In re: Avandia Marketing v.

PRECEDENTIAL

UNITED STATES COURT OF APPEALS
FOR THE THIRD CIRCUIT
_____________

No. 25-2278

In re: AVANDIA MARKETING, SALES PRACTICES and
PRODUCTS LIABILITY LITIGATION,

GlaxoSmithKline LLC,

Appellant

On Appeal from the United States District Court
for the Eastern District of Pennsylvania
(D.C. Civil Action No. 2:07-md-01871)
District Judge: Honorable Cynthia M. Rufe

Argued on February 26, 2026

Before: SHWARTZ, MONTGOMERY-REEVES, and
AMBRO, Circuit Judges

(Opinion filed July 21, 2026)

2

Devora W. Allon (Argued)
Jay P. Lefkowitz
Kevin M. Neylan, Jr.
Kirkland & Ellis
601 Lexington Avenue
New York, NY 10022

Cole Carter
Kirkland & Ellis
333 W Wolf Point Plaza
Chicago, IL 60654

Kyle A. Dolinsky
Troutman Pepper Locke
11682 El Camino Real
Suite 400
San Diego, CA 92130

Robin P. Sumner
Troutman Pepper Locke
3000 Two Logan Square
18th and Arch Streets
Philadelphia, PA 19103

Counsel for Appellant

Hannah W. Brennan
Erin C. Burns
Thomas M. Sobol (Argued)
Hagens Berman Sobol Shapiro
One Faneuil Hall Square
5th Floor
Boston, MA 02109

3

Edwina Bullard Clarke
David Zimmer
Zimmer Citron & Clarke
711 Atlantic Avenue
Sixth Floor
Boston, MA 02111

Joseph H. Meltzer
Jonathan F. Neumann
Terence S. Ziegler
Kessler Topaz Meltzer & Check
280 King of Prussia Road
Radnor, PA 19087

Julia Solomon-Strauss
Zimmer Citron & Clarke
14 Ridge Square NW
Suite 328
Washington, DC 20016

Anne-Marie J. De Bartolomeo
Kaplan Fox & Kilsheimer
1999 Harrison Street
Suite 1501
Oakland, CA 94612

Counsel for Appellees

OPINION OF THE COURT

4

AMBRO, Circuit Judge

We do not presume in law that x caused y merely
because x happened first. The connection might be causal. But
it might be coincidental. Or some z might be responsible for x
and y alike. As statisticians emphasize, correlation alone does
not prove causation.

One method experts have developed for distinguishing
true causation from mere correlation is multiple regression
analysis. It can “define statistically the relationship between a
dependent variable (e.g., salary) and one or more independent
variables (e.g., education or work experience),” enabling us to
conclude with confidence that the latter is the reason for the
former (or is at least one reason). Weisfeld v. Sun Chem. Corp.,
84 F. App’x 257, 261 n.3 (Table) (3d Cir. 2004); see also 1
David L. Faigman et al., Mod. Sci. Evidence § 6.1 (2025-2026
ed.). It can also enable us “to control for other independent
variables” so we can rule out competing explanations (or at
least rule them unlikely). Id. And it can quantify how much of
a difference the cause makes to the effect. Thus, although it
may be that “[t]he only empirical facts . . . we can discover
about the world are facts about correlation,” regression
analysis can justify the “inference” of causation by “testing and
attempted invalidation” of other “causal hypotheses.” See
United States v. Mosley, 454 F.3d 249, 266 (3d Cir. 2006).

This appeal raises this issue and more. Third-party
payors (“TPPs”) who covered prescriptions for the diabetes
medication Avandia brought a putative class action against the
manufacturer, GlaxoSmithKline LLC (“GSK”), for
misrepresenting the drug’s cardiovascular risks and benefits.
GSK’s misrepresentations, they claim, caused more health care

5

providers (we use “physicians,” “prescribers,” and similar
terms interchangeably) to prescribe Avandia than cheaper
alternatives. That, in turn, allegedly caused these TPPs to
reimburse patients for Avandia that otherwise would not have
been prescribed.
GSK challenges the District Court’s certification of the
class. It argues the proposed class is not ascertainable because
there is not enough evidence to identify which TPPs
reimbursed members for the drug. And it contends common
issues do not predominate on causation because the plaintiffs
lack class-wide evidence GSK’s fraud caused them to cover
more Avandia prescriptions than they would have otherwise.
Their evidence, GSK contends, shows only correlation, not
causation.

Though we hold the proposed class is ascertainable, we
part with the District Court’s ruling that common issues would
predominate on causation. We join the other circuits that have
addressed this issue—the First, Second, and Ninth—and
conclude that plaintiffs in a pharmaceutical fraud RICO class
action may use statistical evidence to prove the defendant was
responsible for their injuries when the evidence can establish
causation, not merely correlation. The Plans’ statistical
evidence does not satisfy this standard yet. After laying out the
type of statistical evidence that may be used to prove causation
in a case like this one, we vacate the District Court’s
certification of the class and remand for further fact-finding on
predominance under the clarified standard.

I. BACKGROUND

6

In 1999, the Food and Drug Administration (“FDA”)
approved a new treatment for Type II diabetes: Avandia.
1

GSK, its developer, sold it for a higher price than the drug’s
older rivals, like metformin. But GSK said Avandia was worth
the premium. According to the pharmaceutical company’s
marketing, Avandia would not help patients manage only their
blood sugar. It would also reduce their cardiovascular risks.
To diabetics, that mattered. Roughly two-thirds of diabetes
patients die of cardiovascular conditions. So despite Avandia’s
higher cost, health insurers added it to their formularies and
reimbursed patients. In the years following FDA approval,
Avandia prescriptions soared. In 2006 alone, GSK sold $2.2
billion of the drug in the United States.

Meanwhile, GSK’s own research began to suggest an
inconvenient truth: the drug actually posed distinctive
cardiovascular risks. In 2004, GSK began an internal meta-
analysis, that is, a study of what existing clinical trials
indicated about Avandia’s cardiovascular profile. See In re
Paoli R.R. Yard PCB Litig., 916 F.2d 829, 856 (3d Cir. 1990)
(defining as “meta-analysis” a study “combining the results of
different . . . studies done by other scientists, and re-analyzing
the combined data to see if the data, in toto, renders different
results than the individual studies done with a smaller data
sample”). That September, it completed its initial review of
those trials, the precursor to the study later dubbed ICT-37.

1
We limit our recitation of the facts, as we have recounted
them in two prior precedential opinions in this case. See In re
Avandia Mktg., Sales Practices & Prod. Liab. Litig., 804 F.3d
633 (3d Cir. 2015); In re Avandia Mktg., Sales Practices &
Prod. Liab. Litig., 945 F.3d 749 (3d Cir. 2019), cert. denied,
141 S. Ct. 265 (Mem).

7

GSK did not complete the meta-analysis and present the results
to its safety board as ICT-37 until a year later, September 2005.
The conclusion: Avandia was associated with a statistically
significant increase of serious ischemic events (compromises
of the blood flow to the heart causing heart attack, disability,
or death). GSK nonetheless did not immediately update the
drug’s label or ask the FDA for permission to do so.

In February 2006, GSK completed a new meta-analysis,
reexamining the results of the clinical trials it examined in ICT-
37 and evaluating five additional ones. This study, ICT-42,
reached the same conclusion: Avandia posed serious
cardiovascular risks. In August the same year, the company
asked the FDA for permission to add a warning to Avandia’s
label, notifying patients that ICT-42 found the drug caused a
statistically significant risk of an increase in myocardial
ischemic events. In May 2007, GSK also asked to make that
warning clearer and more prominent.

Then—before the FDA had ruled on either of these
requests—the truth reached the public. The New England
Journal of Medicine published a study of Avandia by Dr. Steve
Nissen (the “Nissen study”). It found Avandia “was associated
with a significant increase in the risk of myocardial infarction
and with an increase in the risk of death from cardiovascular
causes that had borderline significance.” JA 443. The FDA
denied GSK’s requests to change the label, then demanded that
it make more dramatic changes to address the Nissen study and
other emerging research. By the end of 2007, the FDA had
directed GSK to add a “black-box warning”
2
conveying that

2
Though officially called a “boxed warning,” the term refers
to the FDA’s most stringent safety alert communicated through

8

Avandia could cause or exacerbate congestive heart failure and
increase the risk of myocardial ischemic events.

Prescriptions plummeted. Between 2004 and 2006,
doctors prescribed Avandia about 1.2 million times per month.
By the end of 2007, that number was below 500,000
prescriptions per month. After over 50 additional clinical trials,
the FDA enhanced the black-box warning and restricted
distribution of Avandia.
3

Consequences followed. In 2012, GSK pled guilty to
one count of failing to report clinical data to the FDA and
agreed to pay a criminal fine of nearly $250 million. GSK also
reached a civil settlement for misrepresenting Avandia’s
cardiac risks in violation of the False Claims Act, 31 U.S.C. §
3729, and agreed to pay the federal Government over $500
million.

United Food and Commercial Workers Local 1776 and
Participating Employers Health and Welfare Fund (“UFCW”),
along with J.B. Hunt Transport Services, Inc. (“J.B. Hunt”),
provide prescription drug coverage to their members or
employees. In 2010, they sued GSK on behalf of a proposed

a bold, black-bordered notice atop a drug’s label or package
insert. See generally 21 C.F.R. § 201.57(c)(1) (explaining that
black-box warnings reveal “[c]ertain contraindications or
serious warnings, particularly those that may lead to death or
serious injury”).
3
In 2013, the FDA relaxed these restrictions and removed the
black-box warning for myocardial ischemic events (blocked
blood vessels), but it retained the one for myocardial infarction
(heart attack).

9

class of TPPs (the “Plans”) that paid for Avandia, alleging
violations of the Racketeer Influenced and Corrupt
Organizations Act (“RICO”), 18 U.S.C. § 1962(c), and a host
of state laws. Their cases were combined into the In re Avandia
Marketing, Sales Practices and Products Liability multi-
district litigation.

The Plans alleged GSK fraudulently misrepresented
Avandia’s true cardiovascular profile, marketing the drug for
its supposed cardiovascular benefits while failing to disclose
its real cardiovascular risks. In reliance on GSK’s
misrepresentations, they included Avandia on their formularies
even though the drug cost more than the alternatives and
covered more Avandia prescriptions than they would have
absent the fraud.

GSK moved to dismiss the complaint for failure to state
a claim, contending in relevant part that the Plans had failed to
allege plausibly they suffered an economic injury and that
GSK’s alleged fraud was its proximate cause. In re Avandia
Mktg., Sales Practices & Prod. Liab. Litig., 804 F.3d 633, 637
(3d Cir. 2015) (“Avandia I”). The District Court denied the
motion as to the RICO claim, holding the plaintiffs plausibly
alleged GSK’s misrepresentations about Avandia’s cardiac
benefits deprived them of “the substantial savings they would
have experienced had they covered cheaper alternatives to
Avandia.” Id. We affirmed. Id. at 634.

In 2016, GSK moved for summary judgment on the
grounds the federal Food, Drug, and Cosmetic Act, 21 U.S.C.
Ch. 9, preempted the remaining state law claims and that the
Plans had failed to create a genuine dispute of material fact
about the existence of a RICO enterprise. In re Avandia Mktg.,

10

Sales Practices & Prod. Liab. Litig., 945 F.3d 749, 756 (3d
Cir. 2019), cert. denied, 141 S. Ct. 265 (Mem). The District
Court granted summary judgment for GSK. Id. We reversed
the Court’s order on the state law claims and vacated its order
on the RICO claims, holding the Plans were entitled to
discovery on the latter. Id. at 752–53.

In May 2023, the Plans moved to certify the following
class:

All entities in the United States of America and
its territories, which indirectly purchased, and/or
provided reimbursement for some or all of the
purchase price for the drugs Avandia,
Avandamet, and/or Avandaryl from May 25,
1999 until August 14, 2007.
4

Included in the Class are self-insured non-
governmental entities and third-party payers that
offer insured plans to private individuals and
groups. Likewise third-party payers that offer
insured plans to government entities including
the Federal Employee Program, Managed
Medicaid, and Medicare Part D are class
members.

They also proposed five exclusions:
(i) governmental entities other than
municipalities and/or local governments with
self-funded prescription drug plans; (ii) fully

4
We use Avandia throughout to refer to all three drugs.

11

insured health plans, i.e., plans for which the
insurer bears 100% of the risk for the
reimbursement obligations to members; (iii)
pharmacy benefit managers; (iv) natural person
consumers; and (v) employees of [GSK],
including its officers or directors, and
subsidiaries and affiliates.

And they later agreed to narrow the class period to January 1,
2005 to August 14, 2007.

In support of their motion for class certification, the
Plans moved to introduce two expert reports. The first, from
Dr. Meredith Rosenthal, sought to prove causation. She offered
two main opinions: (a) GSK’s misrepresentations of Avandia’s
health benefits were responsible for a significant portion of the
prescriptions the Plans covered; and (b) 41% of prescriptions
in the class period would not have happened but for the fraud.
Dr. Rosenthal based these conclusions on a multiple regression
analysis performed to isolate the causal significance of GSK’s
misrepresentations on prescription volume.

The Plans introduced the second report, from Dr.
Thomas McGuire, to prove damages. Using several models, he
estimated damages totaling between $500 million and over $2
billion. As relevant here, one model—the Scenario 1 Step
Down Adjustment—assumed the earlier publication of ICT-37
would have caused a similar decline in prescriptions as did the
actual publication of the Nissen study. On that assumption,
Dr. McGuire calculated damages by applying the percentage
decline in prescriptions after that study to the earlier time
period and estimated damages of around $1 billion.

12

GSK moved to exclude both reports. In October 2024,
after a Daubert hearing, the District Court granted GSK’s
motion to exclude the Rosenthal report and granted in part and
denied in part its motion to exclude the McGuire report. The
Court found Dr. Rosenthal’s opinions unreliable because her
regression model deviated from the standards of her field and
the deviations were consequential. The model appeared
“results-oriented” and particularly susceptible to false
positives. In re Avandia Mktg., Sales Pracs. & Prods. Liab.
Litig., No. 07-MD-1871, 2024 WL 4582876, at *10–11 (E.D.
Pa. Oct. 25, 2024). For instance, not only did it show a
statistically significant positive causal relationship between
prescription volume and GSK’s promotion of Avandia, it also
indicated a statistically significant positive causal relationship
between prescription volume and obviously irrelevant
variables, like U.S. carbon emissions. So the District Court
granted the motion to exclude it. The Court found one of Dr.
McGuire’s models unreliable because it depended on Dr.
Rosenthal’s opinions and rejected others as unreliable on their
own merits. The Court nonetheless left one standing: the
Scenario 1 Step Down Adjustment, explained above. The Plans
did not appeal the District Court’s evidentiary rulings.

In May 2025, the District Court granted the Plans’
motion to certify the class. First, it held the proposed class
satisfied the elements of Federal Rule of Civil Procedure 23(a):
numerosity, commonality, typicality, and adequacy of
representation. GSK did not contest any of the Rule 23(a)
issues. Second, the Court ruled that the proposed class satisfied
Rule 23(b)(3); in relevant part, it held the Plans established that
common questions would predominate and that the proposed
class was ascertainable.

13

GSK argued the Plans had not proven common
questions would predominate on one element of their RICO
claim: causation. The Plans advanced a quantity-effect theory
of causation: that GSK’s misrepresentations about Avandia’s
cardiac profile led doctors to issue more Avandia prescriptions,
which in turn led the Plans to reimburse more prescriptions for
this expensive drug. GSK claimed that individualized issues
would predominate over common ones. It offered two
arguments on this point.

First, GSK insisted, questions of reliance are typically
individual questions: they turn on why particular doctors
issued particular prescriptions to particular patients. Matters
like that, it said, are not susceptible to proof by class-wide
evidence. The District Court rejected GSK’s argument. It held
that the Plans had introduced enough common evidence to
justify an inference of “class-wide reliance,” emphasizing two
forms of evidence. In re Avandia Mktg., Sales Pracs. & Prods.
Liab. Litig., No. 07-MD-1871, 2025 WL 1479618, at *7 (E.D.
Pa. May 22, 2025). They had common evidence of the alleged
uniform scheme to cover up the drug’s dangers, like marketing
materials touting Avandia’s cardiac benefits and testimony
from GSK employees about the success of that marketing. And
they also had, the District Court found, statistical evidence that
GSK’s marketing mattered: internal GSK studies showed that
some of its marketing campaigns increased prescriptions.
Together, the Court determined, these two forms of common
evidence could prove causation.
5

5
Along the way, it rejected GSK’s argument that the
individualized nature of providers’ decision-making precluded
an inference of reliance. As the Court saw the matter, “even
though the decision to prescribe Avandia is not ‘one-

14

Second, GSK argued, even if causation could be proven
on a class-wide basis by these two forms of common evidence,
the evidence here was insufficient. Specifically, according to
GSK, its internal studies played a key role in the Plans’
common evidence. But those studies did not isolate causation.
And, without causation isolated, they could not serve as class-
wide evidence capable of proving it. While the District Court
acknowledged GSK’s second argument, it did not articulate the
reasons for rejecting it.

On ascertainability, the Plans proposed to identify
members of the class through a combination of drug purchase
records for GSK, a data publisher’s list of entities that paid for
Avandia, and a pharmaceutical database, supplemented by
affidavits to confirm membership. GSK argued there would not
be enough reliable evidence to determine which prospective
class members had actually covered Avandia prescriptions.
After a careful review of the record, the District Court rejected
these concerns. So it held the class was ascertainable and
granted the motion to certify.

dimensional,’ it is not so subjective that a provider’s reliance
cannot be inferred.” Avandia, 2025 WL 1479618, at *8. It
acknowledged “there are many factors that impact a provider’s
decision to prescribe one medication over another.” Id. Yet it
perceived “every provider’s goal when considering whether to
prescribe Avandia is largely the same—to treat a patient’s
diabetes without otherwise causing harm to their health.” Id.
And it thought “providers generally consider the same set of
factors” with minimal patient-by-patient variation. Id.

15

In June 2025, GSK petitioned us for permission to
appeal under Federal Rule of Civil Procedure 23(f). We
granted it, and this appeal followed.

II. JURISDICTION AND STANDARDS OF REVIEW

The District Court had jurisdiction under 28 U.S.C. §
1331. We have jurisdiction under 28 U.S.C. § 1292(e) and
Federal Rule of Civil Procedure 23(f). We review the
certification of a class for abuse of discretion, “which occurs if
the district court’s decision rests upon a clearly erroneous
finding of fact, an errant conclusion of law, or an improper
application of law to fact.” In re Hydrogen Peroxide Antitrust
Litig., 552 F.3d 305, 312 (3d Cir. 2008) (cleaned up).

III. DISCUSSION

GSK challenges the District Court’s holdings that the
class is ascertainable and that common issues would
predominate on causation. We asked the parties to brief a third,
threshold issue: the scope of our review of a Rule 23(f) appeal.
That is where we begin.

A. We may limit the scope of our review under Rule 23(f).

Rule 23(f) provides that a “court of appeals may permit
an appeal from an order granting or denying class-action
certification under this rule.” Fed. R. Civ. P. 23(f). The text
does not expressly address whether a motions or merits panel
may limit review to specific questions. No precedent of ours
directly answers this question either. However, since Rule
23(f)’s inception, we have analogized it to another
interlocutory appeal provision, 28 U.S.C. § 1292(b). Under that

16

subsection, we may choose which issues we review when we
accept and decide an appeal. For this reason, we now hold we
have the same discretion under Rule 23(f).

Rule 23(f) models § 1292(b). Newton v. Merrill Lynch,
Pierce, Fenner & Smith, Inc., 259 F.3d 154, 163 (3d Cir. 2001)
(citing 7B Charles Alan Wright, Arthur R. Miller & Mary Kay
Kane, Fed. Prac. & Proc. § 1802 (West Supp. 2000)). The
statute authorizes us to permit an interlocutory appeal of a non-
final order that “involves a controlling question of law as to
which there is substantial ground for difference of opinion”
where “an immediate appeal from the order may materially
advance the ultimate termination of the litigation.” 28 U.S.C. §
1292(b). As the Supreme Court has explained, “[c]ourts of
appeals wield ‘unfettered discretion’ under Rule 23(f), akin to
the discretion afforded circuit courts under § 1292(b). But Rule
23(f) otherwise ‘departs from the § 1292(b) model,’ for it
requires neither district court certification nor adherence to §
1292(b)’s other ‘limiting requirements.’” Microsoft Corp. v.
Baker, 582 U.S. 23, 31 (2017) (quoting Comm. Note on Rule
23(f)). In short, Rule 23(f) is the class-action version of §
1292(b), save that it confers on courts of appeals even greater
discretion.

Under § 1292(b), we may choose which issues to
review. The Supreme Court has taken care to use permissive
rather than mandatory language in characterizing appellate
courts’ power over the scope of our review in § 1292(b)
appeals. Yamaha Motor Corp., U.S.A. v. Calhoun held that a
court of appeals has appellate jurisdiction over the entire order
certified for appeal because “it is the order that is appealable,”
not the certified question. 516 U.S. 199, 205 (1996) (emphasis
omitted) (quoting 9 J. Moore & B. Ward, Moore’s Federal

17

Practice ¶ 110.25[1], p. 300 (2d ed. 1995). The Court could
have said a court of appeals must consider every question the
parties raise. It did not. Instead, it consistently said that appeals
courts may do that. See, e.g., id. at 205.

We too have taken care to use permissive words like
“may” and “can,” rather than mandatory language like “must”
or “shall,” in the § 1292(b) context. See, e.g., Lundeen v. 10 W.
Ferry St. Operations LLC, 156 F.4th 332, 337 (3d Cir. 2025)
(“Although the District Court certified a single question, our
review may reach any matter ‘fairly included within the
certified order.’” (emphasis added) (quoting Barbato v.
Greystone All., LLC, 916 F.3d 260, 264 (3d Cir. 2019)));
Gruber v. Price Waterhouse, 911 F.2d 960, 963 (3d Cir. 1990)
(“[S]ince we review orders and not isolated legal questions, we
may consider all grounds that might require reversal of the
order.” (emphasis added) (citation omitted)); see also McMunn
v. Babcock & Wilcox Power Generation Grp., 869 F.3d 246,
266 (3d Cir. 2017) (noting “may” is permissive).

This permissive language is no accident. It reflects the
position we adopted around the same time we started writing
on these discretionary terms: on a § 1292(b) appeal, “we are
free to consider ‘all grounds advanced in support of [reversal]
and all grounds suggested for [affirmance].’” See Miller v.
Bolger, 802 F.2d 660, 666 (3d Cir. 1986) (quoting Struble v.
N.J. Brewery Emps.’ Welfare Tr. Fund, 732 F.2d 325, 336 n.10
(3d Cir. 1984)). “In deciding whether to exercise that power,
we are guided by prudential considerations”—not by any
obligation to consider every challenge the appellant raises. Id.

Were there any doubt, in 1988 our full Court adopted
these expressly permissive terms for § 1292(b) appeals,

18

providing that we “may consider all grounds that might require
reversal of the order appealed from.” In re Data Access Sys.
Sec. Litig., 843 F.2d 1537, 1539 (3d Cir. 1988) (en banc)
(emphasis added) (citing Merican, Inc. v. Caterpillar Tractor
Co., 713 F.2d 958, 962 n.7 (3d Cir. 1983)), superseded on other
grounds by statute as stated in In re Exxon Mobil Corp. Sec.
Litig., 500 F.3d 189, 198 (3d Cir. 2007).

We have refused to consider a ground for reversal on a
§ 1292(b) appeal at least three times: in Miller, 802 F.2d at
666–67; Resol. Tr. Corp. v. Cityfed Fin. Corp., 57 F.3d 1231,
1236 & n.6 (3d Cir. 1995), vacated on other grounds sub nom.
Atherton v. FDIC, 519 U.S. 213 (1997); and most recently in
this very case. A decade ago, we heard an appeal from the
District Court’s denial of GSK’s motion to dismiss for failure
to state a claim. See Avandia I, 804 F.3d at 637. GSK had
moved to dismiss all of the Plans’ claims, federal and state. See
id. at 636–37 & n.9. The District Court certified for appeal
three questions about its denial of the motion to dismiss the
federal-law claims. Id. at 637. Even though GSK asked us to
reverse on the state-law claims as well, we refused to consider
them. Id. at 637 n.11.

The lesson of our § 1292(b) practice: we may
circumscribe the scope of our review, refusing to consider
asserted grounds for reversal. See 16 Charles Alan Wright,
Arthur R. Miller & Edward H. Cooper, Fed. Prac. & Proc. §
3931.1 (3d. ed. 2026).

To be sure, in the more distant past we declared that
“[o]n a Section 1292(b) appeal we consider all grounds which
might require a reversal of the order appealed from.” Murphy
v. Heppenstall Co., 635 F.2d 233, 235 n.1 (3d Cir. 1980); see

19

also Merican, 713 F.2d at 962 n.7 (same); In re Sch. Asbestos
Litig., 789 F.2d 996, 1002 & n.5 (3d Cir. 1986) (same). But
these cases predate our 1988 en banc decision in Data Access.
Since then, we have never repeated the arguably mandatory
language. Instead, we have used the permissive version. See,
e.g., Howard Hess Dental Lab’ys Inc. v. Dentsply Int’l, Inc.,
424 F.3d 363, 368–69 (3d Cir. 2005) (“We may ‘consider all
grounds which might require a reversal of the order appealed
from.’” (emphasis added) (quoting Merican, 713 F.2d at 962
n.7)).
6

There are two caveats. First, we cannot set aside
challenges to our subject matter jurisdiction. See Fox v.
Saginaw Cnty., 67 F.4th 284, 292 (6th Cir. 2023). Even if a
motions panel purported to refuse review of a jurisdictional
question, the merits panel not only may, but must, address it.
Cf. Council Tree Commc’ns, Inc. v. FCC, 503 F.3d 284, 292

6
There is also some language in Katz v. Carte Blanche
Corporation, 496 F.2d 747 (3d Cir. 1974) (en banc), cert.
denied, 419 U.S. 885 (1974), that one could read as requiring
us to consider any argument that a district court erroneously
certified a class, see id. at 756. Even if that were the right
reading of Katz, it would not survive our decision in Link v.
Mercedes-Benz of North America, Inc., 550 F.2d 860 (3d Cir.
1977) (en banc). The case presented two questions: whether
the District Court erred in certifying a class and whether it
could bifurcate the trial and use separate juries for the liability
phase and damages phase were liability found. Id. at 861–62.
Even though a motions panel had granted permission to appeal
these matters, id. at 862, the full Court declined to rule on either
one, id. at 861, 864–65; see also Sperling v. Hoffman-La Roche
Inc., 862 F.2d 439, 443–44 (3d Cir. 1988).

20

(3d Cir. 2007). Second, as the history we have recounted
reveals, although a motions panel is free to specify the issues
for the merits panel to review, the merits panel may refuse to
answer them. See, e.g., Link v. Mercedes-Benz of North
America, Inc., 550 F.2d 860, 861, 864–65 (3d Cir. 1977) (en
banc).

In sum, Rule 23(f) essentially mirrors § 1292(b). To the
extent they differ, Rule 23(f) provides more discretion, not
less. Under § 1292(b), we may limit the scope of our review of
an order to select issues. We have exercised this discretion for
decades. There is no reason to think we may not do the same
on a Rule 23(f) appeal. Accordingly, we hold we may
circumscribe review of a Rule 23(f) appeal to the issues we
choose.

B. The proposed class is ascertainable.

Class certification is warranted when the putative class
meets the requirements of Rule 23(a) and Rule 23(b)(1), (2), or
(3). See Reyes v. Netdeposit, LLC, 802 F.3d 469, 482 (3d Cir.
2015). Here, there is no dispute the Plans satisfied Rule 23(a).
What is disputed is whether they satisfy two requirements of
Rule 23(b)(3): predominance and ascertainability. Although
we do not believe the ascertainability issue presented in this
case merits review, we briefly address it out of respect for the
parties.

To establish ascertainability, the plaintiffs must prove
that “(1) the class is defined with reference to objective
criteria” and “(2) there is a reliable and administratively
feasible mechanism for determining whether putative class
members fall within the class definition.” In re Niaspan

21

Antitrust Litig., 67 F.4th 118, 130 (3d Cir. 2023) (quoting
Hargrove v. Sleepy’s LLC, 974 F.3d 467, 469–70 (3d Cir.
2020)). That happens when membership can be determined
from a combination of objective records and verifiable
affidavits from potential members. See Byrd v. Aaron’s Inc.,
784 F.3d 154, 171 (3d Cir. 2015); Hargrove, 974 F.3d at 480;
City Select Auto Sales Inc. v. BMW Bank of N. Am. Inc., 867
F.3d 434, 441 (3d Cir. 2017). The Plans intend to determine
membership using this very combination of evidence:
identifying potential members with Avandia purchase and
reimbursement records, then confirming actual membership
with an affidavit and proof of purchase.

GSK nonetheless advances two arguments that the
Plans failed to prove there is a reliable and feasible mechanism
for identifying class members. First, it contends there are no
records of which putative members reimbursed Avandia
prescriptions in the class period. Second, it asserts that even if
those records existed, they would not indicate whether a
putative member was an end-payor (thus eligible for class
membership) or was fully insured (hence ineligible for class
membership).

The District Court’s finding that there are objective
records of the Plans’ reimbursements for Avandia prescriptions
has no clear error. We note that two named plaintiffs—Allied
Services Division Welfare Fund and United Benefit Fund
(“UBF”)—dropped out of the case when they could not obtain
purchase records from their pharmaceutical benefit managers
(“PBMs”). But, after receiving a subpoena, UBF’s PBM
provided the data, showing it has (or at least had) the records
after all. Even if PBMs could not provide the requisite data, the
District Court’s finding that class members would be able to

22

corroborate their claims was not clearly erroneous because a
potential class member’s affidavit can be corroborated using
multiple forms of documentation.

The District Court’s finding the Plans will be able to
distinguish end-payors from fully insured plans did not clearly
err either. GSK claims that most TPPs will rely on PBM
purchase records to establish their membership in the class.
Generally, those documents do not identify whether a
purchaser was an end-payor. See Niaspan, 67 F.4th at 136. The
Plans, however, do not propose to rely on PBM data alone; the
record contained ample evidence they would be able to use
other resources to distinguish end-payors from fully insured
TPPs.

In a final bid to prevent class certification, GSK frames
the Plans’ proposal as a dilemma. In its view, the proposal
permits potential class members to identify themselves—
violating our admonition that ascertainability cannot rest on
mere say-so—or it depends on individualized evidence, risking
the very mini-trials the ascertainability requirement exists to
prevent. See id. at 130–31.

This dilemma is illusory. The Plans’ proposal for
distinguishing end-payors from fully insured plans does not
rely on mere say-so. Each potential class member’s payor
status would be verified with documentation like receipts,
claims data, plan documents, or public records. And using
these records to confirm a TPP is an end-payor hardly
constitutes a mini-trial. To the contrary, it is the
“straightforward ‘yes-or-no’ review of existing records to
identify class members” we have held “is administratively
feasible even if it requires review of individual records with

23

cross-referencing of voluminous data from multiple sources.”
Kelly v. RealPage Inc., 47 F.4th 202, 224 (3d Cir. 2022).

The Plans’ proposal falls well within the bounds of our
precedents, leaving no abuse of discretion in holding the class
is ascertainable.

C. The record so far does not show that common issues
would predominate.

Rule 23(b)(3) requires “questions of law or fact
common to class members predominate over any questions
affecting only individual members.” Reyes, 802 F.3d at 482
(quoting Fed. R. Civ. P. 23(b)(3)). “An individual question is
one where ‘members of a proposed class will need to present
evidence that varies from member to member,’ while a
common question is one where ‘the same evidence will suffice
for each member to make a prima facie showing [or] the issue
is susceptible to generalized, class-wide proof.’” Tyson Foods,
Inc. v. Bouaphakeo, 577 U.S. 442, 453 (2016) (alteration in
original) (quoting 2 W. Rubenstein, Newberg on Class Actions
§ 4:50 196–97 (5th ed. 2012)).

“[T]he presence of individual questions does not per se
rule out a finding of predominance.” In re Prudential Ins. Co.
Am. Sales Prac. Litig. Agent Actions, 148 F.3d 283, 315 (3d
Cir. 1998). Common questions predominate so long as they
“overwhelm individual issues.” Neale v. Volvo Cars of N. Am.,
LLC, 794 F.3d 353, 371 (3d Cir. 2015) (citing Amgen Inc. v.
Conn. Ret. Plans & Tr. Funds, 568 U.S. 455, 468–69 (2013)).
In other words, “[w]hen one or more of the central issues in the
action are common to the class and can be said to predominate,
the action may be considered proper under Rule 23(b)(3) even

24

though other important matters will have to be tried separately,
such as damages or some affirmative defenses peculiar to some
individual class members.” Tyson, 577 U.S. at 453–54 (cleaned
up).

“Class certification is proper only ‘if the trial court is
satisfied, after a rigorous analysis, that the prerequisites’ of
Rule 23 are met.” Hydrogen Peroxide, 552 F.3d at 309
(quoting Gen. Tel. Co. of Sw. v. Falcon, 457 U.S. 147, 161
(1982)). “Factual determinations necessary to make Rule 23
findings must be made by a preponderance of the evidence.”
Id. at 320. This requires resolving “every dispute that is
relevant to class certification” before it can be granted.
Ferreras v. Am. Airlines, 946 F.3d 178, 183 (3d Cir. 2019).

We “analyze predominance in the context of Plaintiffs’
actual claims” because “the nature of the evidence that will
suffice to resolve a question determines whether the question
is common or individual.” Neale, 794 F.3d at 371–72 (quoting,
in second passage, Hydrogen Peroxide, 552 F.3d at 311). That
means class-wide evidence of each element of the cause of
action must predominate over evidence particular to individual
class members. See Reyes, 802 F.3d at 482–83, 489; see also
Hydrogen Peroxide, 552 F.3d at 311 (“If proof of the essential
elements of the cause of action requires individual treatment,
then class certification is unsuitable.” (cleaned up)).
Accordingly, here predominance is “satisfied if each element
of the alleged RICO violation involves common questions of
law and fact capable of proof by evidence common to the
class.” Reyes, 802 F.3d at 489 (emphasis omitted).

“Establishing liability under [§ 1962(c)] of the RICO
statute requires (1) conduct (2) of an enterprise (3) through a

25

pattern (4) of racketeering activity, plus an injury to business
or property.” Id. (alteration in original) (quoting In re Ins.
Brokerage Antitrust Litig., 579 F.3d 241, 269 (3d Cir. 2009)).
Implicit is a requirement that the racketeering activity was the
but-for cause of the alleged injury. Id. “The simplest statement
of the concept of but-for causation is that event A is a but-for
cause of event B if event B could not happen without event A
happening first.” Mosley, 454 F.3d at 266.

“Although reliance on the defendant’s alleged
misrepresentation is not an element of a RICO mail-fraud
claim, the plaintiffs’ theory of injury in most RICO mail-fraud
cases will nevertheless depend on establishing that someone—
whether the plaintiffs themselves or third parties—relied on the
defendant’s misrepresentation.” Sergeants Benevolent Ass’n
Health & Welfare Fund v. Sanofi-Aventis U.S. LLP, 806 F.3d
71, 87 (2d Cir. 2015) cert. denied, 580 U.S. 825 (2016).

That is because reliance will typically be a
necessary step in the causal chain linking the
defendant’s alleged misrepresentation to the
plaintiffs’ injury: if the person who was allegedly
deceived by the misrepresentation (plaintiff or
not) would have acted in the same way regardless
of the misrepresentation, then the
misrepresentation cannot be a but-for, much less
proximate, cause of the plaintiffs’ injury.

Id.
Because “reliance is nearly always an individualized
question,” Harnish v. Widener Univ. Sch. of L., 833 F.3d 298,
309 (3d Cir. 2016), establishing predominance on causation in

26

a putative RICO class action can be challenging. Yet if a
reasonable jury could find reliance based on class-wide
evidence, then common questions predominate. See Sergeants
Benevolent, 806 F.3d at 92–94.

In this case, the predominance question is whether the
Plans can prove by class-wide evidence that they covered more
Avandia prescriptions by reason of GSK’s fraud. To do so,
they must prove prescribers relied on GSK’s
misrepresentations. That is because the Plans propose a
quantity-effect theory of third-party reliance: they claim
GSK’s fraud caused them to reimburse more prescriptions for
Avandia because the fraud swayed physicians to prescribe it.
And the District Court agreed. But it did not fully address all
disputes relevant to class certification. Even if it had, certain of
the evidence proved correlation not causation, and the
remaining evidence has not yet been analyzed.

We have not addressed whether TPPs in a RICO
pharmaceutical fraud action can prove reliance by statistical
evidence. We hold it is possible, but the Plans have not done it
to date. Statistical evidence can prove reliance. But not just any
data will do. Proof of correlation is not enough. Proof of
causation is required. One form such proof may take is a
regression analysis or comparably robust quantitative study
that is sufficiently reliable to pass muster under Daubert v.
Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993).

1. All disputes relevant to class certification need
resolution, so we must vacate.
The District Court mentioned three possible regression
or statistical analyses that the Plans could have leaned on as

27

class-wide evidence: Dr. Rosenthal’s analysis, Dr. McGuire’s
analysis, and GSK’s own internal statistical analyses. The
Court, however, held that only one of these analyses could be
used to prove reliance: GSK’s internal statistical analyses.

GSK argued to the District Court that these analyses
could not isolate causation and thus could not reliably serve as
a form of class-wide evidence. And the Court acknowledged
that argument in its opinion. JA 16 (“GSK argues that because
its own internal statistical analyses do not isolate the effect of
their marketing regarding Avandia’s cardiovascular profile
specifically, they cannot be used to demonstrate class-wide
reliance on that marketing.”). But, as explored more below, it
never grappled with this concern despite the direct bearing on
the adequacy of key statistical evidence purporting to serve as
class-wide evidence. That unresolved dispute requires us to
vacate.

“Rule 23 requires rigorous consideration of all the
evidence and arguments offered by the parties.” Hydrogen
Peroxide, 552 F.3d at 321; see also Byrd, 784 F.3d at 163
(noting that “a court evaluating a motion for class certification
is obligated to probe behind the pleadings when necessary . . .
to determine whether the Rule 23 certification requirements are
satisfied”). This is because its requirements “must be met,” not
just supported by some evidence or conditionally met.
Hydrogen Peroxide, 552 F.3d at 321 (emphasis added)
(cleaned up). Thus, “[p]rior to certifying a class, a district court
must resolve every dispute that is relevant to class
certification.” Ferreras, 946 F.3d at 183. Not to do so presents
a problem for an affirmance. See id.; Hydrogen Peroxide, 552
F.3d at 320 (“[A] district court exercising proper discretion in
deciding whether to certify a class will . . . hav[e] considered

28

all relevant evidence and arguments presented by the parties.”).
“[T]he party proposing class-action certification bears the
burden of affirmatively demonstrating by a preponderance of
the evidence her compliance with the requirements of Rule
23.” Byrd, 784 F.3d at 163.

2. We part from the District Court’s rationale for holding
common issues would predominate.

As alluded above, the District Court held the Plans
established predominance on the causation issue by
introducing two forms of class-wide evidence that would
justify an inference of class-wide reliance: “evidence of a
common scheme to deceive” and “statistical or econometric
models that show the allegedly unlawful promotion caused an
increase in prescriptions” during the class period. Avandia,
2025 WL 1479618, at *7–9. Despite our deep respect for the
Court’s thoughtful work on this challenging case, we part
ways. As to the first, we take a different view of the law. As to
the second, we take a different view of the record.

a. The Plans did not seek, and the record does
not warrant, an inference of “class-wide
reliance.”

Before proceeding to those two points, we begin by
trying to understand what the District Court meant when it
found that the Plans’ evidence justified an “inference of class-
wide reliance.” See generally Avandia, 2025 WL 1479618, at
*8–9 (repeatedly using variations of the phrase “inference of
class-wide reliance”). It cannot have meant that every member
of the class relied on GSK’s misrepresentations. The Plans
could not have shown every member of the class relied on

29

GSK’s fraud, for they have not even alleged any member relied
on it. They offered a different theory: that every member of the
class was injured because prescribers relied on the
misrepresentations.

Perhaps, then, the District Court meant that the Plans
justified an inference that every physician relied on the fraud
or that every prescription was made in reliance. That would
make sense as to the heart of its reasoning: that it could infer
reliance from the supposed fact that every prescription decision
turned on the same considerations. So it is tempting to think
the Court meant that it could infer all physicians relied, because
all of them would have made their prescribing decisions in
light of GSK’s claims about Avandia’s cardiac profile. But that
doesn’t work either. The Court expressly found the Plans’
evidence could not support an inference that, but for GSK’s
fraud, no prescriptions would have been made. Avandia, 2025
WL 1479618, at *8.

The Plans offer a third interpretation of the opinion: the
District Court meant that many physicians relied on the fraud—
enough to infer every member of the class covered at least one
prescription made in reliance. That would match the Plans’
theory. But the Court did not say that.

b. The evidence of a common scheme, without
more, did not warrant an inference of
reliance in this case.

Next, we turn to the significance of the evidence of a
common scheme to defraud. As the District Court noted,
“courts may infer class-wide reliance when a plaintiff presents
evidence of a common scheme to deceive.” Id. at *7 (citing In

30

re Cmty. Bank of N. Va. Mortg. Lending Pracs. Litig., 795 F.3d
380, 408 (3d Cir. 2015)). But our cases do not support a broad,
uniform rule that courts may infer class-wide reliance any time
plaintiffs propose to prove the defendant committed fraud.
True, we have endorsed the position that “where proof of the
RICO violation is demonstrated through common evidence of
a common scheme, reliance may be inferred on a classwide
basis.” Cmty. Bank, 795 F.3d at 408. But as noted earlier, we
also have warned, in a case where the proof of the violation
came from common evidence of a common scheme, that
“reliance is nearly always an individualized question.”
Harnish, 833 F.3d at 309. So how should we distinguish cases
where common evidence of a unified scheme justifies an
inference of reliance from cases where it does not?

Several considerations distinguish this one. First, there
is no fraud exception to the requirement of proving
predominance. Decades ago, the Supreme Court observed that
“[p]redominance is a test readily met in certain cases alleging
consumer or securities fraud or violations of the antitrust
laws.” Amchem Prods., Inc. v. Windsor, 521 U.S. 591, 625
(1997). But since then we have emphasized that “it does not
follow that a court should relax its certification analysis, or
presume a requirement for certification is met, merely because
a plaintiff’s claims fall within one of those substantive
categories.” Hydrogen Peroxide, 552 F.3d at 322. “‘[A]ctual,
not presumed, conformance’ with the Rule 23 requirements
remains necessary.” Id. (alteration in original) (quoting
Newton, 259 F.3d at 167).

Second, aside from Harnish, the only two cases of ours
the District Court relied on for this part of its decision come up
short. One, In re Warfarin Sodium Antitrust Litig., 391 F.3d

31

516 (3d Cir. 2004), shows that evidence of a common scheme
to defraud can establish predominance when the theory of
causation does not turn on reliance at all. There, “individual
reliance on the misrepresentations was irrelevant to liability.”
Id. at 529. Here, however, the Plans’ theory of causation
requires that physicians relied on GSK’s misrepresentations.

The other, Community Bank, the District Court
described as governing cases where the plaintiffs allege the
fraud concerned the only factor that mattered to the decision
that resulted in their injuries, see Avandia, 2025 WL 1479618,
at *8 n.56. The Second Circuit explained the point well:

In certain factual contexts, it may well be
reasonable to infer that each class member would
only have taken the action leading to its injury if
it had relied on the defendant’s alleged
misrepresentation. Such an inference may be
available if, for example, the class members all
faced “the same more-or-less one-dimensional
decisionmaking process,” such that the alleged
misrepresentation would have been “essentially
determinative” for each plaintiff.

Sergeants Benevolent, 806 F.3d at 88 (quoting Richard A.
Nagareda, Class Certification in the Age of Aggregate Proof,
84 N.Y.U. L. Rev. 97, 121 (2009)). For example, when
plaintiffs allege the defendant fraudulently overbilled them, a
court may presume they paid too much in reliance on the
defendant’s misrepresentation about the amount due because
there is generally no other reason plaintiffs would pay more
than they owed. Id. at 89 (citing In re U.S. Foodservice Inc.
Pricing Litig., 729 F.3d 108, 120 (2d Cir. 2013), cert. denied,

32

572 U.S. 1087 (2014)); see also Klay v. Humana, Inc., 382 F.3d
1241, 1259–60 (11th Cir. 2004), abrogated in part on other
grounds by Bridge v. Phoenix Bond & Indem. Co., 553 U.S.
639 (2008).

It is in this context we last declared that “where proof of
the RICO violation is demonstrated through common evidence
of a common scheme, reliance may be inferred on a classwide
basis.” Cmty. Bank, 795 F.3d at 408. There, the plaintiffs
asserted “class-wide evidence demonstrate[d] that [the
defendant] performed no services in exchange for settlement
charges.” Id. We held that “on this record and in this
context”—where the plaintiffs had alleged the defendant
charged them for services it falsely claimed to have provided—
we “d[id] not believe that the District Court abused its
discretion in accepting the Plaintiffs’ position” that their
evidence of a class-wide scheme justified an inference they
paid for those nonexistent services in reliance on the alleged
misrepresentations. Id. When the decision is one-dimensional,
reliance may be treated as one-dimensional.

Community Bank does not apply here because, as the
District Court acknowledged, whether to prescribe Avandia
was not a one-dimensional decision. The Court thought it could
rely on it anyway because, “even though the decision to
prescribe Avandia is not ‘one-dimensional,’ it is not so
subjective that a provider’s reliance cannot be inferred.”
Avandia, 2025 WL 1479618, at *8. The record does not justify
that inference. The Court’s portrayal of prescribers’ decision-
making process did not cite a single page of the several-
thousand-page record. This speculation about how providers
make decisions is not the “rigorous analysis” required for class

33

certification. See Hydrogen Peroxide, 552 F.3d at 309 (quoting
Gen. Tel. Co., 457 U.S. at 161).

In this context, the Plans could not prove reliance with
their evidence of GSK’s scheme alone. The law did not permit
that inference.

c. The District Court did not adequately
justify its finding that GSK’s marketing
studies warranted an inference of reliance.

Last, we turn to the statistical evidence. To repeat, the
District Court found the Plans could justify an inference of
class-wide reliance because they had introduced “statistical or
econometric models that show the allegedly unlawful
promotion caused an increase in prescriptions.” Avandia, 2025
WL 1479618, at *7, *9. That evidence is crucial because their
theory of but-for causation is that they were injured because
physicians “relied on [GSK’s] misrepresentation[s].” See
Sergeants Benevolent, 806 F.3d at 87. In our view, the District
Court did not provide the necessary analysis of this point to
establish that the prerequisites of Rule 23 are met. See
Hydrogen Peroxide, 552 F.3d at 309.

The Court stated that the Plans introduced internal GSK
studies showing a few of its marketing campaigns caused an
increase in prescriptions. Avandia, 2025 WL 1479618, at *9.
On their own, they may not be evidence that GSK’s
purportedly “unlawful promotion” caused an increase in
prescriptions, id. at *7 (emphasis added), because they did not
isolate the cardiovascular messaging from other messaging
about Avandia as the cause of increased prescriptions.

34

The Court, while acknowledging GSK raised this
concern, defended its determination that the internal marketing
research could establish “class-wide reliance” by citing cases
holding that, in general, such an inference can be justified by
statistical or aggregate evidence a RICO defendant’s fraud
caused the plaintiffs’ injuries. See Avandia, 2025 WL 1479618,
at *9. That is true. But it does not resolve GSK’s concern. If
the GSK studies showed its allegedly fraudulent claims about
Avandia’s cardiac effects caused an uptick in prescriptions,
then the cases the District Court cited would provide a legal
foundation for putting them to work here. But as far as we can
tell from the expert reports describing the studies, they did not
identify the distinctive consequences of GSK’s allegedly
fraudulent marketing as opposed to its marketing generally.
Even if they could show physicians relied on what GSK said
about Avandia, they may not be able to show doctors were
swayed by its alleged misrepresentations about its cardiac
effects in particular. Similarly, even if the problem is that all of
GSK’s marketing omitted the truth about Avandia’s cardiac
profile, the data in the record may not explain how much of its
efficacy arose from that omission.

If the District Court understood the actual scope of the
studies but reasoned they sufficed to justify an inference of
class-wide reliance, an error of law crept in: what the studies
can prove is a legal matter, not a factual one. See Comcast
Corp. v. Behrend, 569 U.S. 27, 36 n.5 (2013) (“[W]hile the
data contained within a . . . model may well be ‘questions of
fact’ in the relevant sense, what those data prove is no more a
question of fact than what our opinions hold.”). Even by the
lights of the cases the District Court cited, only a report that
attempts to isolate the effects of the alleged fraud could warrant
that inference. See In re Celexa & Lexapro Mktg. & Sales

35

Pracs. Litig., 915 F.3d 1, 13–14 (1st Cir. 2019) (discussing In
re Celexa & Lexapro Mktg. & Sales Pracs. Litig., 315 F.R.D.
116, 126–29 (D. Mass. 2016)); Sergeants Benevolent, 806 F.3d
at 91–93; In re Neurontin Mktg. & Sales Pracs. Litig.
(“Neurontin III”), 712 F.3d 60, 68 (1st Cir. 2013); cf. Comcast,
569 U.S. at 32, 35 (holding a damages model must at least
attempt to isolate the effects of each of the plaintiffs’
independent theories of liability, at least where some have been
rejected).

And if the District Court incorrectly understood what
the data concerned, its factual error was a clear one. See Reyes,
802 F.3d at 483. Nothing in the record indicated the studies
isolated the effects of GSK’s fraudulent marketing from the
effects of its marketing in general. And the Plans disavow
using these studies for that purpose. Further, because these
studies do not seem to speak to the issue at hand, combining
them with the common evidence of GSK’s alleged scheme
does not yield an adequate basis for the District Court’s
decision either.

3. The Plans’ alternative basis for holding common
issues would predominate does not work either.

The Plans claim they can show reliance differently: by
introducing statistical evidence of a correlation between the
Nissen study’s exposure of the fraud and the subsequent
decline in prescriptions, then using circumstantial evidence to
show that the earlier exposure of the fraud would have led to
the same decline—thus closing the gap between correlation

36

and causation.
7
In their view, if the public had learned the truth
about Avandia’s cardiovascular profile at the start of the class
period in 2005, prescriptions would have declined in the same
way they did after the Nissen study exposed that information
in 2007.

We disagree. A few courts have permitted TPPs in
similar RICO class actions to prove reliance with class-wide
evidence when the evidence includes both statistical and
circumstantial evidence of causation. But we are not aware that
any court has permitted a putative class of TPPs to prove
providers’ reliance by class-wide evidence without statistical
evidence the defendant’s conduct caused the injuries. The
Plans could have cleared the bar if they had introduced
statistical evidence of causation, like a regression analysis.
They tried. But the District Court struck that evidence after a
Daubert hearing. Without it, the Plans have only statistical
evidence of correlation and circumstantial evidence of
causation. In this context, where their theory depends on
establishing the fraud caused enough prescriptions to make it
likely the fraud injured the entire class, that is not enough. TPP
pharmaceutical fraud class action plaintiffs can prove
causation by class-wide statistical evidence. But doing so takes
more rigorous statistics than the Plans so far have presented.

Every circuit to address this question has held or
suggested that statistical evidence of causation may prove a

7
The Plans claim the District Court considered and endorsed
this theory. Yet they do not—and cannot—point to a single
page of the District Court’s class-certification opinion where it
did. Instead, they cite only a couple of remarks it made during
a prior hearing.

37

drug manufacturer’s fraud injured a class of TPPs. Begin with
the First Circuit. In three related decisions, it held that multiple
regression analysis suggesting a causal relationship between a
drug maker’s illegal marketing and an increase in prescriptions
can help prove the fraud injured a TPP. See In re Neurontin
Mktg. & Sales Pracs. Litig. (“Neurontin I”), 712 F.3d 21, 29–
30, 40, 46 (1st Cir. 2013); In re Neurontin Mktg. & Sales Pracs.
Litig. (“Neurontin II”), 712 F.3d 51, 56–58 (1st Cir. 2013);
Neurontin III, 712 F.3d at 68. These cases were RICO actions
health insurers brought against Pfizer for illegally promoting
its drug Neurontin for off-label use. Neurontin I, 712 F.3d at
25–26; Neurontin II, 712 F.3d at 52–53; Neurontin III, 712
F.3d at 61–62, 68–69. Across these cases, the but-for causation
question was whether Pfizer’s unlawful marketing caused the
insurers to reimburse more Neurontin prescriptions than they
otherwise would have because physicians issued more
prescriptions in reliance on the messaging. Neurontin I, 712
F.3d at 25–26, 40–42; Neurontin II, 712 F.3d at 52–53, 57;
Neurontin III, 712 F.3d at 61–62, 68.

The TPPs tried to answer that question by introducing
class-wide statistical evidence and circumstantial evidence.
Neurontin I, 712 F.3d at 42–45; Neurontin II, 712 F.3d at 56;
Neurontin III, 712 F.3d at 63–64. The primary statistical
evidence was the analysis of expert Dr. Rosenthal, who
performed a regression analysis that “found a causal
connection between the fraudulent marketing and the quantity
of prescriptions written for off-label indications.” Neurontin I,
712 F.3d at 29–30; see also Neurontin II, 712 F.3d at 56;
Neurontin III, 712 F.3d at 63. Her analysis also identified “the
percentage of prescriptions caused by Pfizer’s fraudulent off-
label marketing.” Neurontin I, 712 F.3d at 30.

38

In Neurontin I and Neurontin II, the First Circuit held
this statistical evidence, combined with circumstantial
evidence, sufficed to permit a jury to find Pfizer’s fraud caused
the plaintiff insurers’ injuries. Neurontin I, 712 F.3d at 45–47;
Neurontin II, 712 F.3d at 57–58. The Court observed that
“regression analysis is a well recognized and scientifically
valid approach to understanding statistical data.” Neurontin I,
712 F.3d at 42. It also noted that courts permit plaintiffs to
prove causation with regression analysis in multiple legal
domains, including Title VII claims, Sixth Amendment claims
about whether a jury pool reflects a reasonable cross-section of
the community, and antitrust claims. Id. (citations omitted).
The antitrust precedents were particularly significant because
“RICO has drawn many of its causation principles” from
antitrust law, though antitrust law generally does not require
proof of reliance. Id. at 44. The Court further considered and
set aside Pfizer’s contention that some individual doctors
might have prescribed Neurontin for off-label uses without
relying on the illicit marketing. Id. at 45. “The existence of
some doctors who purportedly were not influenced by Pfizer’s
misinformation would not defeat the inference that this
misinformation had a significant influence on prescribing
decisions which injured” the plaintiff insurers—and that
inference was all the plaintiffs needed. Id. Finally, Dr.
Rosenthal’s regression analysis proved causation, not just
correlation, particularly when combined with the
circumstantial evidence. Id. at 46.

Most relevantly, in Neurontin III the First Circuit
vacated a denial of class certification on the same ground. 712
F.3d at 62, 68, 70. Echoing the prior analysis, it held “that the
Rosenthal report is capable of providing proof of but-for
causation.” Id. at 68. “[R]egression analysis is a widely

39

accepted method of showing causation under several causes of
action,” the Court explained, and it saw “no reason to reach a
different conclusion for a specific subset of RICO claims based
on fraudulent pharmaceutical marketing.” Id. at 69. Because
the trial court’s denial of class certification “pivoted on the
determination that the Rosenthal report could not provide proof
of causation,” that denial could not stand. Id. at 70. Class-wide
statistical evidence capable of distinguishing causation from
correlation, such as a regression analysis, may prove common
questions predominate on RICO but-for causation.

The First Circuit later reaffirmed that this reasoning
applies to the predominance analysis of class certification. In
another RICO class action brought by TPPs against a drug
manufacturer, it explained that the plaintiffs’ “clinical and
statistical evidence, if believed, could establish causation and
injury at least for any TPP who paid for more than a handful of
different patients’ prescriptions.” Celexa, 915 F.3d at 14. The
Court emphasized that the record featured the testimony of two
experts, who opined not only that the drug company’s
promotional spending was correlated positively with sales, but
also that its allegedly unlawful promotions caused 76 percent
and 54 percent, respectively, of the off-label prescriptions for
the relevant two drugs. Id. at 13. From these figures, one of the
experts estimated that if a TPP covered as few as five
prescriptions, there was a 98 percent chance at least one
resulted from the illicit marketing. Id. The plaintiffs bolstered
that statistical evidence with direct and circumstantial evidence
the fraud caused their injuries, like evidence the
pharmaceutical company’s sales representatives had called or
visited physicians who subsequently issued the off-label
prescriptions at issue. Id. In the First Circuit, then, a TPP in a
pharmaceutical fraud RICO case may prove causation by class-

40

wide evidence where that evidence includes statistical analysis
that isolates the effects of the fraud from other variables and
shows it is likely the fraud injured each TPP.

A pair of Second Circuit decisions contemplated the
same approach. The first, UFCW Local 1776 v. Eli Lilly & Co.
(“Zyprexa”), 620 F.3d 121 (2d Cir. 2010), was a RICO class
action brought by a putative class of TPPs against Eli Lilly and
Company for misrepresenting the efficacy and side effects of
Zyprexa. Id. at 123, 129. The TPPs advanced the now-familiar
theory that Eli Lilly’s fraud increased the number of Zyprexa
prescriptions doctors made, which in turn increased the number
members of the putative class covered. Id. at 135. However, as
the Court later explained in a separate case discussing Zyprexa,
“[t]o prove that doctors had, in fact, relied on Lilly’s
misrepresentations in making their prescription decisions, the
[Zyprexa] plaintiffs primarily offered evidence that the number
of Zyprexa prescriptions fell after the drug’s weight- and
diabetes-related side effects were disclosed.” Sergeants
Benevolent, 806 F.3d at 89–90. That evidence of a correlation
between the exposure of the truth and a decline in prescriptions
could not prove the fraud caused more pre-exposure
prescriptions, the Court explained, because there were “any
number[ ]of a multitude of reasons” a doctor might have
prescribed Zyprexa. Id. at 90. “The fact that Zyprexa
prescriptions declined markedly following the disclosure of the
previously concealed information was not sufficient to support
this necessary inference [of causation].” Id. “[B]ecause a
reasonable jury would be unable to find RICO causation
satisfied for each class member based on the generalized proof
offered by the plaintiffs, common questions did not
predominate, and class certification under Rule 23(b)(3) was
therefore inappropriate.” Id.

41

A few years later, the Second Circuit rejected a similar
attempt to establish predominance on causation for the same
reason, while clarifying what would suffice. In Sergeants
Benevolent, 806 F.3d 71 (2d Cir. 2015), the plaintiffs “sought
to certify a class of all [TPPs] that paid for Ketek prescriptions
on the theory that such [TPPs] were injured as a result of
paying for Ketek prescriptions that would not have been
written if Aventis had not concealed Ketek’s safety risks.” Id.
at 74. The Court acknowledged that “it may be possible for a
class of plaintiffs to prove the causation element of a
pharmaceutical fraud claim such as this one with generalized
proof,” but it held the plaintiffs “failed to offer such proof”
there. Id. at 74–75.

Like the Zyprexa plaintiffs, the Sergeants Benevolent
plaintiffs tried to prove causation by presenting evidence
prescriptions “dropped precipitously” after the truth was
exposed. See id. at 91. And like the Zyprexa plaintiffs, they
introduced the testimony of an expert—again, Dr. Rosenthal—
that “this unprecedented drop must have been caused entirely
by the disclosure of Ketek’s post-marketing safety data.” Id.
However, as in Zyprexa, the expert testimony was not
accompanied by statistical proof of causation, like a regression
analysis; it instead assumed the causal link the plaintiffs
needed to prove. Id. at 91–92, 95. “Plaintiffs [thus] made no
attempt to control for [potential confounding variables], or to
supply any other information that might render reasonable the
inference that the drop in sales was actually attributable to the
safety disclosures, as opposed to other factors.” Id. at 92.
Because “mere correlation does not demonstrate causation,”
evidence of “a simple correlation between the safety disclosure
and the decline in prescriptions,” id., “is insufficient to prove

42

class-wide RICO causation on the theory that the defendant’s
withholding of safety information caused doctors to write
excess prescriptions,” id. at 95.

The Court then went out of its way to note “that it may
be possible to demonstrate class-wide RICO causation in a case
such as this one by adducing generalized proof from which a
reasonable jury could conclude that only some prescriptions
paid for by each class member were written based on the
defendant’s alleged misrepresentations.” Id. at 94 (emphasis
omitted). Even if the plaintiffs had advanced such a theory,
evidence of correlation would have been insufficient to prove
causation because it did not “isolate[] the relative causal effect
of the numerous variables bearing on the decline in Ketek’s
sales.” Id. at 95–96. The Court distinguished the evidence
before it from Dr. Rosenthal’s regression analysis in
Neurontin, which controlled for other variables. Id. at 96
(citing Neurontin I, 712 F.3d at 30). Although Dr. Rosenthal
was also an expert in Sergeants Benevolent, she had not
performed a regression analysis. Id. at 95.

The Court pointedly concluded by observing
“Neurontin does indicate that where individual physicians’
reliance on a pharmaceutical company’s misrepresentations
forms a necessary link in the causal chain between those
misrepresentations and the plaintiffs’ injury, such reliance can
be proved to a jury with sufficiently powerful aggregate
evidence.” Id. at 97. In short, what the plaintiffs needed to add
was a regression analysis or some other evidence “firmer than
speculation” that prescriptions were written in reliance on the
fraudulent marketing. Id. at 95.

43

Last year, the Ninth Circuit joined this nascent
consensus. Painters & Allied Trades Dist. Council 82 Health
Care Fund v. Takeda Pharm. Co., No. 23-55742, 2025 WL
1683472 (9th Cir. June 16, 2025). Certifying a class of TPPs in
a RICO class action against a pharmaceutical company, a
district court had held that “a statistical regression . . . can
establish but-for causation for a civil RICO claim, especially
when used in tandem with other circumstantial or direct
evidence.” Painters & Allied Trades Dist. Council 82 Health
Care Fund v. Takeda Pharm. Co., 674 F. Supp. 3d 799, 827
(C.D. Cal. 2023). The Circuit Court affirmed, explaining that
the expert report the TPPs submitted could prove causation
because it “used regression models to show that the decline in
prescriptions was caused by the failure to disclose,” even
accounting for possible confounding variables. Painters, 2025
WL 1683472, at *2 (footnote omitted). The upshot (again):
TPPs can prove providers relied on a drug manufacturer’s
misrepresentations by introducing statistical evidence of
causation—like a regression model—supplemented by
circumstantial evidence.

We too conclude that, in a RICO pharmaceutical fraud
action, TPPs may use class-wide statistical evidence to prove
physicians’ reliance, and hence but-for causation, so long as
the evidence can distinguish correlation from causation. That
evidence may take the form of a regression analysis or any
comparably robust statistical method.
8

8
Whether other types of evidence could prove common
questions would predominate on causation in this case is not
before us.

44

As a general matter, the Supreme Court and our Court
have held plaintiffs may use statistical evidence to prove
causation. See, e.g., Texas Dep’t of Housing and Cmty. Affairs
v. Inclusive Cmtys. Project, 576 U.S. 519, 542–43 (2015);
Watson v. Fort Worth Bank & Tr., 487 U.S. 977, 994 (1988);
Bazemore v. Friday, 478 U.S. 385, 400–03 (1986) (Brennan,
J., concurring in part); In re Linerboard Antitrust Litig., 305
F.3d 145, 153–55 (3d Cir. 2002).

In this context, using statistical evidence to prove but-
for causation would satisfy the Supreme Court’s general
criteria for using such evidence in predominance analysis.
Although the Court has declined to adopt “broad and
categorical rules governing the use of representative and
statistical evidence in class actions,” it has identified two
criteria which, if jointly met, permit us to find common
questions predominate. See Tyson, 577 U.S. at 454–60.

First, statistical evidence may be used to show
predominance only if a reasonable jury could have relied on it
to find each class member proved the element at issue in an
individual action. Id. at 455. “[W]here representative evidence
is relevant in proving a plaintiff’s individual claim, that
evidence cannot be deemed improper merely because the claim
is brought on behalf of a class.” Id. Tyson itself held the
plaintiffs in a Fair Labor Standards Act collective action could
use evidence of how long a representative sample of employees
spent donning and doffing protective gear to prove how long
each worker spent. Id. at 454, 456–59. That was because an
individual worker could have deployed the same representative
evidence to prove the same point in an individual action where
his or her employer had failed to keep adequate records. See id.
at 455–57. In this situation, “[r]ather than absolving the

45

employees from proving individual injury, the representative
evidence . . . was a permissible means of making that very
showing.” Id. at 457.

Second, statistical evidence can support class-wide
liability when the defenses to it are also class-wide. Id. at 457.
Tyson met that criterion because, in the absence of employer
records of how long each employee worked, the “primary
defense” available to the employer “was to show that [the]
study was unrepresentative or inaccurate”—a defense
“common to the claims made by all class members.” Id.

In a case like ours, statistical evidence of causation
could meet both criteria. As the Neurontin Court recognized in
its two non-class-action decisions, an individual TPP may use
statistical evidence, combined with circumstantial evidence, to
prove a pharmaceutical manufacturer’s fraud caused it to
reimburse more prescriptions than it otherwise would have. See
Neurontin I, 712 F.3d at 45–47; Neurontin II, 712 F.3d at 57–
58. GSK cites no authority holding otherwise.

The “primary defense” against this common evidence
would also apply class-wide. Tyson, 577 U.S. at 457. “[T]he
question is whether [the drug maker’s] misrepresentations
caused an injury to each [TPP], and because each [TPP] paid
for numerous [drug] prescriptions, each would have been
injured . . . so long as at least some of the prescriptions for
which it paid were written in reliance on those
misrepresentations.” Sergeants Benevolent, 806 F.3d at 94
(emphasis omitted). Because it is enough for the TPPs to show
“some prescriptions paid for by each class member were
written based on the defendant’s alleged misrepresentations,”
id. (emphasis in text), “[t]he existence of some doctors who

46

purportedly were not influenced by [the defendant’s]
misinformation would not defeat the inference that this
misinformation had a significant influence on prescribing
decisions which injured [the plaintiff],” Neurontin I, 712 F.3d
at 45. The theme: proving any particular doctor prescribed
Avandia without relying on GSK’s misrepresentations would
not disprove the Plans’ theory of the case. Their claim is true
regardless whether any specific physician prescribed Avandia
in reliance on GSK’s fraud, so long as enough doctors relied to
justify the inference each class member covered at least one
prescription that was made in reliance. That is the sort of claim
statistical evidence is all about. As in Tyson, the defendant’s
“primary defense [is] to show that [the] study was
unrepresentative or inaccurate”—a defense “common to the
claims made by all class members.” Tyson, 577 U.S. at 457.

All that said, not just any statistical evidence will do. As
we have recognized, correlation does not imply causation. See
City of Hoboken v. Chevron Corp., 45 F.4th 699, 709–10 (3d
Cir. 2022); Linerboard, 305 F.3d at 153. We do not see how a
reasonable jury could find that a drug maker’s fraud injured a
TPP in an individual action with nothing more than evidence
of a correlation between the exposure of the fraud and a decline
in prescriptions market-wide. Evidence of mere correlation
leaves causation speculative, particularly regarding
quantitative questions like how much of the prescription
volume the fraud caused or the probability that every member
of the class covered at least one relevant prescription. The point
is not that a jury might not believe the data. It is that there is a
limit to what the data, if believed, can prove. See Comcast, 569
U.S. at 36 n.5.

47

Informed by the First, Second, and Ninth Circuits, we
hold that TPPs in a pharmaceutical fraud RICO action may
prove but-for causation with class-wide statistical evidence so
long as that evidence is sufficiently rigorous to show causation,
not just correlation. Statistical evidence has such rigor if, like
a regression analysis, it can distinguish the causal significance
of the variable at issue and justify the rejection of competing
explanations.

We turn to the evidence in this case. It is much more
like the evidence that fell short in Zyprexa and Sergeants
Benevolent than the evidence that courts indicated would
suffice in the Neurontin cases, Celexa, and Painters. The Plans
have about the same circumstantial evidence the plaintiffs had
in all those cases. But that evidence boils down to a graph
showing Avandia prescriptions fell sharply after the Nissen
study exposed the drug’s true cardiac profile. They have no
regression analysis or expert testimony to distinguish causation
from correlation by isolating other potential variables; the
expert concluding that each TPP likely paid for Avandia in
reliance on the fraud was offered only for damages. No
appellate court has yet held TPPs established predominance on
RICO but-for causation with as weak statistical evidence as the
Plans so far put forward here. We will not be the first.

The Plans refer to four forms of statistical evidence of
causation: GSK’s internal marketing studies, the Rosenthal
report, the McGuire report, and the post-Nissen study drop in
Avandia prescriptions. None alone holds up. As discussed in
our review of the District Court’s analysis, GSK’s internal
marketing studies cannot show its fraud caused the Plans’
injuries because it appears they do not even purport to isolate
the causal effects of its allegedly illicit marketing from those

48

of its marketing as a whole. The regression analysis the Plans
commissioned from Dr. Rosenthal, which purported to show
GSK’s fraudulent marketing caused TPPs to pay for Avandia
prescriptions, was excluded after the Daubert hearing. The
Plans have not appealed that ruling. Whatever the McGuire
report may be able to prove, the Plans may not use it to prove
causation in this appeal. They did not offer it to
prove causation in the pre-trial proceedings, and the District
Court expressly declined to consider the report for purposes of
establishing reliance—and thus, causation—in its class-
certification decision. Avandia, 2025 WL 1479618, at *9 n.64.

That leaves the post-Nissen study drop—the heart of the
Plans’ alternative basis for affirmance. They introduced data
showing that when the Nissen study came out in 2007, Avandia
prescriptions sharply declined. Their aim was to use
“Nissen’s publication in 2007 as a natural experiment to show
exactly what would have happened to Avandia prescriptions
had GSK published the same information in 2005.” Answering
Br. 35 (emphasis in text). Just as Avandia prescriptions
plummeted after the Nissen study revealed the drug’s true
cardiac risks, they say, so prescriptions would have fallen had
ICT-37 exposed the danger earlier.

The post-Nissen study drop data is the very sort of
statistical evidence the Second Circuit found insufficient in
Zyprexa and Sergeants Benevolent. Recall what that Court said
about Zyprexa: “To prove that doctors had, in fact, relied on
Lilly’s misrepresentations in making their prescription
decisions, the [Zyprexa] plaintiffs primarily offered evidence
that the number of Zyprexa prescriptions fell after the drug’s
weight- and diabetes-related side effects were disclosed.”
Sergeants Benevolent, 806 F.3d at 89–90 (emphasis added).

49

The Second Circuit described the Sergeants evidence in much
the same way, reporting that the plaintiffs “present[ed]
evidence showing that sales of Ketek dropped precipitously
after the FDA’s public health advisory and Ketek’s label
revisions in 2006,” and argued “this sequence of events
illustrates that doctors must have prescribed Ketek in reliance
on Aventis’s misrepresentations prior to the new safety
disclosures, because they stopped prescribing Ketek upon
learning of that new information.” Id. at 91 (emphasis in first
quotation added; emphasis in second quotation in text). That is
what the Plans have here—evidence that, after the publication
of the Nissen study, Avandia prescriptions declined. In fact, the
Plans have even less rigorous evidence than the plaintiffs who
fell short in Sergeants. At least there the insurers had an
expert’s testimony “that she had never seen anything like”
Ketek’s post-disclosure decline. Id. Here, that is absent.

The Plans’ statistical evidence thus far is also weaker
than that in Painters, Celexa, and the Neurontin cases. In
Painters, the Ninth Circuit held the TPPs’ statistical evidence
was sufficiently rigorous to prove causation class-wide
because the expert’s regression analysis accounted for
alternative explanations to explain prescription volume,
enabling the Court to confirm the exposure of the fraud caused
the ensuing decline. See Painters, 2025 WL 1683472, at *2–3.
Similarly, in Celexa the First Circuit thought the TPPs’
statistical evidence could support causation because two expert
analyses quantified the portion of prescriptions attributable to
the fraud and the odds that a given class member had
reimbursed its insureds for one of those prescriptions. See
Celexa, 915 F.3d at 13. And in the Neurontin cases the
statistical evidence sufficed because the expert’s regression
analysis isolated the causal significance of the fraud. See

50

Neurontin I, 712 F.3d at 30. Here the Plans lack the crucial
evidence that can sift out causation from correlation. Indeed,
the Neurontin and Sergeants Benevolent Courts were so
dismissive toward bare correlation data like the Plans’ Nissen
study drop evidence that they did not even characterize it as
statistical evidence at all, but as circumstantial evidence (that
is, facts that imply, rather than directly prove, a result). See
Neurontin II, 712 F.3d at 57; Sergeants Benevolent, 806 F.3d
at 92. We agree—by the standards of the First, Second, and
Ninth Circuits, the Plans’ Nissen study drop data alone does
not prove causation.

The Plans insist “[t]his case is largely indistinguishable
from Neurontin, with the only distinction being that plaintiffs’
‘aggregate evidence’ is the natural experiment of Nissen
instead of a regression analysis.” Answering Br. 32. But that
distinction makes all the difference. The problem is not with
their natural experiment approach, but with their class-wide
evidence. The Plans’ logic is straightforward: A caused C; B is
relevantly similar to A; so if B had happened, then C would
have happened. According to the Plans, the Nissen study
caused a steep decline in Avandia prescriptions; the
publication of ICT-37 would have exposed the drug’s cardiac
effects in the same way Nissen’s report did; therefore, if ICT-
37 had been published, Avandia prescriptions would have
declined as much as they did after publishing the Nissen study.
This argument is sound only if the first premise is true. And the
bare fact that Avandia prescriptions fell after the Nissen study
came out does not alone prove it is. A “natural experiment” is
no substitute for a regression analysis (or comparably robust
evidence of causation) because its first premise requires one to
establish a causal link.

51

The Plans cite cases where courts accepted natural
experiments as evidence. However, in all of those cases, the
first premise—that A either caused or did not cause C—was
backed up by expert analysis, not presumed from a bare
correlation. See In re Nat’l Collegiate Athletic Ass’n Athletic
Grant-in-Aid Cap Antitrust Litig., 958 F.3d 1239, 1249–50
(9th Cir. 2020); In re Flonase Antitrust Litig., 284 F.R.D. 207,
220, 222–23 (E.D. Pa. 2012); In re Namenda Indirect
Purchaser Antitrust Litig., 338 F.R.D. 527, 562–63 (S.D.N.Y.
2021). If there were evidence the Nissen study caused the post-
Nissen drop, then perhaps we could infer that the exposure of
the same information a few years earlier would have caused a
similar drop. But that argument cannot get off the ground
without it.

The Plans also claim they have put forward the very
case Sergeants contemplated. They have not. When the
Sergeants Court said “reliance can be proved to a jury with
sufficiently powerful aggregate evidence,” it was referring to
evidence capable of distinguishing causation from correlation.
See 806 F.3d at 96–97. While it entertained the possibility that
a precipitous decline in prescriptions alone could justify an
inference of reliance, it expressly cabined that to the
“extraordinary case” of “a drug so dangerous that no physician
would ever prescribe it to treat a non-fatal condition if that
physician were aware of its true risks.” Id. at 92. Avandia is
not that drug.

52

Though but-for causation is generally provable by
statistical evidence, the evidence the Plans have here is not yet
sufficient.
9

* * *

9
Our concurring colleague observes that one District Court in
our Circuit granted class certification in a pharmaceutical case
without a regression analysis. See In re Flonase Antitrust
Litig., 284 F.R.D. 207, 221 (E.D. Pa. 2012). But Flonase
differs in two important respects. First, it was not a RICO case.
Id. at 210–11. That matters, because predominance turns on
whether each element of the cause of action is provable by
common evidence. See Neale, 794 F.3d at 371–72. And in this
context the causation requirement of a RICO fraud claim
requires establishing reliance on the defendant’s
misrepresentations. See Sergeants Benevolent, 806 F.3d at 87.
Because Flonase was not a RICO case, the plaintiffs were not
required to prove reliance—the showing for which we seek a
regression here. Second, predominance in Flonase turned on
whether there was common evidence the defendant’s alleged
conduct—delaying the entry of a generic—led the plaintiffs to
pay higher prices. See Flonase, 284 F.R.D. at 221. As we have
acknowledged before, a price-effect theory like that does not
require proof of reliance, for its premise is that the defendant’s
conduct caused the plaintiffs’ injuries by raising prices across
the market, which the plaintiffs then paid in the ordinary
course. See Harnish, 833 F.3d at 311–12. In sum, our opinion
is consistent with Flonase because our holding concerns the
sort of common evidence that can establish reliance in a RICO
fraud action—a question Flonase did not pose, let alone
answer.

53

We summarize as follows: we may limit the scope of
our review under Rule 23(f); the District Court soundly
determined the class is ascertainable; and the Court has not
resolved all class-certification disputes. Even if it had, the
Plans have yet to satisfy the predominance requirement. The
District Court understood the Plans as seeking an inference that
every member of the class, or every doctor, relied on GSK’s
fraud. We cannot treat GSK’s studies of the effects of its
marketing campaigns in general as studies of the effects of its
fraudulent marketing specifically. And the Court thought the
law permits an inference of reliance under broader
circumstances than it does.

The Plans’ alternative ground for affirmance cannot fix
the problem. Their overarching theory is that the 2007
publication of the Nissen study caused Avandia prescriptions
to decline because it exposed the drug’s real cardiac profile, so
a 2005 publication of similar facts by GSK would have caused
a similar drop. The problem is that the Plans’ statistical
evidence shows only that Nissen’s publication in 2007 was
followed by a decline in prescriptions. That alone does not
prove causation. The First and Ninth Circuits have held TPPs
in a pharmaceutical fraud RICO action may prove common
questions predominate on but-for causation with class-wide
evidence where the statistical evidence speaks to actual
causation rather than mere correlation. The Second Circuit has
suggested it favors the same position. We agree with them. But
as noted, the Plans have not met that requirement at this time.
Without a regression analysis or other causation evidence, they
currently do not have enough to get from correlation to
causation.

54

However, it may yet be possible for the Plans to clear
the bar. They now claim the McGuire report can prove
causation. Answering Br. 29; Oral Argument Tr., Dkt. No. 70,
at 16-17. And the District Court reserved judgment on that
question. Avandia, 2025 WL 1479618, at *9 n.64. Other
evidence the Court may allow might prove relevant as well.

With this understanding of the governing law in our
Circuit, we vacate the District Court’s certification of the class
and remand for proceedings consistent with this opinion. Cf.
Hydrogen Peroxide, 552 F.3d at 326–27; Hargrove, 974 F.3d
at 481 n.8. On remand, the Court may reopen discovery on
predominance under Federal Rule of Civil Procedure 16(b) if
good cause is shown.

1

SHWARTZ, J., concurring in part and concurring in the
judgment.

I agree with my colleagues that a motions panel has the
authority to select the issues worthy of interlocutory review
under Federal Rule of Civil Procedure 23(f) and that the
proposed class is ascertainable, so I join those portions of the
Majority’s opinion in full. As to the predominance issue,
however, I join only Part III(C)(1), which explains that remand
is required to allow the District Court to address specific issues.
Therefore, I concur in the judgment.

I first address my reasons for joining Section III(C)(1)
and then explain why I decline to join my colleagues’
announcement that a regression analysis is almost always
needed in a case like this.

I

In RICO putative class actions, we evaluate whether a
district court fulfilled its obligation when considering
predominance under Rule 23 by determining whether the court
acted within its discretion in finding that the plaintiffs
demonstrated, by a preponderance of the evidence, that their
RICO claim is “capable of common proof at trial.” See In re
Lamictal Direct Purchaser Antitrust Litig., 957 F.3d 184, 190-
92 (3d Cir. 2020); In re Hydrogen Peroxide Antitrust Litig., 552
F.3d 305, 325 (3d Cir. 2008), as amended (Jan. 16, 2009)
(stating that the predominance inquiry asks whether the
elements of the claim are “susceptible to proof at trial through
available evidence common to the class”).

2

Here, the Plans presented evidence that Avandia
prescriptions plummeted after the 2007 Nissen study disclosed
cardiovascular risks that had been known to GSK for about two
years and were particularly consequential to diabetics,
although an expert opined that GSK marketed Avandia as
having cardiovascular benefits.

The Plans also told us that they will offer “GSK
marketing analyses showing the impact of” the marketing that
portrayed Avandia as having cardiovascular benefits. Appellee
Br. at 28. Although those marketing analyses are not in the
record on appeal, the District Court (1) apparently relied on
them to conclude that the Plans had offered “statistical or
econometric models that show the allegedly unlawful
promotion caused an increase in prescriptions,” In re Avandia
Mktg., Sales Pracs. & Prods. Liab. Litig., No. 07-MD-1871,
2025 WL 1479618, at *7, *9 (E.D. Pa. May 22, 2025)
(discussing arguments concerning GSK’s “own internal
statistical analyses”), and (2) rejected GSK’s argument that
those analyses did not isolate the fraudulent marketing as the
cause of increased sales, id. at *9.

The District Court, however, did not explain how (1) the
analyses isolated fraudulent marketing as a cause for the
prescription increase, or (2) if they did not isolate such a cause,
how the analyses still served as common evidence that
fraudulent marketing caused prescription volume to increase.
See id. Our precedent requires the Court to provide the reason
for its conclusion. See Hydrogen Peroxide, 552 F.3d at 321.

Furthermore, the District Court seemed to conclude that
the predominance requirement was met only because evidence
of individual reliance by every prescribing doctor was not

3

required. See Avandia, 2025 WL 1479618, at *9. Although
individualized evidence is not required, the District Court did
not explain why this view shows that the Plans presented
sufficient common evidence from which to decide whether
GSK’s scheme caused the drop.

In short, the District Court must provide its view on the
Plans’ theory that prescriptions would have dropped by the
same amount they did after the Nissen study if GSK had
disclosed what it knew years before and not fraudulently
marketed Avandia. Although the District Court mentions
GSK’s marketing materials and planning documents as
evidence of GSK’s scheme to deceive, id. at *8, it does not
explain how these materials lead to its conclusion.

For those reasons, I join the Majority to the extent it
concludes that a remand is warranted to allow the District
Court to address these matters and others the Majority
identifies in Part III(C)(1).

II

I also write separately to explain why I depart from my
colleagues’ view that a regression analysis is almost always
required to prove causation in a case like this. I view our
responsibility at this phase to determine whether there is
sufficient common evidence in the record from which a fact
finder could make a finding on the elements of the cause of
action.

Here, the record contains common evidence, including
(1) the post-Nissen decline in prescriptions, (2) expert
testimony that the risks disclosed by the Nissen Study were the

4

same ones GSK should have disclosed in 2005, (3) internal
GSK documents acknowledging that the Nissen Study was
consistent with GSK’s own conclusions reached in 2005, (4)
evidence that diabetes patients were particularly vulnerable to
cardiovascular disease, and (5) expert testimony that GSK
marketed Avandia as having cardiovascular benefits. In
addition, and notably, GSK has offered no explanation for the
plummeting prescriptions after the Nissen study’s publication
aside from the revelation of its fraud. See Oral Arg. Tr. at 44-
46. Together, this may constitute sufficient common evidence
upon which a jury could decide whether the Plans have proven
that GSK’s actions caused their injury. While it is
circumstantial, it could support an inference that at least some
Avandia prescriptions written before the drop were written in
reliance on the fraudulent marketing, particularly in view of
GSK’s failure to offer an alternative explanation for the drop.
Through this proof, the Plans (subject to District Court’s
consideration of the items described in Part I) may well have
shown that causation is “susceptible to proof at trial through”
common evidence. Hydrogen Peroxide, 552 F.3d at 318, 325.
1

My colleagues essentially assert that a regression
analysis is necessary to satisfy the predominance element and
certify the class, but this seems to exceed the requirements for

1
The District Court’s determination that the Plans have offered
sufficient common evidence for causation at class certification
does “not bind the fact-finder on the merits,” so a jury would
still be free to find that the Plans have not met their burden if
it is not convinced that this evidence shows that GSK’s fraud
caused the Plans’ injuries. In re Hydrogen Peroxide Antitrust
Litig., 552 F.3d 305, 318 (3d Cir. 2008), as amended (Jan. 16,
2009).

5

class certification. Although I recognize that the evidence must
ultimately prove causation and not correlation, that seems to be
a different question from whether sufficient common evidence
exists concerning a matter each class member, through the
named plaintiffs, must prove, namely whether the Plans have
adduced sufficient common evidence to support an inference
of causation.

Although a regression analysis may be strong evidence
of causation, and the Majority opinion cites many cases where
the presence or absence of a regression analysis was
dispositive, such proof is not required. Indeed, in In re Flonase
Antitrust Litigation, no regression was offered to support the
conclusion that purchasers who switched to a generic version
of a drug would have switched earlier if the generic had been
available earlier. 284 F.R.D. 207, 220-22 (E.D. Pa. 2012).
There, circumstantial evidence, including the fact that those
purchasers switched to the generic when it became available
and an expert’s conclusion that the generic would have been
cheaper than the brand-name if released earlier, provided
common evidence from which a jury could reasonably
conclude that those purchasers would have switched earlier,
which was sufficient. Id.; see also In re Namenda Indirect
Purchaser Antitrust Litig., 338 F.R.D. 527, 562-63 (S.D.N.Y.
2021) (concluding that every TPP class member likely
reimbursed someone who would have switched to a generic
based on (1) expert evidence that people did not often file
claims with one TPP, and (2) “massive switching from brand
to generics following generic launch”). Although Flonase was
not a RICO case and was based on a price-effect theory, it
exemplifies the propriety of drawing an inference about
causation from strong circumstantial evidence in the absence
of statistical evidence directly supporting the inference.

6

The strong common circumstantial evidence here may
similarly provide a basis upon which a factfinder could decide
whether GSK’s conduct caused the Plans’ injury, including
expert reports that Nissen revealed the exact health risks,
which were consequential to diabetic patients, that GSK
concealed. Combined with the stark drop in prescriptions after
the Nissen study revealed Avandia’s cardiovascular risks, for
which GSK offers no alternative explanation, this evidence
may well be sufficient for a jury to find that the fraudulent
marketing caused the Plans to pay for some Avandia
prescriptions that they otherwise would not have.

Because I nevertheless agree that the District Court
should evaluate the items described here in Part I and the
Majority’s opinion in Part III(C)(1), I concur in the judgment
on the predominance issue.