GREGORY, Circuit Judge:

This case concerns the certification of a class of black steel workers who allege endemic racial discrimination at a South Carolina plant owned by Nucor Corporation and Nucor Steel Berkeley (collectively, "Nucor"). Plaintiffs-appellants ("the workers") accuse Nucor of both discriminatory job promotion practices and a racially hostile work environment under Title VII of the Civil Rights Act of 1964 and 42 U.S.C. § 1981. The district court originally denied class certification for both claims, and this Court reversed. See Brown v. Nucor Corp., 576 F.3d 149 (4th Cir.2009) ("Brown I").

The district court has revisited certification and decertified the promotions class in light of the Supreme Court's opinion in Wal-Mart Stores, Inc. v. Dukes, ___ U.S. ___, 131 S.Ct. 2541, 180 L.Ed.2d 374 (2011).1 We thus again confront the question of whether the workers' have presented a common question of employment discrimination through evidence of racism in the workplace. Despite Wal-Mart's reshaping of the class action landscape, we hold that the district court has for a second time erred in refusing to certify the workers' class, where (1) statistics indicate that promotions at Nucor depended in part on whether an individual was black or white; (2) substantial anecdotal evidence suggests discrimination in specific promotions decisions in multiple plant departments; and (3) there is also significant evidence that those promotions decisions were made in the context of a racially hostile work environment.

Against that backdrop, the district court fundamentally misapprehended the reach of Wal-Mart and its application to the workers' promotions class. We thus vacate the district court's decision in part and remand for re-certification of the class.

I.

The Nucor plant encompasses six production departments that work together to melt, form, finish, and ship steel products to customers. See Brown I, 576 F.3d at 151. At the start of this litigation, 611 employees worked at the plant. Seventy-one (11.62%) were black.2 There was, however, at most one black supervisor in the production departments until after the Equal Employment Opportunity Commission ("EEOC") initiated charges that preceded the putative class action.

The workers' promotions claim rests on alternative theories of liability under Title VII, which prohibits employment discrimination because of an individual's "race, color, religion, sex, or national origin." 42 U.S.C. § 2000e-2. The promotions claim first alleges a pattern or practice of racially disparate treatment in promotions decisions. See Teamsters v. United States, 431 U.S. 324, 336, 97 S.Ct. 1843, 52 L.Ed.2d 396 (1977). Second, it charges that Nucor's facially neutral promotions policies and procedures had a racially disparate impact. See Griggs v. Duke Power Co., 401 U.S. 424, 431, 91 S.Ct. 849, 28 L.Ed.2d 158 (1971); Wal-Mart, 131 S.Ct. at 2554.

Both theories are grounded in a statistical analysis of racial disparities in job promotions at the plant combined with anecdotal evidence of discrimination. The workers' statistical evidence spans the four-year period preceding the litigation, between December 1999 and December 2003. Because Nucor destroyed and/or discarded the actual bidding data for the period before 2001, the workers' experts established an alternative benchmark using `change-of-status' forms filed by the company whenever a promotion took place at the plant. The experts extrapolated comparative statistics for that period using an assumption that the racial composition of the bidding pool for those jobs was the same as for the post-2001 jobs analyzed (when Nucor retained actual bidding records).

The workers also presented abundant direct and circumstantial anecdotal evidence of discrimination in promotions, including:

The facts undergirding the workers' separate hostile work environment claim, not directly at issue in this appeal, also bear on the promotions analysis. Those facts are disquieting in their volume, specificity, and consistency. Supervisors allegedly routinely referred to black workers as "nigger" and "DAN (dumb ass nigger)," with one supervisor reportedly stating "niggers aren't smart enough" to break production records, while others tolerated the routine use of epithets like "bologna lips," "yard ape," and "porch monkey." These epithets and others were broadcast over the plant-wide radio system — comprising a network of walkie-talkies used to communicate — along with monkey noises and the songs "Dixie" and "High Cotton." The workers' declarations and depositions further suggest that departmental supervisors and the plant's general manager consistently ignored racial harassment carried out by white workers, including the circulation of racist emails, the prominent display of a hangman's noose, the commonplace showing of the Confederate flag, and an episode when a white employee draped a white sheet over his head with eyes cut out in the form of a KKK hood.

In 2007, the South Carolina district court denied the workers' motion for class certification for both the promotions and hostile work environment claims. In 2009, a divided panel of this Court reversed, concluding that the workers satisfied the threshold requirements of Federal Rule of Civil Procedure 23. We remanded the case "with instructions to certify the appellants' class action." Brown I, 576 F.3d at 160.

On February 17, 2011, the district court followed our instructions to certify the class, concluding that the workers satisfied Rule 23(b)(3)'s requirements that common questions predominate and that the class action was superior to other litigation devices to resolve the dispute. The district court later declined to stay the case pending a ruling in Wal-Mart, and it declined to reconsider its order certifying the class.

The Supreme Court decided Wal-Mart in June 2011, decertifying an unprecedented nationwide class of approximately 1.5 million female employees spread over 3,400 stores. Wal-Mart held that the plaintiffs had failed to present a "common contention" of employment discrimination capable of "classwide resolution," as required by Rule 23(a)(2). Wal-Mart, 131 S.Ct. at 2551. Given the diffuse class and number of employment decisions at issue, the Supreme Court observed that "[w]ithout some glue holding the alleged reasons for all those decisions together, it will be impossible to say that examination of all class members' claims for relief will produce a common answer to the crucial question why was I disfavored." Id. at 2552 (emphasis in original). The plaintiffs, Wal-Mart concluded, failed to meet that standard when they premised liability on a company policy of decentralized subjective decision-making by local managers, combined with statistics showing gender-based employment disparities, limited anecdotal evidence, and expert testimony about a corporate culture that allowed for the transmission of bias. See id. at 2551, 2554-55.

On September 11, 2012, the district court relied on Wal-Mart to decertify the workers' promotions class, invoking the court's authority under Rule 23(c)(1)(C) to amend a certification order at any time before final judgment. Wal-Mart, the court observed, clarified and heightened the commonality requirement of Rule 23(a)(2), requiring the workers to present "significant proof" that Nucor "operated under a general policy of discrimination" and that they suffered a common injury. J.A. 10934 (quoting Wal-Mart, 131 S.Ct. at 2553).

Under that standard, the district court concluded that decertification of the promotions class was required because: (1) this Court's examination of the workers' statistical analysis in Brown I was not sufficiently "rigorous" to assess whether it raised questions common to the class under Rule 23(a)(2); (2) the workers' statistical and anecdotal evidence failed to establish such commonality because it did not provide "significant proof" that there existed both a "general policy of discrimination" and a "common injury"; (3) the delegation of subjective decision-making to Nucor supervisors was not, without more, a sufficiently uniform policy to present "`common' issues appropriate for resolution on a class-wide basis"; and (4) even if the workers had identified a common question of law or fact satisfying Rule 23(a)(2), they failed to independently satisfy Rule 23(b)(3)'s requirements that common issues predominate and that the class action is a superior litigation device.

Although the court decertified the class for the promotions claim, it refused to do so for the hostile work environment claim. The district court reaffirmed that the workers had demonstrated that the "landscape of the total work environment was hostile towards the class." J.A. 10964 (quoting Newsome v. Up-To-Date Laundry, Inc., 219 F.R.D. 356, 362 (D.Md. 2004)). Unlike the promotions claim, the court determined that the hostile environment allegations required no showing of a company-wide adherence to a common policy of discrimination. Still, the court found that "there is significant evidence that management ignored a wide range of harassment" and that the workers "met their burden to present significant proof of a general policy of discrimination." J.A. 10968.

On September 30, 2013, the workers appealed the district court's decertification of the promotions class.

II.

We typically review a district court's certification order for abuse of discretion. Doe v. Chao, 306 F.3d 170, 183 (4th Cir.2002), aff'd on other grounds, 540 U.S. 614, 124 S.Ct. 1204, 157 L.Ed.2d 1122 (2004). We review de novo, however, whether a district court contravenes a prior express or implicit mandate issued by this Court. United States v. Bell, 5 F.3d 64, 66 (4th Cir.1993); S. Atl. Ltd. P'ship of Tenn. v. Riese, 356 F.3d 576, 583 (4th Cir.2004) ("We review de novo ... whether a post-mandate judgment of a district court contravenes the mandate rule, or whether the mandate has been `scrupulously and fully carried out.'" (quoting 2A Fed. Proc., L.Ed. § 3:1016)).

Determining the appropriate standard of review thus requires a two step approach. First, we examine de novo whether the district court's decertification order violated our mandate in Brown I to certify the workers' class. Second, if no such violation occurred, we must determine anew whether the district court abused its discretion in decertifying the promotions class.

As to the first question, an "extraordinary" exception to the mandate rule exists when there is "a show[ing] that controlling legal authority has changed dramatically." Bell, 5 F.3d at 67 (alteration in original). Moreover, Rule 23(c)(1)(C) provides a district court with broad discretion to alter or amend a prior class certification decision at any time before final judgment.

Against that backdrop, the parties disagree about whether Wal-Mart provided sufficient justification for the district court to invoke its powers to revisit certification. Nucor maintains that Wal-Mart represents a "sea change" and that "class actions may proceed only in the most exceptional of cases." Resp'ts' Br. 15, 20. The workers suggest, however, that the Supreme Court instead largely reaffirmed existing precedent. Appellants' Br. 34.

The truth has settled somewhere in between. See Scott v. Family Dollar Stores, Inc., 733 F.3d 105, 113-14 (4th Cir.2013) (discussing limitations on the scope of Wal-Mart's holding); McReynolds v. Merrill Lynch, Pierce, Fenner & Smith, Inc., 672 F.3d 482, 487-88 (7th Cir.2012), cert. denied, ___ U.S. ___, 133 S.Ct. 338, 184 L.Ed.2d 157 (2012) (finding that Wal-Mart provided the basis for a renewed class certification motion); DL v. District of Columbia, 713 F.3d 120, 126 (D.C.Cir.2013) (surveying how Wal-Mart has changed the class action landscape); Elizabeth Tippett, Robbing A Barren Vault: The Implications of Dukes v. Wal-Mart for Cases Challenging Subjective Employment Practices, 29 Hofstra Lab. & Emp. L.J. 433 (2012) (using an empirical analysis to predict Wal-Mart's likely impact on class certifications in the future). At the very least, Wal-Mart recalibrated and sharpened the lens through which a court examines class certification decisions under Rule 23(a)(2), an impact plainly manifested by the number of certifications overturned in its wake. See, e.g., EQT Prod. Co. v. Adair, 764 F.3d 347 (4th Cir.2014); Rodriguez v. Nat'l City Bank, 726 F.3d 372, 376 (3d Cir.2013); M.D. ex rel. Stukenberg v. Perry, 675 F.3d 832, 839, 841-44 (5th Cir. 2012); Ellis v. Costco Wholesale Corp., 657 F.3d 970, 974 (9th Cir.2011).

In that light, we find that the district court's decision to reconsider the certification of the workers' class did not itself violate our mandate in Brown I. Per this Court's original remand instructions, the district court certified both the promotions and hostile work environment classes. Although the court had no discretion to then reconsider questions decided by this Court under then-existing facts and law, Wal-Mart provided a sufficiently significant change in the governing legal standard to permit a limited reexamination of whether the class satisfied the commonality requirement of Rule 23(a)(2).3 There are, however, instances described below when the district court unnecessarily revisited other discrete determinations made by this Court in Brown I, such as whether the Nucor plant should be treated analytically as a single entity, and whether the class independently met the requirements of Rule 23(b)(3). The reconsideration of those determinations was not compelled by Wal-Mart and contravened our mandate in Brown I.

Because the district court could reexamine whether the workers met the requirement of commonality, we review those findings under the abuse of discretion standard that typically applies to certification orders. See Amchem Prods., Inc. v. Windsor, 521 U.S. 591, 630, 117 S.Ct. 2231, 138 L.Ed.2d 689 (1997) ("The law gives broad leeway to district courts in making class certification decisions, and their judgments are to be reviewed by the court of appeals only for abuse of discretion."); Brown I, 576 F.3d at 152; Thorn v. Jefferson-Pilot Life Ins. Co., 445 F.3d 311, 317 (4th Cir.2006). A district court abuses its discretion when it materially misapplies the requirements of Rule 23. See Gunnells v. Healthplan Servs., Inc., 348 F.3d 417, 424 (4th Cir.2003); Thorn, 445 F.3d at 317-18 ("A district court per se abuses its discretion when it makes an error of law or clearly errs in its factual findings."). The decisive question here is whether the district court materially misapplied Rule 23(a)(2) to the facts at hand in light of Wal-Mart.4

III.

Rule 23(a)(2) establishes that a class action may be maintained only if "there are questions of law or fact common to the class." The district court determined that Wal-Mart required decertification of the workers' promotions class insofar as the Supreme Court's interpretation of the rule (1) emphasized the analytical rigor required to evaluate a plaintiff's statistical evidence of commonality at the class certification stage, (2) placed the burden on plaintiffs to provide "significant proof" of a "general policy of discrimination" and "common injury," and (3) relatedly established that a company's policy of discretionary decision-making cannot sustain class certification without a showing that supervisors exercised their discretion in a common way.

Each of these arguments is considered in turn.

A.

Wal-Mart reaffirmed existing precedent that courts must rigorously examine whether plaintiffs have met the prerequisites of Rule 23(a) at the certification stage, an analysis that will often overlap with the merits of a claim. Wal-Mart, 131 S.Ct. at 2551 (citing Gen. Tel. Co. of Sw. v. Falcon, 457 U.S. 147, 160-61, 102 S.Ct. 2364, 72 L.Ed.2d 740 (1982)). But as the Court later clarified, "Rule 23 grants courts no license to engage in free-ranging merits inquiries at the certification stage." Amgen Inc. v. Conn. Ret. Plans & Trust Funds, ___ U.S. ___, 133 S.Ct. 1184, 1194-95, 185 L.Ed.2d 308 (2013). Instead, the merits of a claim may be considered only when "relevant to determining whether the Rule 23 prerequisites for class certification are satisfied." Id. at 1195.5

This Court's precedent and its approach in Brown I are consistent with Wal-Mart and Falcon. See Gariety v. Grant Thornton, LLP, 368 F.3d 356, 366 (4th Cir.2004) (observing that "while an evaluation of the merits to determine the strength of the plaintiffs' case is not part of a Rule 23 analysis, the factors spelled out in Rule 23 must be addressed through findings, even if they overlap with issues on the merits"). In Brown I, this Court expressly invoked Falcon's requirement of a rigorous analysis to determine compliance with Rule 23. 576 F.3d at 152. More important, of course, we actually conducted such an analysis, providing a detailed evaluation of the workers' anecdotal and statistical evidence to ensure that it presented a common question under Rule 23(a)(2). Id. at 153-56.

Contrary to the dissent's assertion, we do not (and Brown I did not) suggest that Rule 23 is a mere pleading standard. See post at 933. Far from it. It is true that Brown I cautioned that "an in-depth assessment of the merits of appellants' claim at this stage would be improper." Id. at 156. Such a statement, however, is consistent with the Supreme Court's dictate in Amgen that a court should engage the merits of a claim only to the extent necessary to verify that Rule 23 has been satisfied. Amgen, 133 S.Ct. at 1194-95. Brown I did precisely that.

1.

Even evaluated in a still more painstaking manner, the workers' statistical evidence is methodologically sound while yielding results that satisfy Wal-Mart's heightened requirement of commonality discussed below. The parties' central dispute concerns the data used to analyze the period from December 1999 to January 2001, when Nucor failed to retain actual bidding records. For that period, the workers' expert developed an alternative benchmark that uses 27 relevant `change-of-status' forms — filled out when an employee changes positions at the plant — to extrapolate promotions data because actual bidding information was unavailable.

Of course, it belabors the obvious to observe that the alternative benchmark is a less precise measure than actual bidding data. It is also clear, however, that plaintiffs may rely on other reliable data sources and estimates when a company has destroyed or discarded the primary evidence in a discrimination case. More than two decades of this Court's precedent affirm as much. See Lewis v. Bloomsburg Mills, Inc., 773 F.2d 561 (4th Cir.1985); United States v. County of Fairfax, 629 F.2d 932, 940 (4th Cir.1980); see generally Ramona L. Paetzold & Steven L. Willborn, The Statistics of Discrimination: Using Statistical Evidence in Discrimination Cases § 4.03 (2014) (describing the use of proxy data when actual data is unavailable or unreliable). In Lewis v. Bloomsburg Mills, Inc., this Court approved the use of Census data to establish a hypothetical available pool of black female job applicants after a company discarded employment applications for the relevant period. 773 F.2d at 568.6 Plaintiffs then compared the "observed" annual rate of hires of black women with the "expected" rates based upon the proportional availability of black females in the labor pool. Id. We endorsed a similar use of proxy data in United States v. County of Fairfax, involving a county government that had destroyed three years of employment applications. 629 F.2d at 940. To analyze hiring during that time, plaintiffs assumed that the proportion of black and women applicants for those years was the same as in the first year for which the county retained records. Id. This Court approved, concluding the alternative benchmark was "the most salient proof of the County's labor market." Id.7

2.

The critical question is thus not whether the data used is perfect but instead whether it is reliable and probative of discrimination. To that end, a court must examine whether any statistical assumptions made in the analysis are reasonable. See Paetzold & Willborn, supra, § 4.16. The district court here identified two assumptions made by the workers' experts as problematic.

The district court first questioned the assumption that the job changes described on the 27 forms represent promotions. See J.A. 10942. As an example of clear factual error committed by the court, it quoted at length from the dissent in Brown I to argue that the forms may represent job changes unrelated to promotions. J.A. 10942 (quoting Brown I, 576 F.3d at 167-68 (Agee, J., concurring in part and dissenting in part)). The forms cited in Judge Agee's original dissent, however, are plainly not among the 27 relied upon by the workers' experts in constructing the alternative benchmark. Compare J.A. 10942 (the district court's decertification order quoting the dissent in Brown I), with J.A. 11005-11032 (copies of the actual change-of-status forms used in the expert analysis). Worse still, the dissent in Brown I reached the question of whether the 27 forms represented promotions without the issue having been raised, much less analyzed, by the district court in its original order denying certification, see J.A. 8979, or by Nucor itself in its briefing before this Court in Brown I.8 The dissent in Brown I thus both engaged in sua sponte fact-finding to divine which forms were used, and then got the facts wrong.9 Using the flawed data, the dissent concluded in Brown I that "[o]n this record, it is difficult, if not impossible to discern whether the 2000 data based on the nebulous change-of-status forms proves those positions were promotion positions available for employee bidding and thus relevant to the formulation of statistical evidence for the appellants' claims." Brown I, 576 F.3d at 168 (Agee, J., concurring in part and dissenting in part). The district court expressly embraced that conclusion in decertifying the promotions class after Wal-Mart. J.A. 10942.

Upon examining the correct change-of-status forms, discerning whether they represent promotions is a relatively straightforward enterprise. Nineteen of the 27 forms expressly state they are for a promotion, for a "successful bidder" on a "higher position," or for a new position that was "awarded" or "earned." Two of the forms describe changes in job classification accompanied by an increase in pay. One form notes that an inspector was a "successful bidder" on a mill adjuster job — a move referred to on another change form as a promotion. Two forms are for a "successful bidder" on a new position where no new pay grade is noted. The remaining three forms appear to involve changes in positions or training that involved a decrease in pay, but there is no indication, or argument by Nucor or the district court, that the exclusion of those forms would substantially undermine the probativeness of the expert analysis.

The second assumption criticized by the district court was that the bidding pools for the 27 positions filled between December 1999 and January 2001 had the same average racial composition as the pools for similar jobs analyzed from 2001 to December 2003, when the company retained actual bidding data. Because of discovery limitations imposed by the district court, the information available regarding the 2001-2003 promotions was restricted to positions similar to ones bid on by the named plaintiffs, where there was at least one black bidder. However, because Nucor failed to retain bidding records for 1999-2000, the data from that period could not be limited to positions where there was a known black bidder. Instead, the alternative benchmark had to assume that there was at least one black worker applying for each promotion analyzed — an assumption that the district court concluded helped render the statistical analysis unreliable. But as we already determined in Brown I, the assumption does not fatally undermine the probativeness of the experts' findings. The workers' experts limited the records they analyzed to the same positions identified in the later period when bidding data was available, positions for which there was a black bidder. J.A. 1161-62. In its original order denying certification, the district court observed that the assumptions regarding bidding "may be reasonable and the statistics based thereon may be relevant to prove discrimination at the plant," but "the necessity of the assumptions diminishes their probative value."10 J.A. 8987; see also Brown I, 576 F.3d at 156. As we previously concluded, an incremental reduction in probative value — which is a natural consequence of the use of proxy data — does not itself render a statistical study unreliable in establishing a question of discrimination common to the class. Brown I, 576 F.3d at 156. Indeed, to conclude otherwise would undermine our prior precedent in cases like Lewis and Fairfax, rendering plaintiffs unable to bring a statistics-based employment discrimination claim after a company has intentionally or inadvertently destroyed actual applicant data.11 See Lewis, 773 F.2d at 568; Fairfax, 629 F.2d at 940.

3.

The dissent points to still more statistical assumptions — assumptions not discussed by either the district court or Nucor — to further question the reliability of the alternative benchmark. Specifically, the dissent suggests that the black workers may not have been qualified for higher paying jobs and that they may have been denied promotions because of disciplinary records that were not themselves the result of racial animus. See post at 940, 941-43. As to the qualifications of the workers, Nucor identifies nothing in the record — or in any factual findings by the district court — to suggest that black workers regularly applied for jobs for which they were not qualified, such that the reliability of the study would be compromised. Indeed, the Nucor job postings explicitly listed the minimum qualifications required, and the workers' experts reasonably assumed that individuals would normally apply only if they believe they met such qualifications. See J.A. 7763 (an example of a job posting); J.A. 1162. That is not to say that patently unqualified workers did not apply in isolated cases. But there is no reason to believe that such incidents would have substantially reduced the reliability of the statistical conclusions. It also bears repeating that it was Nucor that failed to retain or produce records that would have allowed the experts to take other variables like qualifications more precisely into account. See J.A. 1165.

The dissent, however, goes a step further in speculating that black workers may have been denied promotions because of their disciplinary records. See post at 940. Again, Nucor itself does not make this argument. Instead, the argument the dissent constructs is based on the company's self-serving responses to the workers' interrogatories and requests for production — where Nucor asserts that some of the black workers were not chosen for promotions due to disciplinary issues. The record, however, does not include disciplinary records for the named plaintiffs or putative class members. More fundamental, the workers allege that any disproportionate disciplinary action levied against them was itself a product of racial discrimination, with the disciplinary records then used as a pretext in hiring decisions. As worker Ramon Roane has stated:

J.A. 1000; see also J.A. 1024 (Alvin Simmons's statement that a white employee was promoted over him despite the fact that the white employee "had been disciplined less than a year earlier for `not paying attention' when operating equipment"); J.A. 1111 (Earl Ravenell's statement that black workers were disproportionately singled out for disciplinary action); J.A. 6783 (Michael Rhode's description of discrimination in disciplinary action). See generally J.A. 10960-10972 (the district court's factual findings regarding the existence of a racially hostile work environment); Desert Palace, Inc. v. Costa, 539 U.S. 90, 101-02, 123 S.Ct. 2148, 156 L.Ed.2d 84 (2003) (allowing the use of circumstantial evidence to show that race was a motivating factor in a "mixed-motive" case involving both legitimate and illegitimate reasons for an employment decision); Rowland v. Am. Gen. Fin., Inc., 340 F.3d 187, 193-94 (4th Cir.2003) (allowing the use of circumstantial evidence to show that gender was "a motivating factor" in a failure to promote an employee). Given that background, it is easy to see why the district court chose not to advance the arguments that the dissent makes today.

Finally, the dissent criticizes the assumption that the 27 positions identified were actually open for bidding.12 Post at 939-40. That assumption, however, derives directly from Nucor's stated policy that every job vacancy is posted on plant bulletin boards and is open to bidding plant-wide — a policy cited by Nucor's own expert and the district court. See J.A. 5887 (the Report of Finis Welch, observing that "[o]pen positions are posted on bulletin boards and through email," and that "[a]ll employees in the plant are eligible to bid on a posted job"); see also Resp'ts' Br. 9 ("Department managers set the process in motion by sending postings for available promotions to Personnel employees, who performed a purely clerical role and advertised postings plantwide."); J.A. 8979 (the district court's original order denying certification, finding that "[w]hen a position in a department becomes available, the job is posted on the plant's e-mail system, which is accessible to all employees in the plant"). The dissent nonetheless argues that the statistical assumption was unreasonable.13 We disagree.

4.

With the alternative benchmark evidence included, the statistical disparity in promotions is statistically significant at 2.54 standard deviations from what would be expected if race were a neutral factor. See Hazelwood Sch. Dist. v. United States, 433 U.S. 299, 308 n. 14, 97 S.Ct. 2736, 53 L.Ed.2d 768 (1977) (indicating that anything greater than two or three standard deviations in racial discrimination cases is suspicious, at least for large sample sizes); Brown I, 576 F.3d at 156 n. 9 (applying the Hazelwood standard to the workers' statistical evidence); Jones v. City of Boston, 752 F.3d 38, 46-47 (1st Cir.2014) (observing that two standard deviations has become the commonly accepted threshold for social scientists and federal courts "in analyzing statistical showings of disparate impact"). According to the experts' analysis, black employees constitute 19.24% of those who applied for relevant promotions. Yet such employees were only 7.94% percent of those promoted.

Of course, statistical significance is not always synonymous with legal significance. EEOC v. Fed. Reserve Bank of Richmond, 698 F.2d 633, 648 (4th Cir.1983) rev'd on other grounds sub nom. Cooper v. Fed. Reserve Bank of Richmond, 467 U.S. 867, 104 S.Ct. 2794, 81 L.Ed.2d 718 (1984). Indeed, the usefulness of statistical evidence often "depends on all of the surrounding facts and circumstances." Teamsters, 431 U.S. at 340, 97 S.Ct. 1843. Here, the surrounding circumstances and anecdotal evidence of discrimination, as described in greater detail below, are precisely what help animate the statistical findings.14 As we held in Brown I and reaffirm today, "because the appellants' direct evidence alone was sufficient to demonstrate common claims of disparate treatment and disparate impact, their statistical data did not need to meet a two-standard-deviation threshold." Brown I, 576 F.3d at 156-57. Thus it is plain that when the statistical disparity actually exceeded two standard deviations, the district court abused its discretion in decertifying the class.

B.

The district court further concluded that the workers' statistical and anecdotal evidence was insufficient for class certification insofar as the evidence did not demonstrate a uniform class-wide injury that spanned the entire Nucor plant. As the court observed, Wal-Mart instructs that plaintiffs must present a common contention capable of being proven or disproven in "one stroke" to satisfy Rule 23(a)(2)'s commonality requirement. Wal-Mart, 131 S.Ct. at 2551. Thus, a class-wide proceeding must be able to generate common answers that drive the litigation. Id.; see also Jimenez v. Allstate Ins. Co., 765 F.3d 1161, 1165 (9th Cir.2014) (observing that "a class meets Rule 23(a)(2)'s commonality requirement when the common questions it has raised are apt to drive the resolution of the litigation, no matter their number" (internal quotation marks omitted)). For a claim based on discrimination in employment decisions, "[w]ithout some glue holding the alleged reasons for all those decisions together, it will be impossible to say that examination of all the class members' claims for relief will produce a common answer to the crucial question why was I disfavored." Wal-Mart, 131 S.Ct. at 2552 (emphasis omitted); see also Scott v. Family Dollar Stores, Inc., 733 F.3d 105, 113 (4th Cir.2013).

The workers here most generally present two such common contentions capable of class-wide answers under Title VII. Under a disparate treatment theory, the common contention is that Nucor engaged in a pattern or practice of unlawful discrimination against black workers in promotions decisions. See Teamsters, 431 U.S. at 336, 97 S.Ct. 1843. Under the workers' disparate impact theory, the common contention is that a facially neutral promotions policy resulted in a disparate racial impact. See Griggs, 401 U.S. at 429-31, 91 S.Ct. 849. As Wal-Mart observed, however, semantic dexterity in crafting a common contention is not enough. Commonality instead "requires the plaintiff to demonstrate that the class members `have suffered the same injury[.]'" Wal-Mart, 131 S.Ct. at 2551 (quoting Falcon, 457 U.S. at 157, 102 S.Ct. 2364). As such, a court must examine whether differences between class members impede the discovery of common answers. Id. at 2551.

In the absence of a common job evaluation procedure, Wal-Mart held that statistical proof of employment discrimination at the regional and national level, coupled with limited anecdotal evidence from some states, was insufficient to show that the company maintained a "general policy of discrimination" present in each store where class members worked. See Wal-Mart, 131 S.Ct. at 2553. Similarly, the district court here found that the workers' statistical and anecdotal evidence was insufficient to show a general policy in all Nucor departments that caused the class injury.

The district court, however, failed to adequately appreciate three significant differences from Wal-Mart that make the case largely inapposite to the facts at hand.

1.

First, Wal-Mart discounted the plaintiffs' statistical evidence in large part because the statistics failed to show discrimination on a store-by-store basis. See Wal-Mart, 131 S.Ct. at 2555. As such, the plaintiffs could not establish that a store greeter in Northern California, for instance, was subject to the same discrimination as a cashier in New Hampshire. These dissimilarities between class members were exacerbated by the sheer size of the Wal-Mart class — 1.5 million members working at 3,400 stores under "a kaleidoscope of supervisors (male and female), subject to a variety of regional policies that all differed." Id. at 2557 (quoting Dukes v. Wal-Mart Stores, Inc., 603 F.3d 571, 652 (9th Cir.2010) (Kozinski, J., dissenting)). The scale and scope of the putative class, combined with the nature of the evidence offered, was thus essential to Wal-Mart's holding. Had the class been limited to a single Wal-Mart store spanning multiple departments, or had the plaintiffs' evidence captured discrimination at a store level, a very different Rule 23(a)(2) analysis would have been required.

In contrast to Wal-Mart, this litigation concerns approximately 100 class members in a single steel plant in Huger, South Carolina. The class members shared common spaces, were in regular physical contact with other departments, could apply for promotions in other departments, and were subject to hostile plant-wide policies and practices. See Brown I, 576 F.3d at 151. Such differences are not merely superficial. Instead, a more centralized, circumscribed environment generally increases the uniformity of shared injuries, the consistency with which managerial discretion is exercised, and the likelihood that one manager's promotions decisions will impact employees in other departments. That is particularly the case where, as discussed further below, the entire Nucor plant was allegedly infected by express racial bias and stereotypes — a culture that management took few affirmative steps to meaningfully combat.

Nonetheless, the district court analogized to Wal-Mart in finding that the workers' evidence of discrimination was insufficient because it disproportionately concerned a single department — the Beam Mill — and because there was an insufficient showing that all departments operated under a common policy of discrimination. J.A. 10949-54. As such, a class-wide proceeding would not generate "common answers" as Wal-Mart required, the district court found. See Wal-Mart, 131 S.Ct. at 2551.

The district court, however, inappropriately discounted, and often ignored, evidence that establishes discrimination in other Nucor departments. Although 11 of the 16 employees submitting declarations on behalf of the plaintiffs worked in the Beam Mill, the declarants describe frequent instances of alleged promotions discrimination in other departments. See J.A. 1021-24; 1032-35; 1049-51; 1055-56; 1061-63; 1085-86; 1091-92; 1103; 1110-11; 1118-19. Even the additional affidavits obtained by Nucor, discussed in further detail below, present numerous allegations of discrimination in non-Beam Mill departments. See J.A. 5992-95 (discrimination in the Hot Mill and Melt Shop); 6143-45 (discrimination in the Hot Mill); 6174 (general observations of promotions discrimination); 6369-70 (discrimination in the Melt Shop); 6505-07 (discrimination in the Hot Mill); 7036 (discrimination in the Melt Shop). The record additionally indicates numerous complaints of discrimination made to the plant's general manager, who allegedly did little to nothing in response. Such alleged tolerance of discrimination from top management at the plant supports the workers' contention of a class-wide injury that affected them all.15

The district court made a still more fundamental error by choosing to treat the Nucor departments as autonomous operations in the first place instead of part of a single facility, contravening both this Court's instructions in Brown I and the district court's own prior findings. The district court's original order to certify the class recognized that a department-by-department approach had been foreclosed, writing:

J.A. 9705. Wal-Mart provided no grounds for the court to reconsider that finding because nothing in the Supreme Court's opinion suggests that single, localized operations must be analytically dissected into component departments.16 Here, all of the workers' evidence concerns a single connected facility.

Even if not required by our prior ruling, treating the plant as a single entity remains sound. In addition to the direct and circumstantial evidence of discrimination in promotions decisions in multiple departments, racial bias in one Nucor plant department itself diminished the promotional opportunities for black workers in all the departments — including those who wanted promotions into the infected department and those who sought promotions to other departments and needed their supervisors' recommendations. To that end, the workers cogently observe that requirements for dual approvals for promotions — by originating and destination department heads — "carr[ied] the effects of racial discrimination from one department and supervisor to another, either by systemic tolerance, acquiescence or design." Appellants' Reply Br. 24 (citing Smith v. Bray, 681 F.3d 888, 897 & n. 3 (7th Cir.2012)).

Such a conclusion is further strengthened by the workers' hostile work environment claim. As the district court itself found, "the plaintiffs have submitted significant proof that the landscape of the total work environment at the Berkeley plant was hostile towards African-Americans and that the defendants failed to take `remedial action reasonably calculated to end the harassment.'" J.A. 10966; see also Brown I, 576 F.3d at 157-58. That environment, the workers argue, supports their showing of an atmosphere of systemic tolerance of racial hostility by managers and supervisors, forming part of the overall pattern or practice that "infected black employees' promotion opportunities." We agree.

2.

Second, the Wal-Mart plaintiffs' theory of commonality relied, in part, on showing that the company maintained a corporate culture that facilitated the uniform transmission of implicit, or subconscious, bias into the hiring process. See Wal-Mart, 131 S.Ct. at 2548. To that end, the plaintiffs' expert testified the company was "vulnerable" to "gender bias." Id. at 2553. The Court, however, concluded that the expert could not with specificity determine how the culture concretely influenced individual employment decisions. Id. at 2553-54. The testimony was therefore insufficient to show a common policy that produced a common injury.

Here, however, the workers have provided substantial evidence of unadulterated, consciously articulated, odious racism throughout the Nucor plant, including affirmative actions by supervisors and a widespread attitude of permissiveness of racial hostility. The examples in the record are ubiquitous: bigoted epithets and monkey noises broadcast across the plant radio system, emails with highly offensive images sent to black workers, a hangman's noose prominently displayed, a white supervisor stating that "niggers aren't smart enough" to break production records, and abundant racist graffiti in locker rooms and shared spaces. Moreover, no more than one black supervisor worked in the Nucor production departments until after the EEOC charge that preceded this litigation. It strains the intellect to posit an equitable promotions system set against that cultural backdrop, particularly in light of the other evidence presented.

The dissent rejects the idea that evidence of a racially hostile work environment may help establish a claim for disparate treatment in promotions decisions.17 Post at 945-46. Indeed, the dissent goes so far as to observe that "locker rooms and radios bear no relationship to promotions decisions." Id. at 946. Such a perspective, however, is perplexingly divorced from reality and the history of workplace discrimination. It is difficult to fathom how widespread racial animus of the type alleged here, an animus that consistently emphasized the inferiority of black workers, bears no relationship to decisions whether or not to promote an employee of that race. Although the dissent asserts that "nothing in the record supports" making a connection between the work environment and promotions practices, we are not limited to the record in making such elementary judgments. Justice is not blind to history, and we need not avert our eyes from the broader circumstances surrounding employment decisions, and the inferences that naturally follow.

3.

Third, and related, the anecdotal evidence of discrimination in this case is substantially more probative than that in Wal-Mart. The Wal-Mart plaintiffs presented affidavits from about 120 female employees, representing approximately one affidavit for every 12,500 class members. Wal-Mart, 131 S.Ct. at 2556. The affidavits captured only 235 of Wal-Mart's 3,400 stores, and there were no affidavits from workers in 14 states. Id. The evidence thus fell far short of the benchmark for a showing of company-wide discrimination established by Teamsters, 431 U.S. 324, 97 S.Ct. 1843. In Teamsters, the plaintiffs produced statistical evidence of racial bias combined with approximately 40 accounts of discrimination from particular individuals. Id. at 338, 97 S.Ct. 1843. Given the class size of approximately 334 persons, there was roughly one anecdote for every eight members of the class. See id. at 331, 338, 97 S.Ct. 1843; Wal-Mart, 131 S.Ct. at 2556. "[T]he anecdotes came from individuals spread throughout the company who for the most part worked at the company's operational centers that employed the largest numbers of the class members." See Wal-Mart, 131 S.Ct. at 2556 (internal quotation marks omitted). Similarly, this litigation includes anecdotal evidence from more than 16 individuals18 in a class that numbered approximately one-hundred "past and present black employees at the plant" at the time litigation commenced — an approximate ratio of one anecdote for every 6.25 class members.19 See Brown I, 576 F.3d at 151 (describing the class size).

Balanced against such evidence, the district court gave "limited weight" to approximately 80 affidavits from Nucor employees largely disclaiming discrimination at the plant — affidavits taken by company lawyers after the EEOC charges had been filed. See J.A. 10950-51. Common sense and prudence, however, instruct that the affidavits do little to rebut the evidence of discrimination insofar as they were given under potentially coercive circumstances, where the company reserved its ability to use them against other employees in any future lawsuit (a fact that was omitted from the Statement of Participation given to affiants). See J.A. 6003 (the Statement of Participation), 9379 (Nucor's statement that it intended "to use the affidavits for every purpose permitted under the Federal Rules of Evidence," including the opposition to class certification and the impeachment of witnesses); see also Kleiner v. First Nat'l Bank of Atlanta, 751 F.2d 1193, 1202 (11th Cir.1985) (observing that after a class action has been filed, "[a] unilateral communications scheme ... is rife with potential for coercion"); Quezada v. Schneider Logistics Transloading & Distrib., No. CV 12-2188 CAS, 2013 WL 1296761, at *5 (C.D.Cal. Mar. 25, 2013) (finding in a class action context that "[f]ailing to inform the employees of the evidence-gathering purpose of the interviews rendered the communications fundamentally misleading and deceptive because the employees were unaware that the interview was taking place in an adversarial context, and that the employees' statements could be used to limit their right to relief"); Longcrier v. HL-A Co., 595 F.Supp.2d 1218, 1228 (S.D.Ala.2008); Mevorah v. Wells Fargo Home Mort., Inc., No. C 05-1175 MHP, 2005 WL 4813532, at *4 (N.D.Cal. Nov. 17, 2005). Of course, companies may investigate allegations of discrimination and take statements from employees. But when it comes to assessing the probative value of those statements — when weighed against the numerous declarations of employees who took the often grave risk of accusing an employer of a workplace violation — courts should proceed with eyes open to the imbalance of power and competing interests.20 Moreover, as previously observed, the company-obtained affidavits still contain numerous allegations of discrimination in promotions decisions — allegations that carry significant weight given the circumstances in which they were made. See J.A. 5992-95, 6143-46, 6174, 6370, 6506, 7036.

Of course, a plaintiff need not "offer evidence that each person for whom it will ultimately seek relief was a victim of the employer's discriminatory policy." Teamsters, 431 U.S. at 360, 97 S.Ct. 1843; see also EEOC v. Korn Indus., Inc., 662 F.2d 256, 260 (4th Cir.1981). Instead, a bifurcated class action proceeding allows for a "liability" stage to first determine whether an employer engaged in a pattern or practice of discriminatory conduct. Teamsters, 431 U.S. at 360, 97 S.Ct. 1843; Korn, 662 F.2d at 260. Upon a finding of liability, a second damages stage allows for the consideration of which individuals were specifically harmed by the policy. Teamsters, 431 U.S. at 361, 97 S.Ct. 1843; Korn, 662 F.2d at 260.

4.

Here, for a liability determination in a disparate treatment claim, the workers' statistical and anecdotal evidence, especially when combined, thus provide precisely the `glue' of commonality that Wal-Mart demands. See Brown I, 576 F.3d at 156. Such a claim requires proof of a "systemwide pattern or practice" of discrimination such that the discrimination is "the regular rather than the unusual practice." Teamsters, 431 U.S. at 336, 97 S.Ct. 1843; Cooper, 467 U.S. at 875-76, 104 S.Ct. 2794; see also Wal-Mart, 131 S.Ct. at 2552 n. 7. The required discriminatory intent may be inferred upon such a showing. See Teamsters, 431 U.S. at 339-40, 97 S.Ct. 1843; Hazelwood, 433 U.S. at 308-09, 97 S.Ct. 2736 (observing that "[w]here gross statistical disparities can be shown, they alone may in a proper case constitute prima facie proof of a pattern or practice of discrimination").

Whereas there may have been many answers in Wal-Mart to the question of why any individual employee was disfavored, the workers here have sufficiently alleged that there is only one answer to the question of why Nucor's black workers were consistently disfavored.21 Unlike a disparate impact claim, a showing of disparate treatment does not require the identification of a specific employment policy responsible for the discrimination. See Teamsters, 431 U.S. at 336 n. 16, 97 S.Ct. 1843 (discussing the legislative history of Title VII and concluding that the words "pattern or practice" should be interpreted according to their plain meaning). A pattern of discrimination, revealed through statistics and anecdotal evidence, can alone support a disparate treatment claim, even where the pattern is the result of discretionary decision-making.

To hold otherwise would dramatically undermine Title VII's prophylactic powers. As the Supreme Court observed in Griggs, a central purpose of Title VII is "to achieve equality of employment opportunities and remove barriers that have operated in the past to favor an identifiable group of white employees over other employees." 401 U.S. at 429-30, 91 S.Ct. 849; see also Albemarle Paper Co. v. Moody, 422 U.S. 405, 417-18, 95 S.Ct. 2362, 45 L.Ed.2d 280 (1975) (stressing Title VII's prophylactic goals in addition to its purpose "to make persons whole for injuries suffered on account of unlawful employment discrimination"). Here, where substantial evidence suggests a pattern of engrained discriminatory decision-making that consistently disadvantaged black workers at Nucor, to deny class certification would significantly weaken Title VII as a bulwark against discrimination.

C.

Statistics and anecdotes suggesting a pattern of discrimination, however, are not enough alone to sustain a disparate impact claim. See Wal-Mart, 131 S.Ct. at 2555; Watson v. Fort Worth Bank & Trust, 487 U.S. 977, 994, 108 S.Ct. 2777, 101 L.Ed.2d 827 (1988). Disparate impact liability requires the identification of a specific employment practice that caused racially disparate results. See 42 U.S.C. § 2000e-2(k); Watson, 487 U.S. at 986-87, 108 S.Ct. 2777; Griggs, 401 U.S. at 431, 91 S.Ct. 849. Unlike disparate treatment, the disparate impact theory does not require proof of improper intent to sustain a Title VII violation. Teamsters, 431 U.S. at 349, 97 S.Ct. 1843; Griggs, 401 U.S. at 429-31, 91 S.Ct. 849 (finding the use of standardized tests resulted in a disparate impact). Instead, liability is premised on facially neutral policies. Griggs, 401 U.S. at 431, 91 S.Ct. 849.

Under Wal-Mart, a mere showing that a "policy of discretion has produced an overall ... disparity does not suffice." Wal-Mart, 131 S.Ct. at 2556. Instead, plaintiffs who allege such a policy of discretion must demonstrate that a "common mode of exercising discretion" actually existed throughout a company. Id. at 2554; see also Tabor v. Hilti, Inc., 703 F.3d 1206, 1229 (10th Cir.2013) (observing that "after Wal-Mart, federal courts ... have generally denied certification when allegedly discriminatory policies are highly discretionary and the plaintiffs do not point to a common mode of exercising discretion that pervades the entire company" (internal quotation marks omitted)). Given that standard, the district court here found that the workers "failed to identify any factor that unites the manner in which the various decision makers throughout the Berkeley plant exercised their discretion." J.A. 10955.

Wal-Mart recognizes that in certain cases, "giving discretion to lower-level supervisors can be the basis of Title VII liability under a disparate-impact theory," 131 S.Ct. at 2554, because "an employer's undisciplined system of subjective decisionmaking [can have] precisely the same effects as a system pervaded by impermissible intentional discrimination." Id. (alteration in original) (quoting Watson, 487 U.S. at 990, 108 S.Ct. 2777). For a nationwide class, Wal-Mart found that proving a consistent exercise of discretion will be difficult, if not impossible in some circumstances. Id.; see also Davis v. Cintas Corp., 717 F.3d 476, 488 (6th Cir.2013) (noting the difficulties Wal-Mart presents for parties seeking to certify a nationwide class).

But for a localized, circumscribed class of workers at a single facility, a policy of subjective, discretionary decision-making can more easily form the basis of Title VII liability, particularly when paired with a clear showing of pervasive racial hostility. In such cases, the underlying animus may help establish a consistently discriminatory exercise of discretion.

This Court's recent opinion in Scott v. Family Dollar Stores, Inc. specifically provides several ways that such a disparate impact claim may satisfy Rule 23 after Wal-Mart, including: (1) when the exercise of discretion is "tied to a specific employment practice" that "affected the class in a uniform manner"; (2) when there is "also an allegation of a company-wide policy of discrimination" that affected employment decisions; and (3) "when high-level personnel exercise" the discretion at issue. Scott, 733 F.3d at 113-14.

The first and second of Scott's alternatives are most relevant to this case. A specific employment practice or policy can comprise affirmative acts or inaction. Cf. Ellison v. Brady, 924 F.2d 872, 881 (9th Cir.1991) (explaining an employer's responsibility to act to rectify a hostile or offensive work environment under Title VII). Regarding affirmative acts, the district court has established that Nucor's promotions practice provides that "[e]mployees in each of the production departments may bid on positions available in other departments," and that in order to promote one of the bidders, "the supervisor, the department manager, and the general manager must approve a written change of status and then submit the change of status form to the personnel office." J.A. 477-78.

For purposes of class certification, the workers have provided sufficient evidence that such a policy, paired with the exercise of discretion by supervisors acting within it, created or exacerbated racially disparate results. The promotions system, requiring approvals from different levels of management, created an environment in which the discriminatory exercise of discretion by one department head harmed the promotions opportunities for all black workers at the plant by foreclosing on opportunities in that department and generally impeding upward mobility. Moreover, the disproportionate promotions of white workers had to be ratified by the general manager, Ladd Hall, who was thus on notice, or should have been on notice, that there were pronounced racial disparities in department-level promotion practices, as indicated by the statistical and anecdotal evidence presented.

The workers have also presented sufficient evidence of a practice of inaction by the general manager who ignored the evidence of, and complaints regarding, discrimination in promotions at the plant. See, e.g., J.A. 996-97, 1016, 1056, 1087, 1104. Such managerial inaction occurred despite Nucor's status as an "Equal Opportunity Employer" and its claim to have a "plantwide policy barring racial discrimination." Resp'ts' Br. 6. One black worker, Ray Roane, has testified that he complained directly to Hall about discrimination in promotions. J.A. 996-97. Hall threatened his job. J.A. 997. Consistent with that evidence, the workers observe in the context of their hostile work environment claim that despite a policy of investigating complaints of racial harassment, "[n]ot even one of the five department managers has been shown to have lifted a finger to redress the racially hostile work environment found to exist both plant-wide and in each department." Appellants' Br. 25. The workers have sufficiently alleged that such a uniform policy of managerial inaction also contributed to racial disparities in promotions decisions.

Consistent with Scott, the workers have further demonstrated that the exercise of discretion at Nucor was joined by "a company-wide policy of discrimination" that was encouraged, or at least tolerated, by supervisors and managers. See Scott, 733 F.3d at 114. In addition to the evidence of a hostile work environment previously described in detail, one white supervisor has expressly stated in a deposition that he heard the head of the Beam Mill declare, "I don't think we'll ever have a black supervisor while I'm here." J.A. 1885-86. Such facts provide a critical nexus between the racial animus at the plant and promotions decisions that impacted all black workers by foreclosing opportunities for them. Or, using Wal-Mart's language, the evidence of pervasive racial hostility in the working environment provides a "common mode of exercising discretion that pervade[d] the entire company." Wal-Mart, 131 S.Ct. at 2554-55.

In the end, Wal-Mart simply "found it unlikely" that thousands of managers across different regions "would exercise their discretion in a common way without some common direction." Tabor, 703 F.3d at 1222. Here, however, the workers have provided ample evidence supporting their allegation of a common, racially-biased exercise of discretion throughout the plant — demonstrated through alleged incidents of specific discrimination in promotions decisions, statistical disparities, and facts suggesting pervasive plant-wide racism. The district court abused its discretion in finding that such evidence was insufficient to meet the burden that Wal-Mart imposes.

IV.

Nucor further argues that the workers have failed to contest the district court's independent finding that the putative class failed to satisfy Rule 23(b)(3). As the company observes, the district court specifically held that the class failed to meet the rule's requirements for a class action seeking individualized money damages, namely, that common questions predominate over individualized inquiries and that the class action is "superior to other available methods for fairly and efficiently adjudicating the controversy." Fed.R.Civ.P. 23(b)(3). The court remarked that "even if the Fourth Circuit subsequently concludes that the plaintiffs have identified a common issue that satisfies Rule 23(a)(2), this Court nonetheless finds that `common issues,' as that term is defined by Wal-Mart, do not predominate over individual issues with regard to the plaintiffs' promotions claims."22 J.A. 10956.

Nucor contends that nowhere in the workers' opening brief is the Rule 23(b)(3) ruling addressed, and that any challenge to derives from the Federal Rules of Appellate Procedure, which require that the argument section of an appellant's opening brief contain the "appellant's contentions and the reasons for them, with citations to the authorities and parts of the record on which the appellant relies." Fed. R.App. P. 28(a)(8)(A); see also Mayfield v. Nat'l Ass'n for Stock Car Auto Racing, Inc., 674 F.3d 369, 376-77 (4th Cir.2012). "Failure of a party in its opening brief to challenge an alternate ground for a district court's ruling ... waives that challenge." United States ex rel. Ubl v. IIF Data Solutions, 650 F.3d 445, 456 (4th Cir.2011) (quoting Rodriguez v. Hayes, 591 F.3d 1105, 1118 n. 6 (9th Cir.2010)).

The workers contend first, and we agree, that no waiver occurred because their arguments in the opening brief extended to the district court's discussion of both predominance and commonality. The single issue identified by the workers on appeal did not differentiate between the court's findings on either question. The issue, as presented, was this:

Consistent with that framing, the workers' opening brief describes the district court's decision in equally broad terms without distinguishing between commonality and predominance. See Appellants' Br. 28-29 ("The district court erred as a matter of law by declining to follow this Court's mandate that held there is sufficient statistical and non-statistical evidence to certify a class covering all six production departments."); Appellants' Br. 3 (citing to the portion of the district court opinion where predominance is discussed).

Although more explicit separation of the predominance and commonality inquiries would no doubt have been wise, the workers' arguments throughout their brief directly respond to the issues the district court raised in both contexts (issues that, as discussed below, were intertwined by the court). The workers, for instance, specifically cite cases discussing predominance when arguing about the extent to which a court may look to merits in deciding certification. See Appellants' Br. 34-35. Elsewhere, in discussing the sufficiency of the anecdotal evidence presented, the workers argued in favor of our holding in Brown I that "[t]his evidence alone establishes common claims of discrimination worthy of class certification." Appellants' Br. 42 (citing Brown I, 576 F.3d at 153). Certification of the workers' class required a finding that Rule 23(b) was satisfied, in addition to a finding of commonality under Rule 23(a)(2). More generally, without limiting its analysis to the question of commonality, the workers' opening brief observes that "[t]he district court's finding that there is no pattern-or-practice evidence in the non-Beam Mill departments is directly contrary to the evidence and [the Fourth Circuit's] mandate." Appellants' Br. 42-43.

It is true that the workers arguments often focus expressly on the question of commonality, as Wal-Mart focused its analysis. In that regard, however, the workers have merely followed the district court's lead insofar as the court itself raised the same arguments under Rule 23(b)(3) as it did regarding commonality under Rule 23(a)(2).23 See J.A. 10958-59; see also United States v. Goforth, 465 F.3d 730, 737 (6th Cir.2006) (observing that "where an argument advanced in an appellant's opening brief applies to and essentially subsumes an alternative basis for affirmance not separately argued therein, the appellant does not waive that alternative basis for affirmance"). The district court based its conclusion that common issues did not predominate on the observation that because the workers' evidence disproportionately concerns the Beam Mill, "there is no `glue' connecting the promotions decisions in the Beam Mill to the decisions in the other departments." J.A. 10959. That is exactly the same argument raised, and responded to by the workers, in the context of Rule 23(a)(2) commonality. See J.A. 10950-54; Appellants' Br. 42-47. Elsewhere in its Rule 23(b)(3) discussion, the court observes that "[a]lthough there are, to varying degrees, a few allegations of discrimination in promotions in departments other than the Beam Mill, there is nothing to link these allegations to the pattern of behavior alleged in the Beam Mill." J.A. 10959. Again, this argument is also made in the Rule 23(a)(2) context and responded to in detail by the workers there. Indeed, the district court itself acknowledged that it "employ[ed] the language of Wal-Mart" regarding Rule 23(a)(2) in discussing the requirements of Rule 23(b)(3). J.A. 10958-59. In responding directly to the reasons given by the district court for its predominance determination, the workers have thus done far more than take a mere "passing shot at the issue." See Belk, Inc. v. Meyer Corp., 679 F.3d 146, 152 n. 4 (4th Cir.2012) (finding that an issue was waived after a party mentioned the issue in a heading but failed to further develop the argument); see also Williams v. Woodford, 384 F.3d 567, 587 n. 5 (9th Cir.2004) (concluding that an appellant preserved a claim for review even though the argument consisted of "eight sentences in a footnote," where the argument identified the basis of disagreement with the district court, the requested relief, and relevant citations to case law and the record).

Nonetheless, the dissent argues that "many different reasons underlay [the district court's] predominance finding, including several individual questions that could `overwhelm' common ones." Post at 924. But a plain reading of the district court's opinion belies the idea that it made any predominance arguments that were not responded to by the workers. The only specific argument cited by the dissent as unaddressed contends that because of the workers' reliance on anecdotal evidence, a jury "would have to delve into the merits of each individual promotion decision." J.A. 10959; post at 924. Yet, as observed above, the workers specifically argued that the anecdotal evidence establishes "common claims of discrimination" that merit certification, not merely a finding of commonality. Appellants' Br. 42 (quoting Brown I, 576 F.3d at 153). Indeed, such an argument is consistent with the workers' fundamental contention throughout their brief that plant-wide discrimination existed.

As this Court has observed, the purpose of the waiver doctrine is to avoid unfairness to an appellee and minimize the "risk of an improvident or ill-advised opinion being issued on an unbriefed issue." United States v. Leeson, 453 F.3d 631, 638 n. 4 (4th Cir.2006) (citing McBride v. Merrell Dow & Pharm., Inc., 800 F.2d 1208, 1211 (D.C.Cir.1986)). Given the briefing presented, the fully developed record below, and the lack of any showing of unfairness or prejudice, there is simply no reason why we should exercise our discretion to discard years of litigation on appeal because of an inartful opening brief. See A Helping Hand, LLC v. Baltimore Cnty., Md., 515 F.3d 356, 369 (4th Cir.2008) (observing that even when an argument has been waived, this Court may nonetheless consider it if a "miscarriage of justice would otherwise result" (internal quotation marks omitted)); cf. In re Am. W. Airlines, Inc., 217 F.3d 1161, 1165 (9th Cir. 2000) (observing that a court may refuse to find waiver and consider an argument raised for the first time on appeal when the issue "is one of law and either does not depend on the factual record, or the record has been fully developed").

Independent of the adequacy of the workers' opening brief, the district court had no grounds to revisit the question of predominance in the first place given this Court's remand instructions and mandate in Brown I. Unlike the requirement of commonality under Rule 23(a)(2) discussed above, Wal-Mart did not change, nor purport to change, the Rule 23(b)(3) analysis. Indeed, any impact of the Supreme Court's ruling on the question of whether common questions predominate is only incidental insofar as Wal-Mart recalibrated what constitutes a common question in the first place. The majority in Wal-Mart only invoked Rule 23(b)(3) to argue that the rule's well-established procedural protections should apply to the plaintiffs' claims for backpay. See Wal-Mart, 131 S.Ct. at 2559.

Following our instructions in Brown I for the district court to "certify the appellants' class action," the court found that "the putative class satisfied both the predominance and superiority requirements of Rule 23(b)(3)." J.A. 10930. The court then certified the class for those employed in all six Nucor operations departments. The district court cites no new facts or legal precedent after Brown I to justify revisiting that determination once the underlying question of commonality has been resolved.

Nonetheless, the dissent insists that our decision in Brown I "did not prevent the district court in any way from considering predominance because our prior decision did not say anything about predominance." Post at 926 (emphasis added). Such a conclusion misconstrues both the plain language of our original mandate and ignores the district court's equally plain understanding of it. The pivotal question in determining the scope of the mandate is whether the district court was free on remand to find that the workers had not satisfied the predominance requirement. If so, then our mandate did not reach the issue and the district court was free to reconsider it. But if the court did not have such liberty, then we must ask whether "controlling legal authority has changed dramatically" regarding Rule 23(b)(3) such that the court could reconsider the question. See Bell, 5 F.3d at 67. If no such change has occurred, then the district court could not revisit it.

As for the first question, the district court had no discretion to find that the workers' class failed to satisfy Rule 23(b)(3), after we expressly told it "to certify the appellants' class action and to engage in further proceedings consistent with this opinion." Brown I, 576 F.3d at 160; see also Bell, 5 F.3d at 66 (requiring that a district court "implement both the letter and spirit of the ... mandate, taking into account [our] opinion and the circumstances it embraces" (internal quotation marks and citation omitted)); United States v. Pileggi, 703 F.3d 675, 679 (4th Cir.2013) (observing that the mandate rule "forecloses relitigation of issues expressly or impliedly decided by the appellate court" (quoting Bell, 5 F.3d at 66)); S. Atl. Ltd., 356 F.3d at 583 (observing that a mandate must be "scrupulously and fully carried out" (internal quotation marks and citation omitted)).

Indeed, the district court itself recognized that we had "dictate[d] the general outcome to be reached (class certification) while leaving [the district court] to fill in the details." J.A. 9886 (Order Den. Mot. for Recons. 8 n.2). Of course, the court could have, and did, evaluate whether certification was best under Rule 23(b)(2) or (b)(3). But it had no discretion to then find that the prerequisites of either rule were not met. As the court observed, Nucor's argument on remand that the workers had failed to satisfy Rule 23(b) "overlook[ed] the Fourth Circuit's prior holding in this case." J.A. 9704 (Certification Order).24 Thus, the dissent misstates the record when it maintains that our original decision did not "in any way" prevent the district court from considering predominance. Post at 926-27. Indeed, following our instructions and findings in Brown I, the court proceeded to make the only finding it could under Rule 23(b)(3), namely, that "common issues predominate and that a class action is superior to any other method for adjudication of the claims in this case." The dissent is thus also misinformed when it states we are now certifying "a Rule 23(b)(3) class action without any court ever finding that the Rule 23(b)(3) requirements are satisfied." Post at 927.

Given the fact that our prior ruling foreclosed the denial of certification on the basis of Rule 23(b)(3), the district court needed some compelling reason to reconsider the question. Bell, 5 F.3d at 67 (describing the "extraordinary" exception to the mandate rule when there is "a show[ing] that controlling legal authority has changed dramatically"). But the court cited no such reason and, unlike the question of commonality, Wal-Mart provided none. Indeed, as the district court itself acknowledged, Wal-Mart only incidentally narrowed an inquiry into whether common questions predominate by clarifying what constitutes a common question in the first place under Rule 23(a)(2). J.A. 10971-72.

V.

More than seven years have now elapsed since the workers first filed their class certification motion, and the district court twice has refused to certify the class. The nature of the allegations, the evidentiary support buttressing them, and the inherent cohesiveness of the class all demonstrate that the court's failure to certify was an error. Rule 23 provides wide discretion to district courts, in part, to promote the systemic class action virtues of efficiency and flexibility. The realization of such benefits, however, requires that a district court exercise its judgment in a reasoned and expeditious manner.

The dissent rightly observes that the majority presses forward "[o]n the road to its desired result." Post at 956. And that result is simple justice. At bottom, the workers seek nothing more than the chance to speak with one voice about the promotions discrimination they allegedly suffered as one class on account of one uniting feature: the color of their skin. The dissent would deny them that chance while leading this Court down a different road — a road that would further weaken the class action as a tool to realize Title VII's core promise of equality.

We vacate the district court's decertification of the workers' promotions class and remand the case to the district court with instructions to certify the class.

VACATED IN PART, AND REMANDED WITH INSTRUCTIONS.

AGEE, Circuit Judge, dissenting:

We typically tread lightly when reviewing a class certification decision, affording "substantial deference" to the district court, especially when it provides "well-supported factual findings." Ward v. Dixie Nat'l Life Ins. Co., 595 F.3d 164, 179 (4th Cir.2010). Class certification proceedings often call for fact-intensive choices requiring intimate knowledge of the peculiarities of complex litigation. Id. We usually trust that the district court has the better eye for these sorts of questions.

The majority today declines to follow that path. It instead takes issue with almost every aspect of the district court's decision to decertify, reversing that court's determination because of newfound facts on appeal and different notions about the nature of this case. In doing so, the majority creates a split between this Court and another, see Bennett v. Nucor Corp., 656 F.3d 802 (8th Cir.2011), overlooks a plain and decisive waiver from the appellants, and drains a critical Supreme Court decision of much of its meaning, see Wal-Mart Stores, Inc. v. Dukes, ___ U.S. ___, 131 S.Ct. 2541, 180 L.Ed.2d 374 (2011). I respectfully dissent.

I. Predominance

A.

The district court decertified Plaintiffs' promotions classes for two distinct reasons. First, the court found that Plaintiffs had not identified a "question[] of law or fact common to the class," as Rule 23(a)(2) of the Federal Rules of Civil Procedure requires. Second, it held that any questions common to the class members did not "predominate over any questions affecting only individual members," so the class could not be certified under Rule 23(b)(3). Each of these separate reasons — commonality or predominance — provide an independent ground to decertify the class. See, e.g., Thorn v. Jefferson-Pilot Life Ins. Co., 445 F.3d 311, 319 (4th Cir.2006).

Because the district court provided two different bases for its decision, Plaintiffs were required to contest both. They did not. Plaintiffs' opening brief nowhere mentions the topic of predominance. Neither does it refer to Rule 23(b). And even though "the main concern in the predominance inquiry" is "the balance between individual and common issues," Myers v. Hertz Corp., 624 F.3d 537, 549 (2d Cir. 2010), a reader searches in vain for any mention of such a "balancing" in Plaintiffs' submissions. Instead, Plaintiffs' opening brief focuses solely on Rule 23(a) commonality. The brief does not even contain a simple statement that the district court erred as to predominance for the same reasons that it purportedly erred as to commonality — not to say that such a statement would be sufficient, either. See Jimenez v. Allstate Ins. Co., 765 F.3d 1161, 1165 n. 4 (9th Cir.2014) (holding that "cursory statements that the district court's order also incorrectly applied Rule 23(b)(3)'s [predominance] requirement" are "not enough to preserve the issue for appeal").

An appellant must raise every issue that he wishes to press in his opening brief. If the appellant fails to address an issue there, then we will deem the issue waived or abandoned. We have repeated this rule so often that it might rightfully be termed the best-established rule in appellate procedure. See, e.g., Metro. Reg'l Info. Sys., Inc. v. Am. Home Realty Network, 722 F.3d 591, 602 n. 13 (4th Cir.2013); Kensington Volunteer Fire Dep't, Inc. v. Montgomery Cnty., 684 F.3d 462, 472 n. 4 (4th Cir.2012); Mayfield v. Nat'l Ass'n for Stock Car Auto Racing, Inc., 674 F.3d 369, 376 (4th Cir.2012); A Helping Hand, LLC v. Balt. Cnty., 515 F.3d 356, 369 (4th Cir. 2008); French v. Assurance Co. of Am., 448 F.3d 693, 699 n. 2 (4th Cir.2006). As a rule that "all the federal courts of appeals employ," waiver "makes excellent sense." Joseph v. United States, ___ U.S. ___, 135 S.Ct. 705, 705, 190 L.Ed.2d 461 (2014) (Kagan, J., respecting denial of certiorari).

In past cases, we have endeavored to apply our waiver rule consistently, finding waiver whenever a party fails to "develop [his] argument" — even if his brief takes a passing shot at the issue. Belk, Inc. v. Meyer Corp., 679 F.3d 146, 152 n. 4 (4th Cir.2012). We have further found arguments waived even though they might have had merit. See IGEN Int'l, Inc. v. Roche Diagnostics GmbH, 335 F.3d 303, 308-09 (4th Cir.2003); Pleasurecraft Marine Engine Co. v. Thermo Power Corp., 272 F.3d 654, 657 (4th Cir.2001). And we have applied the doctrine despite its potentially significant impact. See, e.g., Carter v. Lee, 283 F.3d 240, 252 n. 11 (4th Cir. 2002) (applying the doctrine in a death penalty case).

Given that Plaintiffs failed to challenge the district court's ruling on predominance, the plain and consistent waiver rule defeats their appeal. "[T]o obtain reversal of a district court judgment based on multiple, independent grounds, an appellant must convince us that every stated ground for the judgment against him is incorrect." In re Under Seal, 749 F.3d 276, 289 (4th Cir.2014); accord Maher v. City of Chi., 547 F.3d 817, 821 (7th Cir.2008); Jankovic v. Int'l Crisis Grp., 494 F.3d 1080, 1086 (D.C.Cir.2007). Appellate courts have repeatedly affirmed district court decisions denying class certification where plaintiffs failed to contest a predominance finding. See, e.g., Little v. T-Mobile USA, Inc., 691 F.3d 1302, 1306-08 (11th Cir.2012); Klay v. Humana, Inc., 382 F.3d 1241, 1268 (11th Cir.2004), abrogated on other grounds by Bridge v. Phoenix Bond & Indemnity Co., 553 U.S. 639, 128 S.Ct. 2131, 170 L.Ed.2d 1012 (2008); Applewhite v. Reichhold Chems., Inc., 67 F.3d 571, 573-74 (5th Cir.1995). Nothing calls for a different result here.

B.

In view of their failure to raise the predominance issue, Plaintiffs now suggest that "[p]redominance and commonality ... are [both] part of Rule 23(b)(3)," such that a challenge concerning one should be treated as a challenge to both. Appellant's Reply Br. 2. They are mistaken.

Commonality, found in Rule 23(a)(2), asks whether the proposed class will "resolve an issue that is central to the validity of each of one of the claims in one stroke." EQT Prod. Co. v. Adair, 764 F.3d 347, 360 (4th Cir.2014). Predominance, found in Rule 23(b)(3), presents a "far more demanding" inquiry, Amchem Prods., Inc. v. Windsor, 521 U.S. 591, 624, 117 S.Ct. 2231, 138 L.Ed.2d 689 (1997), namely whether any common questions "pre-dominate over any questions affecting only individual members," Fed.R.Civ.P. 23(b)(3). Thus, while a "common issue" will establish commonality, that common issue only goes to one part of the predominance inquiry. Consequently, courts and parties must address these requirements separately, rather than muddle them together. See Vega v. T-Mobile USA, Inc., 564 F.3d 1256, 1268-70 (11th Cir.2009); In re Ins. Brokerage Litig., 579 F.3d 241, 277 (3d Cir. 2009); accord Ealy v. Pinkerton Gov't Servs., Inc., 514 Fed.Appx. 299, 305 (4th Cir.2013) ("[T]he Rule 23(a) commonality requirement[] and the Rule 23(b)(3) predominance requirement remain separate inquiries and the inquiries should not be `blended.'").

The majority excuses Plaintiffs' waiver because it believes that Plaintiffs "followed the district court's lead" in combining the two issues. Maj. op. at 919. Thus, even though commonality and predominance are legally distinct, the majority speculates that the district court did not treat them as such here. The majority's analysis mischaracterizes the district court's opinion.

The district court did not just repeat back its commonality findings in determining that Plaintiffs' class failed as to predominance. To the contrary, the court expressly held that it could not find the required predominance "even if the Fourth Circuit subsequently conclude[d] that plaintiffs have identified a common issue that satisfies Rule 23(a)(2)." J.A. 10956. The court then explained — over several pages — that many different reasons underlay its predominance finding, including several individual questions that could "overwhelm" common ones. Amgen Inc. v. Conn. Ret. Plans & Trust Funds, ___ U.S. ___, 133 S.Ct. 1184, 1196, 185 L.Ed.2d 308 (2013). Because Plaintiffs heavily rely on anecdotal evidence, for instance, the district court correctly concluded that a jury "would have to delve into the merits of each individual promotion decision" to determine whether each decision evidenced discrimination. J.A. 10959. Thus, a trial meant to resolve class-wide issues would likely devolve into a series of mini-trials examining each promotion decision made in the Nucor plant. The court further acknowledged that "individual damages determinations," like those that would be required here, can "cut against class certification." J.A. 10956. Although it concluded that such damages determinations did not, standing alone, compel decertification in this case, J.A. 10958, they did provide the district court an additional basis for caution in making its predominance finding. See, e.g., Cooper v. So. Co., 390 F.3d 695, 722-23 (11th Cir.2004), overruled on other grounds by Ash v. Tyson Foods, Inc., 546 U.S. 454, 126 S.Ct. 1195, 163 L.Ed.2d 1053 (2006) (noting that individualized damage issues could swamp the advantages coming from an initial, class-wide liability determination); accord Allison v. Citgo Petroleum Corp., 151 F.3d 402, 421-22 (5th Cir.1998), cited with approval in Gunnells v. Healthplan Servs., Inc., 348 F.3d 417, 445 n. 18 (4th Cir.2003); see also Comcast Corp. v. Behrend, ___ U.S. ___, 133 S.Ct. 1426, 1433, 185 L.Ed.2d 515 (2013) (explaining that individual damage-related questions might destroy predominance); Windham v. Am. Brands, Inc., 565 F.2d 59, 71-72 (4th Cir.1977).

The district court appropriately resolved predominance separately from commonality. Plaintiffs' failure to address the predominance finding in any way ends their appeal.

C.

The majority at least recognizes that Plaintiffs should have been "more explicit" in addressing predominance. Maj. op. at 918; see also id. at 919 (acknowledging that Plaintiffs' "express[]" arguments largely concern commonality). Even so, it concludes that certain oblique references in Plaintiffs' briefs preserved a predominance-related challenge on appeal. They do not.

Plaintiffs' statement of the issue on appeal, for instance, does not help them. See maj. op. at 918. The statement asks only whether "it [was] error or an abuse of discretion for the district court not to follow this Circuit's mandate" when it decertified the class. See Appellant's Br. 1. Here again, Plaintiffs never mention predominance, and the statement does not otherwise indicate any specific complaint with the district court's predominance holding. Even if it had, that reference would not have been enough without some further argument on the matter — an argument that Plaintiffs wholly failed to provide. See Belk, Inc., 679 F.3d at 153 n. 6; 11126 Balt. Blvd., Inc. v. Prince George's Cnty., Md., 58 F.3d 988, 993 n. 7 (4th Cir.1995).

The majority also ignores Plaintiffs' waiver because their brief contains some broadly stated attacks on the district court's decertification decision — attacks purportedly not "limit[ed] to the question of commonality." Maj. op. at 919. But in the usual case, a generalized attack on the lower court's decision does not preserve the specific arguments that might be subsumed within the broader one. Quite the opposite: a "generalized assertion of error" will not suffice to preserve anything. MMG Fin. Corp. v. Midwest Amusements Park, LLC, 630 F.3d 651, 659 (7th Cir. 2011); see also, e.g., Garrett v. Selby Connor Maddux & Janer, 425 F.3d 836, 841 (10th Cir.2005); Norman v. United States, 429 F.3d 1081, 1091 n. 5 (Fed.Cir.2005). Preservation would have little to recommend it if litigants could make nebulous, broadly worded arguments and trust appellate courts to work out the details once the opposing party points out the default.

In much the same way, Plaintiffs did not preserve their predominance challenge by citing a few cases that happen to touch upon the concept. See maj. op. at 918. The traditional rule provides that citations to the "occasional case," without any fuller discussion, do not preserve an argument. Pike v. Guarino, 492 F.3d 61, 78 n. 9 (1st Cir.2007); see also Am. Wildlands v. Kempthorne, 530 F.3d 991, 1001 (D.C.Cir. 2008) ("A fleeting statement in the parenthetical of a citation is no more sufficient to raise a claim than a cursory remark in a footnote[.]"). Similarly, "[m]ere notation of the applicable law, without any argumentation as to how it applies to [this] case, does not raise the issue of its application on appeal." Sou v. Gonzales, 450 F.3d 1, 6 n. 11 (1st Cir.2006) (internal quotation marks and citations omitted here and throughout); accord Johnson v. United States, 734 F.3d 352, 360 (4th Cir.2013).

The majority's analysis casts an inappropriate role for an appellate court. Now, a court must review each decision that an appellant cites and independently consider whether any part of it might undermine the district court's judgment for some reason that the appellant never raised. That concept reconceives the appellate courts' role, as those "courts do not sit as self-directed boards of legal inquiry and research." Nat'l Aeronautics & Space Admin. v. Nelson, 562 U.S. 134, 147 n. 10, 131 S.Ct. 746, 178 L.Ed.2d 667 (2011); see also Walker v. Prince George's Cnty., Md., 575 F.3d 426, 429 n.* (4th Cir.2009) ("Judges are not like pigs, hunting for truffles buried in briefs."). In addition, using the majority's new rule, appellants may now launch late-in-the-day challenges to any part of a district court's certification decision so long as they serendipitously cited a case canvassing Rule 23 in their opening brief. This "preservation-by-citation" approach renders the waiver rule a nullity.

D.

In the end, the majority declares itself unwilling to exercise its "discretion" to "discard years of litigation on appeal because of an inartful brief." Maj. op. at 920. That approach seems to give pro se litigant treatment to a brief crafted by experienced class counsel — counsel that has appeared in our court before. Surely it does not expect too much from veteran counsel to ask them to make their arguments straight up and square. All the more so when these counsel have been specifically cautioned about waiver on previous occasions. See, e.g., Davis v. Coca-Cola Bottling Co. Consol., 516 F.3d 955, 972-73 (11th Cir.2008) (holding that party represented by same counsel had "abandoned" claim by failing to raise it in his opening brief); see also Angles v. Dollar Tree Stores, Inc., 494 Fed.Appx. 326, 330 n. 6 (4th Cir.2012) (same); cf. Bennett, 656 F.3d at 821 (holding that party represented by same counsel had "essentially abandoned" argument by making only a "conclusory challenge"); Anderson v. Cagle's, Inc., 488 F.3d 945, 959 (11th Cir. 2007) (same).

The "purpose" of the preservation rule is also not served by overlooking Plaintiffs' waiver. See maj. op. at 919-20. The rule "ensures that the opposing party has an opportunity to reflect upon and respond in writing to the arguments that his adversary is raising." Hamilton v. Southland Christian Sch., Inc., 680 F.3d 1316, 1319 (11th Cir.2012); see also United States v. Leeson, 453 F.3d 631, 638 n. 4 (4th Cir. 2006) (noting that late arguments are "unfair to the appellee"); Pignons S.A. de Mecanique v. Polaroid Corp., 701 F.2d 1, 3 (1st Cir.1983) ("In preparing briefs and arguments, an appellee is entitled to rely on the content of an appellant's brief for the scope of the issues appealed[.]"). Nucor never had a chance to address Plaintiffs' predominance arguments directly, as Plaintiffs waited until their reply brief to make them. Plaintiffs argued in their reply brief, for example, that no "heightened" predominance standard applies after Wal-Mart Stores, Inc. v. Dukes, ___ U.S. ___, 131 S.Ct. 2541, 2551, 180 L.Ed.2d 374 (2011), and the majority agrees, see maj. op. at 921. There might very well be reason to believe otherwise, though Nucor has never had a chance to make that argument. See, e.g., Andrey Spektor, The Death Knell of Issue Certification and Why That Matters After Wal-Mart v. Dukes, 26 St. Thomas L.Rev. 165, 172 (2014) (suggesting that Wal-Mart rendered it harder for issues to predominate). It must be cold comfort to Nucor, then, to hear that it was not "prejudice[d]" by these and other unanswerable arguments. Maj. op. at 920.

E.

The majority goes on to hold that the mandate rule barred the district court from examining Rule 23(b)(3) predominance. See maj. op. at 920-22. That view is factually and legally incorrect. The decision in the prior appeal in this case did not prevent the district court in any way from considering predominance because our prior decision did not say anything about predominance.

In its original class certification decision in 2007, the district court held that Plaintiffs did not satisfy three of Rule 23(a)'s four requirements. It expressly declined to consider "the remaining requirements of Rule 23(b)." J.A. 8997. On appeal, the parties' submissions focused solely on Rule 23(a). A majority of the Court then reviewed these "Rule 23(a) factors" and found them "satisfied." Brown v. Nucor Corp., 576 F.3d 149, 160 (4th Cir.2009) ("Brown I"). The Brown I majority initially went on to hold, in a single sentence at the end of the opinion, that "the requirements of [Rule] 23(b)(3) ha[d] also been satisfied for these claims." See Brown v. Nucor Corp., 576 F.3d 149, 160 (4th Cir.2009). Nucor then petitioned for rehearing en banc, arguing, among other things, that neither the lower court nor the parties had previously analyzed the Rule 23(b) issue. See Nucor Pet. for Reh'g at 9, Brown I, 576 F.3d 149 (No. 08-1247), ECF No. 53. In response, the Brown I panel amended its opinion and excised any mention of Rule 23(b)(3). See Order, Brown v. Nucor Corp., No. 08-1247 (4th Cir. Oct. 8, 2009). One can easily discern why the opinion was amended: Brown I could not decide a fact-intensive issue — that is, the predominance issue under Rule 23(b)(3) — when the parties had not yet argued it and the district court had not yet addressed it. See Transamerica Leasing, Inc. v. Instit. of London Underwriters, 430 F.3d 1326, 1332 (11th Cir.2005) (explaining that the mandate rule and the broader law of the case doctrine "cannot apply when the issue in question was outside the scope of the prior appeal"). In fact, up to that point, Plaintiffs had never even sought certification under Rule 23(b)(3); they sought to certify only a Rule 23(b)(2) class or, in the alternative, a so-called "hybrid" action.

By removing any reference to Rule 23(b), Brown I left it to the district court to determine in the first instance whether Plaintiffs' class met that provision's requirements. The district court complied with both the letter and the spirit of Brown I, and it correctly took "into account [the] opinion and the circumstances it embrace[d]." United States v. Bell, 5 F.3d 64, 66 (4th Cir.1993); see also, e.g., Lindy Pen Co. v. Bic Pen Corp., 982 F.2d 1400, 1404-05 (9th Cir.1993) (affirming district court's decision not to order accounting or damages, despite appellate court's instructions to "order an accounting and to award damages," where district court acted in line with the "spirit" of the mandate). An appellate mandate "does not reach questions which might have been decided but were not." United States v. Lentz, 524 F.3d 501, 528 (4th Cir.2008). And "[w]hile a mandate is controlling as to matters within its compass, on the remand a lower court is free as to other issues." Sprague v. Ticonic Nat'l Bank, 307 U.S. 161, 168, 59 S.Ct. 777, 83 L.Ed. 1184 (1939). Simply put, the Brown I mandate did not apply to Rule 23(b)(3), nor could it.

On remand after Brown I, the district court initially certified the two promotions classes under Rule 23(b)(3). The court later reconsidered, as it was entitled to do under Rule 23, which provides that "[a]n order that grants or denies class certification may be altered or amended before final judgment." Fed.R.Civ.P. 23(c)(1)(C); see also Fed.R.Civ.P. 54(b). "[C]ertifications are not frozen once made," Amgen, Inc., 133 S.Ct. at 1202 n. 9, and a district court has "considerable discretion to decertify the class," Cent. Wesleyan Coll. v. W.R. Grace & Co., 6 F.3d 177, 189 (4th Cir.1993). See also Prado-Steiman ex rel. Prado v. Bush, 221 F.3d 1266, 1273 (11th Cir.2000). The district court could revisit its interlocutory decision regardless of whether, as the majority puts it, "new facts or legal precedent [arose] after Brown I." Maj. op. at 920.

In effect, the majority today certifies a Rule 23(b)(3) class action without any court ever finding that the Rule 23(b)(3) requirements are satisfied. It cannot genuinely contend that Brown I did the work, as "the Fourth Circuit has never allowed the rigorous Rule 23 analysis to be accomplished implicitly." Partington v. Am. Int'l Specialty Lines Ins. Co., 443 F.3d 334, 341 (4th Cir.2006). And the district court ultimately did not make such a finding either. The majority's decision to certify in part on this illusory mandate, then, substantially damages Rule 23(b)(3)'s "vital prescription." Amchem, 521 U.S. at 623, 117 S.Ct. 2231. The Supreme Court recently reminded us that "plaintiffs wishing to proceed through a class action must actually prove — not simply plead — that their proposed class satisfies each requirement of Rule 23, including ... the predominance requirement of Rule 23(b)(3)." Halliburton Co. v. Erica P. John Fund, Inc., ___ U.S. ___, 134 S.Ct. 2398, 2412, 189 L.Ed.2d 339 (2014). At least as to predominance, Plaintiffs have yet to prove anything.

* * *

Plaintiffs did not challenge the district court's predominance ruling and do not credibly explain why they failed to do so. The district court's decision should therefore be affirmed on that basis alone.

II. Relevant Standards

Even ignoring Plaintiffs' waiver of the predominance issue, they have not established that the district court abused its discretion in finding insufficient commonality. To see why, it is first necessary to recognize the standard that appellate courts use in reviewing a district court's class-certification decision. Then, the standard that the district court used in evaluating the evidence at the certification stage must be considered.

A.

1.

A district court's ultimate class-certification decision — that is, how it applied the Rule 23 factors — is reviewed for an abuse of discretion. See, e.g., EQT Prod. Co., 764 F.3d at 357; Ward, 595 F.3d at 179; Monroe v. City of Charlottesville, Va., 579 F.3d 380, 384 (4th Cir.2009); Gregory v. Finova Capital Corp., 442 F.3d 188, 190 (4th Cir.2006). But reciting the standard is not enough; there must be genuine respect and adherence paid to the limits that it imposes.

The abuse-of-discretion standard does establish some substantial limits, representing "one of the most deferential standards of review." Matthew Bender & Co. v. West Publ'g Co., 240 F.3d 116, 121 (2d Cir.2001). Under it, the appellate court may reverse only when "the [trial] court's exercise of discretion, considering the law and the facts, was arbitrary and capricious." United States v. Mason, 52 F.3d 1286, 1289 (4th Cir.1995). We act only when the decision could not "have been reached by a reasonable jurist," or when we may call it "fundamentally wrong," "clearly unreasonable, arbitrary, or fanciful." Bluestein v. Cent. Wis. Anesthesiology, S.C., 769 F.3d 944, 957 (7th Cir.2014); accord Am. Copper & Brass, Inc. v. Lake City Indus. Prods., Inc., 757 F.3d 540, 543 (6th Cir.2014) (characterizing review of a class certification decision as "very limited").

Of course, deference does not equal blind acceptance. If, for instance, the district court entirely fails to undertake some part of the requisite analysis, then it may be appropriate to reverse. See, e.g., EQT Prod., 764 F.3d at 371 (vacating and remanding a certification order where the district court failed to conduct an appropriately rigorous analysis of Rule 23's requirements). But when our review ventures into intensely factual matters or areas of practical concern, then our deference must be at its greatest — indeed, we must stand aside in those circumstances unless the lower court was "clearly wrong." Windham, 565 F.2d at 65; accord CGC Holding Co., LLC v. Broad & Cassel, 773 F.3d 1076, 1086 (10th Cir. 2014) ("[A]s long as the district court applies the proper Rule 23 standard, we will defer to its class certification ruling provided that decision falls within the bounds of rationally available choices given the facts and law involved in the matter at hand.").

We do not then reverse anytime we disagree with the result that the district court reaches. See First Penn-Pac. Life Ins. Co. v. Evans, 304 F.3d 345, 348 (4th Cir.2002). Rather, "the [abuse-of-discretion] standard draws a line ... between the unsupportable and the merely mistaken, between the legal error, disorder of reason, severe lapse of judgment, and procedural failure that a reviewing court may always correct, and the simple disagreement that, on this standard, it may not." Evans v. Eaton Corp. Long Term Disability Plan, 514 F.3d 315, 322 (4th Cir.2008); see also Cooter & Gell v. Hartmarx Corp., 496 U.S. 384, 405, 110 S.Ct. 2447, 110 L.Ed.2d 359 (1990) (holding that the district court did not abuse its discretion where it "applied the correct legal standard and offered substantial justification for its finding").

These principles might strike some as truisms, but they carry special force in the class-certification context. "Granting or denying class certification is a highly fact-intensive matter of practicality," Monreal v. Potter, 367 F.3d 1224, 1238 (10th Cir. 2004), so much so that "[h]ighly fact-based, complex, difficult matters" arise as a matter of routine, Amchem, 521 U.S. at 630, 117 S.Ct. 2231 (Breyer, J., concurring in part and dissenting in part). Unsurprisingly, then, we give district courts "broad discretion in deciding whether to allow the maintenance of a class action." Roman v. ESB, Inc., 550 F.2d 1343, 1348 (4th Cir. 1976); see also Lowery v. Circuit City Stores, Inc., 158 F.3d 742, 757-58 (4th Cir.1998), vacated 527 U.S. 1031, 119 S.Ct. 2388, 144 L.Ed.2d 790 (1999), reaff'd in relevant part, 206 F.3d 431 (4th Cir.2000). As with any other decision that appellate courts review for abuse of discretion, we should affirm a certification decision even if we are convinced that "reasons clearly existed for taking the other course." Lewis v. Bloomsburg Mills, Inc., 773 F.2d 561, 564 (4th Cir.1985); accord Simmons v. Poe, 47 F.3d 1370, 1382 (4th Cir.1995).

2.

An appellate court must be even more careful in reviewing any factual findings underlying the district court's decision, as we review those only for clear error. Thorn, 445 F.3d at 317-18; see also Fed. R.Civ.P. 52(a)(6). "The clear error standard... protects district courts' primacy as triers of fact." Evans, 514 F.3d at 321. Our opinions have repeatedly emphasized that clear-error review is "narrow," Walker v. Kelly, 593 F.3d 319, 323 (4th Cir. 2010), "highly deferential," Green v. Johnson, 515 F.3d 290, 301 (4th Cir.2008), and "particularly circumscribed," Jiminez v. Mary Washington Coll., 57 F.3d 369, 378 (4th Cir.1995). We may reverse findings reviewed under this standard only when, having reviewed the entire record, we are "left with the definite and firm conviction that a mistake has been committed." United States v. Heyer, 740 F.3d 284, 292 (4th Cir.2014). If the district court chose between "two permissible views of the evidence," or if it otherwise offered a "plausible" account of that evidence, Anderson v. City of Bessemer City, N.C., 470 U.S. 564, 574, 105 S.Ct. 1504, 84 L.Ed.2d 518 (1985), then its factual findings are "conclusive," Walker, 593 F.3d at 323. And as with the abuse-of-discretion standard, we cannot reverse merely because we would have decided the matter differently. See Anderson, 470 U.S. at 573, 105 S.Ct. 1504.

3.

Despite these deferential standards of review, the majority identifies reversible error in virtually every legal and factual judgment that the district court rendered. Yet in searching the majority's opinion for any of the hallmarks of deference — explanations as to how the district court clearly erred, or full analysis of how the district court abused its discretion — we find very little.

In truth, the majority seems to apply just about every standard of review but a deferential one. For the most part, the majority offers bare statements that the district court erred, apparently because the district court decided things differently than the majority would have. For instance, it insists that Plaintiffs' statistical evidence is simply "less precise" and rejects out-of-hand the district court's view that the evidence was "fundamentally unreliable." Maj. op. at 904, 906. Likewise, it draws its own conclusions about the anecdotal evidence, reciting certain portions of certain affidavits and declaring them enough. It makes credibility determinations, categorically rejecting Nucor's evidence as "self-serving," id. at 906, or "coercive," id. at 913, while embracing contrary statements from Plaintiffs because the majority finds them "credible," id. at 913. And it offers its own notions about what is "plain," id. at 908, "elementary," id. at 912, or "common sense," id. at 913. The majority does so even while decrying the dangers of "cherry pick[ing] facts from an 11,000 page record." Id. at 913. In short, the majority opinion shows little respect for a district court that is far more familiar with each page of the record than we are.

Contravening our "axiomatic" rule against factual findings on appeal, Core Commc'ns, Inc. v. Verizon Md. LLC, 744 F.3d 310, 324 (4th Cir.2014), the majority eventually finds in the first instance that "there is only one answer to the question of why Nucor's black workers were consistently disfavored," maj. op. at 913. This adventuresome approach is rather jarring when placed against the more measured methods found in some of our other class certification decisions. See, e.g., EQT Prod., 764 F.3d at 371 (remanding for further consideration of class certification after determining that district court misapplied the relevant standards); Gariety v. Grant Thornton, LLP, 368 F.3d 356, 366 (4th Cir.2004) (same). Making matters worse, the majority offers no good reason for it. Instead, it engages in a rather extended discussion of the Brown I dissent and then declares any attack on the majority's factfinding today "iron[ic]." Maj. op. at 905.

Too often, we fail to give standards of review the attention that they deserve. We see them recited in boilerplate and then dispensed with when the perceived exigencies of a case seem to call for it. But "[s]tandards of review are ... an elemental expression of judicial restraint, which, in their deferential varieties, safeguard the superior vantage points of those entrusted with primary decisional responsibility." Evans, 514 F.3d at 320-21. An appellate court should not be so quick to ignore them.

B.

We must next consider the district court's role in deciding the certification motion in the first place. The majority implies that the district court too readily dismissed Plaintiffs' efforts to certify. But the district court was not just permitted to take a hard look at Plaintiffs' submissions — it was required to.

1.

Although plaintiffs shoulder the burden of demonstrating that a proposed class complies with Rule 23, the district court has an "independent obligation to perform a rigorous analysis to ensure that all of the prerequisites have been satisfied." EQT Prod., 764 F.3d at 358. Among other things, this "rigorous analysis" requires the district court "to resolve a genuine legal or factual dispute relevant to determining the requirements." In re Hydrogen Peroxide Antitrust Litig., 552 F.3d 305, 320 (3d Cir.2008).

"[C]areful attention to the requirements of [Rule] 23 remains ... indispensable" even in cases "alleging racial or ethnic discrimination." E. Tex. Motor Freight Sys., Inc. v. Rodriguez, 431 U.S. 395, 405, 97 S.Ct. 1891, 52 L.Ed.2d 453 (1977). Thus, "a Title VII class action, like any other class action, may only be certified if the trial court is satisfied, after a rigorous analysis, that the prerequisites of [the Rule] have been satisfied." Gen. Tel. Co. of Sw. v. Falcon, 457 U.S. 147, 161, 102 S.Ct. 2364, 72 L.Ed.2d 740 (1982); see also Desert Palace, Inc. v. Costa, 539 U.S. 90, 99, 123 S.Ct. 2148, 156 L.Ed.2d 84 (2003) (noting the "conventional rule[s] of civil litigation ... generally appl[y] in Title VII cases"). And there is no "entitlement to class proceedings for the vindication of statutory rights," Am. Express Co. v. Italian Colors Rest., ___ U.S. ___, 133 S.Ct. 2304, 2309, 186 L.Ed.2d 417 (2013), Title VII included. Thus, the Court must be careful not to bend and twist the "rigorous analysis" that Rule 23 compels merely for the sake of abstract notions of Title VII's objectives and purposes. Cf. Touche Ross & Co. v. Redington, 442 U.S. 560, 578, 99 S.Ct. 2479, 61 L.Ed.2d 82 (1979) ("[G]eneralized references to the `remedial purposes' of [a statute] will not justify reading a provision more broadly than its language and the statutory scheme reasonably permit."). To do so would not only ignore the Supreme Court's warnings; it might also have unforeseen effects in the many other areas of law in which Rule 23 is implicated.

In basic terms, the rigorous-analysis standard tests whether plaintiffs have presented substantial evidence of compliance with Rule 23. Plaintiffs may "not simply plead" that the relevant requirements have been met, but must "actually prove" it. Halliburton, 134 S.Ct. at 2412; accord Monroe, 579 F.3d at 384. To meet that standard, plaintiffs must summon "evidentiary proof," Comcast, 133 S.Ct. at 1432, and "affirmatively demonstrate [their] compliance with the Rule," Wal-Mart, 131 S.Ct. at 2551. "[S]ome evidence" is not enough. In re Initial Pub. Offerings ["IPO"] Sec. Litig., 471 F.3d 24, 33 (2d Cir.2006).

Before certifying a class action, courts will require a plaintiff to establish by a preponderance of the evidence that the action complies with each part of Rule 23. See In re U.S. Foodservice Inc. Pricing Litig., 729 F.3d 108, 117 (2d Cir.2013); Carrera v. Bayer Corp., 727 F.3d 300, 306 (3d Cir.2013); Messner v. Northshore Univ. HealthSystem, 669 F.3d 802, 811 (7th Cir.2012); Ala. Elec. Pension Fund v. Flowserve Corp., 572 F.3d 221, 228 (5th Cir.2009), abrogated in other respects by Halliburton, 134 S.Ct. 2398; accord In re Titanium Dioxide Antitrust Litig., 284 F.R.D. 328, 336 (D.Md.2012); In re Mills Corp. Sec. Litig., 257 F.R.D. 101, 104 (E.D.Va.2009); In re Safety-Kleen Corp. Bondholders Litig., No. 3:00-1145-17, 2004 WL 3115870, at *2 (D.S.C. Nov. 1, 2004); see also Anthony F. Fata, Doomsday Delayed: How the Court's Party-Neutral Clarification of Class Certification Standards in Wal-Mart v. Dukes Actually Helps Plaintiffs, 62 DePaul L.Rev. 675, 681 (2013) (reading Wal-Mart to apply a preponderance-of-the-evidence standard).

2.

"[T]he factors spelled out in Rule 23 must be addressed through findings, even if they overlap with issues on the merits." Gariety, 368 F.3d at 366; accord In re Rail Freight Fuel Surcharge Antitrust Litig., 725 F.3d 244, 249 (D.C.Cir.2013) (recognizing that certification will sometimes "resemble[] an appraisal on the merits"). Obviously, "[a] court may not say something like `let's resolve the merits first and worry about the class later' ... or `I'm not going to certify a class unless I think that the plaintiffs will prevail.'" Szabo v. Bridgeport Machs., Inc., 249 F.3d 672, 677 (7th Cir.2001), cited with approval in Wal-Mart, 131 S.Ct. at 2552. But overlap "cannot be helped," as certification "generally involves considerations that are enmeshed in the factual and legal issues comprising the plaintiff's cause of action." Wal-Mart, 131 S.Ct. at 2551-52. Compare Brown I, 576 F.3d at 156 (citing Eisen v. Carlisle & Jacquelin, 417 U.S. 156, 177, 94 S.Ct. 2140, 40 L.Ed.2d 732 (1974), and refusing to inquire into Plaintiffs' statistics because it would be an impermissibly "in-depth assessment of the merits"), with Wal-Mart, 131 S.Ct. at 2552 & n. 6 (admonishing courts not to "mistakenly cite[]" Eisen for the incorrect idea that merits inquiries are barred).

3.

Contrast these well-defined and rigorous standards with the ambiguous and limitless ones found in the majority opinion. The majority acknowledges the "rigorous analysis" that lower courts must perform, but abandons that standard soon after mentioning it. Instead, it treats the evidentiary standard for certification as one different from that required for a party to prevail on the merits, never acknowledging that this view breaks from the many courts (including those in our Circuit) that apply the preponderance standard. Nor does it even tell us what a "rigorous analysis" might consist of. Instead, it merely invokes Amgen, a case that addresses what questions may be considered on class certification, not what evidence will suffice to answer them. 133 S.Ct. at 1194-95. Having rendered the rigorous analysis less rigorous than other courts' (though to what degree, one does not know), the majority then proceeds to apply its weakened test, repeatedly using mere allegations — or, sometimes, allegations "proven" by allegations — to justify certification. See, e.g., maj. op. at 906, 910, 910, 912, 914, 915, 917, 917, 921. The necessary implication is that the majority's "rigorous analysis" consists of very little.

One finds a further hint at the level of proof that the majority means to apply when it embraces Brown I's metric. Maj. op. at 903. Brown I held that "allegations" of disparate treatment were enough to establish commonality, a conclusion at odds with Wal-Mart. Compare Brown I, 576 F.3d at 153, with Wal-Mart, 131 S.Ct. at 2553 (distinguishing between an "otherwise unsupported allegation" and the "significant proof" required to establish a common policy). The majority in Brown I also said that anecdotes from three employees concentrated in a single department proved a common policy of discrimination. 576 F.3d at 153. And it held that statistical evidence of "relatively weak probative value" was enough, even though problems in that evidence — the statistical evidence seen here — might "very well discredit" it at some later stage. Id. at 156 & n. 10. In short, Brown I required the plaintiffs to summon an exceptionally low, almost nonexistent level of proof at the class-certification stage.

The majority's decision to reanimate Brown I's negligible evidentiary standard leaves this circuit alone on an island. The Brown I majority suggested that its lenient view of the necessary evidence aligned with the Second Circuit's decision in Caridad v. Metro-North Commuter Railroad, 191 F.3d 283, 293 (2d Cir.1999). But by the time Brown I was issued, the Second Circuit had already repudiated any part of Caridad suggesting a lesser burden of proof than a preponderance of the evidence. See In re IPO, 471 F.3d at 42 ("[O]ur conclusions necessarily preclude the use of a `some showing' standard, and to whatever extent Caridad might have implied such a standard for a Rule 23 requirement, that implication is disavowed."). Only one circuit followed Brown I's lead and accepted such a low degree of proof: the Ninth Circuit, in its now-reversed decision in Dukes v. Wal-Mart Stores, Inc. See 603 F.3d 571, 595-96 & n. 17 (9th Cir.2010). (citing Brown I, 576 F.3d at 156). In the meantime, another circuit rejected Brown I outright. See Bennett, 656 F.3d at 816 n. 2 (declining to "follow" Brown I's finding that sufficient evidence established commonality, as "Brown [I] was decided without the benefit of the Supreme Court's recent opinion in Dukes").

All in all, despite assurances otherwise, the majority treats Rule 23 as something akin to a pleading standard. It is not. See Wal-Mart, 131 S.Ct. at 2551. Were the rule written as the majority envisions it, district courts would get to "duck hard questions." West v. Prudential Sec., Inc., 282 F.3d 935, 938 (7th Cir.2002). But framing class certification as a mere pleading standard "amounts to a delegation of judicial power to the plaintiffs." Id. "[A] district court's certification order often bestows upon plaintiffs extraordinary leverage, and its bite should dictate the process that precedes it." Oscar Private Equity Invs. v. Allegiance Telecom, Inc., 487 F.3d 261, 267 (5th Cir.2007), abrogated in other respects by Erica P. John Fund, Inc. v. Halliburton Co., ___ U.S. ___, 131 S.Ct. 2179, 180 L.Ed.2d 24 (2011).

III. Commonality

With the proper standards in mind, it becomes evident that the district court did not abuse its discretion in finding that Plaintiffs failed to establish commonality.

"In this case, proof of commonality necessarily overlaps with [Plaintiffs'] merits contention that [Nucor] engages in a pattern or practice of discrimination." Wal-Mart, 131 S.Ct. at 2552. Plaintiffs must establish a unifying policy of discrimination at certification, or "it will be impossible to say that examination of all the class members' claims for relief will produce a common answer to the crucial question [of] why was I disfavored." Id. In other words, Plaintiffs cannot simply identify a group of people who they allege have suffered some type of Title VII injury. Id. To certify the class, Plaintiffs must be able to trace that injury to a single, common source. Id.; accord Ellis v. Costco Wholesale Corp., 657 F.3d 970, 981 (9th Cir. 2011); see also William B. Rubenstein, Newberg on Class Actions § 3:19 (5th ed.2014) (citing Brown I as an example of a case that approached commonality "loosely" and explaining that Wal-Mart articulated "a more explicit definition of commonality"). Plaintiffs here must identify a common policy with common injury to members of a class spanning more than a decade, covering Nucor's entire South Carolina production facility, and touching upon dozens of relevant decisionmakers. That task can be decidedly difficult, especially given that Plaintiffs premise their class in part on a disparate treatment theory. See Stastny v. S. Bell Tel. & Tel. Co., 628 F.2d 267, 274 n. 10 (4th Cir.1980); see also Garcia v. Johanns, 444 F.3d 625, 633 (D.C.Cir.2006) ("Establishing commonality for a disparate treatment class is particularly difficult where, as here, multiple decisionmakers with significant local autonomy exist.").

A plaintiff who brings a class-wide charge of discrimination must traverse a "wide gap" between his claim of individual mistreatment and a class-wide harm. Falcon, 457 U.S. at 157, 102 S.Ct. 2364. The plaintiff could do so in one of two ways. See Wal-Mart, 131 S.Ct. at 2553. First, he might identify a "biased testing procedure" that is used to evaluate applicants and employees. Id. By all accounts, Plaintiffs do not identify that sort of procedure here. Second, a plaintiff might offer "significant proof" that an employer "operated under a general policy of discrimination... [that] manifested itself in hiring and promotion practices in the same general fashion." Id. This second route forms the focus of this case.

Plaintiffs offer two types of evidence that they say bridge the gap between individual and class-wide claims: statistical evidence and anecdotal evidence. Whether examining these two categories of evidence separately or together, the district court did not abuse its discretion in deeming the Plaintiffs' case insufficient.

A. Statistical Evidence

1.

Plaintiffs first present a statistical study comparing a hypothesized, weighted benchmark of black bidders for promotions to the number of black employees that they assumed Nucor promoted during the relevant period. This evidence performs a double duty, as it goes to Plaintiffs' disparate impact claim and their disparate treatment claim.

As to the disparate impact claim, this sort of statistical evidence should identify disparities that are "sufficiently substantial" to raise "an inference of causation." Anderson v. Westinghouse Savannah River Co., 406 F.3d 248, 281 (4th Cir.2005). Without "substantial" disparities, we cannot be confident that a challenged policy produced an injury common to the class. See Wal-Mart, 131 S.Ct. at 2551.

As to the disparate treatment claim, "gross statistical disparities" "may in a proper case constitute prima facie proof of a pattern or practice of discrimination." Hazelwood Sch. Dist. v. United States, 433 U.S. 299, 307-08, 97 S.Ct. 2736, 53 L.Ed.2d 768 (1977); accord Ardrey v. United Parcel Serv., 798 F.2d 679, 683 (4th Cir.1986). But see Warren v. Halstead Indus., Inc., 802 F.2d 746, 753 (4th Cir.1986) ("[S]tatistics cannot alone prove the existence of a pattern or practice of discrimination[.]"). But not every case will present the truly egregious and unexplained disparities that leave no room for any inference other than intentional discrimination. Moreover, "[i]nferring past discrimination from statistics alone assumes the most dubious of conclusions: that the true measure of racial equality is always to be found in numeric proportionality." Md. Troopers Ass'n, Inc. v. Evans, 993 F.2d 1072, 1077 (4th Cir.1993).

2.

The majority observes that Plaintiffs' evidence is "statistically significant at 2.54 standard deviations from what would be expected if race were a neutral factor." Maj. op. at 908. Statistical significance, however, is a necessary but not sufficient condition to finding a discriminatory practice or policy; statistical significance does not axiomatically equate with legal significance. See EEOC v. Fed. Reserve Bank of Richmond, 698 F.2d 633, 648 (4th Cir. 1983) ("[S]tatistical significance as measured by the standards of acceptable statistical principles will not necessarily be legally significant[.]"), rev'd sub nom on other grounds, Cooper v. Fed. Reserve Bank of Richmond, 467 U.S. 867, 104 S.Ct. 2794, 81 L.Ed.2d 718 (1984). High statistical significance levels might lack practical and legal significance, for instance, because "a high significance level may be a misleading artifact of the study's design." Kadas v. MCI Systemhouse Corp., 255 F.3d 359, 362 (7th Cir.2001). Thus, determining what is legally significant — as opposed to statistically significant — "is a legal determination properly made by the court and not by an expert." Fed. Reserve Bank of Richmond, 698 F.2d at 648; cf. United States v. Philip Morris USA, Inc., 449 F.Supp.2d 1, 706 n. 29 (D.D.C.2006) (criticizing one of Plaintiffs' experts for his undue reliance on statistical significance).

Nevertheless, the majority seems to defer to Plaintiffs' experts and assume legal significance because the statistical evidence crosses the two-standard-deviation threshold, the threshold for statistical significance at a 95% confidence level. Yet "courts of law should be extremely cautious in drawing any conclusions from standard deviations in the range of one to three." EEOC v. Am. Nat'l Bank, 652 F.2d 1176, 1192 (4th Cir.1981); see also Kingsley R. Browne, Statistical Proof of Discrimination: Beyond "Damned Lies", 68 Wash. L.Rev. 477, 503 (1993) ("Random disparities of this magnitude are pervasive in the workplace and are not suggestive of a nonrandom cause, let alone an illegal one."). In specific cases, even higher numbers may not be enough. EEOC v. Western Electric Co., Inc., 713 F.2d 1011 (4th Cir.1983), provides one example. There, we held that a district court clearly erred in finding a policy or practice of discrimination, even though statistics showed overall disparities of 4.7955 and 5.883 standard deviations. Id. at 1018-19.

Similarly, other courts have rejected statistical evidence even though the evidence met the two-standard-deviation threshold. See, e.g., Carpenter v. Boeing Co., 456 F.3d 1183, 1201 (10th Cir.2006) (7.95 and 38.03 standard deviations); Lopez v. Laborers Int'l Union Local No. 18, 987 F.2d 1210, 1213-14 (5th Cir.1993) (3.26 and 3.01 standard deviations); Waisome v. Port Auth. of N.Y. & N.J., 948 F.2d 1370, 1376 (2d Cir.1991) (2.68 standard deviations); EEOC v. Chi. Miniature Lamp Works, 947 F.2d 292, 300 (7th Cir.1991) (20.1 standard deviations); Gay v. Waiters' & Dairy Lunchmen's Union, Local No. 30, 694 F.2d 531, 551 (9th Cir.1982) (2.45 standard deviations). In short, "there is nothing magical about two or three standard deviations." Ramona L. Paetzold & Steve L. Willborn, The Statistics of Discrimination § 4:13 (2014).

3.

Instead of assuming "that any particular number of `standard deviations'" establishes a discriminatory policy, courts must evaluate statistical evidence on a "case-by-case basis." Watson v. Fort Worth Bank & Trust, 487 U.S. 977, 995 n. 3, 108 S.Ct. 2777, 101 L.Ed.2d 827 (1988) (plurality opinion); see also Int'l Bhd. of Teamsters v. United States, 431 U.S. 324, 340, 97 S.Ct. 1843, 52 L.Ed.2d 396 (1977). Neither "courts [n]or defendants [are] obliged to assume that plaintiffs' statistical evidence is reliable." Watson, 487 U.S. at 996, 108 S.Ct. 2777. And we must always keep in mind that we are looking for reliable indications of "gross" or "substantial" disparities that amount to "significant proof." Wal-Mart, 131 S.Ct. at 2551, 2553; Hazelwood, 433 U.S. at 307-08, 97 S.Ct. 2736.

The duty to test the relevant statistical evidence attaches at the class certification stage, Comcast, 133 S.Ct. at 1433, as "reliance on unverifiable evidence is hardly better than relying on bare allegations," Unger v. Amedisys, Inc., 401 F.3d 316, 324 (5th Cir.2005). District courts must probe the validity of statistical evidence, as "any method of measurement" would otherwise become "acceptable so long as it c[ould] be applied classwide, no matter how arbitrary the measurements may be." Comcast, 133 S.Ct. at 1433; accord Rail Freight Fuel Surcharge Antitrust Litig., 725 F.3d at 254; Am. Honda Motor Co., Inc. v. Allen, 600 F.3d 813, 815 (7th Cir.2010).

In this case, the district court evaluated Plaintiffs' statistical evidence, reasonably found it wanting, and explained in detail why that was so. It should not then be said that the district court clearly erred by refusing to give weight to unconvincing evidence. And when one takes a closer look, Plaintiffs' statistical evidence truly is fundamentally unconvincing, not just — as the majority calls it — "less precise." Maj. op. at 904.

4.

"[T]rial judges may evaluate the data offered to support an expert's bottom-line opinions to determine if that data provides adequate support to mark the expert's testimony as reliable." Milward v. Acuity Specialty Prods. Grp., Inc., 639 F.3d 11, 15 (1st Cir.2011). And in any case involving expert testimony, "a court may conclude that there is simply too great an analytical gap between the data and the opinion offered." Gen. Elec. Co. v. Joiner, 522 U.S. 136, 146, 118 S.Ct. 512, 139 L.Ed.2d 508 (1997).

Plaintiffs' own experts conceded that they used problematic data. In support of a motion to compel, one of Plaintiffs' experts affirmed under oath that the information he had received thus far was "incomplete in a number of important ways that ma[d]e it impossible to calculate reliable statistics." J.A. 399. Because of this "inadequate" data, the expert opined that he could not calculate "proper statistics" or perform "any of th[e] three standard forms of statistical analysis." J.A. 403, 409. Without additional data, it was concededly "impossible to calculate ... statistical patterns that might show whether or not a common issue of fact exists in this case." J.A. 403-04. Ultimately, the expert did not receive any of the additional data that he professed to need for a scientifically valid analysis. But, despite his sworn statements that the task was "impossible," he and another expert nevertheless produced statistical analyses based on the "incomplete" and "inadequate" data.

Plaintiffs' experts' report confirms that they used incomplete data to support and reach their conclusions. For instance, even though the experts drew conclusions about positions throughout the Nucor plant, they did not employ any data from either the shipping or maintenance departments. J.A. 1154. They used only a "limited amount of data" for the remaining departments. J.A. 1153. And although Plaintiffs' experts chose to use bidding data to determine an expected number of black promotions, they conceded that incomplete data "undermined" their "ability to use posting and bidding records to analyze [those] promotions." J.A. 1161. Nucor's expert identified other basic issues in Plaintiffs' experts' data that the majority opinion ignores. See J.A. 5892. For instance, Plaintiffs' experts included a promotion won by an external candidate in their pool — even though this case only concerns internally filled promotions. They further overlooked seven selections of black employees for promotions. See J.A. 5891.

The district court did not clearly err in discrediting this incomplete work and deeming it unworthy of evidentiary weight.

5.

a.

To further understand why Plaintiffs' statistical evidence is problematic, it helps to consider how it came about. In discovery, Nucor produced bidding packets and other promotion-related applicant data covering certain promotions from January 2001 to February 2006. Plaintiffs' analysis of the 2001-2006 data indicated that the black selection rate fell only 0.84 standard deviations from the mean — a statistically insignificant result. See J.A. 5872. Fortunately for Plaintiffs, the district court limited the use of the actual data to the January 2001 to December 2003 period. But an analysis of that period's data did not produce a statistically significant disparity, either. At best, analysis of the 2001-2003 data produced disparities falling only 1.53 standard deviations from the mean. See J.A. 1449.

Left with no results from actual records that suggested discrimination, Plaintiffs' experts set about creating extrapolated "benchmark" figures for promotions bidding between December 1999 and January 2001. They began by using so-called "change-of-status" forms plucked from personnel records to identify 27 purported promotions during the period. The experts then constructed a hypothetical bidding pool by essentially guessing that bidders in early years were racially identical to bidders in later ones. See J.A. 1162. With their theoretical promotion and bid figures established, Plaintiffs' experts then calculated an expected black promotion rate and compared it to the "actual" black promotion rate for the same period. Tied with the actual promotions figures from 2001 through 2003, Plaintiffs' extrapolated figures produced the number on which the majority now relies — 2.54 standard deviations.

b.

Plaintiffs' experts, however, based their extrapolations on several erroneous assumptions that render their model unreliable.

It begins with the change-of-status forms, which Nucor used to record any change of employee status. Because the forms also recorded demotions, pay increases, reassignments, and transfers, one cannot and should not assume that every form reflects a posted promotion. But up to the time that the district court decertified the promotions classes, Plaintiffs had never provided the 27 relevant change-of-status forms to the district court. Quite understandably, the district court wanted more concrete assurance that Plaintiffs' selected forms showed actual promotions. The district court never got that assurance, and it was "not inclined" to "take [Plaintiffs'] word for it." J.A. 10943. Plaintiffs did eventually submit the 27 relevant change-of-status forms — but only after the district court decertified the promotions classes. As it turns out, those forms do little to dispel the concern that Plaintiffs misidentified promotions. For example, two forms seem to show transfers, not promotions, J.A. 11006 (Reynolds), 11028 (Forsell), while another just reflects training, J.A. 11029 (Green). Others do not involve pay raises, suggesting no promotion occurred. See J.A. 11006 (Haselden), 11030 (Cooper). Certain other forms are ambiguous, failing to indicate whether pay rates changed or what the nature of the position change was. See, e.g., J.A. 11022 (Anderson), 11024 (Proskine), 11025 (Pope). Most of the forms fail to indicate whether Nucor posted the relevant opening for bidding. See, e.g., J.A. 11006-15, 11019-21, 11023, 11026-32. So, the district court was reasonably concerned that the 27 purported promotions — representing nearly half of the promotions in Plaintiffs' statistical analysis — were suspect and statistically useless.

The problems with Plaintiffs' experts' model continue to mount when the hypothesized bidding pools for the purported promotions are examined. Plaintiffs' experts hypothesized that at least one black employee bid on each of the 27 assumed promotion opportunities. But that approach rejects the prospect of an all-white bidding pool during the projected period, something likely to randomly happen from time to time given Nucor's 11% black workforce. Consequently, Nucor's expert concluded that Plaintiffs' experts' model "overstat[ed] the expected number of African American selections" between December 1999 and January 2001, as the model very likely inflated the number of black bidders. J.A. 5912. And indeed, Plaintiffs' experts calculated that black workers applied to jobs at a substantially higher pace than their actual percentage of the workforce, further suggesting some degree of inflation. Compare J.A. 1157 (noting that workforce was "11.3% African-American"), with J.A. 1162 ("The racial composition of the bidders ... was 19.24% African-American.").

An "inflated pool" like the one that Plaintiffs used "can undermine the validity of a statistical study to determine imbalances." Smith v. Va. Commonwealth Univ., 84 F.3d 672, 677 (4th Cir.1996). When a statistical model overestimates the number of black bidders, for instance, then black bidding rates artificially rise and black selection rates artificially fall. These effects might explain, for instance, why the black bidder selection rate for January 2001 to December 2003 — when actual data was available — was three times higher than the calculated selection rate for December 1999 to January 2001 — when projected data was used. If, during the projected period, the hypothesized number of black bidders in the pool (artificially) rose while the number of black bidder selections stayed the same, then the hypothesized black selection rate would be (artificially) driven down during the projected period.

c.

The majority nevertheless dubs the extrapolated data "sound." Maj. op. at 903. That conclusion, however, reflects an unwillingness to confront genuine concerns over statistical validity.

For instance, although admitting that the change-of-status forms are ambiguous, the majority blames Nucor for not explaining how these ambiguities would affect Plaintiffs' statistical accuracy. Maj. op. at 905. That burden was not Nucor's. Cf. Cooper v. Smith & Nephew, Inc., 259 F.3d 194, 199 (4th Cir.2001) (noting that the "proponent of the testimony" bears the burden of proving that it is reliable). Recently, for example, the Court affirmed a district court's refusal to consider statistical evidence offered to show disparate impact because the evidence contained a number of "mistakes and omissions" in its analysis. EEOC v. Freeman, 778 F.3d 463, 467 (4th Cir.2015). The Court did so even though the plaintiff there raised the very same argument that the majority now embraces: that the employer never "show[ed] that correcting the errors would negate the disparate impact." Brief for Appellant at 26, Freeman, 778 F.3d 463 (No. 13-2365), 2014 WL 320746. The Court appropriately rejected that argument then; it should have done the same now.

Rather than focusing on the reliability of the extrapolated statistics, the majority prefers to revisit the Brown I dissent. See maj. op. at 904-05. That dissent noted some of the concerns mentioned here: not all change-of-status forms used to extrapolate openings reflect promotions, many forms are unclear, and few forms indicate whether positions were posted. See Brown I, 576 F.3d at 168 (Agee, J., concurring in part and dissenting in part). To illustrate these concerns, the dissent examined "the change-of-status forms found in the record for 2000." Id. Bear in mind that, at least up to that point, Plaintiffs had never produced the particular change-of-status forms that they relied upon to guesstimate their statistics. Nor had they informed the Court that the forms in the record were not those upon which they based their statistical evidence. So, the Brown I dissent used the only change-of-status forms that were available to assess whether they could credibly support Plaintiffs' alleged statistical disparities. Id. Although the majority labels this exercise "sua sponte fact-finding," maj. op. at 905, the discussion in the Brown I dissent consisted of nothing more than explication by example.

The majority then attempts to tie the district court's decertification decision to the "error" that the majority mistakenly identifies in the Brown I dissent. According to the majority, the district court committed "clear factual error" by assuming that the change-of-status forms discussed in the Brown I dissent were those that Plaintiffs relied upon to build their statistical model. But here's the rub: the district court expressly disclaimed that very assumption. The district court noted that, at the time of decertification, Plaintiffs still had not produced the relevant forms. So, it had "never seen the 27 change-of-status forms upon which [Plaintiffs'] experts apparently relied." J.A. 10943. Thus, the district court cited the Brown I dissent only to emphasize the potential problems inherent in using the forms and why it needed to see them. See J.A. 10942-43. The majority's protracted discussion of the Brown I dissent therefore does nothing to rehabilitate Plaintiffs' evidence, resting as it does on a twofold misreading of the Brown I dissent and the district court's decertification decision.

Nor does the majority explain why inflated black bidding rates can be excused. Rather than address that obstacle, the majority assures the reader that the problem causes only "an incremental reduction in probative value" that does not "fatally undermine the probativeness of the experts' findings." Maj. op. at 906. But it is hard to minimize these defects so quickly when Plaintiffs' experts offered few explanations for their assumptions or any assessment of the expected impact of those assumptions. The experts did not say, for instance, whether black bidding rates varied during the years for which data was available. If they had shown that the rates remained steady, then one might assume that those same rates applied to the extrapolated years. But if the rates varied, then Plaintiffs' experts' assumptions are not sustainable. Oddly, the majority again blames Nucor for not summoning any evidence going to variation, but that tack once more reverses the burden of proof. "It is the plaintiffs' burden to demonstrate compliance with Rule 23," not Nucor's. EQT Prod., 764 F.3d at 358. The majority further finds that Plaintiffs' experts reasonably assumed that "every" position was posted for bidding. But Plaintiffs themselves submitted testimony identifying several unposted positions. See, e.g., J.A. 1010, 1051, 1091, 1110. Nucor's stated policies also indicated that, at least for a time, "[v]acant supervisory positions [were] not [to] be posted for bidding." J.A. 257.

The majority stresses that, as a general matter, plaintiffs may employ extrapolated data to prove discrimination. Maj. op. at 904. That can be true in some cases, but extrapolated data must still be statistically valid. And the majority ignores a significant and telling distinction between this case and past ones: Plaintiffs' experts extrapolated two data points — the composition of the applicant pool and the success rates — whereas experts in our prior cases only extrapolated one data point. See Lewis, 773 F.2d at 568; United States v. Cnty. of Fairfax, Va., 629 F.2d 932, 940 (4th Cir.1980).

The majority's cited cases also involved defendants who wrongfully destroyed relevant evidence. See Lewis, 773 F.2d at 568 (noting that the defendant "improperly disposed" of applicant records); Cnty. of Fairfax, 629 F.2d at 936 n. 4 (noting that the defendant destroyed applicant data "[i]n violation of the record keeping regulations of [two statutes]"). In a situation involving spoliation of evidence, the Court commonly draws adverse inferences against the spoliators. But this record contains no evidence of spoliation.

Regardless, no authority requires the district court to find extrapolated data convincing in every case. Our precedent holds just the opposite. In Allen v. Prince George's County, 737 F.2d 1299, 1306 (4th Cir.1984), for example, the district court relied solely upon actual applicant flow data "to the exclusion of all [other] statistical evidence," including evidence crafted from alternative benchmarks. We affirmed, emphasizing that we could not "second-guess" a fact-bound decision concerning "the relative weights to be accorded to the parties' respective evidence." Id. The district court here did essentially the same thing as the district court in Allen, giving weight for good reason to the actual data available to the exclusion of the speculative extrapolation evidence. As in Allen, we should not say that the district court clearly erred in doing so.

6.

a.

Plaintiffs' statistical evidence also does not apply controls for non-discriminatory factors that could very well have caused any observed disparities. See Lowery, 158 F.3d at 764. Seniority, for instance, influences promotions decisions at Nucor. See, e.g., J.A. 257. Disciplinary issues also led Nucor to reject certain applicants for promotion — including frequent bidder Jason Guy, who is black. See J.A. 659-67; see also Coates v. Johnson & Johnson, 756 F.2d 524, 544 n. 20 ("[A]n employee's prior discipline record seems likely to be a major, if not the most important, factor in [an employment] decision."). But Plaintiffs' experts admitted that they did not control for these or any other "additional factors beyond the control for each job posting." J.A. 1164. The majority would wish these considerations away, reasoning that Nucor never raised them. But Nucor's expert noted the need to "control for characteristics that would seem to affect the chance of selection," which would include matters like seniority and discipline. See J.A. 5893. Anyway, we could have affirmed the district court's decision here on "any basis supported by the record." Defenders of Wildlife v. N.C. Dep't of Transp., 762 F.3d 374, 392 (4th Cir.2014).

The majority also tries to summon its own justifications for these omissions, implying that records were not available to control for matters like discipline. Maj. op. at 906. Even Plaintiffs' experts conceded that they were. See J.A. 1165 (acknowledging that Nucor had maintained and produced "bidders' training, discipline, and bidding records"); see also J.A. 5893 (Nucor's expert observing that "separate discipline and training files [were] provided to Drs. Bradley and Fox and [him]"). And, based on allegations and personal assessments from Plaintiffs themselves, the majority assumes that potential explanatory variables are themselves racially biased. See maj. op. at 906-07. Yet here again, Plaintiffs' experts do not assume so, perhaps because there is no concrete evidence of such taint in the record. See Ottaviani v. State Univ. of N.Y. at New Paltz, 875 F.2d 365, 375 (2d Cir.1989) (holding that district court correctly required the plaintiffs to account for potential explanatory variable where the plaintiffs alleged but did not prove that the variable was biased). And even if one were to indulge the majority's assumption that discipline at Nucor was itself biased, that outcome would not justify excluding the variable from the statistical model completely. "[T]ainted variables should not be routinely excluded from the regression equation. Instead, the effects of the inclusion of a tainted variable must be assessed and minimized." Paetzold & Willborn, supra, § 6:13. The majority's reasons, then, do not fill the gaps in Plaintiffs' experts' work.

The failure to control for non-race-related explanatory variables "is sufficiently serious so as to weaken the statistical study's probativeness." Lowery, 158 F.3d at 764; see also Smith, 84 F.3d at 676; accord Rodriguez v. Nat'l City Bank, 726 F.3d 372, 384-85 (3d Cir.2013); Morgan v. United Parcel Serv. of Am., Inc., 380 F.3d 459, 468 (8th Cir.2004); Munoz v. Orr, 200 F.3d 291, 301 (5th Cir.2000); Sheehan v. Daily Racing Form, Inc., 104 F.3d 940, 942 (7th Cir.1997); Penk v. Or. St. Bd. of Higher Educ., 816 F.2d 458, 465 (9th Cir. 1987). A trier of fact must determine whether racial discrimination — rather than chance or some other "confounding factor[]" — caused an alleged disparity. In re Navy Chaplaincy, 738 F.3d 425, 429 (D.C.Cir.2013). Only a controlled model can provide that answer, and Plaintiffs' experts' evidentiary model did not meet that definition.

b.

In most every employment case, a valid statistical model must account for one particularly important explanatory variable: the applicant pool's qualifications. "[T]he relevant comparison is between the percentage of minority employees and the percentage of potential minority applicants in the qualified labor pool." Carter v. Ball, 33 F.3d 450, 456 (4th Cir.1994); see also City of Richmond v. J.A. Croson Co., 488 U.S. 469, 501-02, 109 S.Ct. 706, 102 L.Ed.2d 854 (1989); McNairn v. Sullivan, 929 F.2d 974, 979 (4th Cir.1991). If courts were to accept statistical models containing unqualified applicants, then employers could be punished merely because of a "dearth of qualified nonwhite applicants (for reasons that are not [the employers'] fault)." Wards Cove Packing Co. v. Atonio, 490 U.S. 642, 651, 109 S.Ct. 2115, 104 L.Ed.2d 733 (1989). Thus, "statistics based on an applicant pool containing individuals lacking minimal qualifications for the job [are] of little probative value." Watson, 487 U.S. at 997, 108 S.Ct. 2777; see also Paetzold & Willborn, supra, § 4:3 ("[W]hen considering potential discrimination in promotions within an organization, only employees qualified for promotion should be considered in the proxy pool."). Furthermore, "[n]o rational enterprise that has several qualified candidates for a position selects among them by lot; it picks the best qualified." Mason v. Cont'l Ill. Nat'l Bank, 704 F.2d 361, 364 (7th Cir.1983). So, a truly effective statistical model will not just account for minimum qualifications, but should control for the variations in skills even among minimally qualified applicants.

By this point, Plaintiffs and their experts should have known better than to ignore other explanatory factors. In a related case challenging promotions practices at a different Nucor facility, the Eighth Circuit found that similarly substandard work from the same expert did not create a triable question of fact on summary judgment. See Bennett, 656 F.3d at 812. In so holding, the Eighth Circuit emphasized that the expert's statistics had "little force" because they "assumed that all applicants were qualified for promotion to each available position." Id. at 818. The Eighth Circuit is not alone. Other courts have criticized Plaintiffs' principal expert for employing his "warm body hypothesis," which "assumes that every person is just as qualified and skilled and experienced as everyone else." Davis v. Ala. Dep't of Educ. Dep't of Disability Determination Serv., 768 F.Supp. 1471, 1477 (N.D.Ala.1991); accord Adams v. Austal, U.S.A., L.L.C., No. 08-00155-KDN, 2011 WL 1558790, at *8 (S.D.Ala. Apr. 25, 2011); Rollins v. Ala. Cmty. Coll. Sys., No. 2:09cv636-WHA, 2010 WL 4269133, at *8-9 (M.D.Ala. Oct. 25, 2010); Bennett v. Nucor Corp., No. 3:04CV00291 SWW, 2007 WL 2333193, at *3 (E.D.Ark. Aug. 13, 2007); Yapp v. Union Pac. R.R. Co., 229 F.R.D. 608, 619 (E.D.Mo. Aug.5, 2005); Rhodes v. Cracker Barrel Old Country Store, Inc., No. Civ.A. 4:99-CV-217-H, 2002 WL 32058462, at *65 (N.D.Ga. Dec. 31, 2002). We even affirmed a district court's choice to exclude work from the same expert precisely because he did not incorporate adequate controls. See Anderson, 406 F.3d at 262-63 (agreeing with the district court's view that the expert had ignored "actual job performance or job requirements" even though he "conceded" that he could have "use[d] a control factor that would control for the actual job title or the job duties").

Plaintiffs' experts assumed that all persons in each bidding pool were equally qualified because "only persons who decided to bid based on the posted qualifications were included." J.A. 1162. This opaque language obscures another faulty assumption built into the model: the experts assumed that only qualified persons applied for each promotion opportunity. It takes no expertise to comprehend that some people "might be discouraged from applying because of a self-recognized inability to meet the [opening's] standards." Dothard v. Rawlinson, 433 U.S. 321, 330, 97 S.Ct. 2720, 53 L.Ed.2d 786 (1977). But one could hardly assume that every job applicant is so discerning, and even the majority seems unwilling to make that assumption. See maj. op. at 906. The majority prefers to guess that the number of unqualified applicants will be so trivially small as to be statistically irrelevant, and it makes that guess simply because the job announcement includes job requirements. In practical effect, the majority has read the "qualified applicants" limitation found in our prior cases out of the law, as most every job opening provides some minimal description of what skills are required.

"A statistical study that fails to correct for explanatory variables, or even to make the most elementary comparisons, has no value as causal explanation[.]" People Who Care v. Rockford Bd. of Educ., 111 F.3d 528, 537 (7th Cir.1997). Plaintiffs presented just such a study here, and the district court did not clearly err in rejecting it.

7.

Lastly, Plaintiffs' statistical evidence improperly aggregates data in a way that distorts the results.

a.

The objective in a class action — even in a proceeding that alleges disparate treatment — is to identify a common, uniform policy. "While in a case alleging intentional discrimination, such as this one, a plaintiff need not isolate the particular practice and prove that such practice caused the discrimination, plaintiffs must make a significant showing to permit the court to infer that members of the class suffered from a common policy of discrimination that pervaded all of the employer's challenged employment decisions." Love v. Johanns, 439 F.3d 723, 728 (D.C.Cir.2006).

Thus, if the class challenges a policy implemented at the nationwide level, then plaintiffs might use applicable statistics showing nationwide disparities to establish the policy's effects. Conversely, if the class challenges policies implemented on a plant-by-plant or department-by-department basis, then the class must summon statistics showing disparities at that level. Otherwise, non-uniform decisions made by one discriminatory decisionmaker might create disparities that, when aggregated with other, neutral decisions, misleadingly indicate discrimination across the whole group of decisionmakers.

Wal-Mart demonstrates these concepts well. There, the plaintiffs offered statistics purporting to show regional and national disparities in employment decisions at Wal-Mart. Wal-Mart, 131 S.Ct. at 2555. Those decisions, however, were made at the store level. Id. at 2547. Because of that disconnect, the Supreme Court held that plaintiffs' statistics did not establish a common policy. Once again, the broader disparities might have been "attributable only to a small set of Wal-Mart stores" and did not "establish the uniform, store-by-store disparity upon which plaintiffs' theory of commonality depend[ed]." Id. at 2555. In essence, Wal-Mart agreed with our own, earlier cases indicating that statistics should not be aggregated together to create disparities that are not actually representative of the class as a whole. Compare Stastny, 628 F.2d at 279-80 (requiring the plaintiffs' statistics to focus on the "locus of autonomy"), with Elizabeth Tippett, Robbing a Barren Vault: The Implications of Dukes v. Wal-Mart for Cases Challenging Subjective Employment Practices, 29 Hofstra Lab. & Emp. L.J. 433, 447 (2012), cited with approval by Scott v. Family Dollar Stores, Inc., 733 F.3d 105, 113 (4th Cir. 2013) (explaining that Wal-Mart requires that plaintiffs' statistics focus on "the locus of the subjective decision-making").

In requiring the plaintiffs' statistics to be centered at the level of relevant decisionmaking, Wal-Mart did not distinguish between nationwide and other class actions. Rather, Wal-Mart asked whether the plaintiffs there were too dissimilar to bring their claims together, regardless of how many claims there might be. Thus, courts have applied principles from Wal-Mart in cases involving classes of roughly the same size as the class at issue here. See, e.g., Wang v. Chinese Daily News, Inc., 737 F.3d 538, 544 (9th Cir.2013) (200 class members); Ealy, 514 Fed.Appx. at 304-08 (150 class members). Even statisticians agree that Wal-Mart reaches classes big and small. See, e.g., Dr. Mary Dunn Baker, Class Certification Statistical Analysis Post-Dukes, 27 ABA J. Lab. & Empl. L. 471, 479 (2012) ("[T]he size of the putative class or the number of establishments the defendant operates will have little to do with whether the Dukes commonality approach is applicable."). So, even though Plaintiffs here challenge practices in one plant, they still must offer statistics showing disparities among all the relevant decisionmakers, regardless of that one-plant focus. See Rubenstein, supra, § 24:40 ("Courts have certified [only] limited classes when the facts show that no uniform personnel policies are applied among the various plants, departments, or levels of employees.").

b.

Here, as the Brown I majority agreed, the evidence indicates "that each department manager" in each of Nucor's six production departments "has unbridled discretion to make promotions within his department utilizing whatever objective or subjective factors he wishes." Brown I, 576 F.3d at 151. Department managers took full advantage of that discretion, developing processes that they recurrently characterized as unique and independent. See J.A. 7887, 7894-95, 7900, 7906-07. Indeed, these processes were so varied that one supervisor declared that he had "no idea what other departments d[id]." J.A. 8109. Even the decisionmakers varied. In some departments, such as the hot mill and shipping departments, supervisors and the department managers made promotion decisions. In other departments, such as maintenance and the cold mill, promotions decisions were a more collaborative effort involving even lower-level lead men. These different decisionmakers then applied different standards. In the beam mill, for example, the process centered upon interviews alone. In contrast, the melt shop looked to applicants' work history, safety record, psychological interview, job skills, training, attendance, and scores on a job-specific aptitude test. Nucor's general manager quite reasonably described the promotions processes when he said that "each department ha[d] their own way of doing [promotions]." J.A. 1723.

Plaintiffs' own expert found that each department had its own procedures, and at least eight different criteria — not including "numerous other idiosyncratic factors" — might or might not be considered in making any employment decision. J.A. 1518-19. "Different supervisors," he explained, "utilized different criteria weighting schemes with little consistency among the selection officials and among the different hiring/promotion/transfer opportunities." J.A. 1525. Taking all this dissimilarity together, the expert concluded that Nucor's selection process was only "consistent in its inconsistency." J.A. 1519.

Yet Plaintiffs' statistical evidence incorrectly assumed the exact opposite: perfect, plant-wide consistency as to promotions. Given that promotions decisions were made at the department or supervisor level using different and independent criteria, we cannot rightfully assume that a plant-wide disparity resulted from a uniform problem arising in the same way in each Nucor department. See Wal-Mart, 131 S.Ct. at 2555. Put differently, the district court reasonably found that the "locus of autonomy" rested at the departmental level, not a plant-wide one. We cannot then assume that department decisions were made in lockstep, such that plant-wide disparities necessarily reflect common, departmental ones. See Bolden v. Walsh Constr. Co., 688 F.3d 893, 896 (7th Cir. 2012) (rejecting aggregate data because it did not necessarily imply that "all 25 superintendents behaved similarly, so it would not demonstrate commonality").

We have already seen these concepts play out in another employment discrimination action involving a similar Nucor facility. Applying Wal-Mart, the Eighth Circuit rejected statistics — from the same expert — that reflected plant-wide disparities in promotions at an Arkansas Nucor plant. Bennett, 656 F.3d at 815-16. Just as in this case, the statistical evidence there indicated that different departments in the plant applied different criteria for promotions decisions. Id. at 815. The plant-wide evidence therefore "ha[d] little value in the commonality analysis" because it "did not differentiate between the hiring and promotion decisions made in each department." Id. The Eighth Circuit found that, in those sorts of circumstances, "a bottom-line analysis [wa]s insufficient to demonstrate that any disparate treatment or disparate impact present in one department was also common to all others." Id. at 815-16.

As in Bennett, Nucor here provided its own analysis that demonstrated how the statistical disparities varied among the different departments in the plant. Nucor's expert measured how selection rates varied between white and black applicants on a department-by-department basis over the period for which bidding information was available. With proper controls applied, the expert found that race differences between departments could vary by as much as 2.44 standard deviations. J.A. 5894. In other words, some departments experienced decidedly smaller disparities in selection rates, undermining any inference of uniformity and commonality among all departments.

Given the wide variance in promotions practices at the Nucor facility, the district court did not clearly err in rejecting a statistical study that failed to account for that variance.

c.

The majority finds, however, that Nucor's entire plant should be treated "as a single entity" when it comes to promotions decisions. Maj. op. at 911 (alluding to Brown I, 576 F.3d at 158). Although the majority suggests otherwise, Brown I did not decide this issue. Brown I held that the district court should treat Nucor's various production departments as a single facility only for purposes of Plaintiffs' hostile work environment claim. 576 F.3d at 158 ("[T]he affidavits of employees in one department are admissible to prove a plant-wide hostile environment that affected employees in other departments, and the plaintiffs have satisfied the commonality requirement for their hostile work environment claim." (emphasis added)); see also id. at 157 (discussing how a "hostile environment determination" must be made in the context of discussing Plaintiffs' "single entity" argument). It said nothing about the uniformity of promotions decisions across the plant. Id. The Brown I majority did so because Plaintiffs likewise focused their "single entity" argument on only the hostile work environment claim. See Brief for Appellant at 25-35, Brown I, 576 F.3d 149 (No. 08-1247), 2008 WL 2307453. Thus, as with predominance, the district court was not constrained in deciding the "single facility" issue, as no Brown I mandate existed as to that issue.

Nonetheless, the majority concludes that facts establishing a single hostile work environment claim also establish a common promotions policy. Maj. op. at 911-12. Yet "[d]isparate treatment ... is inherently different from hostile work environment. The federal courts treat the two types of cases differently for good reason." See Pollard v. E.I. DuPont de Nemours Co., 213 F.3d 933, 943 (6th Cir.2000), rev'd on other grounds, 532 U.S. 843, 121 S.Ct. 1946, 150 L.Ed.2d 62 (2001). And no court has held that a common hostile work environment establishes that a facility must be treated as a single entity for purposes of every other kind of employment discrimination claim.

In finding a common environment, Brown I focused on shared locker rooms and spaces, plant-wide email, and plant-wide radio systems. 576 F.3d at 158. When it comes to a hostile work environment claim, those facts may matter: racial slurs and "monkey noises" uttered in a common space or transmitted via plant-wide radio can affect whoever hears them. See Burlington Indus., Inc. v. Ellerth, 524 U.S. 742, 762, 118 S.Ct. 2257, 141 L.Ed.2d 633 (1998) ("[A]nyone who has regular contact with an employee can inflict psychological injuries by his or her offensive conduct."). But locker rooms and radios bear no relationship to promotions decisions; certainly nothing in the record supports such a concept. Only supervisors can inflict the "pain" of a denied promotion, and they can do so only when empowered by company structure, not common spaces. We should not assume that dozens of supervisors acted in concert merely because their employees might have changed clothes in the same room. Nor should we assume — in the face of expressly different criteria applied to different groups of employees — that applicants in each department nevertheless suffer the same injury merely because of their physical proximity to one another at some point during a workday. Though the majority insists that "centralized, circumscribed environments" will "generally" increase "consistency" in managerial decisionmaking, maj. op. at 910, Plaintiffs' own expert made clear that this hypothesized general rule cannot apply here, see J.A. 1519 ("The best sentiment I can muster in favor of the [Nucor] selection procedure is that it is consistent in its inconsistency."). See also, e.g., Tabor v. Hilti, Inc., 703 F.3d 1206, 1229 (10th Cir.2013) (affirming denial of class certification where "Plaintiffs challenge[d] a highly discretionary policy for granting promotions").

The majority also notes that the general manager formally approved promotions in the plant. Maj. op. at 917. Without saying so explicitly, the majority seems to propose that the general manager provided some common, plant-wide direction that drove common, plant-wide disparities. Yet even the Brown I majority recognized that the general manager played no genuine role in the promotions decisionmaking process. 576 F.3d at 152 ("Although, by policy, the plant's general manager approves all promotions and handles discrimination and harassment investigations, the record suggests that each department manager has unbridled discretion to make promotions within his department utilizing whatever objective or subjective factors he wishes."). The evidence confirms that proposition. Promotions, the general manager explained, were "not [his] area of responsibility," as he had "department managers that ma[d]e those decisions." J.A. 8163. Nucor instead trained its department managers to make promotions decisions and implement the anti-discrimination policy.

The majority nevertheless says the general manager engaged in "inaction." Maj. op. at 912, 917. The majority's theory — premised on an assumed culture of "odious racism" and passive enabling — resembles a theory that Wal-Mart out-and-out rejected. See 131 S.Ct. at 2553-54 (refusing to credit evidence asserting that a "strong corporate culture," enabled by policies of discretion, permitted bias in pay decisions); accord Davis v. Cintas Corp., 717 F.3d 476, 489 (6th Cir.2013).

Even if one assumes that such a theory were viable and relevant here, it would not prove commonality. "Inaction" — letting supervisors do as they wish — is just discretion by another name. "[I]t is a policy against having uniform employment practices." Wal-Mart, 131 S.Ct. at 2554. "Wal-Mart tells us that local discretion cannot support a company-wide class no matter how cleverly lawyers" (or judges) "may try to repackage local variability as uniformity." Bolden, 688 F.3d at 898; accord In re Navy Chaplaincy, No. 1:07-mc-269 (GK), ___ F.R.D. ___, ___, 2014 WL 4378781, at *15 (D.D.C. Sept. 4, 2014). Were it otherwise, one could find a common policy in most every case, as most every company has a management head at the top that could be accused of not doing enough. Beyond that, Plaintiffs' experts never traced their identified disparities to the general manager, and their reports never even mention him. For good reason. Individual acts of discretion, not the general manager's purported acquiescence, would have caused any disparities and the injuries that they reflect. Thus, the not-very-common common policy does not present a common injury.

Nucor also used a plant-wide "dual-approval" scheme, under which promotions required approval from both "originating" and "destination" department heads. The majority sees this as a case of potential "cat's paw" liability, wherein a non-decisionmaker influences the ultimate decisionmaker's choice in a discriminatory way. Maj. op. at 911-12 (citing Smith v. Bray, 681 F.3d 888, 897 & n. 3 (7th Cir.2012)). But nothing other than speculation indicates that dual approval was used to effect discrimination in any common way, and any cat's paw must be the "proximate cause" of the discriminatory harm to be actionable. Staub v. Proctor Hosp., 562 U.S. 411, 131 S.Ct. 1186, 1192, 179 L.Ed.2d 144 (2011). Not even Plaintiffs' statistical experts attempt to tie their disparities to a dual-approval policy.

The majority surmises that a discriminatory supervisor in one department could have theoretically used dual approval to inflict his animus upon employees outside his own department. But if a racist department head had tried to use the dual-approval scheme to disadvantage black workers, he would not have been able to reach all or even most of the promotions decisions in the plant, dual approval notwithstanding. A discriminatory department head in the beam mill, for instance, would have had no say when it came to a cold mill employee seeking a higher position within the cold mill, hot mill, melt shop, maintenance department, or shipping department. Perhaps, then, the majority's concept — if properly supported with evidence — might justify a class of persons applying in and out of a particularly problematic department. In fact, the district court proposed certifying just such a class as to the beam mill. See J.A. 10953-54 & n.16. But it would not justify the plant-wide class action that Plaintiffs now mean to bring. Cf. Ellis, 657 F.3d at 983 ("A disparity in only 25% of the regions, however, would not show that discrimination manifested in promotions practices in the same general fashion.").

* * *

In sum, the district court did not clearly err in choosing not to rely on Plaintiffs' statistical evidence. Faced with evidence based on questionable data, uncontrolled explanatory variables, and poorly structured methodologies, the district court did not act irrationally in determining that such evidence was of negligible credence. The "troubling effects of statistical inferences require thoughtful consideration in each case," Mister v. Ill. Cent. Gulf R.R. Co., 832 F.2d 1427, 1437 (7th Cir.1987), and that consideration is sorely lacking from the work of Plaintiffs' experts. Thus, Plaintiffs' evidence, with its many deficiencies, does not establish the common policy necessary for class certification. The district court did not abuse its discretion in making that finding.

B. Anecdotal Evidence

Plaintiffs also present affidavits from sixteen employees in support of certifying the promotions classes. The district court did not abuse its discretion in refusing to certify Plaintiffs' proposed class based on this limited evidence.

1.

In their original class certification motion, Plaintiffs never argued that anecdotal evidence, standing alone, could establish a common policy of discrimination. Rather, Plaintiffs presented the anecdotal evidence only to supplement their statistical evidence. See Brown I, 576 F.3d at 164 (Agee, J., dissenting). The Brown I majority constructed its own theory of the case, finding that Plaintiffs could in fact advance their case on anecdotal evidence "alone." Id. at 153. Plaintiffs now take up the Brown I majority's theory in this appeal.

Plaintiffs made the better choice in their initial offering, as anecdotes only help tell the story. They are meant to bring "the cold numbers convincingly to life," Teamsters, 431 U.S. at 339, 97 S.Ct. 1843, providing "texture" for statistical evidence. Robinson v. Metro-North Commuter R.R. Co., 267 F.3d 147, 168 (2d Cir.2001), abrogated on other grounds by Wal-Mart, 131 S.Ct. at 2560-62. But standing alone, "anecdotal evidence ... [will] rarely, if ever,... show a systemic pattern of discrimination." O'Donnell Constr. Co. v. Dist. of Columbia, 963 F.2d 420, 427 (D.C.Cir. 1992); accord Briggs v. Anderson, 796 F.2d 1009, 1019 (8th Cir.1986) (observing that plaintiffs "punished themselves" by choosing to rely on anecdotal evidence); EEOC v. Bloomberg L.P., 778 F.Supp.2d 458, 470-71 & n. 8 (S.D.N.Y.2011) (collecting cases); see also Michael Selmi, Theorizing Systemic Disparate Treatment Law: After Wal-Mart v. Dukes, 32 Berkeley J. Emp. & Lab. L. 477, 501 (2011) ("[A]necdotal evidence is always of marginal significance in a pattern or practice claim.").

In discrimination cases, courts move anecdotal evidence to the background because such evidence does not prove much. "Anecdotal reports ... are ordinarily more helpful in generating lines of inquiry than in proving causation." Federal Judicial Center, Reference Manual on Scientific Evidence 217 (2011). Individual stories say little, for instance, about the frequency of an event's occurrence or the reasons for that occurrence. Without knowing at least those two items, it can hardly be assumed that the stories reflect a broader trend flowing directly from intentional discrimination. See Wessmann v. Gittens, 160 F.3d 790, 805-06 (1st Cir.1998); Coral Constr. Co. v. King Cnty., 941 F.2d 910, 919 (9th Cir.1991). Anecdotes are also more susceptible to mistaken perception, leading to erroneous conclusions — especially when collections of stories are treated as quasi-statistics. See Fisher v. Vassar Coll., 70 F.3d 1420, 1444-45 (2d Cir.1995). And bias can skew anecdotal evidence, as when only those who feel most strongly about an issue offer anecdotes or when the soliciting party has a particular objective in mind. Cf. United States v. Local 560 of Int'l Bhd. of Teamsters, Chauffeurs, Warehousemen, & Helpers of Am., 780 F.2d 267, 277 (3d Cir.1985) (finding that a survey that was meant to show the "reputation" of a particular organization should have been excluded when it only surveyed persons known "to be hostile" to the organization). Because "anecdotes provide no mechanism for assessing truthfulness, typicality, or frequency," courts can and should question their usefulness, just as "[s]cientists and medical researchers" have done for many years. David A. Hyman, Lies, Damned Lies, and Narrative, 73 Ind. L.J. 797, 803 (1998).

2.

The majority finds Plaintiffs' anecdotal evidence sufficient principally because the ratio reflecting the number of affidavits alleging discrimination compared to the number of class members is purportedly small. Maj. op. at 912-13. As of 2006, Plaintiffs' experts determined that "approximately 150 African-Americans" comprised the class. J.A. 1154. Given that the class period extends well into 2011, it is reasonable to assume that Nucor hired additional black applicants since 2006, conservatively setting the present class size at 160 black employees or more. The sixteen affidavits that Plaintiffs provide therefore represent roughly one affidavit for every ten class members — a weak sample from the entire class. "[A] court must be wary of a claim that the true color of a forest is better revealed by reptiles hidden in the weeds than by the foliage of countless free-standing trees." Cooper, 467 U.S. at 879-80, 104 S.Ct. 2794. When ten percent of a class (or less) complains of mistreatment in a discrimination case, a district court does not clearly err in finding that such complaints do not establish a "standard operating procedure" of discrimination, Teamsters, 431 U.S. at 336, 97 S.Ct. 1843, "significant adverse effects" on the relevant class, Watson, 487 U.S. at 986, 108 S.Ct. 2777, or "significant proof" of class-wide discrimination, Wal-Mart, 131 S.Ct. at 2553.

3.

What may matter more than the quantity of a plaintiff's evidence is its quality. If, for instance, the anecdotal evidence is indirect and circumstantial, the district court might justifiably probe whether that evidence truly gives rise to a necessary inference of discrimination. After all, "a district court may properly consider the quality of any anecdotal evidence." Rossini v. Ogilvy & Mather, Inc., 798 F.2d 590, 604 (2d Cir.1986); accord Eastland v. Tenn. Valley Auth., 704 F.2d 613, 625 (11th Cir.1983).

At least as to the promotions-related matters at issue in this appeal, Plaintiffs do not present compelling anecdotal evidence. Byron Turner, for instance, does not address promotions at all. Neither does Walter Joseph Cook. In what might be an employment law first, Kenneth Hubbard complains that Nucor promoted him. See J.A. 1097; cf. Kalamazoo Cnty. Rd. Comm'n v. Deleon, ___ U.S. ___, 135 S.Ct. 783, 784, 190 L.Ed.2d 887 (2015) (Alito, J., dissenting from denial of certiorari) ("Respondent's supervisors did not violate federal law by granting him the transfer that he sought and that they had no reason to believe he did not want."). And Earl Ravenell testifies about a time that he applied for a promotion and was not selected — because another black employee was selected for that opening. He also tells us that he chose not to apply for any other positions because of "the look on his [supervisor]'s face." J.A. 1111. These and other examples are not "cherry pick[ed]," maj. op. at 913, but merely offer some insight into why the district court could reasonably decide differently than the majority does.

Much of the anecdotal evidence also amounts to conclusory and speculative statements of personal belief. For instance, even those employees who do mention job qualifications rely almost exclusively on their personal, subjective, and unsubstantiated views of their own abilities. We usually do not give such testimony much, if any, weight. See Williams v. Giant Food Inc., 370 F.3d 423, 433 (4th Cir.2004); Evans v. Techs. Applications & Serv. Co., 80 F.3d 954, 960 (4th Cir.1996). Other employees assume racism in the process without identifying an objective fact to support that view. Named plaintiff Ramon Roane declares, for example, that he applied for a position that was "suddenly cancelled because Nucor was not ready for an African American to hold a supervisory position." J.A. 996. Yet he does not explain how or why he came to that conclusion, and "[a] plaintiff's self-serving opinions, absent anything more, are insufficient to establish a prima facie case of discrimination." Mackey v. Shalala, 360 F.3d 463, 469-70 (4th Cir.2004).

In addition, Plaintiffs' evidence is often so incomplete that it lacks any probative value. For example, Bernard Beaufort discusses a promotions decision that he believes "was made unfairly." J.A. 6008. But he does not know who eventually received the job, what his or her race was, "what [the decision] was based on," or whether "it was based on [his] race." J.A. 6008. Other employees testify about not receiving promotions, but many of these declarants do not indicate whether they were minimally qualified for the position or whether the selected employee was of another race. Without these fundamental facts, we cannot know whether particular promotions decisions raise even a circumstantial inference of discrimination. See Cline v. Roadway Express, Inc., 689 F.2d 481, 485 n. 4 (4th Cir.1982); accord Tex. Dep't of Cmty. Affairs v. Burdine, 450 U.S. 248, 253, 101 S.Ct. 1089, 67 L.Ed.2d 207 (1981).

4.

The majority concentrates on one anecdotal comment from one supervisor in the beam mill: "I don't think we'll ever have a black supervisor while I'm here." J.A. 1885-86; see also maj. op. at 899, 917. That comment could be compelling evidence in a case hinging on decisions made by that particular decisionmaker. On the other hand, it might not be, as we have discounted "stray or isolated" remarks, even at summary judgment. Brinkley v. Harbour Recreation Club, 180 F.3d 598, 608 (4th Cir.1999); see also Birkbeck v. Marvel Lighting Corp., 30 F.3d 507, 511-12 (4th Cir.1994) (finding that decisionmaker's singular remark did not evidence discriminatory practices at company).

In the end, the question proves academic. A class-wide claim challenging decisions made by many different decisionmakers plainly requires something more than a single comment from just one of them. We see this rule — that sparse comments are not enough for class treatment — illustrated in cases like King v. General Electric Company, 960 F.2d 617 (7th Cir.1992). There, the Seventh Circuit found that the plaintiffs' anecdotal evidence in an age-discrimination case was not enough, even though the record contained testimony from a higher manager that the company was "going to get rid of these old farts and get some new blood in here." Id. at 628 (Cudahy, J., dissenting) (summarizing evidence rejected by the majority). This Court, too, has rejected anecdotal evidence of a similarly "damning character," this time in a racial discrimination case. See Coker v. Charleston Cnty. Sch. Dist., No. 92-1589, 1993 WL 309580, at *6 (4th Cir. Aug. 16, 1993). We found that the plaintiffs had not established a policy or practice of discrimination despite testimony that a black principal was told the community would not "accept" him at a predominantly white school. Id. at *4. All this goes to illustrate that plaintiffs likely cannot prove a class-wide policy with a single comment, no matter how bigoted the comment may be. One comment certainly does not make the showing that Plaintiffs insist they make here: a common, uniform policy of animus inflicted by 55 or more independent supervisors upon more than 150 employees scattered throughout a multi-department plant. Consequently, the district court did not abuse its discretion in refusing to certify Plaintiffs' class based on a single comment.

5.

a.

The district court also gave "limited weight" to almost 80 affidavits from black employees at the Nucor plant. J.A. 10950. The affidavits consistently rejected the idea of discrimination in the promotions process, and the district court did not abuse its discretion in affording them some minimal value. Repeatedly, the affidavits suggest that the promotions process was fair. See, e.g., J.A. 6024, 6042, 6052, 6069, 6078. One such employee specifically remarked that "[n]ot all African-Americans feel like they have been discriminated against at Nucor." J.A. 6109. The same employee was actually "upset by this racial discrimination issue because it is not something that has happened to me or is happening across the board here at Nucor." Id. Another employee explained that "the way things are done ... at Nucor are not influenced by race." J.A. 6164.

The list goes on: black employees approved of management's handling of race-related issues in the plant, see, e.g., J.A. 6109, 6215, 6480-81, 6943, explained that they were treated well, see, e.g., J.A. 6350, 6361, and often reasoned that complaints of racism from other employees were unjustified, see, e.g., J.A. 6566. Even those who felt that promotions were not made fairly often blamed factors other than race, such as a "buddy" system in which supervisors promoted friends. See, e.g., J.A. 6258, 6299, 6438, 6494. Some affidavits also directly contradicted the sixteen declarations that Plaintiffs submitted. In fact, Jacob Ravenell, Kenneth Hubbard, Robyn Spann, and Byron Turner all expressly denied that they had been denied promotions because of their race, even though Plaintiffs cite them as four of their sixteen key witnesses. See J.A. 6400, 6746, 6933, 6964. The district court had every right to weigh such self-contradictory testimony and conclude as it did. See Stevenson v. City of Seat Pleasant, Md., 743 F.3d 411, 422 (4th Cir.2014).

b.

Based on "[c]ommon sense and prudence," however, the majority finds yet again that the district court clearly erred — this time by finding that "potentially coercive" affidavits supported Nucor to some small degree. Maj. op. at 913. The majority's naked credibility determination is exactly the sort of decision we are not meant to undertake on appellate review. "[W]hen a trial judge's finding is based on his decision to credit the testimony of [a witness who] ... has told a coherent and facially plausible story that is not contradicted by extrinsic evidence, that finding, if not internally inconsistent, can virtually never be clear error." Anderson, 470 U.S. at 575, 105 S.Ct. 1504.

The majority nevertheless adopts a self-contradictory credibility rule: statements made in support of an employer must be rejected when the employer obtains them, while statements made against the employer will be given "significant weight given the circumstances in which they were made." Maj. op. at 914. The majority draws this distinction by assuming that an employer exercises coercive power in most any interaction with its employees. "However, it is well settled that not every interrogation of employees by Company officials constitutes coercion[.]" NLRB v. Lexington Chair Co., 361 F.2d 283, 289 (4th Cir.1966). And one must not lose sight of the practical effect of the majority's novel approach: employers now have no incentive to investigate and remedy claims of discrimination. Employers will well understand that investigations can no longer benefit them — at most, facts developed during an investigation will only be used against the employer. Even an employer with a supportive workforce will be unable to defend itself with beneficial employee testimony, lest it be accused of unproven coercion. Informal resolution, Congress' preferred course, will therefore become even more difficult. See West v. Gibson, 527 U.S. 212, 218-19, 119 S.Ct. 1906, 144 L.Ed.2d 196 (1999) (noting Congress's intention that Title VII claims would be resolved informally).

One is further left to wonder where the majority's new imagined-coercion-based rule comes from. Generally, the purportedly "coercive nature of the employer-employee relationship ... is insufficient to demonstrate that ... [employer-employee] interviews were improper." Slavinski v. Columbia Ass'n, Inc., No. CCB-08-890, 2011 WL 1310256, at *4 (D.Md. Mar. 30, 2011) (collecting cases); accord Maddock v. KB Homes, Inc., 248 F.R.D. 229, 237 (C.D.Cal.2007); McLaughlin v. Liberty Mut. Ins. Co., 224 F.R.D. 295, 298 (D.Mass.2004); cf. Gulf Oil Co. v. Bernard, 452 U.S. 89, 104, 101 S.Ct. 2193, 68 L.Ed.2d 693 (1981) ("[T]he mere possibility of abuses does not justify routine adoption of a communications ban[.]"). Certainly it cannot be found in the cases the majority cites, which all raised questions about defendants who contacted putative class plaintiffs after a class action had been filed. Here, Nucor investigated and obtained affidavits before any lawsuit was filed, so it could not have been attempting to break up the class — the class did not even exist yet. The majority's cases also involved a level of egregious misconduct not found in this case, suggesting that those cases were directed at a problem that does not exist here. See, e.g., Kleiner v. First Nat'l Bank of Atlanta, 751 F.2d 1193, 1197-98 (11th Cir.1985) (finding unilateral contacts improper where counsel violated direct court order and conducted a vast "selling job" seeking class opt-outs in "[s]ecrecy and haste" during "the district judge's vacation"); see also Burrow v. Sybaris Clubs Int'l, Inc., No. 13 C 2342, 2014 WL 5310525, at *4-5 (N.D.Ill. Oct. 17, 2014) (summarizing many of the same cases and concluding that they "depict[ed] communications so extreme that they actually cut against [the majority's present] position").

We also need not speculate about "potential" coercion, as the circumstances make plain that Nucor did not coerce its employees into making positive statements. No employee has claimed that the affidavits were coercive. No employee has suggested that Nucor retaliated against employees who complained of discrimination. And the contents of the affidavits do not imply coercion either. Employees evidently felt free to speak honestly, as the affidavits were not universally favorable to Nucor. See, e.g., J.A. 10950 (district court noting that the affidavits "actually bolstered the plaintiffs' claims of a common hostile work environment"). Some employees also chose not to give statements at all. See, e.g., J.A. 6911. And still other employees made handwritten corrections to their typed affidavits, indicating that the employees had complete control over their statements. See, e.g., J.A. 6120.

What is more, Nucor gave each employee a written notice explaining that the interview was voluntary, that the interviews were being taken on behalf of the company, that employees could decline to participate, and that they would not face any retaliation for what they said. See, e.g., J.A. 6003. In other contexts, the Court has said that disclosures like these prevent coercion. See, e.g., Overnite Transp. Co. v. NLRB, 280 F.3d 417, 434 (4th Cir.2002). Each employee who chose to participate then signed an acknowledgement and noted in his or her affidavit that Nucor did not coerce the employee. See, e.g., J.A. 6003.

The majority nevertheless condemns Nucor for not informing the employees that the company might use their statements in litigation. This novel requirement — a sort of "civil Miranda rule" — seems an odd one given that litigation had not been filed. Instead, interviewees were accurately informed that "[t]here ha[d] been a few charges of discrimination filed by African-American employees at Nucor," and the interview was meant to "determine what happened." J.A. 6003.

The district court did not clearly err in affording some weight to these many contrary affidavits.

6.

In addition to the affidavits supporting Nucor's view, Plaintiffs' affidavits must also be weighed against the company's announced anti-discrimination policy. In Wal-Mart, the Supreme Court found that a "general policy of discrimination" was harder to find given the company's "announced policy forbid[ding] ... discrimination and ... impos[ing] penalties for denials of equal opportunity." Id. at 2553. The same holds true here. Nucor is an equal-opportunity employer with an express anti-discrimination policy that harshly penalizes employees engaging in discriminatory conduct. Nucor policies even punish supervisors who fail to put an end to their subordinates' discriminatory conduct. The record also contains accounts of instances in which Nucor's general manager condemned discriminatory acts and punished employees for using offensive language. This countervailing evidence supports the district court's conclusion that, as a whole, the anecdotal evidence favored Nucor rather than Plaintiffs.

7.

a.

Aside from the qualitative and quantitative deficiencies in Plaintiffs' anecdotal evidence, it also does not tell a plant-wide story. In Wal-Mart, plaintiffs' anecdotal evidence failed in part because "[m]ore than half of the[] reports [we]re concentrated in only six States." 131 S.Ct. at 2556. As a result, even if one assumed that "every single one of these accounts [were] true, that would not demonstrate that the entire company operate[d] under a general policy of discrimination." Id.

The lack of dispersion that proved fatal to the class in Wal-Mart presents itself here. Eleven of the sixteen declarations — again, more than half — come from employees in a single department: the beam mill. No cold mill or maintenance employees are represented, while only one shipping employee and one melt shop employee appear. And as the district court recognized, when one examines the individual instances of discrimination alleged in Plaintiffs' declarations, most of them concern just one manager and three supervisors who all worked in the beam mill. See J.A. 10951. As one black employee put it, "Whatever [wa]s happening in the beam mill [wa]s not a plant wide problem." J.A. 6109.

b.

The majority somehow finds clear error in the district court's finding that Plaintiffs' accounts were concentrated in the beam mill. But it proves easy to see why the district court found what it did: Plaintiffs do not cite useful, relevant evidence from outside the beam mill. Some anecdotes fall outside the class period. See, e.g., J.A. 1085. Others involve promotions that did in fact go to a black employee. See, e.g., J.A. 1110-11. Some involve transfers, not promotions. See, e.g., J.A. 1063. Still others trace back to beam mill supervisors, not supervisors in other departments. See, e.g., J.A. 1079-80. Plaintiffs count six other instances twice. See Appellant's Br. 9-10. And some of the cited "instances of alleged promotions discrimination" amount to no evidence at all. See, e.g., id. at 9 (citing J.A. 7237 — an application for transfer — as one instance of "promotion discrimination"). Most incredibly, Plaintiffs' argument — which the majority appears to adopt — assumes that one can find evidence of discrimination in every single instance where a black employee does not receive a promotion for which he applies. That concept finds no support in any part of our jurisprudence. Indeed, it turns the Teamsters framework into a circular absurdity. Plaintiffs presume that each denied promotion evidences a discriminatory policy or practice, even though — under Teamsters — Plaintiffs must prove that a discriminatory policy or practice existed before the court may presume that a particular denied promotion was discriminatorily made. See Teamsters, 431 U.S. at 362, 97 S.Ct. 1843.

The district court recognized, as it should have, that the anecdotal evidence was more substantial when it came to the beam mill. For that reason, the district court explained that it was willing to certify a class of those applying out of and into the beam mill. J.A. 10953-54 & n.16. Plaintiffs never accepted the invitation, so they remain responsible for proving plant-wide commonality. That effort requires a substantial showing beyond a single department. See, e.g., Bennett, 656 F.3d at 816 (holding that the district court properly declined to certify a hostile work environment class where anecdotal evidence was concentrated in a single department).

Outside the beam mill, Plaintiffs at best present a few scattered anecdotes in each department. That's not enough. "[A] class plaintiff's attempt to prove the existence of ... a consistent practice within a given department[] may fail even though discrimination against one or two individuals has been proved." Cooper, 467 U.S. at 878, 104 S.Ct. 2794; accord Ste. Marie v. E. R.R. Ass'n, 650 F.2d 395, 406-07 (2d Cir.1981). The district court might very well have clearly erred had it accepted such evidence. One can hardly say that it clearly erred in doing just the opposite.

8.

In a last effort to save their class-wide claim, Plaintiffs make much of other facts that do not relate directly to promotions. They seem to give special attention to the facts underlying their already-certified hostile work environment claim. The majority agrees that such evidence provides a "cultural backdrop" that renders an "equitable promotions system" essentially impossible. Maj. op. at 912. Notably, that view never appeared in Brown I, but references to Plaintiffs' hostile work environment claims now appear at least a dozen times in the majority opinion. The majority also finds evidence of a "culture" in the alleged fact that Nucor hired only one black supervisor before the EEOC investigation, even though "[t]he mere absence of minority employees in upper-level positions does not suffice to prove [even] a prima facie case of discrimination without a comparison to the relevant labor pool." Carter, 33 F.3d at 457.

We have never held that class plaintiffs may establish a common, classwide policy of discrimination with mere evidence of company "culture." Other decisions, including Wal-Mart, reject the notion that "culture" is enough. See Wal-Mart, 131 S.Ct. at 2553; Davis, 717 F.3d at 487-88. The majority would nevertheless "sweep many individual plaintiffs and sets of facts into one class on the premise that all reflect illegal conduct by the defendant in practice and culture if not in policy" — even though that is "precisely the sort of class that the Supreme Court recently rejected in [Wal-Mart]." Jamie S. v. Milwaukee Pub. Schs., 668 F.3d 481, 504 (7th Cir. 2012) (Rovner, J., concurring in part). Furthermore, simply saying that a company has a "cultural problem" does not identify any particular employment policy or practice, McClain v. Lufkin Indus., Inc., 519 F.3d 264, 274 (5th Cir.2008), let alone a common, uniform policy spanning the class.

We have also never held that facts establishing a hostile work environment unavoidably relate to all other employment decisions made in the same company. Such a connection would be hard to justify, as acts giving rise to a hostile work environment are only distantly related to the discrete acts that underlie disparate treatment and impact claims. "The probative value of other discriminatory acts depends... on the nature of the discrimination charged." Hunter v. Allis-Chalmers Corp., Engine Div., 797 F.2d 1417, 1424 (7th Cir.1986), abrogated on other grounds by Patterson v. McLean Credit Union, 491 U.S. 164, 109 S.Ct. 2363, 105 L.Ed.2d 132 (1989). And "[h]ostile environment claims are different in kind from discrete acts." Nat'l R.R. Passenger Corp. v. Morgan, 536 U.S. 101, 115, 122 S.Ct. 2061, 153 L.Ed.2d 106 (2002). In contrast to acts creating a hostile work environment, discriminatory employment decisions "inflict[] direct economic harm." Burlington Indus., 524 U.S. at 762, 118 S.Ct. 2257. They will often require "the imprimatur of the enterprise and the use of its internal processes." Id.

The "probativeness" of items like comments, jokes, and other acts "is [also] circumscribed if they were made [or done] in a situation temporally remote from the date of the employment decision[s], or if they were not related to the employment decision[s] in question or were made by non-decisionmakers." McMillan v. Mass. Soc'y for Prevention of Cruelty to Animals, 140 F.3d 288, 301 (1st Cir.1998). Here, Plaintiffs' evidence suffers to some degree from all three of these defects. For instance, Plaintiffs' statements often do not tell us when the offensive conduct occurred, so we have no way of assessing temporal proximity. None of the "cultural" evidence pertains specifically to promotions. And most all of the relevant hostile-work-environment conduct came from non-decisionmakers, even though it "is the perception of the decisionmaker that is relevant" in claims like Plaintiffs'. Smith v. Flax, 618 F.2d 1062, 1067 (4th Cir.1980); accord Mateu-Anderegg v. Sch. Dist. of Whitefish Bay, 304 F.3d 618, 623 (7th Cir.2002) ("[S]tatements are only relevant if they come from a decisionmaker, someone involved in the adverse employment decision[s]."). Lastly, to the limited extent that supervisors did involve themselves in the incidents that Plaintiffs described, those supervisors chiefly worked in the beam mill — undermining any inference of a common, plant-wide policy.

At bottom, the majority concludes that we should permit Plaintiffs to pursue two class claims pertaining to promotions because they have successfully established their right to pursue a separate, distinguishable hostile-work-environment claim. Title VII does not work that way, and, rhetoric aside, the majority is unable to identify a single decision to support that kind of proposition. "In the law, the absence of precedent is no recommendation." Dukes v. Wal-Mart, Inc., 509 F.3d 1168, 1200 (9th Cir.2007) (Kleinfeld, J., dissenting). Moreover, to assume that a plaintiff establishes a right to class treatment for his discrete-act class merely because he has established such a right as to a hostile-work-environment class is to reinstate a suspect revision of the "across-the-board" rule that the Supreme Court rejected three decades ago. See Falcon, 457 U.S. at 153, 157-59, 102 S.Ct. 2364 (rejecting the idea that "an employee complaining of one employment practice" may automatically "represent another complaining of another practice" merely because both alleged discrimination based on the same protected trait). The district court did not abuse its discretion in refusing to exhume that long-dead idea.

The district court did not clearly err in declining to give dispositive weight to evidence going to Plaintiffs' hostile-work-environment claim when deciding whether to certify Plaintiffs' separate promotions-related classes.

* * *

When closely examined, Plaintiffs' anecdotal evidence proves to be just as unconvincing as their statistical proof. "Because [Plaintiffs] provide no convincing proof of a companywide discriminatory ... promotion policy, ... they have not established the existence of any common question." Wal-Mart, 131 S.Ct. at 2556-57. The district court therefore did not abuse its discretion in declining to certify the class because of its lack of commonality.

IV.

On the road to its desired result, the majority undermines well-established judicial processes, causes a rift between this Court and a co-equal circuit court without explanation, and brings substantial uncertainty to an area of law that begs for clarity.

As to judicial processes, the majority opinion evidences little respect for the role of the district court and the standard of review. The district court has lived with this matter for several years now, and it best understands how the case has developed. Its actions bespeak a court striving to scrupulously apply Rule 23's requirements. The district court complied with our mandate, rejected more than one request to decertify from Nucor, and continually endeavored to respect findings that this Court has (actually) made. Yet the majority shows no concern for that effort. And it shows just as little concern for this Court's well-established waiver rule, which should plainly apply here.

As to our sister circuits, the majority opinion begets a circuit split. The Eighth Circuit affirmed the denial of class certification in a case involving the same claims, the same experts, and the same defendant. As should be clear by now, that decision cannot be reconciled with this one. The majority never even tries to do so.

And as to cases to come, the majority's decision will offer far more questions than answers. What standard of review really applies in this context? How much evidence must a plaintiff summon to comply with Rule 23? Does appellate waiver matter? Does class treatment of one cause of action necessarily warrant class treatment for another? Must statistical evidence prove to be reliable? Does Wal-Mart reach only nationwide class actions? Can a sufficiently "common" policy result from inaction? These are only some of the questions that the majority opinion leaves unresolved.

We should hardly take this troubled road in the name of "simple justice." Maj. op. at 922. "`Simple justice' is achieved when a complex body of law developed over a period of years is evenhandedly applied." San Remo Hotel, L.P. v. City & Cnty. of San Fran., Cal., 545 U.S. 323, 345, 125 S.Ct. 2491, 162 L.Ed.2d 315 (2005). Evenhandedness is nowhere to be found here, so justice remains unserved.

Perhaps the Supreme Court will act to rectify the problems that are sure to follow from today's opinion. One can only hope that it will do so soon. In the meantime, I respectfully dissent. The district court did not abuse its discretion, and its judgment to decertify should be affirmed.