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 WalMart 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., Thom 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 *923reasons 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 appéllant 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 balls for a different result here.
B.
In view of their failure to raise the predominanee issue, Plaintiffs now suggest that “[p]redominanee and eommonality ... 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 *924Rule 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 conelude[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.
*925c.
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] ease, 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 *926majority’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 28 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!)]”). Nu-cor 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 “prejudiced” 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 *927Rule 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 ease 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), Broum 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 Broum 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 Broum 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(e)(1)(C); see also Fed.R.Civ.P. 54(b). “[Certifications 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 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 Broum I did the work, *928as “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 *929must 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 “[hjighly 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), reajfd 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 dear-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 re*930verse 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 “[standards 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 *931take -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, thé 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) (“[Generalized 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 Stan*932dards 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 plaintiffs cause of action.” Wal-Mart, 131 S.Ct. at 2551-52. Compare Brown I, 576 F.3d at 156 (citing Plisen 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 “im-depth assessment of the merits”), with Walr-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 non*933existent 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 (“[0]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.” Wok-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 *934class 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 Sen., 798 F.2d 679, 683 (4th Cir.1986). But see Warren v. Halstead Indus., Inc., 802 F.2d 746, 753 (4th Cir.1986) (“[Statistics 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, “[inferring 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 axiomatieally equate with legal signifi*935canee. See EEOC v. Fed. Reserve Bank of Richmond, 698 F.2d 633, 648 (4th Cir. 1983) (“[Statistical significance as measured by the standards of acceptable statistical principles will not necessarily be legally significan^.]”), 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. 80, 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. Pafetzold & 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,
*9362553; 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, ho 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.
“[Tjrial 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. Nu-cor’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 *937black 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 (Pros-kine), 11025 (Pope). Most of the forms fail *938to 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 reasonábly 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 composi- ' tion 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 thé 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.
*939Rather 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-ofl-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.
*940The 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 *941discipline 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.
*942By 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-KD-N, 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 treat*943ment — 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-depart- ■ ment 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 *944among 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 Nu-cor’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 *945value 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 Nu-cor’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.Sd 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 “[djisparate 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-*946wide 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^] 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-mc269 (GK), 306 F.R.D. 33, 52, 2014 WL 4378781, at *15 (D.D.C. Sept. 4, 2014). Were it otherwise, one could find a com*947mon 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-deci'sionmaker influences the ultimate decision-maker’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.
*9481.
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 *949small. 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 plaintiffs 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 [supervisores face.” J.A. 1111. These and other examples are not “cherry piek[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 conclu*950sion, and “[a] plaintiffs 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 decision-maker’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 decision-makers 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 *951employees 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 “[cjommon 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 coereion[.]” 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, *952Congress’ 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 *4r-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 state*953ments 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 Nu-cor,” 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 ... imposing] 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 dis*954crimination” 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 plaintiffs 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. Mane 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. *9552012) (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 “[hjostile 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 decisions], or if they were not related to the employment decisions] in question or were made by nondecisionmakers.” 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) (“[Statements are only relevant if they come from a decisionmaker, someone involved in the adverse employment decisionfs].”). 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 *956one 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 *957not abuse its discretion, and its judgment to decertify should be affirmed.