper
    prediction limit is unlawful as applied to existing incinerator
    datasets because it does not represent the “average” emissions
    limitation achieved by the best-performing 12 percent of
    incinerators, a fact inconsistent with EPA’s statements that
    the upper prediction limit is “based on” an average. Sierra
    Club asserts that the upper prediction limit is unlawful as
    applied to new sources as well, stating that “the upper
    prediction limit is not an estimate of what the best unit
    actually ‘achieved in practice.’” Sierra Club Br. at 25.
    Sierra Club also argues that EPA’s use of the upper
    prediction limit is arbitrary and capricious because EPA does
    not provide an explanation of how the upper prediction limit
    represents the emissions level achieved by the best
    performing units under the worst reasonably foreseeable
    conditions. To support its arbitrary and capricious argument,
    Sierra Club points to the fact that, by applying the upper
    prediction limit, EPA calculated a MACT floor for the best-
    controlled unit that was higher than the floor based on the
    average emissions limitations achieved by the top 12 percent
    of units. See 76 Fed. Reg. at 15,388–89. Sierra Club also
    48
    alleges that EPA has been inconsistent with its use of the
    upper prediction limit, using the upper prediction limit to set
    the MACT floors but not using it to identify incinerators in
    the first place.
    NACWA does not challenge EPA’s use of the upper
    prediction limit as broadly as Sierra Club, explaining that
    “using the 99 [percent] [upper prediction limit] method to
    account for variability in emission performance is not
    prohibited by statute, nor is it unreasonable, provided EPA
    uses representative data from the congressionally required 12
    [percent] of units.” NACWA Intervenor Br. at 7. Instead,
    NACWA challenges EPA’s justification that the upper
    prediction limit adequately accounts for variability in a
    dataset that itself is not reflective of the variety of conditions
    in which sewage sludge incinerators operate, asserting that
    “the [upper prediction limit] cannot account for variability
    among [sewage sludge incinerators] unless the underlying
    data are representative of the category as a whole.” In other
    words, NACWA argues that because EPA failed to take into
    account regional and seasonal variability in sewage sludge
    pollutant concentrations and failed to use the Part 503 data
    showing variability in sewage sludge characteristics in its
    analysis, EPA produced a limited dataset that was not
    representative of the “emission rates achieved by the best-
    performing sources under the full range of operating
    conditions.” NACWA Br. at 41.
    EPA’s responses to both Sierra Club’s and NACWA’s
    arguments are somewhat conclusory, relying primarily on its
    explanation in the Revised MACT Floor Memo rather than
    addressing the petitioners’ legal arguments. Responding to
    Sierra Club, EPA asserts that it did not “simply ‘pick a
    number’” in setting the upper prediction limit, and explains
    that the reason why some new incinerator MACT floors were
    49
    higher than existing incinerator MACT floors was because of
    the smaller datasets with greater variation. EPA Br. at 52.
    EPA argues that NACWA’s criticism of the upper prediction
    limit is “overly simplistic,” explaining that “EPA’s use of the
    [upper prediction limit] to account for variability also
    addressed any emissions variability due to differences in
    sludge content,” and that if NACWA’s approach were
    accepted, “it would account for variability in sludge content
    on top of the variability in emission levels that are already
    accounted for through the [upper prediction limit].” EPA Br.
    at 50–51.
    Before stepping into the morass of arguments on the
    upper prediction limit, we take a moment to revisit the
    statutory source of EPA’s obligation to set MACT floors.
    Section 129(a)(2) requires existing MACT floors to be no less
    stringent than “the average emissions limitation achieved by
    the best performing 12 percent of units.” As we noted in
    Sierra Club, “this phrase on its own says nothing about how
    the performance of the best units is to be calculated.” 167
    F.3d at 661. In the past, we have primarily relied on Sierra
    Club when reviewing MACT standards to state that EPA can
    reasonably estimate the performance of the best units. But
    underlying Sierra Club’s holding is also the proposition that,
    because EPA can interpret ambiguous statutes under Chevron,
    EPA can decide what value the MACT floors are supposed to
    represent, as long as that decision is a reasonable
    interpretation of the statute. See id. at 661–62.
    In this case, EPA has explained that its “long-standing
    interpretation [of section 129] is that the combination of
    section 129(a)(4), requiring numerical standards for each
    enumerated pollutant, and section 129(a)(2), requiring that
    each such standard be at least as stringent as the MACT floor,
    supports that floors be derived for each pollutant based on the
    50
    emissions levels achieved for each pollutant.” 75 Fed. Reg. at
    63,269. And in the final rule, EPA clarified that it was “using
    lowest emissions limitation as the measure of best
    performance.” 76 Fed. Reg. at 15,389.
    But even with these explicit interpretations of § 129(a)(2)
    as guidelines, the phrase “average emissions limitation
    achieved by the best performing 12 percent of units” could be
    interpreted several different ways, with several different
    variations of what the MACT floor is supposed to represent.
    For example, and without prejudging the legality of these
    different interpretations, EPA could interpret the “average
    emissions limitation achieved by the best performing 12
    percent of units” to mean the average of the emission levels
    achieved by the best-performing units based on EPA’s short-
    term emission test data, without regard to the range of
    conditions under which the incinerators operate. In that case,
    the MACT floors would simply be an average of the
    emissions data EPA collected from the best-performing units.
    The phrase could also mean the average of the emission
    levels that each best-performing unit achieved under the worst
    foreseeable circumstances. If EPA were to take this
    interpretation, it would seem sensible to determine the
    population of the best-performing units based on which units
    achieve the lowest emission levels under the worst
    foreseeable conditions, and then average those emission
    levels.
    Based on EPA’s method of determining MACT floors in
    this case and its response to Sierra Club’s comments,
    however, it seems EPA has adopted yet another interpretation
    of the phrase “average emissions limitation achieved by the
    best performing 12 percent of units.” In the final rule, EPA
    cited Sierra Club’s comments that § 129(a)(2)
    51
    “unambiguously requires EPA to set floors reflecting the
    ‘average’ emission level achieved by the best sources” and
    that “although EPA may consider variability in estimating an
    individual source’s actual performance over time, nothing in
    the [Clean Air Act] or the case law even suggests that EPA
    may account for differences in performance between sources
    except as § 129 provides, by averaging the emission levels
    achieved by the sources in the top 12 percent.” 76 Fed. Reg.
    at 15,389. EPA responded that “[b]ecause the [upper
    prediction limit] represents the value which we can expect the
    mean (i.e., average) of three future observations (3-run
    average) to fall below, based upon the results of the
    independent sample size from the same population, the [upper
    prediction limit] reflects average emissions.” Id.
    It is not clear to us, however, that the “average emissions
    limitation achieved by the best performing 12 percent” would
    refer to the future average of a 3-run test that EPA predicts a
    source in the best-performing 12 percent will fall below with
    99 percent confidence. Instead, the word “average” as
    referred to in the standard for existing unit MACT floors
    seems to mean the average emissions limitation that the
    existing population of the best-performing 12 percent of
    incinerators has achieved, not the average of a future 3-run
    test conducted for compliance purposes.
    This is not to say that the upper prediction limit, which
    EPA applied to the average of the emission levels recorded
    while testing the best-performing 12 percent, would violate
    the statutory standard established in § 129. Under its method,
    EPA seems to have effectively modeled an imaginary
    incinerator based on the short-term emissions test data from
    incinerators in what EPA considers the best-performing 12
    percent of units. Using the average, the standard deviation of
    the dataset, and the number of data points, EPA claims its
    52
    statistical formula models, with 99 percent confidence, the
    upper limit of what that imaginary incinerator would achieve
    based on the distribution of the dataset. Given that EPA
    believes that it must set MACT floors “that the best
    performing sources can meet ‘every day and under all
    operating conditions,’” it seems plausible to state that this
    predicted emission level represents the “average emissions
    limitation achieved by the best performing 12 percent of
    units.” See 75 Fed. Reg. at 63,269 (quoting Mossville
    Environmental Action Now, 
    370 F.3d at
    1241–42).
    EPA has not interpreted “average emissions limitation
    achieved” this way, however, and because we “may not
    supply a reasoned basis for the agency’s action that the
    agency itself has not given,” we can only adjudge EPA’s
    interpretation based on essentially one sentence from the
    Federal Register. Bowman Transportation, Inc. v. Arkansas-
    Best Freight System, Inc., 
    419 U.S. 281
    , 285–86 (1974); see
    76 Fed. Reg. at 15,389. The need for further explanation is
    especially acute when EPA’s approach of using the upper
    prediction limit and its interpretation of “average emissions
    limitation achieved” are both departures from the approaches
    EPA has taken in setting MACT floors in earlier cases. As to
    the interpretation of “average” specifically, EPA has not
    previously interpreted the phrase “average emissions
    limitation achieved” to refer to the average of a future 3-run
    test, but instead the average emissions levels of the best-
    performing 12 percent of sources for which EPA had data (or
    the average of the proxies EPA used to estimate those
    emission levels). See Sierra Club, 
    167 F.3d at 661
     (“[EPA]
    selected the 12 percent of the incinerator population subject to
    the strictest controls and set the floor level for the subcategory
    by averaging the emissions limitations governing those
    incinerators.”); National Lime, 
    233 F.3d at 630
     (“To set
    existing source emission floors, EPA . . . . identified the
    53
    technology used by the median plant out of the best twelve
    percent of plants for which it had information and set the
    existing source emission floor at the emission level of the
    worst performing plant in its database using that
    technology.”); Cement Kiln, 
    255 F.3d at 859
     (explaining that
    for existing sources, EPA identified the best-performing 12
    percent of sources, then identified the emission control
    technology used by sources with emission levels equivalent to
    or lower than the median of the best-performing 12 percent,
    and then set the MACT floor at the worst emission level
    achieved by any source using that technology); Northeast
    Maryland Waste Disposal Authority, 
    358 F.3d at 953
     (“For
    each pollutant, EPA calculated the MACT floor by averaging
    the most stringent 12% of state permit limits in each class.”).
    Although EPA may be able to justify its novel
    interpretation that “average” means the average of a future 3-
    run compliance test, one sentence in the Federal Register is
    not enough of a basis to uphold EPA’s new approach to
    incorporating variability against arbitrary and capricious
    review. Accordingly, on remand, we expect EPA to clarify
    how the upper prediction limit represents the “average
    emissions limitation achieved by the best performing 12
    percent.” See Sierra Club, 
    167 F.3d at 664
     (remanding EPA’s
    MACT floor determination, because “[i]t is possible that EPA
    may be able to explain it”).
    We now turn to Sierra Club’s related challenge to the
    upper prediction limit as arbitrary because “EPA provides no
    support for the notion that the upper prediction limit for the
    single best-performing unit reflects that unit’s actual
    performance, even under the ‘worst reasonably foreseeable’
    circumstances.’” Sierra Club Br. at 26. While it is true that
    EPA did not even use the phrase “worse foreseeable
    circumstances” in its rulemaking, it did explain its belief that
    54
    it must set MACT floors that the “best performing sources can
    expect to meet ‘every day and under all operating
    conditions.’” 75 Fed. Reg. at 63,269 (quoting Mossville
    Environmental Action Now, 
    370 F.3d at
    1241–42). Not only
    are these standards approximately equivalent, EPA is not
    wedded to our “worse foreseeable circumstances”
    interpretation of “achieved in practice.” See Sierra Club, 
    167 F.3d at 665
     (explaining that “EPA would be justified in
    setting the floors at a level that is a reasonable estimate of the
    performance of the ‘best controlled similar unit’ under the
    worst reasonably foreseeable circumstances,” adding “we use
    the subjunctive because it is not clear from the record that the
    agency was doing this”). More substantively, EPA’s citation
    to Mossville Environmental Action Now in the same section as
    EPA’s description of the upper prediction limit may be
    enough to reasonably discern EPA’s justification for the upper
    prediction limit, even if EPA has not directly stated how it
    justifies the upper prediction limit as a method for accounting
    for variability in light of the Clean Air Act and our case law.
    See Bowman Transportation, 419 U.S. at 286.
    Although it may be sufficiently clear that EPA’s
    prediction of the best-performing incinerators’ upper limit
    represents standards these incinerators can “meet every day
    and under all operating conditions,” EPA has not clearly
    explained how the upper prediction limit itself operates to
    predict this value with sufficient accuracy. In the Brick
    MACT case, we held that EPA’s use of the MACT approach
    violated the Clean Air Act because “it . . . failed to show that
    the emission levels achieved by the worst performers using a
    given pollution control device actually predict the range of
    emission levels achieved by the best performers using that
    device.” 
    479 F.3d at 882
    . Similarly, EPA provides little
    explanation on how the upper prediction limit can actually
    predict the upper limit EPA expects the best-performing unit
    55
    or units to achieve. As Sierra Club pointed out, the upper
    prediction limit’s predictive ability does appear somewhat
    doubtful when it produces a higher MACT floor for an
    incinerator with raw test data ranging from 0.31 to 2.26 than
    for a group of incinerators with raw test data ranging from
    0.31 to 40.32 and a mean of 9.38. See Revised MACT Floor
    Memo Attachment B-8 (Sulfur dioxide emissions test data for
    existing multiple hearth incinerator floor); 
    id.
     Attachment D-8
    (Sulfur dioxide emissions test data for new multiple hearth
    incinerator floor).      As EPA stated in explaining this
    apparently illogical result, a smaller dataset may have greater
    variability, and thus a higher upper prediction limit. But if the
    upper prediction limit can vary so much depending on the size
    of the dataset, EPA should explain on remand why the upper
    prediction limit is a reasonable estimate of what an incinerator
    would achieve under the worst foreseeable conditions for
    incinerators with smaller datasets.         Put differently, if
    collecting more data has such a significant effect on the upper
    prediction limit, presumably producing a more accurate
    estimate of what that incinerator would “achieve in practice,”
    EPA should explain why the upper prediction limit could still
    be considered accurate given a small dataset.
    While it is true that we “owe particular deference to EPA
    when its rulemakings rest upon matters of scientific and
    statistical judgment within the agency’s sphere of special
    competence and statutory jurisdiction,” American Coke &
    Chemicals Institute v. EPA, 
    452 F.3d 930
    , 941 (D.C. Cir.
    2006), EPA must still articulate a “rational connection
    between the facts found and the choice made.” Burlington
    Truck Lines, Inc. v. United States, 
    371 U.S. 156
    , 168 (1962).
    Because we are already remanding the upper prediction limit,
    we encourage EPA to elaborate how the statistical formula it
    uses can predict the upper limit of incinerator emissions. We
    are hesitant to rubber-stamp EPA’s invocation of statistics
    56
    without some explanation of the underlying principles or
    reasons why its formulas would produce an accurate result,
    particularly when the “facts found”—the MACT floor
    datasets—demonstrate flaws in the formula.
    We now turn to NACWA’s arguments that EPA failed to
    account for variability in sewage sludge characteristics,
    beginning with NACWA’s argument that the upper prediction
    limit cannot account for this sort of variability. See NACWA
    Br. at 39; EPA Br. at 51 (“EPA’s use of the [upper prediction
    limit] to account for variability also addressed any emissions
    variability due to differences in sludge content.”). Before
    discussing NACWA’s challenges to EPA’s method of
    accounting for variability, however, we take a moment to note
    a distinction NACWA appears to draw in its brief, in which it
    argues that its analyses on the Part 503 data “demonstrate that
    there is significant variability in metals concentrations among
    [publicly-owned treatment works] and within a [publicly
    owned treatment work].” NACWA Br. at 42. As we will
    explain below, we agree with NACWA that under the
    rationale EPA expressed during rulemaking regarding
    variability, EPA should have accounted for variability in
    sewage sludge characteristics within a publicly-owned
    treatment work or better explained why that variability was
    irrelevant. It is unclear, however, whether EPA needed to
    account for variability among publicly-owned treatment
    works.
    First, we agree with NACWA that EPA has not
    adequately explained how the upper prediction limit can
    address variability in sewage sludge characteristics,
    particularly given the context in which it described the upper
    prediction limit in the proposed rule and the Revised MACT
    Floor Memo. EPA explained in the proposed rule and in its
    Revised MACT Floor Memo that “[t]he types of variability
    57
    that EPA attempts to account for include operational
    distinctions between and within tests at the same unit.” 75
    Fed. Reg. at 63,269; Revised MACT Floor Memo at 4–5. As
    to existing incinerator MACT floors, EPA stated that “[b]y
    including multiple emissions tests from units with a test
    average in the top 12 percent, EPA can evaluate intra-unit
    variability of emissions tests over time, considering variability
    in control device performance, unit operations, and fuels fired
    during the test. . . . [T]he [upper prediction limit] was used for
    the [sewage sludge incinerator] MACT floor variability
    analysis.” 75 Fed. Reg. at 63,271; Revised MACT Floor
    Memo at 11–12. These statements appear to contradict EPA’s
    implied position in the final rule (and express position on
    petition for review) that the upper prediction limit can account
    for more than intra-unit variability. See 76 Fed. Reg. at
    15,391 (explaining, in response to NACWA’s comments that
    EPA failed to account for variability in sewage sludge
    characteristics, that in addition to collecting emissions
    information from facilities in nine different states, “[w]e have
    also incorporated variability using the [upper prediction
    limit].”); EPA Br. at 51.
    Not only is EPA contradictory on whether the upper
    prediction limit accounts for more than intra-unit variability,
    NACWA’s argument that the underlying dataset must already
    be representative of variability in sewage sludge
    characteristics before applying the upper prediction limit is
    persuasive, at least as applied to variability within a publicly-
    owned treatment work. Assuming that NACWA’s summary
    of monthly average sewage sludge pollutant concentrations in
    its comments is accurate, short-term emissions testing done in
    January when the monthly average lead concentration is
    lowest may not be representative of incinerators’ performance
    in July, when the monthly average lead concentration is
    almost double.        See Joint Appendix 671 (NACWA
    58
    Comments). If sewage sludge incinerator emissions increase
    based on sewage sludge pollutant concentrations, then it
    would seem to follow that the average emission levels of
    sewage sludge incinerators in July would be higher than the
    average in January. Cf. Sierra Club, 
    167 F.3d at 666
     (“The
    EPA does not deny that the waste stream reductions the Sierra
    Club calls for would reduce pollution. The less mercury in,
    the less mercury out, and the less chlorinated plastic in, the
    less HCl out.”). And assuming that the standard deviation
    does not change from month-to-month, it would also follow
    that the upper prediction limit, and thus the MACT floor,
    would be higher for a dataset based on July emission testing.
    As with the other aspects of the upper prediction limit,
    however, EPA may be able to explain and clarify on remand
    its position on whether the upper prediction limit can account
    for variability in sewage sludge content. EPA may also be
    able to explain why NACWA is incorrect in asserting that
    EPA needs to base its upper prediction limit on a
    representative dataset. Alternatively, EPA could adopt an
    interpretation of “average emissions limitation achieved by
    the best performing 12 percent of units” that does not require
    EPA to determine what the best-performing units achieve
    under the worst foreseeable conditions.
    Having determined that EPA did not make clear whether
    the upper prediction limit accounts only for intra-unit
    variability, we now turn to NACWA’s remaining and related
    arguments on EPA’s variability analysis. We start first with
    NACWA’s argument that EPA failed to demonstrate that its
    MACT floor dataset represented the best-performing 12
    percent of incinerators because it failed to adequately account
    for variability in sewage sludge characteristics. NACWA
    maintains that sewage sludge characteristics can vary not only
    by geographic region, but also by seasons, differences in
    59
    wastewater treatment technologies, and the sanitary wastes
    received from the communities served by publicly-owned
    treatment works. See Joint Appendix 671 (summarizing
    month-to-month variability in sewage sludge pollutant
    concentrations based on Part 503 data); id. at 1075
    (summarizing the minimum, maximum, median, and 25th and
    75th percentile sewage sludge pollutant concentration for
    several different publicly-owned treatment works operating
    sewage sludge incinerators). Moreover, NACWA criticizes
    EPA for not factoring sewage sludge variability into its
    MACT floor methodology, noting that EPA has had twenty
    years of data on sewage sludge from the Part 503 regulations,
    which require regulated entities to report information about
    pollutant concentrations to EPA. See 
    40 C.F.R. § 503.48
    .
    NACWA argues that this limited testing is problematic,
    implying that if EPA used a limited dataset that is not
    representative of the full range of conditions experienced by
    sewage sludge incinerators, and set MACT floors on that
    dataset without accounting for the significant variation in
    sewage sludge, it could underestimate the emissions limitation
    achieved by the best-performing incinerators. For example, if
    the concentration of a pollutant in sewage sludge is lower in
    the winter but higher in the summer, basing MACT floors
    only on a dataset of emission tests taken during the winter
    would underestimate the emissions limitation achieved under
    the worst foreseeable conditions. Relatedly, if sewage sludge
    pollutant concentrations are naturally higher at a sewage
    sludge incinerator in one region versus another incinerator in
    a different region, NACWA appears to assert that EPA must
    account for these variations in its dataset.
    EPA, in addition to responding that the upper prediction
    limit accounts for variability, asserts that NACWA “fails to
    explain how the alleged variations in the metals content in
    60
    sludge affects the emissions performance of the best
    performing units.” Instead, EPA contends, “NACWA simply
    points to data submitted under the [Clean Water Act] Part 503
    regulations without demonstrating whether and how the
    pollutant content of sludge affects emissions.” EPA Br. at 51.
    EPA further explains that “air pollution control devices
    generally used by the best performing sewage sludge
    incinerators result in significant pollutant reductions between
    the incoming sludge feed and outgoing emissions,” and thus
    that “any differences in pollutant concentrations in the sludge
    feed should have minimal impact on emissions because the
    pollutants contained in the sludge feed itself are removed by
    air pollution control devices.” 
    Id.
    EPA did not cite to the record in its brief for these factual
    assertions. Nor are these assertions in the final rule’s section
    responding to NACWA’s comments on this point, in which
    EPA responded not that there was a low correlation between
    emissions and sewage sludge pollutant concentrations, but
    instead that “[w]e requested additional information in the
    [notice of proposed rulemaking], but did not receive adequate
    sampling data from the best-performing sources.” 76 Fed.
    Reg. at 15,391. Upon review of the record, the only place we
    could find support for EPA’s factual assertion was in its letter
    denying NACWA’s petition for reconsideration. There, EPA
    explained that it had collected sludge content data at the same
    time as the emissions tests, and that because it found a “high
    reduction in pollutant levels between incoming sludge and
    emissions due to add-on controls, the variation in the lead
    content in the sludge . . . did not affect the emissions
    performance of those sources.” Joint Appendix 1095 (Letter
    Denying NACWA’s Petition for Reconsideration).
    But even assuming that EPA intended to rely on this
    assertion in defending its rulemaking on the current petition
    61
    for review, it is not clear to us that a rationale offered for the
    first time in a petition for reconsideration is sufficient to be a
    ground upon which we can judge the propriety of EPA’s
    action. We could find no case discussing the propriety of
    judging an agency’s action based on a statement made for the
    first time in a denial of a petition for reconsideration, but it
    seems to be a weak basis for upholding agency action. This is
    particularly true when a petitioner has raised an objection
    during the comment period that gave an agency the
    opportunity to respond to the objection before the denial of
    reconsideration. See 76 Fed. Reg. at 15,391 (summarizing
    several statements from commenters, including that
    “emissions from [sewage sludge incinerators] are affected not
    just by control technology but also by other factors including
    the contents of the sludge that a unit is burning,” and that “the
    proposed standards does [sic] not take into account that
    [mercury, cadmium, lead, hydrogen chloride and sulfur
    dioxide] emissions are a function of the sludge content of
    [mercury, cadmium, lead], chlorine, and sulfur”). A purpose
    of notice-and-comment provisions under the APA (and
    presumably of the more elaborate procedural safeguards in §
    307 of the Clean Air Act) is “to ensure that affected parties
    have an opportunity to participate in and influence agency
    decision making at an early stage, when the agency is likely to
    give real consideration to alternative ideas.” New Jersey,
    Department of Environmental Protection v. EPA, 
    626 F.2d 1038
    , 1049 (D.C. Cir. 1980). By waiting until the petition for
    reconsideration to respond to a comment that had been raised
    during the comment period, EPA deprives the affected party
    of the opportunity to respond to EPA’s rationale and influence
    agency action at an earlier stage. Thus, just as we will not
    entertain an argument raised for the first time in a reply brief
    to prevent sandbagging of appellees and respondents, we are
    reluctant to affirm based on a factual assertion raised for the
    first time in an agency’s denial of a petition for
    62
    reconsideration when the agency had an opportunity to raise
    that point at an earlier point in the rulemaking process. See
    Mohamad v. Rajoub, 
    634 F.3d 604
    , 608 (D.C. Cir. 2011). On
    remand, however, EPA may elaborate upon and explain the
    data that led it to conclude a low correlation between sewage
    sludge concentrations and emissions, using that conclusion to
    support the reasonableness of its estimate if it finds that
    approach appropriate.
    Because EPA did not provide evidence during
    rulemaking that there was a low correlation between sewage
    sludge pollutants and actual emissions, we address EPA’s
    argument that NACWA did not establish a correlation
    between sewage sludge pollutant contents and emissions.
    EPA is mistaken in putting the burden of establishing this
    correlation on NACWA. While EPA could arguably have
    interpreted § 129(a)(2) in a way that does not require it to
    account for variability, in which case NACWA would need to
    argue why EPA’s interpretation is not “based on a permissible
    construction of the statute,” Chevron, 
    467 U.S. at 843
    ,
    having decided to account for variability, and having decided
    to estimate that variability, EPA bears the burden of
    demonstrating with substantial evidence that its estimate is
    reasonable.     See Northeast Maryland Waste Disposal
    Authority, 
    358 F.3d at 954
    . If EPA is arguing its estimated
    MACT floors are reasonable based on the assumption that its
    limited dataset allows it to account for the worst foreseeable
    conditions because sewage sludge variability will have a
    negligible effect on emissions, EPA, and not NACWA, bears
    the burden of demonstrating that this assumption is correct.
    EPA’s related argument that NACWA “failed to
    demonstrate why data from more units across more states
    during different times of the year would have led to a better
    determination of the best performers, why the representative
    63
    data that EPA did collect prevented EPA from reasonably
    determining the best performers, or how consideration of
    [Clean Water Act] Part 503 data or stack test data would have
    changed the results” suffers from similar flaws. EPA Br. at
    52. Although EPA did not have to consider emissions stack
    test data that commenters submitted without appropriate
    documentation, see supra at 42, we disagree with EPA that
    NACWA has the burden of showing why more data would
    better determine best performers. It seems self-evident that
    more data from a broader span of time would have helped
    support (or defeat) EPA’s assumptions about the extent to
    which sewage sludge variability affects emission levels, and if
    EPA wanted to justify a limited dataset from a smaller
    timespan as being representative, it, and not NACWA, bears
    the burden of making that demonstration. Moreover, EPA’s
    argument that NACWA did not demonstrate how more data
    would have prevented EPA from reasonably determining the
    best performers is unavailing. The representative data EPA
    collected did not prevent or assist it from determining the best
    performers, as EPA had already determined the best-
    performing incinerators based on control technology with no
    additional input from NACWA required. In fact, calling the
    data “representative” implies that EPA believes it targeted the
    best performers even regardless of what the data actually
    showed.
    While we agree with NACWA that EPA may have been
    unduly dismissive of the fact that there is significant sewage
    sludge variability within a publicly-owned treatment work,
    NACWA’s claim that EPA must account for variability
    among publicly-owned treatment works raises statutory issues
    beyond the scope of the issues NACWA raised in its petition
    for review. Specifically, by asserting that EPA must account
    for variability among publicly-owned treatment works
    because the publicly-owned treatment works have limited
    64
    control over the pollutant concentration in their sewage
    sludge, NACWA appears to be urging a different
    interpretation of the “average emissions limitation achieved
    by the best performing 12 percent of units” than the one EPA
    has adopted.
    In this rulemaking EPA explained that it “is using lowest
    emissions limitation as the measure of best performance.” 76
    Fed. Reg. at 15,389. Thus, by arguing that EPA must account
    for variability in sewage sludge content over which a
    publicly-owned treatment work has no control, NACWA is
    essentially requesting that EPA adopt a different
    interpretation of the phrase “average emissions limitation
    achieved by the best performing 12 percent of units.” For
    example, if an incinerator, through bad luck, receives 100
    units of pollutants in its sewage sludge, and manages to emit
    only 50 units of pollutants, it could be said to be a better
    performer than an incinerator that receives, through good
    luck, only 30 units but emits 25 units of pollutants. But under
    EPA’s interpretation of “best performing,” the reduction in
    emissions from uncontrolled conditions to controlled
    conditions is irrelevant—the best-performing incinerators are
    those that emit the lowest levels of pollutants, and so the
    incinerator with an emission level of 25 would be the better
    performer. Where EPA explains that a best performer is
    determined by its emission level, apparently in absolute terms,
    the logical consequence of NACWA’s argument is that the
    “best performing” incinerators must be those that are the most
    effective at removing pollutants from incinerated waste before
    emitting pollutants from the stack. In fact, NACWA
    expressly states its desire for a different interpretation in its
    comments on the proposed rule, explaining that “[w]ithout the
    use of long-term data to support the level of emission
    standards, this variability makes numeric technology-based
    limits impractical and infeasible and should provide EPA
    65
    strong motivation to look to other regulatory options.” Joint
    Appendix 671. Relatedly, NACWA’s comments also urged
    EPA to apply a “variability factor,” based on the variability in
    sewage sludge characteristics, to the stack test data.
    While it may be true that a publicly-owned treatment
    work’s obligation to manage all sewage that enters into the
    sanitary sewer system distinguishes it from commercial or
    industrial incinerators that have more control over what waste
    they combust, this fact does not present a unique scenario in
    setting MACT floors. As a concurring opinion noted in the
    Brick MACT case, there seems to be a paradox in § 112’s
    directions on setting MACT standards (which is almost
    identically worded to § 129). See 
    479 F.3d at
    884–85
    (Williams, J., concurring). Specifically, the statute “calls for
    emissions standards that are the most stringent that EPA finds
    to be ‘achievable,’ taking into account a variety of factors
    including cost,” while also requiring that “the standards ‘shall
    not be less stringent’ than the emission control that have been
    ‘achieved in practice.’” 
    Id. at 884
     (internal citations omitted).
    Thus, implicit in the statute’s requirements is that the standard
    for what is achievable will be more stringent than the floors
    that are based on past achievement. 
    Id.
     But if meeting the
    floors is prohibitively costly for an incinerator “because of
    conditions specific to those plants,” for example, because “the
    required technology cannot, given local inputs whose use is
    essential, achieve the floor,” then it would seem that what
    some plants have achieved is not achievable for other plants.
    As applied here, where incinerators have limited control over
    sewage sludge that can vary significantly in its pollutant
    concentrations, this criticism of the MACT standards seems
    especially pointed.
    Even if this were true, however, NACWA has not argued
    that § 129 requires EPA to account for variability in sewage
    66
    sludge inputs among publicly-owned treatment works, or that
    EPA’s interpretation of the statute as basing “best performers”
    on the “lowest emission levels” is impermissible. Nor has
    NACWA argued that EPA acted arbitrarily or capriciously by
    not basing MACT floors on a more “sewage sludge-diverse”
    dataset or by not applying a variability factor to account for
    variations in sewage sludge characteristics.         Moreover,
    NACWA’s urging that EPA adopt a variability factor may put
    EPA in conflict with its own interpretation of what “best
    performing” means. If EPA applies a variability factor, based
    on the heaviest sewage sludge pollutant concentrations
    experienced by publicly-owned treatment works, to an
    incinerator for which it has already estimated emission levels
    under the worst foreseeable circumstances, the MACT floor
    could no longer be said to reflect what that incinerator
    “achieved.” Instead, applying a variability factor would be
    akin to EPA’s approach in other cases to set MACT floors
    based on the emissions of the worst performer using MACT
    technology, which EPA attempted to justify “by claiming that
    floors must be achievable by all sources using MACT
    technology.” Cement Kiln, 
    255 F.3d at 861
    . While we have
    recognized the paradox in requiring all incinerators to comply
    with a floor based on what some incinerators achieved in the
    past but which may be unachievable to other incinerators, we
    have roundly rejected an interpretation of § 129 that would
    require EPA to set the MACT floors at levels achievable by
    all sources. See Brick MACT, 
    479 F.3d at 880
     (“‘EPA may
    not deviate from section 7412(d)(3)’s requirement that floors
    reflect what the best performers actually achieve by claiming
    that floors must be achievable by all sources using MACT
    technology.’”) (quoting Cement Kiln, 
    255 F.3d at 861
    ).
    But while we conclude that NACWA has not argued a
    legal basis for why EPA should account for variability in
    sewage sludge characteristics among publicly-owned
    67
    treatment works, we are somewhat confused by EPA’s
    defense of its variability analysis in the final rule, particularly
    given its statement that it is using the “lowest emissions
    limitation as the measure of best performance.” 76 Fed. Reg.
    at 15,389. In the proposed rule, EPA made no mention of
    deriving a dataset representative of variability among
    publicly-owned treatment works. See 75 Fed. Reg. at 63,269–
    72. But in the final rule, EPA responded to comments that
    EPA did not adequately account for variability by explaining
    that it had “gathered [emissions information] from nine
    different facilities located in nine different states” and that the
    facilities surveyed, when “combined together,” “represent
    sufficient variations in regions, climates and populations that
    adequately incorporates variability in wastewater treatment
    systems across the U.S.”            76 Fed. Reg. at 15,389.
    “[V]ariations in regions, climates and populations,” however,
    seem irrelevant in the hunt for the lowest emitting
    incinerators. See id. If EPA did consider sewage sludge
    incinerators to be a unique type of incinerator and did seek to
    develop a dataset that was geographically and
    demographically diverse, then it should reconcile that goal
    with its statement that the best-performing units are the lowest
    emitting. And if EPA does seek to develop a dataset
    representative of variability among sewage sludge
    incinerators, it needs to address NACWA’s contentions that
    some of its MACT floors are not actually geographically
    representative. See NACWA Br. at 33–34 (demonstrating
    that fluidized bed incinerator MACT floors were based only
    on publicly-owned treatment works in Michigan, Minnesota,
    Pennsylvania, and North Carolina).
    To sum, while we determine that EPA’s use of the upper
    prediction limit may be lawful, we are remanding this portion
    of its rulemaking for further explanation on the issues of how
    the upper prediction limit represents the “average emissions
    68
    limitation achieved,” how the upper prediction limit is a
    reasonable method of predicting the upper limit of the best-
    performing incinerators, and how the upper prediction limit
    accounts for variability in incinerator performance when it is
    not based on a dataset representative of the best-performing
    incinerators under the worst-performing conditions.         We
    further conclude that NACWA’s arguments that EPA must
    consider variability in sewage sludge characteristics among
    publicly-owned treatment works are meritless because
    NACWA has not argued any legal basis invoking either the
    statutory language or arbitrary-and-capricious review why
    EPA is compelled to account for that sort of variability. But
    because EPA’s discussion of its efforts to create a
    representative dataset seems in conflict with its assertion that
    the best-performing incinerators are those with the lowest
    emission levels, we also remand for EPA to reconcile this
    point.
    3. Adequacy of the MACT floor dataset
    We now address Sierra Club’s and NACWA’s argument
    that EPA unlawfully and arbitrarily set certain MACT floors
    on datasets comprising less than 12 percent of the population
    of sewage sludge incinerators. In its rulemaking, EPA
    candidly admitted its data collection efforts yielded a dataset
    that comprised less than 12 percent of existing incinerators for
    certain pollutants. See 75 Fed. Reg. at 63,270 (“EPA does not
    have actual emissions test data for the population of units that
    represent the best-performing 12 percent . . . .”). EPA
    nevertheless concluded that the lack of data for at least 12
    percent of incinerators did not prevent it from setting MACT
    floors, as it had “conducted a statistical analysis to verify the
    minimum number of observations needed to accurately
    characterize the distribution of the 12 percent of units in each
    category,” and had determined that the data it used “m[et] or
    69
    exceed[ed] the number of observations necessary to provide
    an accurate representation of that data.” 76 Fed. Reg. at
    15,387. EPA’s only explanation for why its use of this
    technique was appropriate was that “emission data are
    normally distributed [i.e., on a bell curve], or can be
    transformed to be normally distributed.” Revised MACT
    Floor Memo at 8.
    On petition for review, NACWA argues that EPA’s
    failure to collect sufficient data violates § 129, which, unlike
    § 112, requires EPA to set the MACT floor on the top 12
    percent of performing units without the qualifier “for which
    the Administrator has emissions information.” Compare 
    42 U.S.C. § 7429
    (a)(2), with 
    id.
     § 7412(d). Thus, NACWA
    asserts, EPA may not extrapolate information about the “best
    performing units” from less than 12 percent of such units in
    light of Congress’s unqualified directive on setting solid
    waste incinerator MACT floors.
    NACWA is incorrect that EPA’s decision to set MACT
    floors on less than 12 percent of data is per se unlawful. In
    concluding that the law allows EPA to estimate the “average
    emissions limitation achieved by the best performing 12
    percent of units,” we have not determined any requirement
    that EPA have at least a representative sample of 12 percent
    of the population of incinerators. Instead, we have explained
    that the existing incinerator MACT floor standard “does not
    by its plain meaning exclude estimation either by sampling or
    by some other reliable means.” Sierra Club, 167 F.3d at 662
    (emphasis added). Thus, the fact that EPA does not possess
    data directly collected from 12 percent of incinerators does
    not make its estimate inherently unlawful. Cf. Mossville
    Environmental Action Now, 
    370 F.3d at 1241
     (allowing EPA
    to use one data point—EPA’s preexisting emission
    standard—to establish an existing source MACT floor,
    70
    because the preexisting standard was “just barely satisfied by
    the plant with the lowest overall long term [emission of the
    pollutant at issue]”).
    While NACWA simply asserts EPA cannot lawfully set
    existing incinerator MACT floors on fewer than 12 percent of
    incinerators, Sierra Club goes further and argues that EPA
    acted arbitrarily and capriciously in setting MACT floors
    representative of the best performing 12 percent on less than
    12 percent of data. Sierra Club contends that EPA fails to
    demonstrate or even claim that “emissions from the units for
    which it has data are representative of emissions from the
    ones for which it lacks data,” repeating its argument that
    incinerators for which EPA lacks data may be achieving
    lower emission levels than those for which it has data. Sierra
    Club. Br. at 29. This appears to be the equivalent of arguing
    that EPA cannot compound its error in estimating the best
    performing 12 percent by assuming that an imperfect dataset
    can represent other potentially superior incinerators that
    should be included in the top 12 percent.
    The point that errors in one estimate will be further
    compounded by another estimate is a fair one. If the MACT
    technology approach to selecting best performing incinerators
    was 75 percent accurate, and EPA’s statistical equation
    represented a larger sample size with 80 percent accuracy,
    each estimate, alone, may be sufficiently reasonable. But if
    EPA combined the two and applied an 80 percent accurate
    formula to a 75 percent accurate estimate of best performers
    based on technology, the underlying result may be too
    imprecise to be considered reasonable. And this does not even
    account for the fact that EPA’s MACT floor methodology
    layers another estimate—the upper prediction limit—to
    account for variability.
    71
    Sierra Club is also correct that EPA has “base[d]
    estimates of the performance of one group of units on the
    performance of a different group without demonstrating that
    this approach yields accurate estimates.” Sierra Club Br. at
    29. EPA did so in this case by using a statistical formula for
    determining the minimum number of observations necessary
    to adequately characterize the population of the best
    performing 12 percent of units, which is as follows:
    1
    Revised MACT Floor Memo at 8. In this formula, n
    represents the minimum number of observations required, N
    represents the population size, Z represents a value associated
    with a specific confidence level, E represents the level of
    precision or error tolerance, p represents the degree of
    variability in observations, and q represents one minus p. We
    note that none of the variables in this formula are fixed or
    based on the dataset, aside from N, the population size,
    leaving EPA to select the value for most variables; in this
    case, a 90% confidence level, a precision level of 20%, and a
    degree of variability of 0.5. Revised Mact Floor Memo at 8–
    9.
    The flexibility in defining variables in this formula is of
    some concern, as the reasonableness of EPA’s statistical
    extrapolation depends on variables for which EPA sets values.
    Our confidence in this statistical methodology is hardly
    heartened by the fact that the minimum number of
    observations EPA calculated (11 for fluidized bed incinerators
    and 14 for multiple hearth incinerators), were both just one
    shy of the number of observations EPA actually collected for
    the pollutant with the least amount of test data (12 for
    fluidized bed incinerators and 15 for multiple hearth
    72
    incinerators). Revised MACT Floor Memo at 9. To illustrate
    how much the result can vary, had EPA chosen a precision
    level of 15% instead of 20%, the minimum number of
    observations required for fluidized bed incinerators would
    have increased from 11 to 14. Changing the specific
    confidence level from 90% to 95%3 increases the minimum
    number of observations from 11 to 13.4 To EPA’s credit, it
    chose a value for variability that would maximize the number
    of observations required, and explained it chose that value to
    overestimate the number of minimum observations needed.
    Revised MACT Floor Memo at 8–9. But for the rest of the
    variables, EPA selected values which could have been
    determinative of the validity of its dataset without explaining
    why it selected those numbers. That EPA could have
    determined the statutory sufficiency of its dataset by choosing
    values does not mean that EPA did so, but at a minimum EPA
    must explain why it chose the values it did. “[A]n agency
    may not pluck a number out of thin air when it promulgates
    rules in which percentage terms play a critical role.” WJG
    Telephone Co. v. FCC, 
    675 F.2d 386
    , 388–89 (D.C. Cir.
    1982).
    3
    In entering values into EPA’s statistical equation, we assumed a
    Z-score for a 95% confidence level as 1.960.                    See
    http://people.richland.edu/james/lecture/m170/ch08-int.html
    (noting the 90% confidence level Z-score is 1.645, the same used
    by EPA to calculate the minimum number of observations required
    in its MACT floor analysis); see Revised MACT Floor Memo at 8.
    4
    As in Sierra Club, we note that “[o]ur observations are based on
    our own analysis of EPA’s data, and we may have omitted some
    crucial step in the process.” 167 F.3d at 664. But as we also noted
    in Sierra Club, and will explain in more detail, this “exercise
    highlights the need for additional explanation even if our
    calculation is wrong.” Id.
    73
    In addition to not explaining why it chose the values it
    did for its statistical equation, EPA has not clarified how this
    statistical method can allow a limited dataset to approximate a
    larger portion of the population. In laying out the equation,
    EPA cited a study titled “Sample Size Requirements for
    Studying Small Populations in Gerontology Research” from
    the journal “Health Services Outcomes Research
    Methodologies.” See Revised MACT Floor Memo at 8, 18.
    But EPA provided no justification for how this equation could
    allow it to extrapolate the best performing 12 percent from an
    insufficient dataset, other than to state, “Because the emission
    data are normally distributed, or can be transformed to
    normally distributed . . . a statistical technique can be
    employed to determine the minimum number of observations
    needed to accurately characterize the distribution of the best
    performing 12 percent of units.” Revised MACT Floor Memo
    at 8. As we explained with the upper prediction limit, while it
    is true that we “owe particular deference to EPA when its
    rulemakings rest upon matters of scientific and statistical
    judgment within the agency’s sphere of special competence
    and statutory jurisdiction,” EPA must still articulate a
    “rational connection between the facts found and the choice
    made.” Burlington Truck Lines, 
    371 U.S. at 168
    ; American
    Coke & Chemicals Institute, 
    452 F.3d at 941
    . If EPA chooses
    to use statistics as a shortcut for meeting the Congressional
    mandate to set MACT floors, it must justify its statistical
    analysis with greater detail than “a statistical technique can be
    employed” when “data [are] normally distributed.” Although
    EPA does not need to fill the Federal Register with treatises
    on statistics, it must specify in greater detail why the equation
    it is using can accomplish the purpose for which EPA is using
    the equation. This is not only required as part of EPA’s
    obligation to demonstrate the reasonableness of its estimates
    with substantial evidence, but also to prevent an agency from
    74
    using opaque statistical justification to cover a deficiency in
    its dataset.
    As with the upper prediction limit, EPA’s use of this
    statistical technique is not unlawful as long as EPA can
    demonstrate with substantial evidence why it reasonably
    estimates the performance of incinerators for which it has no
    data. Because this demonstration requires more detail than
    EPA gave here, we remand this portion of the rulemaking for
    further explanation on why EPA can use this formula to
    estimate gaps in its data and an explanation of why EPA
    chose the variables it did.
    4. Incorporating non-detect data
    Sierra Club challenges EPA’s method of accounting for
    certain emissions data that was not quality-assured. When
    collecting data on certain pollutants from sewage sludge
    incinerators, EPA encountered “non-detect data”—i.e.,
    emission levels too low to register in an emissions test. See
    Revised MACT Floor Memo at 14–15. In the rulemaking,
    EPA explained that it would use a two-part test based on the
    method detection level, which is the “minimum concentration
    of a pollutant that can be measured with confidence that the
    level is greater than zero.” EPA Br. at 61 n.20; Revised
    MACT Floor Memo at 14–15.
    Because the method detection level varies depending on
    several factors, EPA first established a value it termed the
    “representative method detection level,” which it defined as
    “the highest test-specific method detection level reported in a
    data set that is also equal to or less than the average emission
    calculated for the data set.” Revised MACT Floor Memo at
    14–15. In other words, if a specific emissions test registered a
    non-detect at a value higher than the average emission level,
    75
    EPA excluded it as a candidate for the representative method
    detection level to “minimiz[e] the effect of a test[] with an
    inordinately high method detection level.” 
    Id.
    The second step in incorporating non-detect data was to
    multiply the representative method detection level by three
    and compare it to the calculated floor emission limit for that
    pollutant. 
    Id.
     EPA’s rationale for choosing to multiply the
    representative method detection level by three, rather than
    using the method detection level itself, was to reduce the
    effect of measurement imprecision. EPA explained that at
    values around the method detection level, measurement
    imprecision is around 40 to 50 percent. Id. at 14. As values
    increase above the method detection level, the testing
    becomes more accurate, with around 10 to 15 percent
    measurement imprecision at three times the method detection
    level. Id.
    If the calculated emission limit was greater than the
    representative method detection level times three, EPA
    concluded that its calculation adequately addressed
    measurement variability, and thus would set the calculated
    emission limit as the MACT floor. Id. If the calculated
    emission level was less than the representative method
    detection level, EPA concluded that its calculation did not
    adequately account for measurement variability, and the
    representative method detection level would become the
    MACT floor. Id. EPA used this method to set the four out of
    the forty MACT floors it established for sewage sludge
    incinerators. See Revised MACT Floor Memo at 19.
    Sierra Club argues that EPA’s method of incorporating
    non-detect data is unlawful because it does not reflect what
    the best performers actually achieve, as required by §
    129(a)(2). See Brick MACT, 
    479 F.3d at 880
     (interpreting §
    76
    112). Sierra Club also contends that the three times
    representative method detection level value is not a
    reasonable estimate of the emission levels achieved by an
    incinerator when a test produces non-detect data.
    We agree with EPA that its method of incorporating non-
    detect data is reasonable, and not arbitrary or capricious. We
    do not expect EPA to perform the impossible, see Cement
    Kiln, 
    255 F.3d at 871
    , and that includes recording emission
    levels that are not accurately detectable with its current
    emissions testing technology. As EPA explains the issue,
    emission levels from zero up to some value above the method
    detection level cannot be stated with accuracy. Because any
    emission level EPA selects at that point will necessarily be an
    estimate, EPA adopted a method to account for measurement
    imprecision that has a rational basis in the correlation between
    increased emission values and increased testing precision.
    Although Sierra Club argued in its comments that EPA
    should have at the very least assumed that non-detect data was
    at the detection limit, it did not offer any evidence that EPA
    was incorrect in explaining why, given the measurement
    imprecision at the method detection level, a non-detect test
    run would always yield emissions data below the method
    detection level. Because we owe significant deference to
    EPA in areas of its technical expertise, we reject Sierra Club’s
    challenge to EPA’s method of addressing non-detect data.
    77
    C. SETTING   CERTAIN  NEW   MULTIPLE HEARTH
    INCINERATOR MACT FLOORS AT EXISTING MULTIPLE
    HEARTH INCINERATOR MACT FLOOR EMISSION
    LEVELS
    Although EPA had explained in its proposed rule that it
    was proposing new incinerator MACT floors for all sewage
    sludge incinerators based only on fluidized bed incinerator
    emission data, in the final rule it set separate new incinerator
    MACT floors for both subcategories in response to industry
    commenters. See 76 Fed. Reg. at 15,384. When EPA applied
    the upper prediction limit to the best performing multiple
    hearth incinerator for each pollutant, however, it yielded new
    multiple hearth incinerator MACT floors for hydrogen
    chloride and sulfur dioxide that were higher than than the
    existing multiple hearth incinerator MACT floors. See 76
    Fed. Reg. at 15,388–89. Reasoning that new incinerator
    MACT floors could not be less stringent than existing
    incinerator MACT floors, EPA set the new multiple hearth
    incinerator MACT floors for hydrogen chloride and sulfur
    dioxide at the existing multiple hearth incinerator MACT
    floors. See id.
    Sierra Club challenges this decision, arguing, among
    other things, that this decision does not at all attempt to satisfy
    § 129(a)(2)’s requirement that the new incinerator MACT
    floor be set at the level the best-controlled units actually
    achieved. But because we are remanding the portions of
    EPA’s rulemaking establishing the upper prediction limit, in
    part to further explain why the upper prediction limit is a
    reasonable estimate given that it can fluctuate so much
    depending on variability, we decline to consider Sierra Club’s
    challenge at this time.
    78
    D. BEYOND-THE-FLOOR STANDARDS
    In the final rule, EPA explained that it chose not to adopt
    beyond-the-floor standards for existing incinerators, primarily
    based on its determination that additional control technologies
    would not be cost-effective, and mentioned nothing about
    setting beyond-the-floor standards for new incinerators. 76
    Fed. Reg. at 15,394.         Sierra Club challenges EPA’s
    determination not to set beyond-the-floor standards for
    existing units based on cost-effectiveness considerations.
    Sierra Club also challenges EPA’s decision not to set beyond-
    the-floor standards for new multiple hearth incinerators even
    though it provided no comment on this issue, explaining that
    because EPA did not adopt new multiple hearth incinerators
    MACT floors until the final rule it was impracticable to do so.
    See 75 Fed. Reg. at 63,272 (deciding to set all new sewage
    sludge incinerator MACT floors at the level of the best-
    performing fluidized bed incinerators); 
    42 U.S.C. § 7607
    (d)(7)(B).
    1. Deciding not to set beyond-the-floor standards for
    existing units
    In challenging EPA’s decision not to set beyond-the-floor
    standards for existing units, Sierra Club argues that §
    129(a)(2) “unambiguously requires the ‘maximum’ degree of
    reduction that can be achieved considering cost and other
    statutory factors.” Sierra Club Br. at 36. Allowing EPA to
    determine whether a cost-per-ton reduction is appropriate,
    Sierra Club asserts, would give EPA greater discretion than §
    129 allows, as Congress requires more from EPA in § 129
    than to undertake a cost-benefit analysis. But in arguing that
    EPA abused its discretion in determining the maximum
    degree of reduction in emissions that is achievable for sewage
    sludge incinerators, “taking into consideration the cost of
    79
    achieving such emission reduction,” Sierra Club must clear a
    high bar, as we are at our most deferential when an agency is
    “making predictions, within its area of special expertise, at the
    frontiers of science.” See Husqvarna AB v. EPA, 
    254 F.3d 195
    , 199 (D.C. Cir. 2001) (quoting Baltimore Gas & Electric
    Co. v. NRDC, 
    462 U.S. 87
    , 103 (1983)).
    EPA argues that § 129(a)(2) does not require it to
    establish a beyond-the-floor standard regardless of costs,
    explaining that we have upheld a similar cost-effectiveness
    analysis in the past based on a similarly-worded statute. See
    Husqvarna, 
    254 F.3d at 200
     (requiring EPA to promulgate
    standards that “shall achieve the greatest degree of emission
    reduction achievable through the application of technology
    which [EPA] determines will be available for the engines or
    vehicles to which such standards apply, giving appropriate
    consideration to the cost of applying such technology within
    the period of time available to manufacturers and to noise,
    energy, and safety factors associated with the application of
    such technology”) (quoting 
    42 U.S.C. § 7547
    (a)(3)). We
    agree. In Husqvarna, we explained that because the similarly
    worded statute did not “mandate a specific method of cost
    analysis, we find reasonable the EPA’s choice to consider
    costs on the per ton of emissions removed basis.” 
    Id.
     (citing
    
    65 Fed. Reg. 24,300
    ). This applies equally here to EPA’s
    decision to consider costs on the per pound of mercury
    emissions removed basis.
    Sierra Club also asserts that EPA’s cost-effectiveness
    analysis is arbitrary and capricious because EPA only
    considered the cost of proposed beyond-the-floor technology
    options in reductions of mercury, without also considering the
    benefit that these proposed technology options would have in
    reducing other pollutants. But in the proposed rule, EPA
    estimated the emission reductions to both dioxins and
    80
    mercury from different control technology options, attributing
    the cost solely to mercury because “99.9 percent of the
    emissions reduction [from applying beyond-the-floor
    technologies] is associated with [mercury].” 75 Fed. Reg. at
    63,277. In the final rule, EPA evaluated adding a fabric filter
    in combination with the beyond-the-floor technologies it
    discussed in its proposed rule, and again attributed reductions
    solely to mercury. 76 Fed. Reg. at 15,393–94. Although EPA
    did not explain in the final rule why it did not consider the
    cost in terms of pounds removed of other pollutants, its failure
    to do so was not arbitrary and capricious, particularly given
    that Congress gave EPA broad discretion in considering
    whether to go beyond-the-floor. See 
    42 U.S.C. § 7429
    (a)(2).
    Finally, Sierra Club argues that EPA violated § 129
    because it did not consider “methods and technologies for
    removal or destruction of pollutants before . . . combustion”
    as required by § 129(a)(3). As evidence that pre-combustion
    methods can significantly reduce emission levels, Sierra Club
    again cites Palo Alto’s comments in which the city described
    its source control program.          This program involved
    “assist[ing] in authoring California legislation that eliminated
    mercury in thermometers, certain switches, and novelty
    items”; “requir[ing] amalgam separators at dental facilities”;
    and operating “an ongoing drop-off program for all types of
    mercury-containing equipment.” Joint Appendix 628–29
    (Palo Alto Comments).
    EPA responds that § 129 “does not authorize EPA to
    regulate the sources of sewage sludge under the [Clean Air
    Act], and no commenter cited any authority to the contrary,”
    which was the same position it explained to commenters.
    EPA Br. at 73; see Joint Appendix 1011. Because § 129(a)(3)
    is ambiguous as to the extent of pre-combustion emission
    reduction methods EPA should consider, and because Palo
    81
    Alto’s program goes far beyond the scope of activities that
    occur at publicly-owned treatment works and their sewage
    sludge incinerators, we uphold EPA’s decision not to consider
    source control in its beyond-the-floor analysis as a reasonable
    interpretation of § 129.
    2. Deciding not to set beyond-the-floor standards for new
    multiple hearth incinerators
    EPA asserts that Sierra Club has waived the issue of
    beyond-the-floor standards for new multiple hearth
    incinerators by failing to comment on this issue. We agree.
    The Clean Air Act’s judicial review provision limits judicial
    review to objections “raised with reasonable specificity
    during the period for public comment.” 
    42 U.S.C. § 7607
    (d)(7)(B). If it was “impracticable to raise such
    objection within such time” and the “objection is of central
    relevance to the outcome of the rule,” an objecting party may
    petition for reconsideration and we may review a denial of
    that petition. 
    42 U.S.C. § 7607
    (d)(7)(B). If a petitioner has
    not satisfied 
    42 U.S.C. § 7607
    (d)(7)(B)’s exhaustion
    requirements in raising its objections before EPA, we do not
    have jurisdiction to hear that objection on a petition for
    review. Portland Cement Ass’n v. EPA, 
    665 F.3d 177
    , 185
    (D.C. Cir. 2011).
    In promulgating the sewage sludge incinerator rule, EPA
    explained in the proposed rulemaking that while it was
    proposing that all new incinerator MACT floors be based on
    data from the best-performing fluidized bed incinerator, it was
    “aware that owners and operators with modified [multiple
    hearth] units may have concerns regarding meeting the new
    source limits.” 75 Fed. Reg. at 63,272. In light of that, EPA
    “request[ed] comment on th[e] proposed approach,” even
    providing a proposal of potential MACT floor emission limits
    82
    “[t]o assist commenters with their evaluation of the proposal.”
    Id.
    EPA’s request for comment and proposed new multiple
    hearth incinerator MACT floors put Sierra Club on notice that
    EPA was seriously considering deviating from its proposed
    new incinerator MACT floors for new multiple hearth
    incinerators. Although the potential MACT standards for new
    multiple hearth incinerators were more stringent than the
    MACT standards EPA ultimately adopted, Sierra Club had
    EPA’s MACT floor dataset, which included raw data from
    emissions test and a list of the control devices used by the
    units EPA considered the best performing. Compare 75 Fed.
    Reg. at 63,272 (tabulating proposed new multiple hearth
    incinerator MACT standards), with 76 Fed. Reg. at 15,388–89
    (tabulating final new multiple hearth incinerator MACT
    standards); see Joint Appendix 607 (Sierra Club comments
    citing EPA’s MACT floor analysis and commenting on EPA’s
    dataset).
    While it is true that “we do not require telepathy,” and
    are reluctant to require advocates for affected groups to
    anticipate every contingency lest we encourage strategic
    vagueness by agencies, we nevertheless “require some degree
    of foresight on the part of commenters.” Portland Cement
    Ass’n, 
    665 F.3d at 186
    . Because Sierra Club was on notice
    that EPA was considering setting new multiple hearth
    incinerator MACT floors and because it had access to the
    dataset EPA would use in setting new multiple hearth
    incinerator floors, we conclude that it was practicable for
    Sierra Club to comment on beyond-the-floor standards for
    new multiple hearth incinerator MACT floors. Accordingly,
    we will not consider Sierra Club’s argument regarding new
    multiple hearth incinerator beyond-the-floor standards. See
    
    42 U.S.C. § 7607
    (d)(7)(B).
    83
    E. SUBCATEGORIZING SEWAGE SLUDGE INCINERATORS
    NACWA challenges EPA’s decision to create only two
    subcategories for sewage sludge incinerators. In its
    rulemaking, EPA proposed subcategorizing sewage sludge
    incinerators into multiple hearth and fluidized bed
    incinerators, and proposed no other categories or
    subdivisions. See 75 Fed. Reg. at 63,268. EPA invited
    comment on whether other combustor designs were used at
    sewage sludge incinerators, requesting emissions information
    from stack tests conducted on those designs. Id.
    Commenters responded and requested that EPA further
    subcategorize “based on size of the [sewage sludge
    incinerator], type of sewage sludge incinerated, limited use
    units, and distance over which the [incinerator] would need to
    transport its sludge for disposal.” 76 Fed. Reg. at 15,384. But
    EPA adopted only the two subcategories it proposed,
    explaining that it did “not have data to support distinguishing
    units based on class, type, or size,” and that “[w]ithout such
    information,” it did “not have a basis for concluding that these
    types of units should be based in a different subcategory.” Id.
    NACWA challenges EPA’s decision not to subcategorize
    further, asserting that it had identified “back-up and
    emergency” sewage sludge incinerators, and other unique
    categories, that would have a difficult time meeting the
    sewage sludge incinerator rule’s testing obligations. As its
    legal basis, NACWA asserts that EPA failed to respond
    adequately to its substantive comment, as required under 
    42 U.S.C. § 7607
    (d)(6)(B), and that EPA’s claim that it did not
    have data is arbitrary because commenters had submitted
    information about these special-use sewage sludge
    incinerators.
    84
    We agree with EPA that its decision to create
    subcategories only for multiple hearth and fluidized bed
    incinerators was not arbitrary or capricious. We have held
    that EPA has authority to subcategorize within
    Congressionally mandated categories under § 129(a)(2). See
    Northeast Maryland Waste Disposal Authority, 
    358 F.3d at
    946–47. We have also held that EPA’s subcategorization
    authority under § 112 involves an expert determination,
    placing a heavy burden on a challenger to overcome
    deference to EPA’s “articulated rational connection between
    the facts found and the choice made.” NRDC v. EPA, 
    489 F.3d 1364
    , 1375 (D.C. Cir. 2007).
    EPA’s decision here appears well within its expert
    determination. Perhaps recognizing the deference we owe
    EPA in its decision to subcategorize, NACWA does not
    challenge EPA’s authority to do so, but instead asserts
    procedural challenges. We do not find these challenges
    meritorious. EPA rationally stated its policy to require
    emission information from stack tests on the combustion
    designs that commenters wanted EPA to accommodate into a
    separate subcategory. Although NACWA identified different
    classes of incinerators and discussed differences in sewage
    sludge variability that it felt justified further
    subcategorization, it does not cite any emissions information
    from stack tests it submitted for the special-circumstance
    sewage sludge incinerators for which it desired
    subcategorization. EPA “was not obligated under its policy”
    to create new subcategories or to offer a further response on
    NACWA’s request for further subcategorization, and we will
    uphold its decision not to create additional subcategories
    beyond the multiple hearth and fluidized bed incinerators.
    See Edison Electric Institute, 
    2 F.3d at 449
    .
    85
    F. MONITORING
    Section 129(c) requires incinerator operators “to monitor
    emissions from the unit at the point at which such emissions
    are emitted into the ambient air . . . and at such other points as
    necessary to protect public health and the environment,” and
    “to monitor such other parameters relating to the operation of
    the unit and its pollution control technology” as EPA deems
    appropriate. In the final rule, EPA required all new and
    existing sewage sludge incinerators to “demonstrate initial
    and annual compliance with the emission limits using EPA-
    approved emission test methods.” 76 Fed. Reg. at 15,377.
    EPA gave existing incinerators the option of continuous
    emissions monitoring in lieu of initial and annual tests, and
    required continuous parameter monitoring. Id. For new
    incinerators, EPA made continuous emissions monitoring
    mandatory for carbon monoxide, and optional for other
    pollutants in lieu of initial and annual emissions testing, and
    required continuous parameter monitoring. Id.
    Sierra Club argues that EPA violates the statute by only
    requiring parameter monitoring and not mandating continuous
    emissions monitoring for all pollutants on all incinerators.
    EPA responds that Congress gave it the discretion to require
    parameter monitoring as appropriate, and that its combination
    of initial and annual emissions testing combined with
    parameter monitoring serves to meet § 129’s monitoring
    requirement. We agree with EPA that § 129(c)(1) does not
    require continuous emissions monitoring and that its
    monitoring requirements in the sewage sludge incinerator rule
    satisfy § 129’s statutory requirement.
    Under 42 U.S.C. § 7661c, EPA “may by rule prescribe
    procedures and methods for determining compliance and for
    monitoring and analysis of pollutants regulated under this
    86
    Act, but continuous emissions monitoring need not be
    required if alternative methods are available that provide
    sufficiently reliable and timely information for determining
    compliance.” 42 U.S.C. § 7661c(b). Although this section
    appears to clarify the Clean Air Act’s mandate that EPA’s
    permit programs include monitoring and reporting
    requirements, it is evidence supporting EPA’s interpretation
    of § 129’s monitoring requirement as requiring assurance of
    compliance with emission standards, but not continuous
    emissions monitoring. Sierra Club provides no legal authority
    to the contrary. Determining whether specified testing and
    monitoring requirements assure compliance with EPA’s
    emission standards “requires a high level of technical
    expertise,” and because Sierra Club has not given us any
    reason to doubt that EPA’s requirements will effectively
    assure compliance, “we must defer to the informed
    discretion” of EPA. National Lime, 
    233 F.3d at 635
     (quoting
    Marsh v. Oregon Natural Resources Council, 
    490 U.S. 360
    ,
    377 (1989)).
    III.   MAXWEST’S INTERVENTION
    MaxWest Environmental Systems, the developer of a
    “proprietary biosolids management process that converts
    biosolids into syngas,” used in units MaxWest terms as
    “gasifiers,” challenges EPA’s treatment (or lack thereof) of
    gasifiers in the final rule. See Intervenor MaxWest Br. at 1–2,
    9–21. EPA asserts several grounds why we need not reach
    MaxWest’s arguments, including that MaxWest lacks
    standing and that the issues it raises are outside the scope of
    those raised by the petitioners.
    “[B]ecause Article III standing is a prerequisite to a
    federal court’s exercise of jurisdiction,” we “cannot proceed
    at all in any cause” unless we first determine that a party
    87
    seeking to be heard has satisfied the three-part test of Lujan.
    Sinochem International Co. v. Malay International Shipping
    Corp., 
    549 U.S. 422
    , 430 (2007); New England Power
    Generators Ass’n v. FERC, 
    707 F.3d 364
    , 368 (D.C. Cir.
    2013). But while we cannot assume hypothetical jurisdiction
    to decide the merits of a case, we have leeway “to choose
    among threshold grounds for denying audience to a case on
    the merits” because “jurisdiction is vital only if the court
    proposes to issue a judgment on the merits.” Sinochem, 
    549 U.S. at 431
     (internal alteration omitted) (quoting Intec USA,
    L.L.C. v. Engle, 
    467 F.3d 1038
    , 1041 (7th Cir. 2006)). Thus,
    we need not resolve MaxWest’s standing to intervene if we
    can dispose of its intervention on another threshold ground
    that does not require us to reach the merits of MaxWest’s
    arguments. See 
    id.
     (explaining that a federal court need not
    determine whether it has jurisdiction when deciding, for
    example, not to exercise supplemental jurisdiction over state
    law claims on discretionary grounds or to abstain under
    Younger v. Harris).
    The alternative threshold ground for rejecting MaxWest’s
    intervention is that MaxWest’s issues are outside the scope of
    those raised by NACWA. None of the six petitioners’ briefs
    filed in this case mention MaxWest’s gasification process.
    When we allowed MaxWest to intervene out of time, we
    warned it that “an intervening party may join issue only on a
    matter that has been brought before the court by another
    party.” NACWA v. EPA, No. 11-1131, Doc. 1344244 at 2
    (D.C. Cir. filed Nov. 28, 2011) (citing Beethoven.com LLC v.
    Librarian of Congress, 
    394 F.3d 939
    , 946 (D.C. Cir. 2005)).
    MaxWest has not heeded our warning. Its bare assertion that
    it modeled its statement of issues after NACWA’s does not
    change the fact that its substantive arguments are unrelated to
    the issues raised by NACWA and Sierra Club. See MaxWest
    88
    Reply Br. at 8–10.         Therefore, we will not consider
    MaxWest’s arguments.
    IV.    CONCLUSION
    For the foregoing reasons, we remand to EPA portions of
    its rule for further explanation without vacating the current
    MACT standards. Specifically, we direct EPA to clarify why
    its Clean Water Act Part 503 regulations control for other
    non-technology factors. We also direct EPA to clarify issues
    related to its upper prediction limit and variability analysis.
    In particular, EPA should explain why the upper prediction
    limit represents the “average emissions limitation achieved by
    the best performing 12 percent of” incinerators; why the
    upper prediction limit reasonably estimates the worst
    foreseeable operating conditions; and why the upper
    prediction limit can account for more than intra-unit
    variability, as EPA claimed it could on petition for review.
    Finally, we direct EPA to elaborate on how it can use a
    statistical method to determine whether a limited dataset is
    representative of incinerators for which it has no data, and to
    explain why it chose the variables it did for that statistical
    analysis.5
    In all other respects, we uphold EPA’s rule against the
    petitioners’ challenges. Because the issues MaxWest raised in
    its intervenor brief are outside the scope raised by the
    petitioners, we do not consider its arguments.
    So ordered.
    5
    We do not, of course, mean to suggest that EPA is bound to reach
    the same conclusion upon reexamination of the record for purposes
    of explanation. Should EPA find itself unable to support its
    conclusions, it is, of course, free to reach different ones.