PRECEDENTIAL


      UNITED STATES COURT OF APPEALS
           FOR THE THIRD CIRCUIT
                _____________

                   No. 15-3435
                  _____________

   RUDOLPH A. KARLO; MARK K. MCLURE;
WILLIAM S. CUNNINGHAM; JEFFREY MARIETTI;
         DAVID MEIXELSBERGER,
                      Appellants

                         v.

       PITTSBURGH GLASS WORKS, LLC
               _____________

   On Appeal from the United States District Court
        for the Western District of Pennsylvania
            District Court No. 2-10-cv-01283
  District Judge: The Honorable Terrence F. McVerry

             Argued November 9, 2016

Before: SMITH, Chief Judge, McKEE, and RESTREPO,
                  Circuit Judges

          (Opinion Filed: January 10, 2017)
Samuel J. Cordes                  [ARGUED]
245 Fort Pitt Boulevard
Pittsburgh, PA 15222

Bruce C. Fox
Andrew J. Horowitz
Obermayer Rebmann Maxwell & Hippel
500 Grant Street
Suite 5240, One Mellon Center
Pittsburgh, PA 15219
       Counsel for Appellant

Rachel E.A. Atterberry
David S. Becker                    [ARGUED]
Jennifer L. Fitzgerald
Tina C. Wills
Freeborn & Peters
311 South Wacker Drive
Suite 3000
Chicago, IL 60606

Robert B. Cottington
Cohen & Grigsby
625 Liberty Avenue
Pittsburgh, PA 15222
       Counsel for Appellee



                              2
Neal D. Mollen                [ARGUED]
Paul Hastings
875 15th Street, N.W.
Suite 1000
Washington, DC 20005
       Counsel for Amicus Appellee Chamber of
       Commerce of the United States of America

Michael P. Bracken
NT Lakis
1501 M Street, N.W.
Suite 400
Washington, DC 20005
       Counsel for Amicus Appellee Equal Employment
       Advisory Council

Anne N. Occhialino           [ARGUED]
Equal Employment Opportunity Commission
5th Floor
131 M Street, N.E.
Washington, DC 20507
      Counsel for Amicus Appellant Equal Employment
      Opportunity Commission

                  ________________

                      OPINION
                  ________________

                           3
SMITH, Chief Judge.

       The Age Discrimination in Employment Act
(“ADEA”) protects only those individuals who are at
least forty years of age. The question in this case is
whether a disparate-impact claim is cognizable where a
“subgroup” of employees at the upper end of that
range—in this case, employees aged fifty and older—
were alleged to have been disfavored relative to younger
employees.

       We answer in the affirmative. Our decision is
dictated by the plain text of the statute as interpreted by
the Supreme Court. In particular, the ADEA prohibits
disparate impacts based on age, not forty-and-older
identity. A rule that disallowed subgroups would ignore
genuine statistical disparities that could otherwise be
actionable through application of the plain text of the
statute. Although several of our sister circuits have ruled
to the contrary, their reasoning relies primarily on policy
arguments that we do not find persuasive.

       We will therefore reverse the judgment of the
District Court based on its interpretation of the ADEA.
We will also vacate the District Court’s order excluding
the testimony of plaintiffs’ statistics expert and remand
for further Daubert proceedings. We will affirm in all
other respects.


                            4
                            I

      Defendant Pittsburgh Glass Works, LLC (“PGW”)
manufactures automotive glass in Harmarville,
Pennsylvania. PGW also owns (1) GTS Services, a
software business, (2) PGW Auto Glass, an automotive
replacement-glass distribution business, (3) LYNX
Services, an insurance claims administrator, and
(4) Aquapel, a glass treatment supplier.

       In 2008, the automobile industry began to falter.
PGW engaged in several reductions in force (“RIFs”) to
offset deteriorating sales. The RIF of relevance to this
case occurred on March 31, 2009, and terminated the
employment of approximately one hundred salaried
employees in over forty locations or divisions. Individual
unit directors had broad discretion in selecting whom to
terminate. PGW did not train those directors in how to
implement the RIF. Nor did PGW employ any written
guidelines or policies, conduct any disparate-impact
analysis, review prospective RIF terminees with counsel,
or document why any particular employee was selected
for inclusion in the RIF.

     Plaintiffs Rudolph A. Karlo, William S.
Cunningham, Jeffrey Marietti, David Meixelberger,
Mark K. McLure, Benjamin D. Thompson, and Richard



                            5
Csukas1 worked in PGW’s Manufacturing Technology
division. They were terminated as part of the March 2009
RIF by their supervisor, Gary Cannon. Each was over
fifty years old at the time.

       In January 2010, plaintiffs filed charges of
employment discrimination with the Equal Employment
Opportunity Commission (“EEOC”). Thereafter, they
received a Dismissal and Notice of Rights from the
EEOC, and this lawsuit followed. Plaintiffs brought a
putative ADEA collective action, asserting three claims:
(1) disparate treatment, (2) disparate impact, and
(3) retaliation as to only Karlo and McLure.

       On plaintiffs’ motion for conditional certification,
the District Court ruled that ADEA subgroups are
cognizable, and conditionally certified a collective action
to be comprised of employees terminated by the RIF who
were at least fifty years old at the time. See Karlo v.
Pittsburgh Glass Works, LLC, 

(W.D. Pa. 2012). In addition to the named plaintiffs,
eleven individuals opted in. Three voluntarily dismissed
their claims and four settled. Four opt-ins remained:
Michael Breen, a former production supervisor at a plant
in Crestline, Ohio; Matthew Clawson, a former Project
Engineer in Evansville, Indiana; Stephen Shaw, a former

      1
        McLure, Thompson, and Csukas settled their
claims prior to this appeal.
                            6
marketing manager in Pittsburgh, Pennsylvania; and John
Titus, a former Area Services Manager in Irving, Texas.

       On June 26, 2013, the case was transferred to
another district judge. PGW filed a motion to decertify
the collective action. On March 31, 2014, the District
Court granted the motion, concluding that the collective
action should be decertified because the opt-in plaintiffs’
claims are factually dissimilar from those of the named
plaintiffs. See Karlo, 

.

      PGW then filed motions to exclude plaintiffs’
experts. Of relevance to this appeal, PGW sought to
exclude three areas of expert testimony. First, Dr.
Michael Campion was prepared to offer statistical
evidence in favor of plaintiffs’ disparate-impact claim.
Second, Dr. Campion intended to offer his expert opinion
on “reasonable” human-resources practices during a RIF.
And third, Dr. Anthony G. Greenwald proposed to testify
as to age-related implicit-bias studies. By Order dated
July 13, 2015, the District Court excluded the testimony
of each. See Karlo, 

.

      PGW moved for summary judgment on each
claim. On September 3, 2015, the District Court ruled on
the motions, granting them in part and denying them in
part. See Karlo, 

. As to plaintiffs’
disparate-impact claims, the District Court granted
summary judgment on two grounds: (1) plaintiffs’ fifty-
and-older disparate-impact claim is not cognizable under
                            7
the ADEA; and (2) plaintiffs’ lack of evidence to support
their claim of disparate impact following the exclusion of
Dr. Campion’s statistics-related testimony. The District
Court also granted summary judgment as to plaintiffs’
disparate-treatment claims. That ruling has not been
appealed. Finally, the District Court denied summary
judgment as to Karlo’s and McLure’s individual
retaliation claims.

       On October 2, 2015, the District Court certified the
disparate-impact and disparate-treatment claims for final
judgment pursuant to Rule 54(b) of the Federal Rules of
Civil Procedure. See Karlo, 

. This
appeal followed. Plaintiffs seek reversal of the District
Court’s summary judgment decision and statistics-related
Daubert ruling regarding their disparate-impact claims.
Plaintiffs also appeal the District Court’s other Daubert
rulings and its order decertifying the collective action.

                            II

      The District Court had jurisdiction pursuant to 28
U.S.C. § 1331. We have jurisdiction pursuant to 28
U.S.C. § 1291.

      The parties dispute whether our jurisdiction
extends to one or all named plaintiffs. PGW concedes
that Karlo perfected an appeal, but argues that the other
remaining named plaintiffs—Cunningham, Marietti, and
Meixelberger—were not identified in the Notice of
                            8
Appeal, and therefore did not preserve their appellate
rights under Rule 3(c) of the Federal Rules of Appellate
Procedure. See Torres v. Oakland Scavenger Co., 

, 317 (1988).2 We conclude that plaintiffs
complied with Rule 3(c) with respect to all named
plaintiffs.

        Rule 3(c)(1)(A) requires a notice of appeal to
“specify the party or parties taking the appeal by naming
each one in the caption or body of the notice,” but that
rule is relaxed where “an attorney [is] representing more
than one party.” Fed. R. App. P. 3(c)(1)(A). The attorney
“may describe those parties with such terms as ‘all
plaintiffs,’ ‘the defendants,’ ‘the plaintiffs A, B, et al.,’ or
‘all defendants except X.’” 

of Appeal here states, “Plaintiffs in the
above-captioned case hereby appeal . . . an order . . .
entering judgment against Plaintiffs . . . on Plaintiffs’
discrimination claims . . . .” A.1 (emphases added). The

       2
         Although it does not influence our analysis, the
Notice of Appeal’s caption has a perfectly innocent
explanation. Following Rule 54(b) certification, the
District Court amended the caption by order to identify
only Karlo and McLure as plaintiffs because their
retaliation claims were the only claims remaining after
summary judgment. The Notice of Appeal used the
District Court’s updated caption.
                               9
use of “Plaintiffs” is equivalent to “the defendants” in the
example provided by the Rule.3 We have observed that
“[t]he purpose of Rule 3(c)’s identification requirement is
to provide notice to the court and the opposing parties of
the identity of the appellants.” In re Cont’l Airlines, 

, 129 (3d Cir. 1997). Because all of the
remaining named plaintiffs were identically situated as to
this appeal, were represented by the same counsel, and
were each identified by name in the District Court’s
“order . . . entering judgment against [all named]
Plaintiffs,” as referenced on the face of the Notice, Rule
3(c)’s purpose is amply served, and “the intent to appeal
is otherwise clear from the notice.” Fed. R. App. P.
3(c)(4); see United States v. Carelock, 

, 441
(3d Cir. 2006) (“The Supreme Court has stated that
courts should ‘liberally construe the requirements of Rule
3.’” (quoting Smith v. Barry, 

, 248 (1992))).

                            III

       The central question in this case is whether so-
called “subgroup” disparate-impact claims are cognizable
under the ADEA. We hold that they are.


      3
         The phrase “in the above captioned case” does
not change our interpretation. We read the Notice to
mean “Plaintiffs in [Civil Action No. 10-1283] hereby
appeal.”
                            10
      Disparate-impact claims in ADEA cases ordinarily
evaluate the effect of a facially neutral policy on all
employees who are at least forty years old—that is, all
employees covered by the ADEA. In this case, plaintiffs
claim to have identified a policy that disproportionately
impacted a subgroup of that population: employees older
than fifty. But because the policy favored younger
members of the protected class, adding those individuals
to the comparison group washes out the statistical
evidence of a disparity.

        Plaintiffs’ claim is cognizable under the ADEA.
Specifically, we hold that an ADEA disparate-impact
claim may proceed when a plaintiff offers evidence that a
specific, facially neutral employment practice caused a
significantly disproportionate adverse impact based on
age. Plaintiffs can demonstrate such impact with various
forms      of     evidence,    including   forty-and-older
comparisons, subgroup comparisons, or more
sophisticated statistical modeling, so long as that
evidence meets the usual standards for admissibility. A
contrary rule would ignore significant age-based
disparities. Where such disparities exist, they must be
justified pursuant to the ADEA’s relatively broad
defenses.

                            A

     We begin with an overview of the statutory
scheme. The Age Discrimination in Employment Act of
                           11
1967, 81 Stat. 602, as amended, 29 U.S.C. § 621 et seq.,
makes it unlawful for an employer:

      (1) to fail or refuse to hire or to discharge any
      individual or otherwise discriminate against
      any individual with respect to his
      compensation,        terms,     conditions,    or
      privileges of employment, because of such
      individual’s age;

      (2) to limit, segregate, or classify his
      employees in any way which would deprive
      or tend to deprive any individual of
      employment opportunities or otherwise
      adversely affect his status as an employee,
      because of such individual’s age; or

      (3) to reduce the wage rate of any employee
      in order to comply with this chapter.

29 U.S.C. § 623(a). “Except for substitution of the word
‘age,’ for the words ‘race, color, religion, sex, or national
origin,’ the language of that provision in the ADEA is
identical to that found in § 703(a)(2) of the Civil Rights
Act of 1964 (Title VII).” Smith v. City of Jackson, 

, 233 (2005). But unlike Title VII, which
protects individuals of every race, color, religion, sex,
and national origin, the ADEA’s protection is “limited to
individuals who are at least 40 years of age.” 29 U.S.C.
§ 631(a).
                             12
       ADEA claims may proceed under a disparate-
impact or disparate-treatment theory. See 

–32. Disparate treatment is governed by
§ 623(a)(1); disparate impact is governed by § 623(a)(2).

(plurality opinion); cf. Watson v. Fort Worth
Bank & Trust, 

, 991 (1988); Connecticut v.
Teal, 

, 446–47 (1982).

       The disparate-impact theory of recovery was first
recognized in Griggs v. Duke Power Co., 

(1971), a Title VII case. Unlike claims of disparate
treatment, disparate-impact claims do not require proof
of discriminatory intent. Disparate impact redresses
policies that are “fair in form, but discriminatory in
operation.” 

To that end, disparate-impact
claims “usually focus[] on statistical disparities . . . .”

.

       To state a prima facie case for disparate impact
under the ADEA, a plaintiff must (1) identify a specific,
facially neutral policy, and (2) proffer statistical evidence
that the policy caused a significant age-based disparity.
Cf. NAACP v. N. Hudson Reg’l Fire & Rescue, 

, 476–77 (3d Cir. 2011). Once a plaintiff establishes a
prima facie case, an employer can defend by arguing that
the challenged practice was based on “reasonable factors
other than age”—commonly referred to as the “RFOA”
defense. 29 U.S.C. § 623(f)(1); 29 C.F.R. § 1625.7.


                             13
       “[T]he scope of disparate-impact liability under the
ADEA is narrower than under Title VII” because of
“[t]wo textual differences” between the statutes. 

. First, the RFOA defense imposes a
lighter burden on the employer than its Title VII
counterpart, the “business necessity” defense. Under the
ADEA, the employer only needs to show that it relied on
a “reasonable” factor, not that “there are [no] other ways
for the employer to achieve its goals . . . .” 

. Congress’s decision to impose a relatively
light burden on employers “is consistent with the fact that
age, unlike race or other classifications protected by Title
VII, not uncommonly has relevance to an individual’s
capacity to engage in certain types of employment.” 

The second textual difference requires ADEA
plaintiffs to “isolat[e] and identify[] the specific
employment practices that are allegedly responsible for
any observed statistical disparities.” 

(quoting
Wards Cove Packing Co. v. Atonio, 

, 656
(1989)). Congress stripped that requirement from Title
VII when it amended the statute in 1991, but it remains
operative under the ADEA. 

see 42 U.S.C. §
2000e–2(k).

                             B

     The ADEA’s disparate-impact provision makes it
unlawful for an employer “to adversely affect [an
employee’s] status . . . because of such individual’s age.”
29 U.S.C. § 623(a)(2). This plain text supports the
                            14
viability of subgroup claims. See Hardt v. Reliance
Standard Life Ins. Co., 

, 251 (2010) (“We
must enforce plain and unambiguous statutory language
according to its terms.”). Two aspects of the text guide
our decision in this case: (1) the focus on age as the
relevant protected trait, as interpreted by O’Connor v.
Consolidated Coin Caterers Corp., 

(1996),
and (2) the focus on the rights of individuals, as
interpreted by Connecticut v. Teal, 

(1982).
Our interpretation is further supported by the ADEA’s
remedial purpose.

                           1

       We begin with the Supreme Court’s unanimous
opinion in O’Connor v. Consolidated Coin Caterers
Corp., 

(1996), an ADEA disparate-
treatment case. O’Connor clarified that the ADEA
proscribes age discrimination, not forty-and-over
discrimination. The same interpretation applies to
identical operative language in the ADEA’s disparate-
impact provision.

      The plaintiff in O’Connor was fifty-six years old
when he was fired and replaced with a younger 

. The plaintiff’s replacement, however,
was over the age of forty, and therefore within the class
of individuals protected by the ADEA. 

Circuit held that the ADEA prima facie case requires the

                           15
replacement to be younger than forty years old. 

The Supreme Court reversed.

       The Supreme Court began its analysis with the
plain text of the statute: “The discrimination prohibited
by the ADEA is discrimination ‘because of [an]
individual’s age,’ though the prohibition is ‘limited to
individuals who are at least 40 years of 

(alteration in original) (citations omitted). On the
basis of that text, the Court held that the ADEA

       does not ban discrimination against
      employees because they are aged 40 or
      older; it bans discrimination against
      employees because of their age, but limits
      the protected class to those who are 40 or
      older. The fact that one person in the
      protected class has lost out to another person
      in the protected class is thus irrelevant, so
      long as he has lost out because of his age.

ADEA protects a class of individuals at
least forty years old, it “prohibits discrimination on the
basis of age and not class membership . . . .” 

It
is therefore “utterly irrelevant” that the beneficiary of age
discrimination was also over the age of forty. 

Accordingly, the proposed limitation on the prima facie
case—replacement by an employee younger than forty—
lacked a “logical connection” to the plain text of the
ADEA. 

As the Supreme Court later
                             16
reaffirmed, “[it] is beyond reasonable doubt[] that the
ADEA was concerned to protect a relatively old worker
from discrimination that works to the advantage of the
relatively young.” Gen. Dynamics Land Sys., Inc. v.
Cline, 

, 590–91 (2004).

       The Supreme Court’s reasoning ineluctably leads
to our conclusion that subgroup claims are cognizable.
Simply put, evidence that a policy disfavors employees
older than fifty is probative of the relevant statutory
question: whether the policy creates a disparate impact
“because of such individual[s’] age.” 29 U.S.C.
§ 623(a)(2). Requiring the comparison group to include
employees in their forties has no “logical connection” to
that prohibition. 

.

      The key insight from O’Connor is that the forty-
and-older line drawn by § 631(a) constrains the ADEA’s
general scope; it does not modify or define the ADEA’s
substantive prohibition against “discriminat[ion] . . .
because of such individual’s age.” § 623(a)(1). The
ADEA protects against “age discrimination []as opposed
to ‘40 or over’ discrimination . . . .” 

.

      The disparate-impact provision uses the same
operative phrase, “because of such individual’s age.” 29
U.S.C. § 623(a)(2). Our interpretation of it, therefore,
should be consistent with our interpretation of the
disparate-treatment provision, § 623(a)(1). See, e.g.,
                           17
Dep’t of Revenue of Or. v. ACF Indus., Inc., 

, 342 (1994) (“[I]dentical words used in different
parts of the same act are intended to have the same
meaning.”). Thus, “adversely affect . . . because of such
individual’s age” must mean adversely affect based on
age, not adversely affect based on forty-and-older status.4

       O’Connor’s applicability is not diminished by the
fact that it addressed a disparate-treatment claim. As
demonstrated by the identical operative phrasing of
§ 623(a)(1) and § 623(a)(2), the two types of claims share
the same “ultimate legal issue . . . .” 

; see also 

–56 (discussed infra
Section III.B.2). Disparate-impact claims are primarily
distinguished by “the factual issues that typically
dominate”—namely, whether a facially neutral policy is

      4
         In Cline, the Supreme Court interpreted the word
“age” to take different meanings in different parts of the

. The Court distinguished
between two alternative definitions: “any number of
years lived, or . . . the longer span and concurrent aches
that make youth look good.” 

determined
that “[t]he presumption of uniform usage thus relents”
when comparing § 623(a)(1) and § 623(f). 

In
this case, the presumption holds because § 623(a)(1) and
§ 623(a)(2) employ “age” in virtually the same context.
Both use “age” as in Cline’s second definition, to mean
“old 

.
                            18
discriminatory in operation. 

(emphasis added); see Tex. Dep’t of Cmty. Affairs v.
Burdine, 

, 252 n.5 (1981). A disparate
impact “may in operation be functionally equivalent to
intentional discrimination.” 

; see
Tex. Dep’t of Hous. & Cmty. Affairs v. Inclusive Cmtys.
Project, Inc., 

, 2522 (2015) (“[D]isparate-
impact liability . . . plays a role in uncovering
discriminatory intent[.]”). Our holding restores the parity
described in Watson. Under the ADEA, both disparate
impact and disparate treatment address the same ultimate
legal issue: age discrimination.

       We conclude that the Supreme Court’s analysis in
O’Connor answers the question now before us. A
specific, facially neutral policy that significantly
disfavors employees over fifty years old supports a claim
of disparate impact under the plain text of § 623(a)(2).
Although the employer’s policy might favor younger
members of the forty-and-over cohort, that is an “utterly
irrelevant factor,” 

, in
evaluating whether a company’s oldest employees were
disproportionately affected because of their age.

                            2

       Our decision is further supported by the Supreme
Court’s opinion in Connecticut v. Teal, 

(1982), a Title VII disparate-impact case. Teal confirms
that, even under a disparate-impact theory, the plain text
                            19
of the statute is designed to protect the rights of
individual employees, not the rights of a class.

        In Teal, a Connecticut state agency used a two-step
process to determine eligibility for promotions. First,
Connecticut required applicants to take a written test.
Second, Connecticut selected the employees for
promotion out of the pool of candidates that passed the
test. 

Black applicants who failed the test sued,
advancing evidence that black employees failed the
written test at a significantly higher rate than white
employees. In response, Connecticut argued that, at the
second step of the process, the black employees who
passed were given preferential treatment through an
affirmative action program, counterbalancing the
discriminatory effect of the written test. Connecticut
argued that its two-step process promoted black
employees at an overall higher rate than white
employees. 

       The Supreme Court rejected this so-called
“bottom-line” defense and held that the purpose of Title
VII “is the protection of the individual employee, rather
than the protection of the minority group as a whole.” 

“[F]avorable treatment of . . . members of
these respondents’ racial group” did not justify
discrimination against other members of the protected
class. 

see El v. Se. Pa. Transp. Auth., 

, 239–40 (3d Cir. 2007) (“Title VII operates not
primarily to the benefit of racial or minority groups, but
                            20
to ensure that individual applicants receive the
consideration they are due . . . .”).

       This case presents a similar issue. The ADEA, like
Title VII, protects individuals who are members of a
protected class, not a class itself. See 29 U.S.C.
§ 623(a)(1) (proscribing forms of discrimination
“because of such individual’s age”); 

(same); 

(limiting the ADEA’s scope to
“individuals who are at least 40 years of age”). Such
protection under the statute does not disappear when a
plaintiff advances a disparate-impact claim. Teal
prohibits the use of a bottom-line statistic to justify
ignoring a disproportionate impact against individuals
that would otherwise be actionable under the plain text of
the statute. That is precisely the problem subgroups are
meant to address here.

     As a result, Teal answers PGW’s argument that
employees older than forty were, as a class, favored to
keep their jobs. That is equivalent to Connecticut’s
argument that black employees were collectively favored
for promotions. The Supreme Court rejected that
argument in Teal, and we reject it here.

       Similar to the position of PGW and its amici in this
case, the dissenting Justices in Teal accused the majority
of “confus[ing] the distinction—uniformly recognized
until today—between disparate impact and disparate

                            21


(Powell, J., dissenting). The
majority responded as follows:

      The fact remains . . . that irrespective of the
      form taken by the discriminatory practice, an
      employer’s treatment of other members of
      the plaintiffs’ group can be of little comfort
      to the victims of . . . discrimination. Title VII
      does not permit the victim of a facially
      discriminatory policy to be told that he has
      not been wronged because other persons of
      his or her race or sex were hired. That answer
      is no more satisfactory when it is given to
      victims of a policy that is facially neutral but
      practically discriminatory. Every individual
      employee is protected against both
      discriminatory treatment and practices that
      are fair in form, but discriminatory in
      operation.

(internal citations and quotation marks
omitted). The same reasoning applies to this case. The
ADEA “does not permit the victim of a facially
discriminatory policy to be told that he has not been
wronged because other persons” aged forty or older were
preferred. 

“That answer is no more
satisfactory when it is given to victims of a policy that is
facially neutral but practically discriminatory.” 

       PGW and its amici maintain that disparate-impact
claims generally rely on comparisons between entire
classes. Even in Teal, for example, plaintiffs’ evidence
showed that the written test caused a disparate impact on
black employees as a 

. That general
focus on groups, however, is explained by the fact that
Title VII protects group identities like race and sex. The
trait protected by the ADEA, age, is qualitatively
different.

       “The term ‘age’ employed by the ADEA is not . . .
comparable to the terms ‘race’ or ‘sex’ employed by Title
VII.” 

. Age is a continuous
variable, whereas race and sex are treated categorically in
the mine-run of Title VII cases. See Bienkowski v. Am.
Airlines, Inc., 

, 1506 (5th Cir. 1988) (“The
ADEA does not lend itself to a bright-line age rule and in
this respect differs from racial or sex discrimination
cases . . . .”); Goldstein v. Manhattan Indus., Inc., 

, 1442 (11th Cir. 1985) (observing that age
discrimination is “qualitatively different from race or sex
discrimination” because “the basis of the discrimination
is not a discre[te] and immutable characteristic of an
employee which separates the members of the protected
group indelibly from persons outside the protected
group”).


                            23
         On account of that difference, the statistical
techniques common in Title VII cases are not perfectly
transferable to ADEA cases. If, for example, the
comparison group in Teal omitted some black employees
who took the written test, the statistics would likely have
failed to address whether there was a disparate impact
“because of . . . race . . . .” 42 U.S.C. § 2000e–2(a)(2);
see also 

It would be unclear
whether the test’s effects fell more harshly on individuals
of a particular race without looking at how the test
affected all members. But with the ADEA, by contrast, a
comparison group that omits employees in their forties is
fully capable of demonstrating disparate impact “because
of . . . age.” 29 U.S.C. § 623(a)(2).

        The forty-and-older line established in § 631(a)
does not convert age into a binary trait. By its own terms,
it imposes a “limit[ation]” on the “individuals” covered
by “[t]he prohibitions in this chapter . . . .” 29 U.S.C.
§ 631(a). It simply establishes “the age at which ADEA
protection begins.” Maxfield v. Sinclair Int’l, 

, 792 (3d Cir. 1985). The appropriate disparate-
impact statistics should be guided by the trait protected
by the statute, not the population of employees inside or
outside the statute’s general scope. In fact, when the
Supreme Court recognized ADEA disparate-impact
liability in Smith, nothing in its reasoning turned on the



                            24
existence or purpose of § 631(a). That provision was not
cited once.5

       PGW and its amici would have us rewrite 29
U.S.C. § 623(a)(2) to proscribe “adverse[] effect[s] . . .
because of such individual’s [membership in the 40-and-
older class].” That interpretation would bring the ADEA
closer to more familiar Title VII territory, but “[w]e have
to read it the way Congress wrote it.” Meacham v. Knolls
Atomic Power Lab., 

, 101–02 (2008). The
continuous, non-categorical nature of age cannot be
adequately addressed by simply aggregating forty-and-
older employees. More exacting analysis may be needed
in certain cases, and subgroups may answer that need.

                            4

       Finally, our decision is supported by the ADEA’s
remedial purpose. Refusing to recognize subgroup claims
would deny redress for significantly discriminatory
policies that affect employees most in need of the
ADEA’s protection.



      5
        To be sure, the plaintiffs in Smith happened to
rely on forty-and-older 

.
Nothing in our opinion should be read to rule out such
evidence. Nonetheless, Smith’s reasoning does not
foreclose subgroup claims.
                            25
       Mandating a forty-and-older comparison group
“would allow an employer to adopt facially neutral
policies which had a profoundly disparate impact on
individuals over age 50 or 55,” so long as younger
individuals within the protected class received
sufficiently favorable treatment. Finch v. Hercules Inc.,

, 1129 (D. Del. 1994). Such policies
“reflect the specific type of arbitrary age discrimination
Congress sought to prohibit,” but would nonetheless
evade judicial scrutiny. Id.; see also Graffam v. Scott
Paper Co., 

, 4 (D. Me. 1994).

        We have also acknowledged in the disparate-
treatment context that “[i]f no intra-age group protection
were provided by the ADEA, it would be of virtually no
use to persons at the upper ages of the protected
class . . . .” 

. The same rationale
applies to the disparate-impact context. The older the
employees affected by a policy, the more confounding
favoritism would be included in the rigid forty-and-older
sample. Thus, an impact on employees in their seventies
may be easier to average out of existence compared to an
impact that also affects younger employees. Mandating
forty-and-older comparisons would predominantly harm
“those most in need of the statute’s protection.” Lowe v.
Commack Union Free Sch. Dist., 

, 1379
(2d Cir. 1989) (Pierce, J., dissenting in relevant part).
“[I]t would indeed be strange, and even perverse, if the
youngest members of the protected class were to be

                            26
accorded a greater degree of statutory protection than
older members of the class.” Id.6

      Accordingly, our interpretation of the ADEA is
supported not only by the statute’s text and Supreme
Court precedent, but also by the ADEA’s purpose.

                            C

      Our holding in this case is at odds with decisions
from three of our sister circuits. See Lowe v. Commack
Union Free Sch. Dist., 

(2d Cir. 1989);
Smith v. Tenn. Valley Auth., 

, 

(6th Cir. 1991) (table opinion); E.E.O.C. v.
McDonnell Douglas Corp., 

(8th Cir.



      6
        Ironically, mandating a forty-and-older sample
has the potential to harm employers in certain
circumstances. For example, if a substantial disparate
impact is experienced only by individuals sixty-five and
older, the effect can show up in the forty-and-older
aggregate statistic, creating the misimpression that forty-
year-old plaintiffs were disparately impacted. See
Ramona L. Paetzold & Steve L. Willborn, The Statistics
of Discrimination: Using Statistical Evidence in
Discrimination Cases § 7:2, at 340 (2016–2017 ed. 2016)
(noting that “[t]he errors can occur in either direction”
when relying on forty-and-older comparisons).
                            27
1999).7 Those decisions have primarily relied on policy
considerations that we do not find persuasive. In short,
they are contradicted by O’Connor and Teal, confuse
evidentiary concerns with statutory interpretation, and
incorrectly assume that recognizing subgroups will
proliferate liability for reasonable employment practices.



      7
        While we are generally reluctant to create circuit
splits, we do so where a “compelling basis” exists.
Wagner v. PennWest Farm Credit, ACA, 

,
912 (3d Cir. 1997). For the reasons discussed in this
opinion, we think a compelling basis exists in this case.
Even so, we note that (1) the Second Circuit and Sixth
Circuit cases predate the Supreme Court’s decisions in
O’Connor and Smith; (2) the Sixth Circuit case is non-
precedential; and (3) the Eighth Circuit case predates
Smith. One circuit has noted the issue but declined to
rule. See Katz v. Regents of the Univ. of Calif., 

, 835–36 (9th Cir. 2000).
       PGW and its amici argue that we already decided
this question in Massarsky v. General Motors Corp., 

, 121 (3d Cir. 1983). We did not. The plaintiff in
Massarsky failed to advance any evidence of disparate
impact. 

contemplation of a specific age
group is dicta not binding on this panel. In any event,
Massarsky predates the Supreme Court’s decisions in
both O’Connor and Smith.
                           28
                              1

       The United States Court of Appeals for the Second
Circuit addressed disparate-impact subgroups in Lowe v.
Commack Union Free Sch. Dist., 

(2d Cir.
1989). See also Criley v. Delta Air Lines, Inc., 

, 105 (2d Cir. 1997). Because Lowe predates
O’Connor, it gives improper significance to the forty-
and-older line drawn by § 631(a), and fails to compare
the textual similarities between § 623(a)(1) and
§ 623(a)(2). Lowe also rejects subgroup claims because
specific types of evidence could be misleading. We do
not find Lowe persuasive.

       The Second Circuit’s legal analysis begins with the
premise that disparate-treatment analysis in Title VII
cases “generally has focused . . . on the protected group
of which plaintiff is a member.” 

.
But Lowe does not address the text of § 623(a)(2).
Divorced from that text, the Second Circuit allows the
“general[] . . . focus[]” of a different statute to limit what
this statute plainly permits. Lowe does not, and cannot,
explain why forty-and-older group membership is
“utterly irrelevant” to discrimination based on age,




                             29


, but is the sine qua non of an
adverse effect based on age.8

      Lowe is primarily concerned with the practical
implications of subgroup claims.9 Its main objection is

      8
          Lowe’s treatment of Teal is similarly
unpersuasive. See 

. We addressed
the same argument in Section 

.
      9
          For example, Lowe offers a flawed reductio ad
absurdum: “an 85 year old plaintiff could seek to prove a
discrimination claim by showing that a hiring practice
caused a disparate impact on the ‘sub-group’ of those age
85 and above, even though all those hired were in their
late seventies.” 

. This argument
relies on the false assumption that an employer who
favors 70-year-old employees could not possibly be
liable under the ADEA. “If an 85-year old person . . .
fails to attain that position for no reason other than age,
s/he has suffered age discrimination under the Act . . . .”

(Pierce, J., dissenting in relevant part). In any
event, we can be reasonably assured that such a
hypothetical would never arise due to the demographic
characteristics of the workforce, which limit the
statistical power to compare impacts on seventy- and
eighty-year-old employees. See Sandra F. Sperino, The
Sky Remains Intact: Why Allowing Subgroup Evidence Is
Consistent with the Age Discrimination in Employment
Act, 90 Marq. L. Rev. 227, 263 (2006).
                             30
evidentiary: “any plaintiff can take his or her own age as
the lower end of a ‘sub-protected group’ and argue that
said ‘sub-group’ is disparately impacted.” 

. Here, PGW and its amici similarly argue
that plaintiffs will be able to “gerrymander” arbitrary age
groups in order to manufacture a statistically significant
effect. We disagree.

       Essentially, PGW and its amici argue that a
particular form of evidence carries such a high risk of
manipulation that we should interpret the ADEA to
preclude the entire claim. That is a thoroughly
unsatisfactory justification for ignoring statutory text and
Supreme Court precedent.10 Our interpretation of the
ADEA is based on text, not evidentiary gatekeeping.
That function is capably performed by district judges
who routinely apply the Federal Rules of Evidence and
Daubert jurisprudence. We consider that to be a
sufficient safeguard against the menace of unscientific
methods and manipulative statistics.


      10
          The Eighth Circuit, which ultimately agreed
with Lowe’s outcome, did criticize the Second Circuit’s
reasoning on this point. See E.E.O.C. v. McDonnell
Douglas Corp., 

, 950–51 (8th Cir. 1999)
(“The fact that a particular interpretation of a statute
might spawn lawsuits is not a reason to reject that
interpretation.”).
                            31
       Preliminarily, PGW’s “gerrymandering” objection
only applies to the kind of statistical studies that compare
subgroups selected by an expert. Some scholars have
proposed the use of statistical models that treat age as a
continuous variable and thus avoid the need to draw
“arbitrary” age groups. Options discussed in the literature
include proportional hazards models and logistic
regression. See Ramona L. Paetzold & Steve L. Willborn,
The Statistics of Discrimination: Using Statistical
Evidence in Discrimination Cases § 7:11, at 372 (2016–
2017 ed. 2016) [hereinafter Paetzold & Willborn]; see
also, e.g., George Woodworth & Joseph Kadane, Age-
and Time-Varying Proportional Hazards Models for
Employment Discrimination, 4 Annals Applied Statistics
1139 (2010); Michael O. Finkelstein & Bruce Levin,
Proportional Hazard Models for Age Discrimination
Cases, 34 Jurimetrics J. 153 (1994).

        We have no need today to bless any one approach.
“Statistics ‘come in infinite variety and . . . their
usefulness depends on all of the surrounding facts and
circumstances.’” 

n.3 (quoting
Teamsters v. United States, 

, 340 (1977)).
Our purpose is rather to demonstrate that the
gerrymandering objection exposes a weakness in one
particular research method, not a cause of action. “The
continuous nature of the age variable need not be a
statistical problem under disparate-impact analysis;
existing statistical procedures can be adapted to the

                            32
specific needs of disparate-impact analysis.” Paetzold &
Willborn § 7:11, at 373.

      Even if the statistical evidence in an ADEA case
uses age groups selected by the expert, PGW and its
amici overstate the risk of manipulation. “The claim can
be analyzed, of course, to determine if the result is robust
across various age breaks and whether the age breaks can
be justified independently of the data . . . .” 

at
344. In fact, some courts have long permitted statistical
subgroup evidence in the context of disparate-treatment
claims. See, e.g., Barnes v. GenCorp Inc., 

,
1466–67 (6th Cir. 1990). We see no reason why that
same evidence would be any less workable in a disparate-
impact case. See MacNamara v. Korean Air Lines, 

, 1148 (3d Cir. 1988) (“[T]he statistical
evidence supporting a claim of disparate impact often
resembles that used to help establish disparate
treatment.”).

       So-called “age-break” analysis has well-
understood limitations. See Paetzold & Willborn § 7:3, at
343–46. For example, if an expert does not devise the age
breaks independently of the data, and instead cherry-
picks groups to manufacture a particular result, that “may
invalidate the usual tests of statistical significance.” 

In addition, “the appropriate inference for plaintiffs
near a selected age break is always likely to be
problematic.” 

Without more, this challenge
may undermine the claims of plaintiffs who “take [their]
                              33
own age as the lower end of a ‘sub-protected group’ and
argue that said ‘sub-group’ is disparately impacted.”

; see also 

–30 (“If a plaintiff attempts to define the subset too
narrowly, he or she will not be able to obtain reliable
statistics upon which to prove a prima facie case.”).

       The EEOC and plaintiffs have only argued in favor
of subgroups with “lower boundaries,” not “upper
boundaries.” Oral Arg. Tr. 23:24–24:3. That rule would
preclude, for example, a “banded” 50-to-55 subgroup.
We think that limitation is well founded. A plaintiff
would benefit from introducing an upper boundary if a
policy favored employees older than that limit. But in
Cline, the Supreme Court interpreted the term “age” in
§ 623(a)(1) to mean “old 

. Under
Cline, the ADEA protects only “relatively old worker[s]
from discrimination that works to the advantage of the
relatively young.” 

If a facially neutral
policy systematically favors a company’s oldest
employees,11 that fact may be fatal to a claim that
      11
          We note that the ability to draw inferences about
the treatment of a company’s oldest employees may be
limited by sample size. In this case, for example,
plaintiffs’ expert argues that there is no statistically
significant effect on employees age sixty and older
because “[t]here are only 14 terminations, which means
the statistical power to detect a significant effect is very
low.” A.244–45.
                            34
members of a younger subgroup were disparately
impacted because of their “old age.” 

see Tex.
Dep’t of Hous. & Cmty. Affairs v. Inclusive Cmtys.
Project, Inc., 

, 2523 (2015) (describing
the importance of “[a] robust causality requirement” in
disparate-impact cases); Gross v. FBL Fin. Servs., Inc.,

, 176 (2009). Thus, a banded subgroup
would be self-defeating under Cline, further limiting
plaintiffs’ ability to gerrymander age groups.

       We reject the notion that the risk of gerrymandered
evidence is so great that it can override what the text of
the statute otherwise permits. District courts should, as in
any other case, ensure that plaintiffs’ evidence is reliable
under Daubert and provides more than the “mere scintilla
of evidence” needed to survive summary judgment. S.H.
ex rel. Durrell v. Lower Merion Sch. Dist., 

,
256 (3d Cir. 2013) (internal quotation marks omitted)
(quoting Jakimas v. Hoffman-La Roche, Inc., 

, 777 (3d Cir. 2007)). Accordingly, we are not
persuaded by Lowe’s legal or practical groundings.

                             2

      The United States Court of Appeals for the Sixth
Circuit addressed disparate-impact subgroups in a non-
precedential opinion, Smith v. Tennessee Valley
Authority, 

, 

(6th Cir.
1991) (table opinion). This decision also predates
O’Connor. Its reasoning contradicts both O’Connor and
                            35
Teal, and conflicts with a precedential Sixth Circuit
opinion that allows subgroup analysis in disparate-
treatment cases.

       In Smith, the Sixth Circuit asserts by citation to
Lowe that “[a] plaintiff cannot succeed under a disparate
impact theory by showing that younger members of the
protected class were preferred over older members of the
protected class.” 

As we have discussed, Lowe’s
reasoning is explicitly rejected by O’Connor and Teal.
Teal held that a plaintiff can succeed under a disparate-
impact theory if other members of the protected class
were 

, and O’Connor held that
forty-and-older status is irrelevant to evaluating the
application of a protection based on 

.

       As we have also noted, the Sixth Circuit has long
recognized statistical subgroup evidence in disparate-
treatment claims. In a precedential opinion, Barnes v.
GenCorp Inc., 

(6th Cir. 1990), the Sixth
Circuit specifically rejected the defendant’s argument
that “the only valid statistics would necessarily divide the
employees into groups age 40-and-over and those under
40.” 

The Sixth Circuit suggests in a footnote,
by citation to Lowe, that “[s]uch sub-group analysis may
not apply to discriminatory impact cases[.]” 

n.12. With the exception of that speculative footnote, we
find the Sixth Circuit’s decision in Barnes more
persuasive than its decision in Smith.
                            36
                           3

      Finally, the United States Court of Appeals for the
Eighth Circuit addressed disparate-impact subgroups in
E.E.O.C. v. McDonnell Douglas Corp., 

(8th Cir. 1999). The Eighth Circuit’s analysis is also
unpersuasive because it contradicts Teal and ignores
important limitations on the scope of disparate-impact
claims.

      First, the Eighth Circuit argued that if subgroup
claims were cognizable,

      a plaintiff could bring a disparate-impact
      claim despite the fact that the statistical
      evidence indicated that an employer’s RIF
      criteria had a very favorable impact upon the
      entire protected group of employees aged 40
      and older, compared to those employees
      outside the protected group. We do not
      believe that Congress could have intended
      such a result.

This is no more than an endorsement of the
bottom-line defense that the Supreme Court rejected in
Teal. The State of Connecticut tried a similar argument
by suggesting that black employees were favored for
promotions as an overall class. But that bottom-line
outcome concealed individual rights violations. Far from
being a result “Congress could [not] have intended,” 

Supreme Court’s ruling in Teal vindicated Title VII’s
plain text and purpose. The same applies to the ADEA.

      Second, the Eighth Circuit panel wrote:

      [T]he consequence would be to require an
      employer engaging in a RIF to attempt what
      might well be impossible: to achieve
      statistical parity among the virtually infinite
      number of age subgroups in its work force.
      Adoption of such a theory, moreover, might
      well have the anomalous result of forcing
      employers to take age into account in making
      layoff decisions, which is the very sort of
      age-based decision-making that the statute
      proscribes.

McDonnell Douglas 

.12

       Even without the prospect of subgroups, it has
always been the case that “a completely neutral practice
will inevitably have some disproportionate impact on one
group or another.” DiBiase v. SmithKline Beecham
Corp., 

, 731–32 (3d Cir. 1995) (quoting City
of L.A., Dep't of Water & Power v. Manhart, 

        PGW’s amici make an opposite argument—that
employers already fine-tune employment decisions to
avoid creating a disparate impact, and our decision will
make it more costly or complicated for them to do so.
                            38
702, 710 n.20 (1978)). That is precisely why deviating
from statistical parity is not, by itself, enough to incur
disparate-impact liability. Just last Term, the Supreme
Court recognized that “disparate-impact liability has
always been properly limited in key respects” so that it is
not “imposed based solely on a showing of a statistical
disparity.” Inclusive Cmtys. Project, 

; see also Wards 

(showing a
statistical disparity alone “will not suffice to make out a
prima facie case of disparate impact”); 

(plurality opinion) (“[P]laintiff’s burden in
establishing a prima facie case goes beyond the need to
show that there are statistical disparities in the
employer’s work force.”).

        To make out a prima facie case, plaintiffs must
first identify a specific employment practice that causes
the disparity. See Wards Cove, 

; see also

(noting that the Wards Cove
holding remains in effect under the ADEA). The
Supreme Court has recognized that this requirement
guards against “the myriad of innocent causes that may
lead to statistical imbalances.” 

(quoting Wards 

). “Identifying a
specific practice is not a trivial burden . . . .” Meacham v.
Knolls Atomic Power Lab., 

, 101 (2008); see
Inclusive Communities 

(“[A]
disparate-impact claim that relies on a statistical disparity


                             39
must fail if the plaintiff cannot point to a defendant’s
policy or policies causing that disparity.”).

       Furthermore, not just any disparity will make out
the prima facie case; the disparity must be significant.
See 

(“[S]tatistical disparities
must be sufficiently substantial that they raise such an
inference of causation.”); 

(“[T]he
facially neutral employment practice [must have] had a
significantly discriminatory impact.”); Wards 

(requiring a “significantly disparate impact”);
Hazelwood Sch. Dist. v. United States, 

,
307–08 (1977) (requiring “gross statistical disparities”).
We have not adopted a uniform rule for what this
requirement entails; it must be evaluated “on a case-by-
case basis.” 

n.3.

       Finally, even if plaintiffs make out a prima facie
case, the RFOA defense imposes a relatively light burden
on employers. See 

. If a company’s
oldest employees are inadvertently disadvantaged by a
merit-based policy, for example, the RFOA defense is
designed to address just such a scenario. See 

(observing that Congress included the RFOA defense
because age “not uncommonly has relevance to an
individual’s capacity to engage in certain types of
employment”). But if an employer can provide no
reasonable justification for a policy that creates a
significant age-based disparity, the ADEA prohibits that
policy.
                            40
       In sum, the limitations applicable to any ADEA
disparate-impact claim preclude liability for reasonable
employment practices, regardless of subgroups.
Nonetheless, as amici argue, our decision may very well
require employers to be more vigilant about the effects of
their employment practices. “But at the end of the day,
amici’s concerns have to be directed at Congress, which
set the balance where it is . . . . We have to read it the
way Congress wrote it.” 

–02;
see 

–99 (plurality opinion)
(explaining why “disparate impact theory need [not] have
any chilling effect on legitimate business practices”).

                     *       *       *

       We conclude that ADEA disparate-impact claims
are not limited to forty-and-older comparisons. While
claims based on subgroups present unique challenges, the
limitations applicable to any other disparate-impact
case—evidentiary gatekeeping, the prima facie case, and
affirmative     defenses—are     adequate   safeguards.
Accordingly, we will reverse the District Court’s
determination that PGW is entitled to summary judgment
on this ground.

                            IV

      We now address the District Court’s second
ground for granting summary judgment in favor of PGW:
the exclusion of plaintiffs’ statistics expert under Daubert
                            41
and Rule 702 of the Federal Rules of Evidence. For the
reasons that follow, we will vacate and remand for
further Daubert proceedings regarding plaintiffs’
statistical evidence. We then turn to plaintiffs’ other
expert reports, concluding that the District Court did not
err in excluding each.

                            A

       Pursuant to Daubert v. Merrell Dow
Pharmaceuticals, Inc., 

(1993), district
courts perform a gatekeeping function to ensure that
expert testimony meets the requirements of Federal Rule
of Evidence 702. That function extends not only to
scientific testimony, but also to other forms of
“technical” or “specialized” knowledge. Fed R. Evid.
702(a); Kumho Tire Co., Ltd. v. Carmichael, 

, 141 (1999). “Rule 702 embodies three distinct
substantive restrictions on the admission of expert
testimony: qualifications, reliability, and fit.” Elcock v.
Kmart Corp., 

, 741 (3d Cir. 2000). This case
presents issues of reliability and fit.

       “In order for expert testimony to meet Daubert’s
reliability standard, it must be based on the methods and
procedures of science, not on subjective belief and
unsupported speculation.” In re TMI Litig., 

,
703–04 (3d Cir. 1999), amended, 

(3d Cir.
2000). “The test of admissibility is not whether a
particular scientific opinion has the best foundation, or
                            42
even whether the opinion is supported by the best
methodology or unassailable research.” 

Instead, the court looks to whether the expert’s testimony
is supported by “good grounds.” 

(quoting In re
Paoli R.R. Yard PCB Litig. (Paoli II), 

, 745
(3d Cir. 1994)). The standard for reliability is “not that
high.” 

“lower than the merits standard of
correctness.” 

of the expert’s opinion
“must be evaluated practically and flexibly without
bright-line exclusionary (or inclusionary) rules.” ZF
Meritor, LLC v. Eaton Corp., 

, 291 (3d Cir.
2012) (quoting Heller v. Shaw Indus., Inc., 

,
155 (3d Cir. 1999)).

      The “fit” requirement ensures that the evidence or
testimony “[helps] the trier of fact to understand the
evidence or to determine a fact in issue.” 

(quoting Fed. R. Evid. 702(a)). “This condition
goes primarily to relevance.” 

509
U.S. at 591).

       “We review a district court’s decision to admit
expert testimony for abuse of discretion and exercise
plenary review over a district court’s legal interpretation
of Rule 702 of the Federal Rules of Evidence.” United
States v. Walker, 

, 175–76 (3d Cir. 2011).




                            43
                             B

      Plaintiffs’ expert, Dr. Michael Campion, proposes
to offer statistical evidence in support of the disparate-
impact claims. Specifically, Dr. Campion would testify
that employees older than forty-five, fifty, and fifty-five
years old were likelier to be fired in the March 2009 RIF
than were younger employees.

      The District Court identified three grounds13 for
exclusion: (1) Dr. Campion used facts or data that were
not reliable; (2) he failed to use a statistical adjustment
called the Bonferroni procedure; and (3) his testimony
lacks “fit” to the case because subgroup claims are not
cognizable. For the reasons that follow, we will vacate
the District Court’s order and remand for further Daubert
proceedings.14

      13
         The District Court assumed that Dr. Campion is
qualified. We need not, then, address that issue.
      14
          Dr. Campion based his opinion on two analyses:
one using the EEOC’s “four-fifths” test, see 29 C.F.R.
§ 1607.4 (1987), and another using a more traditional
statistical method, a z-score test. The District Court noted
that the four-fifths test “has been criticized,” but may be
“used in conjunction” with other statistical evidence.
Karlo, 

, at *13 (citation omitted). That
determination is not disputed on appeal. We therefore
focus on the reliability of Dr. Campion’s z-score test.
                            44
                            1

      First, the District Court concluded that Dr.
Campion’s report should be excluded because it is not
based on reliable data, contrary to Federal Rule of
Evidence 702(b). Specifically, Dr. Campion’s dataset
included certain “Evart Terminees” who were not part of
the “Agreed Data Set” to which the parties stipulated. We
conclude that the District Court abused its discretion to
the extent that it excluded Dr. Campion’s testimony on
this basis because the District Court ignored, without
explanation, Dr. Campion’s subsequent analysis.

       Plaintiffs argue that Dr. Campion cured this
deficiency, and the District Court’s opinion provides no
reason to doubt their argument. Specifically, plaintiffs
claim that Dr. Campion excluded the Evart Terminees
and determined that it did not affect his conclusions. At
oral argument, plaintiffs explained that the Evart
Terminees “skewed the data actually in favor of more of
the defendants,” Oral Arg. Tr. 5:6–7, whereas PGW
insists that the Evart Terminees “skewed the data to favor
Plaintiffs’ theory of the case.” Br. Appellee 29.

       It is appropriate for the District Court to address
this issue in the first instance. But the District Court
noted plaintiffs’ counterargument without addressing it.
To the extent that the District Court excluded Dr.
Campion’s testimony based on problems that were cured
by subsequent analysis, it abused its discretion. To the
                            45
extent that the subsequent analysis was deficient, the
District Court also abused its discretion because it failed
to provide any justification for discrediting that analysis.
Because we will remand for further Daubert proceedings,
as described below, the District Court will have the
opportunity to revisit this issue.

                             2

        Next, the District Court determined that Dr.
Campion “does not apply any of the generally accepted
statistical procedures (i.e., the Bonferroni procedure) to
correct his results for the likelihood of a false indication
of significance. This sort of subgrouping ‘analysis’ is
data-snooping, plain and simple.” Karlo, 

, at *13. We conclude that the District Court
applied an incorrectly rigorous standard for reliability.

       The Bonferroni procedure makes it more difficult
to find statistical significance where a researcher tests
multiple comparisons using the same data. In theory, a
researcher who searches for statistical significance in
multiple attempts raises the probability of discovering it
purely by chance, committing Type I error (i.e., finding a
false positive). See Ballew v. Georgia, 

, 234
(1978) (describing Type I and Type II errors). The
Bonferroni procedure adjusts for that risk by dividing the
“critical” significance level by the number of
comparisons tested. In this case, PGW’s rebuttal expert,
Dr. James L. Rosenberger, argues that the critical
                            46
significance level should be p < 0.01, rather than the
typical p < 0.05, because Dr. Campion tested five age
groups (0.05 / 5 = 0.01).15 Once the Bonferroni
adjustment is applied, Dr. Campion’s results are not
statistically significant. Thus, Dr. Rosenberger argues
that Dr. Campion cannot reject the null hypothesis and
report evidence of disparate impact.16

       Dr. Campion responds that adjusting the required
significance level is generally required in a “data
snooping” scenario where a researcher conducts “a huge
number of analyses of all possibilities to try to find
something significant.” A.239. In contrast to “data
snooping,” Dr. Campion calls his methodology
“hypothesis driven”; he evaluates the likelihood of
termination on a small number of groups based on logical
increments in age to discover “evidence that increasing
age relates to increased likelihood of termination . . . .”

      15
          Dr. Campion notes that he tested only four age
groups, not five. Dr. Rosenberger tests a subgroup of
sixty-and-older employees, which Dr. Campion did not
include in his analysis because “[t]here are only 14
terminations, which means the statistical power to detect
a significant effect is very low.” A.244–45.
      16
         The relationship between statistical significance
and admissibility is currently before this Court in In re
Zoloft (Sertralinehydrochloride) Prod. Liab. Litig., No.
16-2247, appealing 

(E.D. Pa. 2016).
                            47
A.240. He also points out that “nearly all the tests are
significant,” which makes his method analogous to
“cross-validating the relationship between age and
termination at different cut-offs,” or “replication with
different samples.” A.241. And finally, Dr. Campion
includes supplemental results that he claims “control for
the error rate by conducting only one analysis . . . .”
A.242.

       We conclude that the District Court erred by
applying a “merits standard of correctness,” a higher bar
than what Rule 702 demands. 

.
After identifying a potential methodological flaw, the
District Court did not proceed to evaluate whether Dr.
Campion’s opinion nonetheless rests on good grounds.
Instead, it applied a “bright-line exclusionary . . . rule[],”
ZF 

, based on Dr. Campion’s
failure to perform a specific arithmetical adjustment. As
we have observed, there could be good grounds for an
expert’s conclusion “even if the judge thinks that . . . a
scientist’s methodology has some flaws such that if they
had been corrected, the scientist would have reached a
different result.” Paoli 

.

        In certain cases, failure to perform a statistical
adjustment may simply diminish the weight of an
expert’s finding. See Paetzold & Willborn § 6:7, at 308
n.2 (describing the Bonferroni adjustment as “good
statistical practice,” but “not widely or consistently
adopted” in the behavioral and social sciences); E.E.O.C.
                             48
v. Autozone, Inc., No. 00-2923, 

, at *4
(W.D. Tenn. Aug. 29, 2006) (“[T]he Court does not have
a sufficient basis to find that . . . the non-utilization [of
the Bonferroni adjustment] makes [the expert’s] results
unreliable.”). The question of whether a study’s results
were properly calculated or interpreted ordinarily goes to
the weight of the evidence, not to its admissibility. See
Leonard v. Stemtech Int’l Inc., 

, 391 (3d Cir.
2016). “Vigorous cross-examination, presentation of
contrary evidence, and careful instruction on the burden
of proof are the traditional and appropriate means of
attacking shaky but admissible evidence.” 

; cf. Bazemore v. Friday, 

, 400
(1986) (“Normally, failure to include variables will affect
the analysis’ probativeness, not its admissibility.”).

       “That is not to say that a significant error in
application will never go to the admissibility, as opposed
to the weight, of the evidence.” In re Scrap Metal
Antitrust Litig., 

, 530 (6th Cir. 2008). An
expert’s failure to use a statistical adjustment may, in
certain cases, present a “flaw . . . large enough that the
expert lacks ‘good grounds’ for his or her conclusions.”
Paoli 

; see Erica P. John Fund, Inc. v.
Halliburton Co., 

, 266–67 (N.D. Tex.
2015) (applying a less conservative adjustment, Holm-
Bonferroni, based on “the substantial number of
comparisons” made by the expert, and citing an article
explaining that the risk of finding false significance is

                             49
prevalent where at least twenty to forty comparisons are
tested). Nonetheless, “[t]he grounds for the expert’s
opinion merely have to be good, they do not have to be
perfect.” Paoli 

. “So long as the
expert’s testimony rests upon ‘good grounds,’ it should
be tested by the adversary process . . . rather than
excluded from jurors[’] scrutiny for fear that they will not
grasp its complexities or satisfactory [sic] weigh its
inadequacies.” 

(quoting Ruiz–
Troche v. Pepsi Cola of P. R. Bottling Co., 

,
85 (1st Cir. 1998)). Accordingly, we will remand for
further Daubert proceedings as to Dr. Campion’s
statistics-related testimony to allow the District Court to
apply the correct standard for reliability.

                             3

       Finally, the District Court determined that Dr.
Campion’s statistics lacked fit to the case. “[T]he
subgrouping analysis would only be helpful to the
factfinder if this Court held that Plaintiffs could maintain
an over-fifty disparate impact claim.” Karlo, 

, at *13 n.16. Having held that plaintiffs’ over-
fifty disparate-impact claim is cognizable, we conclude
that this ground for exclusion fails as well. Because each
ground fails, we will vacate the District Court’s order




                            50
excluding Dr. Campion’s statistics-related testimony and
remand for further Daubert proceedings.17

                           C

       Dr. Campion offered a second expert report on a
different subject: reasonable human-resources (“HR”)
practices. We conclude that the District Court did not
abuse its discretion in excluding this testimony.

       Dr. Campion intended to testify as to twenty
“reasonable” HR practices that PGW could have, but did
not, employ when conducting its RIFs. Plaintiffs aver
that this testimony is necessary to rebut PGW’s RFOA
defense. The District Court disagreed. It concluded that
Dr. Campion’s HR testimony lacked relevance to the
case because “plaintiffs c[ould] rebut Defendants’ RFOA
defense only by demonstrating that the factors offered by
Defendants [we]re unreasonable.” Karlo, 

, at *15 (quoting Powell v. Dallas Morning
News L.P., 

, 247 (N.D. Tex. 2011),
aff’d 486 F. App’x 469 (5th Cir. 2012)) (alterations in
original).

       We agree. When a defendant proffers a RFOA, the
plaintiff can rebut it by showing that the factor relied

      17
        We do not reach the issue of whether the District
Court abused its discretion by declining to hold a
Daubert hearing.
                           51
upon is unreasonable, not by identifying twenty other
practices that would have been reasonable instead. See

(“While there may have been
other reasonable ways for the City to achieve its goals,
the one selected was not unreasonable.”).

       Plaintiffs also argue that PGW’s proffered RFOA
fails as a matter of law. If true, that would eliminate the
need for Dr. Campion’s HR testimony under plaintiffs’
own explanation for its relevance. But because the
District Court did not grant summary judgment on the
basis of PGW’s RFOA defense, the question of that
defense’s legal sufficiency is not before us.

                            D

       Finally, the District Court excluded the testimony
of Dr. Anthony G. Greenwald. Dr. Greenwald proposed
to testify as to his experience with Implicit Association
Tests (IAT), a type of test designed to measure “the
strength of a mental association that links a social
category (such as race, gender, or age group) with a trait
(i.e., a stereotype) . . . .” A.405. Specifically, Dr.
Greenwald reports that 80% of research participants hold
an implicit bias based on age. He also evaluated the
deposition transcripts of certain PGW employees and
determined that their RIF procedures were susceptible to
implicit biases.


                            52
       The District Court concluded that Dr. Greenwald’s
testimony lacks fit to this case because his population-
wide statistics have only speculative application to PGW
and its decision-makers. The District Court also observed
that disparate-impact claims do not inquire into the
employer’s state of mind. We agree. Plaintiffs are not
required to prove that any particular psychological
mechanism caused the disparity in question; they are
only required to demonstrate that the disparity itself is
“sufficiently substantial that [it] raise[s] such an
inference of causation.” 

. That is
not to say, however, that implicit-bias testimony is never
admissible. Courts may, in their discretion, determine
that such testimony elucidates the kind of headwind
disparate-impact liability is meant to redress. We are
simply unable here to conclude that the District Court
abused its discretion in excluding this evidence.

                            V

       The final issue presented in this appeal is whether
the District Court committed clear error in decertifying
the collective action.18 We hold that it did not.


      18
          Defendants argue that we should not reach this
issue for two reasons. First, they argue that the four opt-
in plaintiffs were not named in the notice of appeal. We
have rejected that argument in Section II of this opinion.
Second, they argue that the Notice of Appeal failed to
                            53
        The collective action19 “is a form of group
litigation in which a named employee plaintiff or
plaintiffs file a complaint ‘in behalf of’ a group of other,
initially unnamed employees who purport to be ‘similarly
situated’ to the named plaintiff.” Halle v. W. Penn
Allegheny Health Sys. Inc., 

, 223 (3d Cir.
2016). Courts in this circuit use a two-step certification

specify that plaintiffs sought review of the decertification
order. We reject that argument as well. We exercise
jurisdiction over orders not specified in a notice of appeal
if: “[1] there is a connection between the specified and
unspecified order, [2] the intention to appeal the
unspecified order is apparent and [3] the opposing party
is not prejudiced and has a full opportunity to brief the
issues.” Lusardi v. Xerox Corp., 

, 972 (3d
Cir. 1992). Each prong is met. The District Court’s 54(b)
memorandum stated that “[T]he same reasons that
warrant the certification of a final judgment on the
disparate impact and disparate treatment claims also fully
justify an immediate appeal on the decertification
ruling.” Karlo, 

, at *4 n.2. Thus,
defendants had full notice and opportunity to brief the
issue.
      19
           “[T]he ADEA incorporates enforcement
provisions of the [Fair Labor Standards Act], including
the collective action provisions of 29 U.S.C. § 216(b).”
Halle v. W. Penn Allegheny Health Sys. Inc., 

, 224 n.8 (3d Cir. 2016).
                            54
process. The first step, so-called conditional certification,
requires the named plaintiffs to make a “modest factual
showing” to demonstrate “a factual nexus between the
manner in which the employer’s alleged policy affected
him or her and the manner in which it affected the
proposed collective action members.” 

       The second step, final certification, is what is at
issue here. “[T]he named plaintiffs bear the burden of
showing that the opt-in plaintiffs are ‘similarly situated’
to them for FLSA purposes.” 

“Being
‘similarly situated’ . . . means that one is subjected to
some common employer practice that, if proved, would
help demonstrate a violation of the FLSA.” 

v. Wal Mart Stores Inc., 

, 538 (3d
Cir. 2012)). In determining whether plaintiffs are
similarly situated, relevant factors include:

      whether the plaintiffs are employed in the
      same corporate department, division, and
      location; whether they advance similar
      claims; whether they seek substantially the
      same form of relief; and whether they have
      similar salaries and circumstances of
      employment. Plaintiffs may also be found
      dissimilar based on the existence of
      individualized defenses.

–37 (emphases added). A district
court’s determination as to whether plaintiffs are
                             55
similarly situated is a finding of fact that we review for
clear error. 

.

       In this case, the District Court properly relied on
Zavala in determining that plaintiffs did not meet their
burden to show that they are similarly situated.
Specifically, the District Court observed that the nine
plaintiffs “held seven different titles with varied job
duties in two separate divisions of PGW and across five
locations in which no less than six decision-makers
independently included them in the RIF.” Karlo, 

, at *18. The District Court also based its
opinion on “[t]he existence of individualized defenses
and procedural concerns . . . .” 

Those
considerations fall squarely within the factors listed in
Zavala.

       To be sure, the named plaintiffs and opt-in
plaintiffs were each terminated in a single RIF that left
full discretion in the hands of local managers. But the
District Court did not clearly err when it concluded that
“[t]he similarities among the proposed plaintiffs are too
few, and the differences among the proposed plaintiffs
are too many” for the case to proceed as a collective
action. 

–38. Such differences
may undermine the “efficiencies for the judicial system
through resolution in one proceeding of common
issues . . . .” 

; see 

(“[T]hese common links are of minimal utility in
streamlining resolution of these cases.”).
                            56
       Plaintiffs essentially concede this point, but argue
that the “small class size” makes the class “easily
manageable even with the presence of potentially
individualized defenses and damages evidence.” Br.
Appellant 34. We decline to read the statutory phrase
“similarly situated” differently depending on the size of
the collective action.

       Plaintiffs are correct that the existence of separate
defenses or damage calculations “does not vitiate
automatically” the collective action. Lockhart v.
Westinghouse Credit Corp., 

, 52 (3d Cir.
1989) (emphasis added); cf. Tyson Foods, Inc. v.
Bouaphakeo, 

, 1045 (2016) (quoting 7AA
Charles Alan Wright, Arthur Miller & Mary Kay Kane,
Federal Practice and Procedure §1778, at 123–24 (3d
ed. 2005)). But under the guidance we have provided in
Zavala, a district court may determine that such
differences are too pronounced for the case to proceed as
a collective action. Under our deferential standard of
review, we are simply unable to conclude that the District
Court committed clear error.

                            VI

        We conclude that plaintiffs’ disparate-impact
claims are cognizable under the ADEA. We will
therefore vacate the District Court’s orders granting
summary judgment in favor of PGW and excluding the
statistics-related testimony of Dr. Campion. We will
                            57
remand for further Daubert proceedings consistent with
this opinion. We will affirm the District Court in all other
respects.




                            58