McReynolds v. Sodexho Marriott Services, Inc.

349 F. Supp. 2d 1, 2004 U.S. Dist. LEXIS 25516, 2004 WL 2943233

• Court: District Court, District of Columbia
• Decided: December 20, 2004
• Docket: CIV.A.01-0510 ESH
• Status: Published

Opinion

MEMORANDUM OPINION

HUVELLE, District Judge.

BACKGROUND

This is a class action alleging race discrimination in employment under Title VII of the Civil Rights Act of 1964, as amended, 42 U.S.C. § 2000e et seq. (“Title VII”), and 42 U.S.C. § 1981. Plaintiffs claim that defendant’s promotion practices throughout the company are discriminatory insofar as they relates to managerial positions at or above the level of an above-the-unit manager. Plaintiffs rely on both disparate treatment and disparate impact theories. In support of their claim, plaintiffs assert that most of the managerial slots are filled without the jobs ever being posted; when jobs are posted, preselection occurs that disadvantages African Americans because the decision-makers are primarily white; and that when there is competition for a managerial position, these decision-makers have unfettered discretion in making their decision because they are not given job-related or objective criteria, and there is no required documentation explaining their reasons, thereby rendering them unreviewable.

In McReynolds v. Sodexho, 208 F.R.D. 428 (D.D.C.2002), this Court certified a class pursuant to Fed.R.Civ.P. 23(b)(2) for liability purposes only. 1 After defendant failed to gain review of this decision by means of an interlocutory appeal (petition for leave to appeal denied, No. 02-8008, 2003 WL 22299806 (D.C.Cir. Feb. 25, 2003), cert. denied, 540 U.S. 818, 124 S.Ct. 84, 157 L.Ed.2d 36 (2003)), the parties engaged in almost two years of heated discovery, involving the well-known phenomenon of “statistical dueling” between two highly-paid experts, both of whom produced a continuing series of reports modifying, correcting or refining their statistical analyses. 208 F.R.D. at 434 (internal quotation marks omitted). This lengthy and often contentious process has now culminated in defendant’s filing of three motions seeking dismissal of plaintiffs’ case: (1) a motion for summary judgment; (2) a motion to decertify the class; and (3) a motion to exclude plaintiffs’ expert under Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579, 113 S.Ct. 2786, 125 L.Ed.2d 469 (1993). 2

In its motion for summary judgment, defendant raises three main challenges to plaintiffs’ case. First, with respect to the disparate treatment claim, it contends that plaintiffs have failed to make a prima facie case of a pattern or practice of company-wide discrimination with respect to promotions. In this regard, Sodexho asserts that the only proper mode of analysis, as both a factual and legal matter, is to disaggregate to the “RVP” (Regional Vice President) level. Defendant therefore argues that plaintiffs’ aggregated statistics must be rejected because statistically significant results occurred in only 9 of 155 purported RVPs. (Def.’s Mot. at 2.) Defendant also challenges plaintiffs’ nonstatistical evidence as being too limited and therefore not probative of company-wide discrimination. (Id.)

Second, defendant claims that even if plaintiffs have demonstrated a statistically significant disparity, it cannot be attributed to race. Defendant’s rebuttal consists of a double-barreled attack: (1) it offers its own multiple regression analysis by its statistical expert, Dr. Joan Haworth, to argue that any disparities in promotions are due to factors other than race; and (2) it argues that the analyses done by plaintiffs’ expert, Dr. Bernard Siskin, are inaccurate and lacking in probative value primarily because he failed to take account of the major variables of education and experience, which must, according to defendant, be controlled for both as a matter of law and fact. (Id.)

Third, defendant attacks plaintiffs’ disparate impact claim, arguing that plaintiffs have failed to identify a particular employment practice that caused significant racial disparities in promotions. 3 (Id. at 3.)

Before addressing these three arguments, the Court will first review the governing principles of law regarding a disparate treatment claim (Section 1(A)). It will then address defendant’s arguments attacking the validity of plaintiffs’ statistical evidence in support of their prima facie case (Section 1(B)) and their anecdotal evidence (Section 1(C)). Next, it will consider the rebuttal evidence presented by defendant from its own expert, as well as its criticisms of plaintiffs’ statistician (Section 1(D)). Finally, the arguments relating to the disparate impact claim will be discussed. (Section II(A)-(D).)

ANALYSIS

I. Disparate Treatment Claim

A. Legal Principles

Plaintiffs’ disparate treatment claim, raised under both Title VII and § 1981, finds its genesis in the Supreme Court’s decision of Int’l Bhd. of Teamsters v. United States, 431 U.S. 324, 97 S.Ct. 1843, 52 L.Ed.2d 396 (1977) and in Circuit precedent. See, e.g., Davis v. Califano, 613 F.2d 957 (D.C.Cir.1979); Segar v. Smith, 738 F.2d 1249 (D.C.Cir.1984); and Palmer v. Shultz, 815 F.2d 84 (D.C.Cir.1987). A disparate treatment claim alleges that the defendant intentionally based an employment decision on the race of the plaintiffs. See Teamsters, 431 U.S. at 335 n. 15, 97 S.Ct. 1843. It can involve an isolated incident of discrimination against an individual, or, as in this case, allegations of a “pattern or practice” of discrimination affecting an entire class of individuals. Id. Thus, to succeed on a pattern-or-practice claim, plaintiffs must prove more than isolated acts of discrimination; rather, they must establish that intentional discrimination was the defendant’s “standard operating procedure.” Id. at 336, 97 S.Ct. 1843.

To establish a prima facie pattern or practice case, “[t]he plaintiffs must, by statistical evidence, individual testimony, or a combination of the two, make a showing adequate to raise the inference that employment decisions were predicated on an illegal criterion.” McKenzie v. Sawyer, 684 F.2d 62, 71 (D.C.Cir.1982). But statistics alone can make out a prima facie case of discrimination. Palmer, 815 F.2d at 90. And while “[exactness is not required at the prima facie stage,” De Medina v. Reinhardt, 686 F.2d 997, 1009 (D.C.Cir.1982), plaintiffs must nonetheless demonstrate to the court’s satisfaction that their statistical comparisons are meaningful, Trout v. Lehman, 702 F.2d 1094, 1101 (D.C.Cir.1983), and in particular, plaintiffs’ statistics must compare the promotion rates of class members with the rates of similarly situated whites within the company. See Palmer, 815 F.2d at 91. If there is a statistically significant disparity (usually defined as at least 1.96 standard deviations), the evidence is sufficient to raise an inference of discrimination. Id. at 99.

If plaintiffs satisfy their prima facie requirement, “[t]he burden [of production] then shifts to the employer to defeat the prima facie showing ... by demonstrating that [plaintiffs’] proof is either inaccurate or insignificant.” Teamsters, 431 U.S. at 360, 97 S.Ct. 1843. The defendant may, for instance, show that the disparity resulted from some legitimate, nondiscriminatory factor, or alternatively, by challenging the statistical calculations upon which the inference of discrimination is premised by showing that the statistics are based on faulty data, flawed computations or improper methodologies. Palmer, 815 F.2d at 99. But “defendant cannot rebut statistical evidence by mere conjectures or assertions, without introducing evidence to support the contention that the missing factor can explain the disparities as a product of a legitimate, nondiscriminatory selection criterion.” Id. at 101 n. 13 (citing Bazemore v. Friday, 478 U.S. 385, 400, 106 S.Ct. 3000, 92 L.Ed.2d 315 (1986)) (Brennan, J, concurring in part, joined by all Justices). See also Segar, 738 F.2d at 1267-70 (explaining that the respective burdens on the parties will typically be much higher than is required in an individual disparate treatment case).

With these guiding principles in mind, the Court now turns to defendant’s arguments as to the statistical data relied on by plaintiffs to support their prima facie case.

B. Plaintiffs’ Prima Facie Showing Based on Statistical Evidence

1. Pools Analysis

The principal type of statistical analyses at issue with respect to plaintiffs’ prima facie case is referred to as a pools analysis. 4 This Circuit has held that a pools analysis, standing alone, that demonstrates a statistically significant adverse result for a protected class relative to what would be expected if discrimination were not in play, is sufficient circumstantial evidence to make out a prima facie pattern or practice Title VII claim. Palmer, 815 F.2d at 91 (citing Segar v. Smith, 738 F.2d at 1278). A pools analysis looks to pools of similarly situated employees to determine how the promotion success of a certain pool — here, African American would-be promotees — stacks up against that of a control group, such as all other would-be promotees. Of course, it is important that a statistician compare “similarly situated employees” and that the pools must be properly defined by controlling for a variety of factors. See Sec’y of Labor v. Keystone Coal Min. Corp., 151 F.3d 1096, 1109 (D.C.Cir.1998). In a promotion ease, such as this one, the organization in which the class members are “employee[s] is appropriately utilized as the relevant labor market,” and “it would be expected that, absent discriminatory promotion practices, the proportion of the protected group in each of the job classifications and grade levels would approximate the proportion of the protected group with the minimum necessary qualifications for promotion in the employer’s labor force as a whole.” Davis, 613 F.2d at 964-65 (citing Teamsters, 431 U.S. at 339-40 n. 20, 97 S.Ct. 1843).

Plaintiffs’ expert, Dr. Bernard Sis-kin, analyzed the data provided by Sodexho and identified highly-significant statistical disparities by race at the company-wide level. Using a pools analysis that did not control for unit, but that did consider division, Siskin found a i;aee effect during the class period of 5.3/4/ standard deviations, 5 which is far in excess of the 1.96 threshold. (See Pls.’ Ex. 160 [11/01 Siskin report] at ¶¶ 4, 6 & Table S-2). 6

Significantly, defendant’s expert also did a pools analysis that analyzed 7,018 promotions for the class period 7 and found a company-wide statistically significant disparity of 3.47 standard deviations. 8 (Def.’s Ex. 8 [10/01 Haworth report] at Table 18.)

2. Defendant’s Use of RVPs

To rebut this showing of a statistically significant disparity at the company-wide level, defendant argues that plaintiffs’ aggregation of the statistical evidence is, as a matter of fact and law, unsupportable, since the only proper approach is to analyze the data at the “RVP” level, and if the statistics are disaggregated to that level, only 9 of Sodexho’s approximately 155 RVPs (fewer than 6%) had statistically significant disparities in promotions of African Americans. (Def.’s Mot. at 5, citing Def. Ex. 7 [3/02 Haworth decl.] at ¶ 4 & at 4; Def. Ex. 8 [10/01 Haworth report] at 56.) See Cooper v. Fed. Reserve Bank of Richmond, 467 U.S. 867, 878, 104 S.Ct. 2794, 81 L.Ed.2d 718 (1984); see also Carpenter v. Boeing, No. 02-1019, 2004 WL 2661691 at *3, *5 (D.Kan. Feb. 24, 2004) (decertifying class for lack of commonality where plaintiffs’ statistical evidence showed that disparate impact was limited to a few “pockets”).

Given the parties’ starkly different approaches to the data, the Court must decide whether plaintiffs can sustain their burden of demonstrating a prima facie case based upon aggregated statistics. This dispute hinges on whether RVPs— Regional Vice Presidents — are, as defendant claims, Sodexho’s “principal decision-making units” that are “functionally and independently responsible for the vast majority of the promotion decisions at issue” (Def.’s Mot. at 2, 17), or whether, as plaintiffs contend, they have no operational significance, they do not function independently with separate promotion procedures, and they are nothing but a “hoax.” (Pls.’ Opp. at 10-13.)

Before the Court can take on the issue of whether disaggregation, as defendant suggests, or aggregation, as plaintiffs argue, should be the operative mode of analysis, it is necessary to grapple with the meaning of RVPs. As even defense counsel conceded at the hearing on this motion, the concept is “admittedly confusing,” since an RVP is not necessarily synonymous with a particular geographic location or region, 9 does not necessarily equate with a particular person who holds this title, and does not represent the actual decision-maker involved in promotions. (Tr. at 35-37, 56-58.) Also, while it appears undisputed that RVPs are an accounting code in the company’s MARRP-AY database, they do not otherwise appear in any corporate organizational chart or corporate documents describing Sodexho’s organizational structure. 10 Thus, the Court turns to Haworth’s explanation, since according to defense counsel, she determined that RVP was a proper functional unit based on her interviews of company personnel. (Tr. at 68.) In her most recent May 17, 2004 declaration (Def.’s Ex. 70), she sets out her explanation of this term in ¶¶ 3-5 and her rationale for why this subset should be used for analyzing the data. There, she states in part:

[We] relied on the RVP .code (number) because it consistently defines the specific portion of the corporate and financial reporting structure at Sodexho to which a unit has been assigned. For example, there are RVP areas defined in each division which separate the operations activities from the functional activities, such as human resources and finance. Some units are operations units and assigned to RVP numbers reflecting the operations activities in a particular reporting structure. Other units contain employees serving functional (or support) activities with RVP numbers contained in the MARRPAY database for those functional areas.

* * * * * *

The Vice President of a region may oversee a group of operational units or a group of functional units {e.g., human resources, strategy, etc.) within the division. In some cases, a Vice President oversees units that provide a particular product line, and another Vice President oversees units in the same geographic area which provide a different product line in the same division {e.g., facilities management for one RVP and food services for another RVP in the same division and possibly same geographic area).

(Id. at ¶¶ 3, 4.)

According to defendant, Haworth’s conclusion that RVPs reflect the functional and operational structure of the company is based on her interviews of five current Vice Presidents of Finance (see Def.’s Ex. 70 [5/04 Haworth decl.] at ¶ 3 n. 4), and therefore, the disaggregation of statistical analyses at the RVP level is factually supportable. To further bolster its argument for RVPs, defendant points to various tables prepared by Haworth that purport to prove that 80% of managers, both white and African American, who are promoted stay within their original RVP, and argues that given the absence of evidence that there is substantial mobility among RVPs, the only proper methodological approach is the one offered by Haworth. (See Def.’s Ex. 1 [7/03 Haworth report] at Tables 12-15 and 20.)

In response to defendant’s assertions that RVPs are factually supportable and thus are the only proper unit for purposes of a pools analysis, plaintiffs have offered more than enough contrary evidence to raise a material issue of fact regarding the underlying premise of Ha-worth’s RVP analysis. (See Pls.’ Opp. at 9-13; Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 5-18; Pls.’ Stmt of Disputed Facts ¶ 26; Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶¶ 1-2.) For instance, Siskin also interviewed three former or current corporate employees, and their statements contradict the company’s position that RVPs reflect Sodexho’s functional and operational structure. (Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 7-10.) Plaintiffs also note that two defense experts, when detailing defendant’s organizational structure in their reports, fail to even mention RVPs. (Pls.’ Ex. 164 [Bloom report] at ch. 2 (“Corporate History and Structure”); Def.’s Ex. 41 [Bier- maim report] at § III (“The Sodexho Corporate Structure”)-)

Similarly, with respect to the ostensible lack of mobility across RVPs, plaintiffs point to statistics that at least raise a factual issue as to this assertion. For instance, the actual promotion data demonstrates that those who were in fact promoted frequently move across RVP codes. 11 Moreover, it may be argued that Tables 12-15 of Haworth’s 7/03 report (Def.’s Ex. 1) provide little support for defendant’s argument regarding the usefulness of using RVPs to analyze the data, since these charts provide little, if any, information specifically about promotions. 12 By contrast, Table 20 does speak to the issue of promotional activity within a subgroup and shows that a similar percentage of African Americans and other employees did not move out of their division or RVP following a promotion. This statistic does not, however, mean that disaggregation at the RVP level is required, especially given plaintiffs’ contention that a similar subjective decision-making process pervades the entire company. For, as argued by plaintiffs, whether managers actually move from one division or RVP to another “is not material to a lawsuit in which plaintiffs allege both discrimination in promotions decisions and segregation. Because the promotion decisions may reflect the discrimination inherent in the promotion process at issue in this case, studying the resulting promotions may mask the discrimination and segregation that occurs at Sodexho.” (Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶ 3, citing Pls.’ Ex. 160 [11/01 Siskin report] at ¶ 14.) Thus, as argued by plaintiffs, the relevant focus is not on whether people actually move, but whether candidates sought promotions outside their RVP or division. (Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶ 3.) And, in this regard, plaintiffs point to the fact that Haworth’s data shows that under the posting system in effect from March 1997 to September 2000, 24.9% of all applicants bid for promotions in different divisions (31.9% of African American applicants bid for promotions in different divisions), and under the subsequent Career Center posting system from June 2000 to June 2001, 37.1% of all applicants bid for promotions in different divisions (38.5% of African American applicants bid for promotions in different divisions). 13 (Def.’s Ex. 1 [7/03 Haworth report] at Table 18.) Thus, as recognized in this Court’s certification decision, given the figures regarding the percentage of applicants who come from outside a division (which is obviously far larger than an EVP), “defendant’s statistics about the lack of promotions among divisions may actually support plaintiffs’ allegations of preselection for open positions.” McReynolds, 208 F.R.D. at 431 n. 6.

As this discussion reveals, whether RVPs have a functional or operational role within the company and whether they have any factual relevance to promotion decisions, is a hotly contested issue, and it therefore necessarily follows that whether the data should be aggregated or disaggregated is an issue properly left to the jury at trial. For “[w]hen, as here, the parties’ experts rely on conflicting sets of facts, it is not the role of the trial court to evaluate the correctness of facts underlying one expert’s testimony.” Micro Chemical, Inc. v. Lextron, Inc., 317 F.3d 1387, 1392 (Fed.Cir.2003) (holding that the role of the court is limited to determining whether an expert’s methodology and principles are proper); rather, that determination is the province of the finder of fact. See Pipi tone v. Biomatrix, Inc., 288 F.3d 239, 249-50 (5th Cir.2002) (holding that factfinder was entitled to hear expert testimony and decide whether to accept or reject it after considering whether predicate facts on which expert relied were accurate). See also Mozee v. Am. Comm. Marine Service Co., 940 F.2d 1036, 1045 (7th Cir.1991) (“the trial judge’s selection [as factfinder at trial] of the appropriate variables required to determine a suitable pool for comparison purposes is generally a question of fact....”); Carpenter, 2004 WL 2661691 at *5 (decertifying a class after finding “no evidence that Defendant somehow gerrymandered the groupings to isolate adversely impacted employees. The sites and job aggregation groups at issue are found in Defendant’s actual operations and, in contrast to cases cited by Plaintiffs, Plaintiffs’ own expert used these same categories in generating his report.”) (emphasis added); Eng’r Contractors Ass’n of South Fl. Inc. v. Metro. Dade County, 122 F.3d 895, 919 (11th Cir.1997) (the trier of fact is entitled to determine how much weight to give to aggregated or disaggregated statistics); Butler v. Home Depot, Inc., No. C-94-4335, 1997 WL 605754, at *7-9, 1997 U.S. Dist. LEXIS 16296, at *27-34 (N.D.Cal. Aug. 29, 1997) (declining to grant summary judgment in a class-wide promotions case because of disputed issues of fact). 14

3. The Law Regarding Aggregation Versus Disaggregation

Proceeding from the notion that only 6% of the RVPs show statistically significant results, defendant argues that this case is comparable to the many decisions where even more impressive statistical disparities were rejected in the context of a class decertification or summary judgment motion. (See Def.’s Mot. at 12-15.) But the factual dispute relating to the validity of using RVPs precludes any such argument. Moreover, as recognized by this Circuit:

Because the fate of disparate treatment claims is heavily dependent on the setting, facts, and circumstances of the particular case, decisions in such cases have limited precedential value.

Valentino, 674 F.2d at 69 (internal citation and quotation omitted).

Even a cursory review of the cases relied on by defendant shows that courts that have found that plaintiffs’ pools or their statistics were not supportable made this finding based on the existence of units that are clearly operationally and functionally distinct (unlike defendant’s RVPs), such as certain manufacturing plants, see Beck v. Boeing Co., 203 F.R.D. 459, 464 (W.D.Wash.2001); gas stations, see Hill v. Amoco Oil Co., No. 97-C7501, 2003 WL 262424, at *3, 2003 U.S. Dist. LEXIS 1795, at *9-10 (N.D.Ill. Jan. 27, 2003); 15 geographical districts for a delivery company, see Morgan v. United Parcel Service of America, Inc., 143 F.Supp.2d 1143, 1146, 1148 (E.D.Mo.2000), aff'd, 380 F.3d 459, 471 (8th Cir.2004); or divisional sites, see Carpenter, 2004 WL 2661691 at *1, *2 n. 2, *5 (finding separate job sites and 25 job aggregation groups). See also Cooper v. Southern Co., 390 F.3d 695 (11th Cir.2004) (certification denied where employees worked in widely diverse job types spread throughout different facilities, job locations and four independent subsidiaries); Grosz v. Boeing Co., No. SACV02-71, 2003 WL 22971025, 2003 U.S. Dist. LEXIS 25341 (C.D.Cal. Nov. 7, 2003) (facilities in five locations belonging to separate business units and engaging in different lines of business); and Moore v. Boeing Co., No. 4:02CV80, 2004 U.S. Dist. LEXIS 5959 (E.D.Mo. Mar. 31, 2004) (four facilities, including four independent business units and numerous job aggregation groups). 16

But more importantly, to sanction the disaggregation of data in this case, especially given the disputed issues of fact, would necessarily mean that plaintiffs could not satisfy their prima facie burden of showing the type of pattern or practice ease that was recognized by the Supreme Court in General Telephone Co. v. Falcon, 457 U.S. 147, 159 n. 15, 102 S.Ct. 2364, 72 L.Ed.2d 740 (1982) (“Significant proof that an employer operated under a general policy of discrimination conceivably could justify a class of both applicants and employees if the discrimination manifested itself in hiring and promotion practices in the same general fashion, such as through entirely subjective decisionmaking processes”) See also Wagner v. Taylor, 836 F.2d 578, 589 (D.C.Cir.1987). For, as recognized in both the case law and in statistical treatises, the more that data is disaggregated, the more difficult it is to demonstrate statistical significance. In this regard, the Fifth Circuit’s analysis is particularly instructive:

In our view, [defendant’s attempt to break down the data by city or year or both] was an unfair and obvious attempt to disaggregate that data to the point where it was difficult to demonstrate statistical significance. By fragmenting the data into small sample groups, the statistical tests became less probative. Indeed, it became impossible in many instances, since even a record of hiring or promoting zero women would not yield statistically significant results.... Aggregating the data as the plaintiff did was a much more reasonable approach, since liability under Title VII depends on whether the EEOC demonstrated a “systemwide pattern or practice” of disparate treatment, rather than “the occurrence of isolated or ‘accidental’ or sporadic discriminatory acts. It had to establish ... [that] discrimination was the company’s standard operating procedure — the regular rather than the unusual practice.”

Capaci v. Katz & Besthoff, Inc., 711 F.2d 647, 654-56 (5th Cir.1983) (quoting Teamsters, 431 U.S. at 336, 97 S.Ct. 1843) (other citation and footnote omitted) (emphasis in original). See also Shafer v. Commander, Army & Air Force Exch. Serv., 667 F.Supp. 414, 427-28 (N.D.Tex.1985).

Moreover, if an expert isolates units or groups and runs separate analyses for them, such methodology may mask whether “the overall decision-making process” produces a discriminatory result, whereas analyzing an entire group will indicate whether the identified employment practice was the cause of the disparity. Smith v. Xerox, 196 F.3d 358, 368-69 (2d Cir.1999) (discussing disparate impact). For, as noted by the Seventh Circuit:

Pooling data is sometimes not only appropriate but necessary, since statistical significance becomes harder to attain as the sample size shrinks. See D. Baldus & J. Cole, Statistical Proof of Discrimination § 9.221, at 309 (1980) (“All other things being equal, the test statistic and level of significance rise as the sample size increases.”); Fisher, Multiple Regression in Legal Proceedings, 80 Colum. L.Rev. 702, 717 (1980) (“In small samples, t-statistics must be larger for a given significance level.”).

Coates v. Johnson & Johnson, 756 F.2d 524, 541 (7th Cir.1985). See also Segar, 738 F.2d at 1286 (affirming the district court’s discounting of defendant’s rebuttal statistics that were excessively disaggregated). In sum, “the plaintiff should not be required to disaggregate the data into subgroups which are smaller than the groups which may be presumed to have been similarly situated and affected by common policies,” for where “plaintiffs’ theory is that the employment practices have identical discriminatory effect upon members of all minority groups, ... [r]ight or wrong, they are entitled to prove their case.” Paige v. State of California, 291 F.3d 1141, 1148-49 (9th Cir.2002) (quoting Baldus & Cole §§ 7.0-7.2) (other citation and internal quotation marks omitted).

These words of caution are particularly appropriate in this case. As explained by Siskin (Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 19-25), disaggregating data to the RVP level could mask the fact of discrimination because many of the RVPs are simply too small or have too few African Americans to yield any statistically significant results. 17 The validity of this argument is borne out by Haworth’s admission that “approximately 40% of all operating units at Sodex-ho are one person units and more than 60% of the operating units have two or fewer employees; [o]nly 25% of the units have four or more employees,” (Def.’s Ex. 1 [7/03 Haworth report] at xii; see also id. at 9 n. 7), as well as her recognition that “[ojbviously, this distribution of unit size does affect the number of units that might have an African-American salaried employee when they are about 12% of the salaried exempt workforce during the time period at issue.” (Id. at xii.) It is also confirmed by Siskin who, when he repeated Haworth’s studies but restricted them to RVP codes in which promotions exist and in which African American promotions are possible, found that 73.7% (84 out of 114) of the RVP codes show a disparity (although not a statistically significant one) adverse to African Americans. (Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶ 24.)

Finally, the Court is unwilling to accept defendant’s invitation to rule as a matter of law that plaintiffs cannot aggregate statistics on a company-wide basis based on the Fourth Circuit’s ruling in Stastny v. Southern Bell Tel. & Tel. Co., 628 F.2d 267, 279 (4th Cir.1980). (Def.’s Reply at 7.) In that case, the court rejected the aggregation of statistics covering defendant’s facilities statewide in the absence of a “showing of the extent to which, if at all, the overall disparities were paralleled in the separate facilities or even a statistically reliable sample of them.” Id. The court found fault in aggregation where there was “no showing by direct evidence that a statewide practice or policy, either objectively or subjectively administered, existed.” Id. There was neither evidence that spoke to “the locus of autonomy in making the challenged employment decisions, [nor] ... whether the management positions in question were filled from an essentially statewide work force labor pool.” Id. Because of these deficiencies, the court held that it was impossible to infer from the statistically significant statewide disparities that discrimination was common to defendant’s numerous “presumably autonomous” facilities throughout the state. Id. at 279-80.

To the extent that Stastny may be read so broadly as to foreclose a claim of a company-wide pattern or practice case based on decentralized, 18 subjective deci-sionmaking, 19 this would clearly be contrary to the Supreme Court’s dicta in Falcon (which was decided after Stastny) and to precedents in this Circuit, as well as others, which have recognized the validity of Falcon’s, at 159 n. 15, 102 S.Ct. 2364. See, e.g., Wagner, 836 F.2d at 589; Hartman v. Duffey, 19 F.3d 1459, 1471 (D.C.Cir.1994); Caridad, 191 F.3d at 292; Griffin v. Dugger, 823 F.2d 1476, 1487, 1490-91 n. 22 (11th Cir.1987). Moreover, the very argument defendant raises here was addressed by this Court in McReynolds:

[Defendant's view of the commonality test would preclude class certification in an action against any company that has decentralized its personnel practices. Sodexho contends that a class cannot be certified “without identifying any centrally unifying general policy of discrimination,” and that there can be no such policy when decision-making procedures are decentralized. This position would permit companies to escape Title VII class actions by minimizing the amount of control they exercise over individual managers. Such a holding would run afoul of the purpose of Title VII, which is “not to provide redress but to avoid harm,” by encouraging employers to adopt antidiscrimination policies and to educate their personnel on Title VII’s prohibitions.

208 F.R.D. at 433 (internal citation omitted). See also Mozee, 940 F.2d at 1050.

For these same reasons, the Court rejects defendant’s argument that “unless the actual decision-making process is national, aggregation is impermissible irrespective of the potential labor pool.” (Def.’s Reply at 7.)

C. Plaintiffs’ Showing Based on Nonstatistical Evidence

In addition to evidence of statistical disparities between African Americans and whites with respect to promotions, plaintiffs rely on anecdotal evidence, as well as other evidence, to bolster their prima facie case. While such anecdotes bring “the cold numbers convincingly to life,” Teamsters, 431 U.S. at 339, 97 S.Ct. 1843, and “provide!] ‘texture’ to the statistics,” Robinson v. Metro-North Commuter R.R. Co., 267 F.3d 147, 168 (2d Cir.2001), plaintiffs may sustain their burden at the prima facie stage exclusively on statistical evidence, for “no sound policy reason exists for subjecting the plaintiff to the additional requirement of either providing anecdotal evidence or showing gross disparities.” Segar, 738 F.2d at 1278; see also Wagner, 836 F.2d at 592. Given this lack of need for anecdotal evidence, much less need be said on this topic.

Defendant challenges plaintiffs’ anecdotes on the basis that plaintiffs have only-offered a limited number of racially offensive comments, their anecdotes relate only to a limited number of RVPs, and many of the comments fall outside the class period. (Def.’s Mot. at 19-21.) Plaintiffs vigorously contest these assertions (see Pls.’ Stmt. of Disputed Facts ¶ 13; Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶¶ 18-19), but more importantly, it is legally irrelevant if some of the comments precede the class period, see McReynolds, 208 F.R.D. at 434 n. 8, 436 n. 16 (citing Bazemore, 478 U.S. at 396, 106 S.Ct. 3000); Mozee, 940 F.2d at 1043 n. 7 (citing United Air Lines, Inc. v. Evans, 431 U.S. 553, 558, 97 S.Ct. 1885, 52 L.Ed.2d 571 (1977)); Paige, 291 F.3d at 1149, or if plaintiffs have failed to offer anecdotes pertaining to many areas of the company. 20 Moreover, with respect to defendant’s claim that plaintiffs’ anecdotes are too limited to show a nationwide problem, plaintiffs have cited numerous statements by Sodexho employees and executives, as well as corporate documents, indicating an awareness of a nationwide problem. 21 (See generally Pls.’ Stmt. of Disputed Facts ¶¶ 10-16.) See also Palmer, 815 F.2d at 102 (noting the admissions of an Undersecretary of State regarding the fact the Foreign Service “had a long way to go” to correct biased attitudes). While the Court need not repeat the many sources cited by plaintiffs in their Statement of Disputed Facts, suffice it to say that this evidence reflects policies and the state of affairs throughout the company, rather than at one individual unit, and it thereby lends powerful support to plaintiffs’ statistical showing of significantly low levels of African American promotions. 22

Defendant additionally argues that plaintiffs have not adequately supported their theory that decisionmaking was decentralized and subjective. As a result of this alleged lack of proof, defendant contends that statistical evidence of company-wide racial disparities is irrelevant. (Def.’s Mot. at 34-35). To support this claim, defendant cites the two-page declarations of nine human resources officials representing three of Sodexho’s six divisions, which, in substantially verbatim language, assert that Sodexho has a “structured” process for promotion decisions, all job openings are posted on the internal company-wide job database and interview questions are available to “provide guidance” to promotion decision-makers. They describe what is “generally” done, while failing to state unequivocally what, if any, procedures are actually required. (See, e.g., Def.’s Ex. 32.) In addition, some of the procedures appear to have been instituted after the class period and are thus irrelevant to the issue of liability. (Id.) And although each declaration attaches sample questions and rating sheets, only one of these sheets is actually filled out (Def.’s Ex. 31), notwithstanding plaintiffs’ discovery request three years ago for such materials.

These declarations stand in stark contrast to numerous negative responses from top Sodexho officials to a 2001 email from defendant’s in-house counsel seeking “rating or ranking sheets” and “forms for interviewers to record interview notes or impressions.” (Pls.’ Ex. 116.) They are further rebutted by the findings of plaintiffs’ industrial psychologist expert, Dr. Erich Prien, who reported that most managers contacted had no such documents. (Pls.’ Ex. 163 at ¶ 14.) They also conflict with defense expert’s analysis of Sodexho’s internal data, which showed that only one quarter of Sodexho promotions were ever posted on the Marriott Center Management System (“MOMS”) or the subsequent Career Center system (Def.’s Ex. 1 [7/03 Haworth report] at ¶ 101), as well as with the company’s internal documentation of serious problems with its posting system. (See, e.g., Pls.’ Ex. 86 (Q & A’s from the First Open Forum, Sodexho Marriott Services Flash-The Integration at SDH13386 (Oct. 6, 1998) (Statement of Randy Harris, Senior V.P. for Human Resources)).) And, of course, they cannot be reconciled with the testimony of numerous company employees who assert that, during many years of employment at Sodexho, the promotions process they encountered was neither structured nor guided by documents or requirements imposed by some central office. (See, e.g., Pls.’ Ex. 98 [Cipollini decl.] at ¶¶ 5-6; see also Pls.’ Opp. at 6-9.)

But perhaps most tellingly, these declarations are contradicted by Sodexho’s own Rule 30(b)(6) witness on its promotion pro cess, Corporate Services Division Director of Staffing William Anstee, who testified that:

[T]he hiring manager has complete discretion to decide whom to interview for each position. There are no company-rules or guidelines regarding the promotion process, and there are no requirements to take any notes during the interview process, to rank applicants, or to explain the selection. After a hiring manager makes his selection, he ‘would take that decision to whoever their boss was [and] say, hey, I’ve selected candidate A .... And then the district manager [puts] together an offer if that, indeed, was what it was.’ In short, So-dexho has ‘no company guidelines or fixed criteria or anything ... to follow in making those decisions.’

McReynolds, 208 F.R.D. at 432-33 (citing Anstee dep. at 81, 90-92, 182).

Given the inconsistencies in Sodexho’s own evidence, and in light of the substantial testimony offered by plaintiffs, the Court must, for summary judgment purposes, credit plaintiffs’ contention that So-dexho used a decentralized, subjective promotions process. The Supreme Court has made clear that “Title VII prohibits discriminatory employment practices, not an abstract policy of discrimination.” Falcon, 457 U.S. at 159 n. 15, 102 S.Ct. 2364. “ ‘[E]ntirely subjective decisionmaking processes,’ rather than ... entirely subjective hiring criteria,” McReynolds, 208 F.R.D. at 442 (quoting id.), is all that is required, in a case such as this one, to make out a pattern or practice Title VII claim. For, even if the Court were to credit defendant’s declarations, the fact that some Sodexho depai'tments may have had proposed interview questions, forms or other procedures that purportedly created a “structured” process does not suffice at this stage to overcome the mountain of evidence showing that Sodexho’s decision-makers exercised unfettered discretion in determining whom to interview and what questions to ask.

Further, the cases defendant cites for the proposition that “excessive subjectivity” does not constitute a company-wide practice or policy, see Grosz, 2003 WL 22971025, 2003 U.S. Dist. LEXIS 25341; Sperling v. Hoffmann-La Roche, Inc., 924 F.Supp. 1346, 1361 (D.N.J.1996) (see Def.’s Mot. at 34-35), are at odds with binding precedent in this Circuit, which embraced the theory underlying this suit twenty-five years ago:

Appellant’s statistical Prima facie case is bolstered by the subjective and ad hoc nature of Appellee’s promotion decisions. No objective criteria were established to guide the promotion decisions of supervisors, branch chiefs and ad hoc promotion panels, who were predominantly male. This Court agrees with the Eighth Circuit Court of Appeals in Rogers v. International Paper Co., ... which stated:

Greater possibilities for abuse ... are inherent in subjective definitions of employment selection and promotion criteria. Yet they are not to be condemned as unlawful per se, for in all fairness to applicants and employers alike, decisions about hiring and promotion in supervisory and managerial jobs cannot realistically be made using objective standards alone. Thus, it is especially important for courts to be sensitive to possible bias in the hiring and promotion process arising from such subjective definition of employment criteria.

Appellee’s promotion procedures are highly suspect and must be closely scrutinized because of their capacity for masking unlawful bias. The “lack of meaningful standards to guide the pro motion decision, whereby there is some assurance of objectivity ... encourage(s) and foster(s) discrimination.”

Davis, 613 F.2d at 965-66 (internal citation and footnotes omitted).

For these reasons, the Court finds that plaintiffs have offered sufficient evidence to carry their burden of establishing a prima facie case of a pattern or practice discrimination claim under Title VII and § 1981. It will now turn to Sodexho’s rebuttal evidence.

D. Defendant’s Rebuttal Evidence

The burden having now shifted to defendant to rebut the inference of discrimination created by plaintiffs’ prima facie case, defendant follows the familiar pattern by arguing that its statistical evidence shows that explanations other than race account for the statistical disparity, and plaintiffs’ statistical evidence is fundamentally flawed because of their expert’s failure to properly control for variables that would show that there are nondiscriminatory reasons for the disparity.

1. The Law Regarding Multivariable Regression Analyses

In rebutting plaintiffs’ prima facie case, which is generally based on a pools analysis, defendants typically rely on a multivariable regression analysis, which is a wholly different statistical methodology. 23 A regression can account for more factors simultaneously than a pools analysis. Baldus & Cole § 8.022[1], at 247. Hence, regressions are better suited to consider “major variables,” which in turn may help to explain the decision-making process. The importance of including the major variables is clear; failure to do so may mask the true cause of a statistical disparity between races. “The mathematical technique measures the probability that the calculated disparity could occur randomly — but the analysis in no way validates the calculation of the disparity itself. If the tested disparity is based on erroneous assumptions or suffers from flaws in the underlying data, then standard deviation analysis is foredoomed to yield an equally faulty result.” Maddox v. Claytor, 764 F.2d 1539, 1552 (11th Cir.1985). See also Keystone Coal, 151 F.3d at 1109 (“[Statistics must reasonably control for a variety of factors to properly define similarly situated employees, and in any event may be counterbalanced by evidence providing an alternate explanation of the pattern or of the particular action in question. The weight given to statistical evidence in such cases is not absolute, but depends on the degree to which it rules out legitimate explanations and how the statistics factor into the balance with the other available evidence.”) Therefore, analyses that exclude major factors may be “so incomplete as to be inadmissible as irrelevant.” Coward, 140 F.3d at 274 (quoting Bazemore, 478 U.S. at 400 n. 10, 106 S.Ct. 3000).

But what constitutes a major variable for purposes of any given analysis will vary depending on the facts and theory of the case. Coward, 140 F.3d at 274; see also id. at 276-77 (Sentelle, J., concurring). While “[t]he plaintiffs’ analysis must take into account the major variables one might expect to cause a statistical disparity,” Mozee, 940 F.2d at 1047, there is a danger that “the ‘other factor’ [may] be something that is a euphemism or proxy” for discrimination. Adams v. Ameritech Serv., Inc., 231 F.3d 414, 428 (7th Cir.2000). Because of this delicate balancing act, courts have erred on the side of caution in deeming inadmissible regressions to be the exception:

While the omission of variables from a regression analysis may render the analysis less probative than it otherwise might be, it can hardly be said, absent some other infirmity, that an analysis which accounts for the major factors must be considered unacceptable as evidence of discrimination. Normally, failure to include variables will affect the analysis’ probativeness, not its admissibility.

Bazemore, 478 U.S. at 400, 106 S.Ct. 3000 (citations and quotation marks omitted). See also Koger v. Reno, 98 F.3d 631, 637 (D.C.Cir.1996) (the “omission of variables from a regression analysis” does not make the analysis “unacceptable as evidence” but courts need not “aceept[ ] regressions from which clearly major variables have been omitted”). In its most basic form, the inquiry is whether the statistical analysis is reliable enough to make “the existence of any fact that is of consequence more or less probable.” Adams, 231 F.3d at 425 (quoting Fed.R.Evid. 401) (internal quotation marks omitted).

Moreover, even if a regression contains the necessary major variables, it may nonetheless carry inadequate weight to make a party’s case, depending on the analysis’ underlying data, methodology and results. See Koger, 98 F.3d at 637 (defective statistical regression analysis was not probative evidence of disparate treatment); De Medina, 686 F.2d at 1010 (“Certain defects in statistical evidence may, of course, be fatal to a plaintiffs case ....”); EEOC v. Sears, Roebuck & Co., 839 F.2d 302, 327 (7th Cir.1988) (“The court recognized that these characteristics are not easily quantified and this data would not be generally available from the application forms, but found that to the extent the EEOC’s regression analyses did not incorporate these factors, they were entitled to less weight. We cannot say that this finding, going to the probative value rather than the relevancy of the regression analy-ses, is clearly erroneous.”); see also Valentino, 674 F.2d at 70-71 (deficiencies in expert’s analysis renders it inadequate to raise an inference of discrimination).

But material questions of fact may complicate the determination of what constitutes a “major variable.” Cf. Coward, 140 F.3d at 274 (“Depending on the theory of the case, some variables may be entirely unsuitable.”); Mozee, 940 F.2d at 1048 (vacating and remanding for resolution of a factual dispute involving whether a major factor accounted for a disparate impact). Thus, if a party claims that a certain factor had to be included in a regression, but that factor’s necessity remains in dispute, summary judgment may not be appropriate.

2. The Parties’ Regression Analyses

Defendant relies on Haworth’s regression analyses to argue that it has conclusively demonstrated nondiscriminatory explanations for perceived racial disparities so as to preclude any finding of a pattern or practice of discrimination. 24 (Def.’s Mot. at 22.) Her methodology can best be summarized by reference to defendant’s motion. (See Def.’s Mot. at 24-25 & n. 7-8.) According to defendant, Haworth gathered information from personnel folders, applications for posted positions and succession planning documents for a stratified sample of Sodexho employees randomly selected across division, salary, grade band ranges and race groups. 25 Her sample included approximately 4,000 of approximately 20,500 salaried exempt employees, of whom 3,800 were African American or white. (Def.’s Mot. at 25 & n. 7-8.) She then coded for education, considering highest level of education and field of study, and coded for job experience based on occupations defined by the Census Bureau, “capturing experience at all prior positions, as well as the nature of the industry of employment.” (Id. at n. 8.) Certification and licensing information was gathered, and Sodexho tenure was measured by distinguishing between seniority in salaried exempt and hourly employees. (Id.) With this information, she found, af ter controlling for education, job experience prior industry, licensing and memberships, tenure, division, RVP, and grade/ band, that no statistically significant patterns emerged that were adverse to African Americans. (Def.’s Mot. at 24-25, citing Def.’s Ex. 1 [7/03 Haworth report].) And to blunt any debate over the meaning of promotion (see note 7 supra), Haworth re-ran her regression analysis using Sis-kin’s definition of promotion, but still found no statistically significant effect of race. (Def.’s Ex. 12 [10/03 Haworth report] at Table 2.) Based on these findings, defendant argues it is entitled to summary judgment, for “[w]hen a defendant conducts its own regression analysis that measures key variables that plaintiffs’ expert had omitted, and where that refined regression demonstrates only statistically insignificant racial disparities, the defendant has more than met its burden of disproving a pattern or practice of discrimination under Teamsters.” (Def.’s Mot. at 25.)

While this is admittedly an accurate statement of the law, it remains to be decided: (1) whether defendant’s analyses did “measure key variables,” or, as plaintiffs argue, they included “tainted” or unnecessary variables that are not required for a valid regression analysis and that the inclusion of these variables affected the results; and (2) whether plaintiffs’ regression analyses omitted key variables and that this omission renders their analyses either inadmissible under Daubert or insufficient to stave off summary judgment as argued by defendant. (Def.’s Mot. at 24.) The resolution of these two issues will determine whether defendant has successfully rebutted plaintiffs’ prima facie ease.

a. The Parties’ Use and Definition of Variables

Plaintiffs devote many pages to criticizing Haworth’s analyses (see Pls.’ Opp. at 31-33; Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶¶ 20-29), and it may well be that there are a variety of arguments that would, at least, preclude the granting of summary judgment in defendant’s favor based on the flaws in her analyses. But as the Court has said, the weight to be accorded to the statistical findings of each party’s expert is an issue properly reserved for the factfinder at trial and not the Court on a motion for summary judgment.

That being said, the Court will explore the issue of variables to clarify the differences between the experts and the effect of these differences on their results. To do this, it is helpful to recall that for purposes of multiple regression analyses, only the “major variables” must be controlled for, but how they are defined is far from clear, since they depend on the facts and theory of the case. Coward, 140 F.3d at 274. One must be cautious not to include variables that could conceivably serve as a proxy for discrimination. See Valentino, 674 F.2d at 73 n. 30. Because of the lack of any bright-line test (see Coward, 140 F.3d at 274 and 276-77 (Sentelle, J., concurring)), the parties launch into a major debate over whether education and experience are “major variables” in this case. 26 Although it may be arguable that as a matter of law these two variables are not so necessary as to cause an analysis that fails to include them to be inadmissible (see McReynolds, 208 F.R.D. at 443-44; see also Coward, 140 F.3d at 275 (declining to reach the question of whether education and experience are always required)), the Court need not resolve this dispute, since it will proceed on the assumption that they are major factors, and they should be included in any regression analysis.

But reliance on that presumption does not end the matter, since there remains a legitimate debate between the parties as to how “experience” should be defined and what other variables, if any, need be included. And the importance of this debate is clear, for what variables are included or excluded and how a variable is defined have a significant impact on whether the expert finds a statistically significant disparity.

With respect to the factor of “experience,” plaintiffs criticize Haworth for not effectively controlling for experience by failing to consider when a person gained the experience and how long they had that experience, and by failing to differentiate between current occupation and a job that someone had 20 years ago or a prior job that is of no relevance to their current Sodexho employment. Therefore, according to plaintiffs, Haworth, by not distinguishing between current occupation and past experience, over-emphasizes past experiences, improperly causing past experience and current occupation to have equal weight. Further, plaintiffs complain that Haworth does not properly control for current occupation and MRR because her analyses only consider employees’ bands and not their MRRs, and therefore, she includes employees for whom a new position would not be a promotion (but rather it might well be a demotion because it would entail a decrease in salary). (See Tr. at 89-93.) This flaw would also, in plaintiffs’ view, operate to deflate the percentage of African Americans available for any position, since African Americans are found' disproportionately at the lower levels of the bands and whites are found at the higher levels. (See id.; Pls.’ Opp. at 31-32; Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶¶ 21, 24, 26.)

While it is difficult (but not necessary) to resolve whether these many criticisms have merit, Haworth did in fact re-run her regression analyses. She used both her definition of promotion, as well as Siskin’s, and she continued to include variables such as RVPs, licenses and certifications, memberships, industry experience and specific unit experiences, 27 but she attempted to modify her analysis to address plaintiffs’ criticisms regarding her failure to properly account for current occupation. The results of her efforts appear in Table 4 to her 3/31/04 declaration. (See Def.’s Ex. 46.) But as argued by plaintiffs, this table reveals that “all things being equal, one factor — the color of an applicant’s skin— lowers that applicant’s chance of promotion by more than 20%,” since that table shows that for the set of 5,116 promotions, the African American promotion probability is 78% of the promotion probability of white employees (or 1.89 units of standard deviation). (See Pls.’ Opp. at 25-26; Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶ 38.)

While the figure plaintiffs cite from Haworth’s Table 4 may not satisfy the usual standard of 1.96, this Circuit recognizes that a showing of 1.89, which is in the “intermediate zone” (defined as standard deviations of 1.65 to 1.96), in conjunction with other relevant evidence, can be sufficient at least to create an issue of fact as to whether discrimination has been shown. Palmer, 815 F.2d at 97 n. 10. Moreover, Table 4 presents a vivid illustration of how “malleable” statistics can be, Vaughan v. MetraHealth Cos., 145 F.3d 197, 203 (4th Cir.1998), for a change in the definition of promotion (Haworth’s 6,984 promotions versus Siskin’s 5,116) or a change in how current occupation is accounted for dramatically affects the statist tical results. 28

As the above demonstrates, how one defines a promotion or how one reflects current occupation can affect the results of a regression analysis. But a similar effect can be demonstrated depending on whether one includes or excludes variables that by all accounts do not fit within the law’s definition of “major factors.” For instance, Haworth controls for RVP, which defendant concedes is not a major variable and need not be included (Tr. at 58-59), as well as licenses and certifications, industry experience, memberships, and “specific unit experiences.” Plaintiffs object to the inclusion of these variables as being unnecessary, contrary to good statistical practice, and as having the potential for skewing the study by masking the effect of race on promotions. (See Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 39-41.) While the Court is not prepared to categorize all of these variables as “tainted” (see note 27 supra), there is at least a serious question whether RVPs are a legitimate explanatory factor to predict the probability of promotion, or whether they may be masking a discriminatory process, thus qualifying under Valentino as an “inappropriate” or “tainted” variable. 674 F.2d at 73 n. 30 (citing Finkelstein, The Judicial Reception of Multiple Variable Regression Studies in Race and Sex Discrimination Cases, 80 Colum. L.Rev. 737, 738-42 (1980)). This could well be the case, because, following Finkelstein’s criteria for inappropriate variables, RVPs could be affected by the employer, are not related to an employee’s productivity, and are correlated with group status. Id. at 738-39. (See also Pls.’ Ex. 162 [12/03 Siskin report] at 3.) This would be particularly true, if, as plaintiffs argue, African Americans are clustered or segregated in RVPs not by choice, and if there is a low probability of promotion out of these RVPs regardless of race. (See note 22 supra.)

If these variables are eliminated because they are unnecessary, “inappropriate,” or “nonsense” variables, see Valentino, 674 F.2d at 71 n. 21, the results of an expert’s regression analyses will necessarily change. For instance, when plaintiffs’ expert purports to re-run Haworth’s regression analysis, controlling for education and experience, but eliminating these allegedly unnecessary variables and using his definition of promotion, he finds a statistically significant race effect for the class period of at least 2.50 standard deviations. 29 (Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 34, 44-45 and Tables 5 and 6.) Moreover, when Siskin represents that he has re-run the analysis done by Haworth in her Table 4 on a reweighted basis as she recommends, but has used his definition of promotion, as well as her RVPs, but excludes allegedly unnecessary variables and unit size but includes MRRs, the results are striking insofar as they show a more statistically significant effect than found in any prior regression analysis (as high as 3.24 standard deviations). 30 (See Pls. Ex. 158 [5/04 Siskin decl.] at ¶ 48 and Table 8.)

While the Court’s analyses may well be oversimplified or insufficiently detailed, there can be no doubt that defendant’s regression analyses are insufficient to defeat plaintiffs’ prima facie case. For, as the above discussion demonstrates, the inclusion or exclusion of non-major variables and how one defines promotion, as well as past job experience and current occupation, can have a profound effect on whether a statistically significant race disparity is found. Furthermore, there is no validity to defendant’s argument that Sis-kin failed to control for the major variables (education and past experience), since at the end of the day, he did just that and he still found statistically significant results. 31 (See Pls.’ Ex. 158 [5/04 Siskin decl.] at ¶¶ 34, 44-48 and Tables 5-8.) While each side may well have valid criticisms of how the opposing expert executed various methodologies, these highly-credentialed statisticians have reached starkly different conclusions based on re-runs of each other’s multiple regression analyses, and the propriety of these different approaches cannot be resolved at the summary judgment stage because they involve issues of credibility and disputed issues of fact.

II. Disparate Impact Claim

Plaintiffs’ second claim alleges that class members have suffered a disparate impact on the basis of race as a result of “Sodex-ho’s entirely subjective decisionmaking process, the elements of which are not capable of separation for analysis.” (Compl. ¶ 27.) This claim arises under Title VII, 42 U.S.C. § 2000e-2(k)(1).

A. Legal Standards

Disparate impact claims, though related to disparate treatment claims, have a different focus. Whereas the latter seek to determine the existence of discriminatory intent, the former/do not focus on intent, but rather consider policies or practices that are facially neutral but that nonetheless have a disparate effect on a protected group. See Griggs v. Duke Power Co., 401 U.S. 424, 432, 91 S.Ct. 849, 28 L.Ed.2d 158 (1971). Like disparate treatment claims, disparate impact claims are governed by a three-part test. First, plaintiffs must make out a prima facie case by pointing to a policy or practice that is neutral on its face but which has a discriminatory effect on fa particular group. Anderson, 180 F.3d at 338-39 (quoting Teamsters, 431 U.S. at 349, 97 S.Ct. 1843). Second, the employer bears the burden both of production and of persuasion in demonstrating that the challenged practice or policy is “job related for the position in question and consistent with business necessity.” Id. at 338-39, 342, 97 S.Ct. 1843 (quoting 42 U.S.C. § 2000e-2(k)(1)(A)(i) and citing Griggs, 401 U.S. at 431, 91 S.Ct. 849). It is a fact question for trial whether a qualification required by an employer is actually “necessary to succeed.” Id. at 433, 91 S.Ct. 849 (citing Goodrich v. Int’l Bhd. of Elec. Workers, 712 F.2d 1488, 1493-94 (D.C.Cir.1983)). Third, if the defendant establishes business necessity, plaintiffs may show that a different practice would meet the employer’s legitimate needs without a similar discriminatory effect. Id. at 339, 97 S.Ct. 1843 (citing 42 U.S.C. § 2000e-2(k)(1)(A)(ii) and Albemarle Paper Co. v. Moody, 422 U.S. 405, 425, 95 S.Ct. 2362, 45 L.Ed.2d 280 (1975)).

“ ‘[A]n important point of convergence’ between disparate treatment and disparate impact claims exists in class actions such as this one. Because both ... claims ‘are attacks on the systemic results of employment practices ... proof of each claim will involve a showing of disparity between the minority and majority groups in an employer’s workforce.’ ” Moore v. Summers, 113 F.Supp.2d 5, 19 (D.D.C.2000) (quoting Segar, 738 F.2d at 1267.) Indeed, a disparate treatment claim may morph into a disparate impact claim, depending on the employment practices involved and the defendant’s arguments. Segar, 738 F.2d at 1298 (Edwards, J., concurring). Thus, as with disparate treatment claims, statistics that reveal a significant disparity play a central role in most disparate impact claims. See Watson v. Fort Worth Bank & Trust, 487 U.S. 977, 994-95, 108 S.Ct. 2777, 101 L.Ed.2d 827 (1988) (O’Connor, J., concurring); EEOC v. Joint Apprenticeship Comm., 186 F.3d 110, 117 (2d Cir.1999). “[T]he statistics must be of a kind and degree sufficient to reveal a causal relationship between the challenged practice and the disparity.” Robinson, 267 F.3d at 160 (citation omitted). Plaintiffs must therefore show that the impugned practice “caused at least some members of the class to be deprived of a promotion, or at least the opportunity of being considered for a promotion.” Koger, 98 F.3d at 639.

Suits alleging subjective practices that result in disparate effects may proceed in order to counter the “unfairness in permitting an employer to perpetuate discriminatory effects by relying for discriminatory results on the individual biases of its managers.” Mozee, 940 F.2d at 1050 (citing Watson, 487 U.S. at 990-91, 108 S.Ct. 2777). “In order to sustain a subjective practice through a business necessity defense, therefore, an employer would have to argue that the subjectivity of the step is job-related — in other words, that something about the position requires the selector to make a subjective evaluation of the applicant’s abilities.” Id. (citing Crawford v. Western Elec. Co., 745 F.2d 1373, 1385-86 (11th Cir.1984)). In the Civil Rights Act of 1991, Congress modified Title VII, in response to the Supreme Court’s decision in Wards Cove Packing Co. v. Atonio, 490 U.S. 642, 656-57, 109 S.Ct. 2115, 104 L.Ed.2d 733 (1989), to permit plaintiffs to focus on a defendant’s overall decisionmak-ing process if plaintiffs can “demonstrate to the court that the elements of respondent’s decisionmaking process are not capable of separation for analysis.” 42 U.S.C. § 2000e-2(k)(1)(B)(i). See Smith, 196 F.3d at 368 (noting that the 1991 amendments “softened the holding of Wards Cove ”).

B. Plaintiffs’ Prima Facie Case

In its motion, defendant argues that plaintiffs cannot satisfy their burden to show a prima facie case because they have failed to identify a specific selection process or criterion that causes a disparate impact on African Americans, and that even if they can show such a practice, they have not satisfied their statistical burden of showing the practice caused a disparate impact. (Def.’s Mot. at 38-44). For many of the reasons already discussed with respect to plaintiffs’ disparate treatment claim, defendant is mistaken.

First, as previously explained, plaintiffs have proffered sufficient evidence at least to create an issue of fact as to whether Sodexho employs an entirely subjective de-cisionmaking process in awarding promotions that adversely impacts African American managers. While defendant obviously has cited evidence that rebuts plaintiffs’ showing, it is not possible for the Court to agree that plaintiffs have “virtually no evidence in support of their disparate impact claim.” (Def.’s Mot. at 39). 32 (See, e.g., Pls.’ Stmt. of Disputed Facts ¶¶ 4, 6, 29, and Pls.’ Rebuttal to Def.’s Stmt. of Facts ¶¶ 20, 31-41.)

Having reviewed the wealth of evidence before it, the Court is more than satisfied that plaintiffs have presented sufficient evidence that the elements of defendant’s decision-making process are not capable of separation for analysis, and this process is subjective and discretionary. Given this conclusion, there is substantial support in the law to justify plaintiffs’ claim. See, e.g., Watson, 487 U.S. at 990-91, 108 S.Ct. 2777, Mozee, 940 F.2d at 1050 and cases collected therein; Cook v. Billington, No. 82-0400, 1992 WL 276936, at *5 (D.D.C. Aug. 14, 1992); McClain v. Lufkin Industries, Inc., 187 F.R.D. 267, 275 (E.D.Tex.1999); Ste nder v. Lucky Stores, 803 F.Supp. 259, 335 (N.D.Cal.1992); Home Depot, 1997 WL 605754, at *13-14, 1997 U.S. Dist. LEXIS 16296, at *46-48. 33

Finally, defendant argues that plaintiffs have not submitted sufficient evidence of causation. While defendant devotes its time to attacking plaintiffs’ industrial psy- ehologist expert, the primary evidence of causation has been provided by Siskin. As already discussed (see Section I(B)-(D)), his analyses provide statistical evidence of a kind and degree from which a jury could reasonably find causation. Moreover, the wealth of anecdotal evidence, when coupled with the testimony of Siskin and Prien, is certainly enough to raise a jury question as to the causal link between Sodexho’s allegedly subjective practices and the race disparity at the managerial level.

CONCLUSION

In sum, there are substantial questions of fact as to almost all material issues, and any conclusion regarding the probativeness of the testimony of the two experts who have appeared in this case (as well as many of the other well-known Title VII class actions that have been brought across the country) depends largely on underlying factual disputes and questions of credibility. Having carefully reviewed the parties’ pleadings and their arguments at the hearing on this motion, as well as the voluminous record in this case, the Court finds that defendant’s motion for summary judgment must be denied except that plaintiffs’ § 1981 disparate impact claim is dismissed with prejudice. It will therefore be for the jury to determine whether plaintiffs can sustain their ultimate burden of showing company-wide discrimination based on a theory of subjective, decentralized decision-making with respect to promotions of managers at So-dexho.

1 . The certified class consists

of all African-Americans who are or were salaried employees of Sodexho [] at any time from March [27], 1998, to [July 1, 2001], and who have held or sought to obtain (1) an upper-level managerial, supervisory, or professional position (above-the-unit or comparable level of responsibility) [for which a promotion decision was made before June 23, 2001] or (2) a job that would lead to such a position, and who have been, or continue to be, or may in the future be adversely impacted by Sodexho’s racially discriminatory policies and practices affecting promotions or advancement.

Id. at 433 (as modified by the Court’s Order dated January 14, 2003).

2 . The Court previously denied defendant's motion to decertify the class, except that it reserved judgment on Sodexho’s arguments regarding commonality and typicality. See McReynolds v. Sodexho, No. 01-510 (D.D.C. May 27, 2004). Defendant's argument in its motion to decertify the class — that faulty statistical evidence offered by plaintiff's expert, Dr. Bernard Siskin, violates Daubert-is addressed in the Court’s Memorandum Opinion, also issued this date, denying defendant's Daubert motion. The other remaining argument in the motion to decertify — that the class members' claims are not sufficiently common to or typical of the class because statistically significant adverse effects appear in only 9 of 155 RVPs — is addressed in this Memorandum Opinion. The Court will therefore not issue any further opinions with respect to the decertification motion.

3 . Defendant correctly argues that plaintiffs cannot bring a disparate impact claim under 42 U.S.C. § 1981, since purposeful discrimination is required under § 1981. See General Bldg. Contractors Ass’n, Inc. v. Pennsylvania, 458 U.S. 375, 391, 102 S.Ct. 3141, 73 L.Ed.2d 835 (1982). Therefore, summary judgment will be entered as to plaintiffs’ § 1981 disparate impact claim.

4 . The use of multiple regression analysis, used by experts for both sides, becomes relevant with respect to defendant’s rebuttal case. (See Section 1(D) infra.)

5 . Siskin analyzed the data for a given month, and considered as “similarly situated” all employees in the same grade or band, occupation code (which he adjusted for Market Reference Rate ("MRR”) differences), and division as the promoted employee prior to promotion. (Pls.’ Ex. 160 [11/01 Siskin report] at ¶ 4.) He ran his analysis for both the class period (March 1998 to June 2001), as well as January 1995 to June 2001. As noted, the former yielded a 5.34 standard deviation disparity, whereas the latter resulted in a disparity of 8.87. (Id. at Table S-2.) His analysis involved 4,276 promotions for the class period, whereas he analyzed 10,002 promotions for the 1995-2001 time period. See also McReynolds, 208 F.R.D. at 435-36 (describing Siskin’s studies).

6 . Siskin's pools analysis shows lower levels of race-based disparities at the upper echelons of Sodexho. (Pls.’ Ex. 159 [10/01 Siskin report] at Table 9.) However, as Siskin explains, "the expected number of promotions to the highest level jobs will be very low if African American managers cannot even reach the positions that lead to those highest level jobs. Thus, ... [t]he results for the highest job levels should not be reviewed in isolation.” (Id. at ¶ 32.) This identification of a mid-level bottleneck is in accord with trends previously identified by this Circuit. ”[I]t is quite plausible that a discriminatory attitude about [African Americans] and their potential for further advancement might affect promotions only at a mid-level step,” because the lower level barrier to promotions could mean that those who were able to overcome it may "have demonstrated such superior skill and aptitude that they would encounter less resistance to advancement in upper level positions.” Palmer, 815 F.2d at 104. Therefore, the Court rejects defendant's insinuation (see Def.'s Mot. at 25) that these results show that there is not a company-wide pattern of discrimination at Sodexho.

7 . Plaintiffs strongly object to defendant’s definition of promotion on the grounds that it is inconsistent with the company’s own definition of a promotion, which requires that a candidate change jobs. (See Pls.’ Ex. 105 at 8.) According to plaintiffs, Haworth’s overly-broad definition overstates the number of African American promotions because it includes lateral moves that are not really promotions, as well as pay increases that resulted from external factors, like an increase in the Market Reference Rate ("MRR”). Her definition produces approximately 2,000 additional promotions during the class period, and according to plaintiffs, this overstates the number of African American promotions in particular because the broadest bands (75, 76 and 77) are in the lower levels where African Americans have been disproportionately stuck due to Sodexho's racially discriminatory promotion process. (See Pls.' Rebuttal to Def.’s Stmt of Facts ¶ 2.) This disputed issue of fact (see Def.’s Mot. at 8, 28; Def. Ex. 12 [10/03 Haworth report] at 3-7) has a direct effect on both the pools analyses of the parties, as well as their multiple regression analysis. (See Section 1(D) infra.) However, for purposes of this motion, it is proper (as defendant appears to have agreed (see 12/3/04 Hearing Tr. [“Tr.”] at 65)) for the Court to apply Siskin’s definition.

8 . Siskin also re-ran his pools analysis using Haworth's definition of promotion and found statistically significant results: 2.97 for the class period involving 7,018 promotions and 6.94 for the period January 1995 to June 2001 for 12,756 promotions. (Pls.’ Ex. 160 [11/01 Siskin report] at Table S-1.) The difference between Haworth's result of 3.47 and Siskin’s result of 2.97 appears to relate to the fact that he did not control for district or unit and did not use weighted feeder pools. But regardless of the exact methodology, both experts found statistically significant results at the company-wide level.

9 . Defendant often compounds the confusion by using the terms RVP and region synonymously. However, it is readily apparent that, regardless of whether RVPs are indeed a functional unit within Sodexho, they bear lit-tie correlation with the 54 geographic regions that exist within Sodexho. The Court, therefore, will refer to RVPs and not to regions. (See Pls.’ Opp. at 11 ("Some RVPs consist of a single employee in one office and other RVPs have over 150 employees working in 25 states in every section of the country. The same geographic location can also be in four or five different RVPs.") (citations omitted); Def.’s Reply at 10 (conceding that "RVP areas are not necessarily geographically contiguous and may overlap one another in a geographic sense”).)

10 . In fact, the only organizational chart reflecting RVPs that defendant has provided to the Court was included in Def. Ex. 2 at the motions hearing and is entitled "Sodexho Or ganizational Structure as of June 2001.” This chart, however, is not a corporate record, but was created after the fact by defendant based on Haworth's 3/02 declaration and her 10/03 report. (Tr. at 58.)

11 . Of 4,276 promotions during the class period that Siskin studied, 1,446 (or 34%) resulted in a change in the promoted employee’s RVP code. (Pls.' Ex. 158 [5/04 Siskin decl.] at ¶ 17.) By contrast, Haworth found a smaller percentage given her more expansive definition of what constitutes a promotion. (Id. at ¶ 17 n. 15; see also note 8 supra.)

12 . These tables show the percentage of African Americans and the percentage of other employees who remained in a division, RVP, district or unit during the class period. It is impossible to determine whether the people who remained were not promoted or whether they were promoted and stayed. Likewise, the employees who left may not have left with a promotion. Hence, these tables do not show the promotion rate within the subgroup, nor do they provide any useful information regarding where a person went in the event of a promotion. In fact, it is even unclear from the tables if the figures reflect attrition in the subgroup for reasons unrelated to a promotion.

13 . Siskin reported a similar figure, finding that 28% of all bidders for open positions came from outside the division. (Pls.’ Ex. 160 [11/01 Siskin report] at ¶ 12 n. 8.) Plaintiffs also buttress their argument with testimony from plaintiffs and corporate employees who attest to the fact that managers often apply for and move across divisions and geographic regions. (See record citations in Pls.’ Rebuttal to Def.'s Stmt. of Facts ¶ 3 at 36-37.)

14 . Similarly, to the extent that defendant may be taking issue with Siskin’s failure to account for minimum objective qualifications in his pools analysis (as opposed to his multiple regression analysis) (see note 26 infra), the Court notes that the weighted feeder pool analysis of Sodexho’s own expert also failed to account for what defendant now appears to contend are minimum objective qualifica tions — i.e., education and experience. (See Def.'s Ex. 8 [10/01 Haworth report] at Table 18.)

Moreover, although at the prima facie stage, minimum objective qualifications must be accounted for in constructing a pool, Davis, 613 F.2d at 964, "minimum” means just that; "not every conceivable factor relevant to a promotion decision must be included in the statistical presentation in order to make out a Prima facie case.” Id. "Rather, what the case law means by 'minimum objective qualifications’ are those objective qualifications that can be shown to be truly required to do the job at issue.” Anderson v. Zubieta, 180 F.3d 329, 342 (D.C.Cir.1999). The prerequisite for inclusion as a minimum objective qualification is that the qualification is actually considered by the employer. Valentino v. United States Postal Serv., 674 F.2d 56, 61 (D.C.Cir.1982) (citing Hazelwood Sch. Dist. v. United States, 433 U.S. 299, 308 n. 13, 97 S.Ct. 2736, 53 L.Ed.2d 768 (1977)). Thus, "if a particular number of years of work experience were established as a minimum job criterion, then that would need to be reflected in the proffered statistics. If, however, work experience were only a factor to be considered in promotion decisions, a plaintiff's statistical data need not take that into account.” Davis, 613 F.2d at 964.

Applying these principles to the case at hand, it is a disputed issue of fact whether there are any minimum necessary objective qualifications for the positions at issue. (See Pls.' Rebuttal to Def.'s Stmt. of Facts ¶ 20.). See also id. at 964-65 n. 42 (finding no minimum objective qualifications necessary for the promotions at issue and thus finding that plaintiffs' prima facie case was not deficient because of its failure to account for the relevant factors of education and experience). Thus, whether plaintiffs' pools analysis can be faulted for failing to account for education and experience (as opposed to their multiple regression analysis) is for the trier of fact, for what constitutes a minimum objective qualification depends on the facts and theories of the case. Bazemore, 478 U.S. at 400, 106 S.Ct. 3000; Coward v. ADT Sec. Sys., 140 F.3d 271, 274 (D.C.Cir.1998). See also Mozee, 940 F.2d at 1045 ("Where subsequently discovered 'qualifications' are not obvious and are far from uniformly applied, we think their use in determining the promotion pool must be carefully restricted.”); Townsend v. Washington Metro. Area Transit Auth., 746 F.Supp. 178, 186 (D.D.C.1990) (because the defendant did not engage in an analysis of the applicants’ qualifications at the time of its decision, reliance on those qualifications to rebut discrimination amounts to a post hoc rationalization that "carries the seeds of its own destruction” (internal quotation omitted)).

15 . AMOCO is particularly irrelevant, since the court was not evaluating any statistical evidence, but only anecdotal evidence. 2003 WL 262424, at *4-5, 2003 U.S. Dist. LEXIS 1795, at *16-19.

16 . Of course, as noted in McReynolds, 208 F.R.D. at 442-43 & n. 21, it is not hard to find cases granting certification or denying summary judgment where claims of company- wide discrimination based on decentralized, subjective decision-making were supported by company-wide aggregated statistics. See, e.g., Butler, 1997 WL 605754, 1997 U.S. Dist. LEXIS 16296; Dukes v. Wal-Mart Stores, Inc., 222 F.R.D. 137, 147-48 (N.D.Cal.2004) (rejecting at the class certification stage Ha-worth’s statistics disaggregating to the store or store sub-unit level and discussing case law relating to aggregation); and cases collected in Moore, 2004 U.S. Dist. LEXIS 5959, at *32. See generally Tristin K. Green, Discrimination in Workplace Dynamics: Toward a Structural Account of Disparate Treatment Theory, 38 Harv. C.R.-C.L. L.Rev. 91, 151-56 (2003).

17 . According to Siskin, there were so few promotion decisions in 125 of the 163 RVP codes identified by Haworth in her March 2002 declaration that no matter what promotion decisions were made, these 125 RVP areas could not show statistically significant adverse results. (Id. at ¶ 22.)

18 . While the analysis at p. 279 of Statsny has rarely been cited, the Court is aware that the Honorable James Robertson of this Court has relied on it in recent decisions. See Garcia v. Veneman, 224 F.R.D. 8, 13 (D.D.C.2004); Love v. Veneman, 224 F.R.D. 240, 243 (D.D.C.2004). But as Judge Robertson acknowledged, this case is distinguishable from those that were before him. See Garcia v. Veneman, 211 F.R.D. 15, 22 (D.D.C.2002).

19 . Arguably Stastny has no applicability here, since plaintiffs' theory is that there was a centralized decision at Sodexho to allow its managers unfettered discretion in making promotion decisions. See Caridad v. Metro-North Commuter R.R., 191 F.3d 283, 291-92 (2d Cir.1999) (crediting plaintiffs' theory, for class certification purposes, that commonality stems from defendant’s centralized decision to grant substantial discretion to local decision-makers).

20 . It also bears noting that the legal authority cited by defendant for the proposition that plaintiffs must "demonstrate a clear nexus between their anecdotes and their pattern-or-practice claim'' (Def.'s Mot. at 19) is inappo-site. The court in Mooney v. Aramco Services Co., 54 F.3d 1207, 1220-21 (5th Cir.1995), affirmed the district court's decision, in a non-class action, to exclude as irrelevant the testimony of witnesses relating to parts of the defendant where the six individual plaintiffs did not work. The case most certainly cannot be read to exclude plaintiffs’ anecdotal evidence that does not address nationwide issues or that does not cover all the 155 RVPs. Moreover, such a requirement in a class action would be at odds with binding precedent, which credits individual testimony even in class cases. See McKenzie, 684 F.2d at 71; Palmer, 815 F.2d at 102.

21 . For instance, Sodexho’s chief EEO officer told the Senior Vice President for Human Resources "there, was a serious need to get more African Americans in above-the-unit jobs,” and discussed "the lack of minority representation” with other senior executives. (Pls.’ Ex. 78 [Watkis dep.] at 114, 142-43.) The EEO officer told other officials "since the day [he] started with the company” there needed to be more African-Americans in above-the-unit jobs. (Id. at 115). According to him, Sodexho's deficiencies included "[v]ery few minorities and women in district manager and large-account general manager positions” and the "[l]ack of minority and female representation in key sales roles.” (Pls.' Ex. 147 at SDX49624.) Consistent with this testimony, Sodexho’s internal "utilization analyses” show that African Americans’ share of above-the-unit positions actually fell during the class period. In 1998, whites were four times as likely as African Americans to hold such a job; by 2000, the ratio was five to one. (See Pls.’ Exs. 150, 153, 78 at 145-46.) And while these types of statistics may be subject to criticism (see McReynolds, 208 F.R.D. at 444 n. 26), they arguably can be considered as evidence in support of plaintiffs' statistical case.

22 . In addition to the above utilization figures, plaintiffs have bolstered their statistical case with evidence of segregated accounts at So-dexho. For instance, in Siskin's segregation analysis, he found that 5.2% of black managers report to black district managers, while only 2.2% of white managers report to a black district manager. (Pls.’ Ex. 161 [6/03 Siskin report] at Table SI.) Siskin determined that the likelihood of this level of clustering is less than 2 in 1,000. (Id. at ¶ 7.) Siskin also determined that 18 out of 25, or, controlling for division, 21 of 30, geographic areas show a higher degree of segregation than would be expected from a racially neutral selection process. (Id. at ¶¶ 14-16 & Tables S2-S3.) Insofar as Siskin’s methodology is disputed because of his failure to account for employee preferences governing which unit a worker winds up in (see Def.'s Reply at 28-29), those charges affect the study's probativeness, which is to be determined at trial. See Palmer, 815 F.2d at 106, 110 (an analysis that ignores individual preferences as a possible explanatory factor is still probative of discrimination). But even in this regard, plaintiffs have offered anecdotal evidence to support their claim of segregation and its purported career-limiting effects. (See, e.g., Pls.' Ex. 12 [Brown dep.] at 95-96, 100 (being sent to a black account is “being sent to the slaughter”); Ex. 18 at 220; Ex. 20 at 261 (accounts were listed in the MCMS job posting system as "predominantly black").)

23 . "Multiple regression analysis is a statistical tool commonly used by social scientists to determine the influence that various independent, predetermined factors (so-called ‘independent variables') have on an observed phenomenon (the so-called ‘dependent variable'). In disparate treatment cases involving claims of [race] discrimination, plaintiffs typically use multiple regression analysis to isolate the influence of [race] on employment decisions relating to a particular job or job benefit .... The first step in such a regression analysis is to specify all of the possible 'legitimate' (i.e., nondiscriminatory) factors that are likely to significantly affect the dependent variable and which could account for disparities in the treatment of [employees of different races]. By identifying those legitimate criteria that affect the decision making process, individual plaintiffs can make predictions about what job or job benefits similarly situated employees should ideally receive, and then can measure the difference between the predicted treatment and the actual treatment of those employees. If there is a disparity between the predicted and actual outcomes for [minority] employees, plaintiffs in a disparate treatment case can argue that the net 'residual' difference represents the unlawful effect of discriminatory animus on the allocation of jobs or job benefits.” Ottaviani v. State University of New York, 875 F.2d 365, 366-67 (2d Cir.1989) (citations omitted).

24 . To bolster its rebuttal, defendant also relies on various other types of evidence, including statistics showing that African American representation is higher in four of Sodexho’s six divisions than "the relevant labor market African American availability.” (Def.’s Mot. at 30.) Predictably, plaintiffs respond that these statistics are meaningless, since the relevant comparison in a promotions case is not the external labor market, but the internal availability. (See Pls.' Rebuttal to Def.’s Stmt. of Facts ¶ 6 and citations therein.) Thus, in response plaintiffs offer their own underutili-zation figures (see note 21 supra), as well as anecdotal evidence regarding underrepresen-tation of African Americans in managerial positions. (See Pls.' Rebuttal to Def.’s Stmt. of Facts ¶ 6.)

Defendant also points to Haworth's two analyses of Sodexho’s job posting systems to argue that, when analyzed separately, neither the earlier MCMS system nor the later Career Center program, shows a statistically significant difference in the number of promotions between African Americans and whites. (Def.’s Mot. at 32 (citing Def. Ex. 8 [10/01 Haworth report]; Def. Ex. 1 [7/03 Haworth report]; and Def. Ex. 12 [10/03 Haworth report] ).) Plaintiffs counterattack by challenging that there is no factual basis for analyzing the two systems separately, since they operated in a similar fashion, and when analyzed together, Siskin found a shortfall of 2.21 standard deviations. (See Pls.’ Rebuttal to Def.'s Stmt. of Facts ¶¶ 42-43 (citing Pls.’ Ex. 157 [2/02 Siskin decl.].))

Suffice it to say that the above is merely illustrative of the substantial issues of fact upon which Sodexho relies in its rebuttal, and discussion of further examples would be superfluous.

25 . Plaintiffs challenge her methodology because Haworth's team was only able to code data for an employee where the résumé affirmatively listed a trait, such as a higher degree or experience in a given field. (Pls.' Opp. at 38 n. 27.) Thus, an employee with a certain kind of experience or education, who did not happen to put such information on the re-sumé that was submitted to Sodexho, would not be accurately accounted for. Plaintiffs liken the resulting data and corresponding experience and education factors to the type of "nonsense” referred to in Valentino, where a “special degree" factor was coded for only if an employee indicated a specialty on a personnel form. In the Circuit's view, the result was too arbitrary for inclusion as a factor in the analysis, since, for instance, "[a] college history major who put 'history’ on the form would be recorded as holding a 'special degree' while one who did not fill in his or her major would not be recorded as a 'special degree’ holder.” 674 F.2d at 71 n. 21. But courts have also admitted statistics that rely on data mined from paper records. See, e.g., Sears, 839 F.2d at 349-50. Given this issue, the Court will not reject Haworth’s analyses on this basis, although the debate may be relevant at trial.

26 . Plaintiffs argue that major variables are the same as "minimum objective qualifications,” as defined in Davis, and since they claim that defendant did not consider factors such as education and experience when making promotion decisions, they should not be considered in a regression analysis. (Pls.' Opp. at 37; see also note 14 supra.) This argument is overbroad and confuses the law. While plaintiffs' contention may be correct at the prima facie stage, it does not apply to regression analyses used thereafter, which seek to identify other factors that could cause the disparity initially identified. Therefore, even if education and experience were not "minimum objective qualifications” actually required for a promotion, they may nonetheless hold significant explanatory value when considered as part of a regression analysis. See Koger, 98 F.3d at 637.

27 . Siskin argues in his 5/04 declaration at ¶¶ 39-43 (Pls.’ Ex. 158) that the addition of at least some of these variables is improper and that they should be deleted as "tainted” variables, because they "inexplicably reduce[] a person’s chances of getting a promotion.” (Pls.' Opp. at 33.) However, just because some variables indicate a negative effect on promotion relative to the qualifications of the average would-be promotee does not necessarily make them suspect in and of themselves. (Def.’s Reply at 25 n. 15.) Therefore, it is not imperative that they be removed, as Siskin suggests. (Pls.’ Ex. 158 at ¶ 41; Def.’s Ex. 70 [5/17/04 Haworth decl.] at ¶ 6 (discussing the danger of introducing specification bias by removing supposedly collinear variables).) On the other hand, factors such as prison experience or the possession of various licenses or memberships do not qualify as major variables that must be included in a regression analysis.

28 . For instance, even if no factor except the definition of promotion is changed, Ha-worth’s promotion definition yields 6,984 promotions and a race effect of 1.06 standard deviations, whereas Siskin's definition yields 5,116 promotions and a race effect of 1.89. (See Def.'s Ex. 46 at Table 4.) Similarly, when Haworth uses Siskin's approach to current occupation and his number of promotions, her estimated race effect changes from 0.90 to 1.89 standard deviations. (Id.)

29 . Not surprisingly, Haworth asserts in response to these results that “[i]f we choose to include division as the decision-making structure, rather than the RVP area in the Ha-worth model estimated in my July 2003 Report, there is still no statistically significant race effect.” (Def.'s Ex. 70 [5/17/04 Haworth decl. at ¶ 5 n. 6.]) But since she provides absolutely no further explanation or documentation in support of this assertion, the Court can hardly reject Siskin's tables as having no probative value. (See also Tr. at 58-59, 135.)

30 . Admittedly, Haworth criticizes Siskin's work for adding some variables and omitting others and for doing more than just replacing RVP code with division code (see Def.’s Ex. 70 [5/17/04 Haworth decl.] at ¶ 10), but these criticisms again create issues of fact that do not result in the exclusion of Siskin's analy-ses.

31 . While defendant argues that Siskin failed to re-run Haworth’s analyses properly, because he did not fully control for past experience, it does admit that he controlled for education and, at least in part, for past experience. (See Tr. 69-70.)

32 . Even defendant concedes that plaintiffs’ industrial psychologist, Dr. Prien, does provide such evidence, but it dismisses his report as lacking any “hard evidence to support his position.” (Id. at 39.)

33 . The only cases from this jurisdiction that defendant cites, Koger, 98 F.3d at 639, and Batson v. Powell, 912 F.Supp. 565, 572 (D.D.C.1996), provide no guidance with respect to the instant question. The latter focuses on a specific practice, the Uniform Dress Code, that allegedly had a disparate impact; however, plaintiffs failed to present significant statistical evidence. The former likewise involved a specific practice, a seven-phase scoring system, the intermediate phases of which purportedly caused a promotion-related disparate impact. The Court rejected plaintiffs' claim, because they had failed to show that any impact was caused by low scores. The other cases relied on by defendant also involve specific employment practices, such as a specific selection criterion that purportedly created a disparate impact, rather than promotion practices that are incapable of separation. (See Def.’s Mot. at 38, citing Smith, 196 F.3d at 367; Munoz, 200 F.3d at 299.)