HOT ISSUES CONCERNING USE OF STATISTICAL … · HOT ISSUES CONCERNING USE OF STATISTICAL EVIDENCE IN EMPLOYMENT DISCRIMINATION ... ofstatistics in employment discrimination class

HOT ISSUES CONCERNING USE OF STATISTICAL EVIDENCE

IN EMPLOYMENT DISCRIMINATION

CLASS ACTION LITIGATION

Michael D. Lieder

September 2009

Sprenger ~~ Lang PLLeATTORNEYS AT LAW

www.sprengerlang.com

Washington, DC1400 Eye Street N.W.Suite 500Washington, DC 20005Tel! 202-265-8010Fax/202-332-6652

Minneapolis, MN310 Fourth Avenue S.Suite 600Minneapolis, MN 55415Tel! 612-871-8910Fax/612-871-9270

Statistical evidence generally has little importance in individual employmentdiscrimination cases and, to the extent it is presented at all, often consists of little morethan demographic information concerning employees in the plaintiffs business unit orjob group. In class action employment discrimination cases, by contrast, statisticalevidence is critical. Counsel for both parties typically engage statistical experts. Theseexperts prepare reports and are deposed in connection with plaintiffs' class certificationmotion, and if the case proceeds to a liability trial, the statistical experts may be eachparty's most important witness. As the Supreme Court has stated, "our cases make itunmistakably clear that '[s]tatistical analyses have served and will continue to serve animportant role' in cases in which the existence of discrimination is a disputed issue."Int'l Bhd. a/Teamsters v. United States, 431 U.S. 324, 339 (1973).

The importance of statistics in employment discrimination class action litigationinevitably results in vigorously contested disputes about the use ofboth statistical expertsand evidence. In this paper I discuss four issues that have been arising in multipleemployment discrimination class actions over the past few years: what weight should begiven to the EEOC's "four-fifths" test, how to deal with missing data or variables omittedfrom an analysis, what type of statistical showing supports or undermines certification ofa class, and how to account (if at all) for supposed differences between members of theprotected group and other employees that are purportedly the product of social forces orinherent differences and not the product of any discrimination by an employer. 1

I. THE EEOC'S FOUR-FIFTHS TEST AND/OR STANDARD DEVIATIONS

Federal employment laws do not state that statistical analyses should or may beused in deciding employment cases, let alone identify an appropriate statistical standardfor evaluating whether an employer's decisions concerning hiring, promotions, payorterminations give rise to an inference of illegal discrimination. In that vacuum, courtshave most often referred to two types of standard.

The first is derived from an EEOC regulation, 29 C.F.R. § 1607.4(D), whichprovides in part:

A selection rate for any race, sex, or ethnic group which is less than fourfifths ( 4/5 ) (or eighty percent) of the rate for the group with the highestrate will generally be regarded by the Federal enforcement agencies asevidence of adverse impact, while a greater than four-fifths rate willgenerally not be regarded by Federal enforcement agencies as evidence ofadverse impact. Smaller differences in selection rate may neverthelessconstitute adverse impact, where they are significant in both statistical andpractical terms or where a user's actions have discouraged applicantsdisproportionately on grounds ofrace, sex, or ethnic group. Greaterdifferences in selection rate may not constitute adverse impact where thedifferences are based on small numbers and are not statistically

I Parts II and III of this paper are edited and updated versions of a paper I prepared for aseminar hosted by ERS Group in the spring of 2009.

Sprenger <> Lang PLLCATTORNEYS AT lAW

1

significant, or where special recruiting or other programs cause the pool ofminority or female candidates to be atypical of the normal pool ofapplicants from that group.

Twenty years ago, the Supreme Court referred to this test somewhat dismissively:"This enforcement standard has been criticized on technical grounds, and it has notprovided more than a rule of thumb for the courts." Watson v. Fort Worth Bank & Trust,487 U.S. 977, 995 (1988) (citations omitted). Presumably because of these critiques,federal courts tended for most of the past twenty years to give this test short shrift in theiranalyses, ifthey mentioned it at all. See, e.g., Isabel v. City ofMemphis, 404 F.3d 404,412-13 (6th Cir. 2005) (rejecting employer's argument that plaintiffs had not establisheddisparate impact when, even though test produced racial disparity in excess of twostandard deviations, results did not fail the four-fifths test).

Instead, at least until a year or two ago, courts generally evaluated differences inthe treatment of two groups of employees in terms of standard deviations. A standarddeviation analysis is more sophisticated than the EEOC's 80% test. When there are twogroups and a binary result (yes/no), a standard deviation is "the square root of the productof the total number in the sample ... times the probability of selecting [a member ofgroup A] times the probability of selecting [a member of group B]." Castaneda v.Partida, 430 U.S. 482, 496 (1977). Thus, it takes into account both the size ofthepopulation and the percentages of each group in that population.

The Supreme Court established a rough standard more than thirty years ago whenit explained that "a fluctuation ofmore than two or three standard deviations wouldundercut the hypothesis that decisions were being made randomly with respect to race."Hazelwood School Dist. v. United States, 433 U.S. 299, 311 n.l7 (1977). This is veryrough because the difference between two and three standard deviations is huge: a twostandard deviation would occur about one time in 20 as a result of random chance,whereas a difference of three standard deviations would occur about one time in 500.Most appellate courts have resolved this murkiness by holding that a disparity of twostandard deviations is sufficient to create an inference that the decision-making was notrandom as to an illegal factor such as race. See, e.g., Brown v. Nucor Corp., 2009 U.S.App. LEXIS 17643, at *18-19 n.9 (4th Cir. Aug. 7,2009).

Indications are that, for reasons unknown because it is still inferior technically toa standard deviation analysis, the EEOC's four-fifths test is being viewed with new favorby the federal courts. A recent decision of the Supreme Court is Exhibit A. In Ricci v.deStefano, 129 S. Ct. 2658 (2009), the Court quoted only the last part of Watson'scharacterization of the test in a manner that completely changed the thrust of the quote -the EEOC's 80-percent standard is "a rule of thumb for the courts." Id. at 2678. TheCourt also stated that "[t]he pass rates ofminorities, which were approximately one-halfthe pass rates for white candidates, fall well below the 80-percent standard set by theEEOC to implement the disparate-impact provision ofTitle VII." Id. The Court nevereven mentioned "standard deviations."

Sprenger 4~ Lang PLLCATTORNEYS AT LAW

2

In at least one decision this year, a district court, as in Ricci, looked primarily tothe four-fifths test in evaluating the statistics. See Mo.C.HA. Society, Inc. v. City ofBuffalo, 2009 U.S. Dist. LEXIS 20070, at *32-34 (W.D.N.Y. Mar. 9,2009) (althoughacknowledging criticisms of four-fifths test on technical grounds, district courtcharacterizes it as "providing useful guidance for determining the significance orsubstantiality of numerical disparities on a case-by-case basis" and relies primarily onthis test in concluding that promotional examination had a disparate impact on AfricanAmerican firefighters). In another case, a district court primarily evaluated the data interms of whether there was a two standard deviation disparity, but also used the fourfifths test as corroboration for its conclusion. See Stagi v. National Railroad PassengerCorp., 2009 U.S. Dist. LEXIS 71207, at *45-46 (E.D. Pa. Aug. 12,2009) (aftercriticizing plaintiffs' expert analysis that yielded disparities in excess of two standarddeviations, the court concluded that even if the methodology were sound, the disparitiesdo not have practical significance because the gender differences did not reach the 80%level). And in a third case, a district court, while ultimately rejecting an employer'sargument that plaintiffs had not proven disparate impact to the extent that disparities didnot reach the 80% level and concluding that plaintiffs had met their burden based on astandard deviation analysis, explained that neither test "always answers the question." Ineffect, it suggested that the two tests had roughly equal value. See United States v. CityofNew York, 2009 U.S. Dist. LEXIS 63153, at *26-27 (E.D.N.Y. July 22,2009).

Hopefully courts will continue to use standard deviation analysis almostexclusively. The four-fifths test is a blunt instrument, and there are no circumstancesunder which it gives more accurate results than a standard deviation analysis. Its primaryutility should be in helping lawyers evaluate a case before a standard deviation analysishas been performed.

II. MISSING VARIABLES AND MISSING DATA

Twenty years ago, the Supreme Court stated in Bazemore v. Friday, 478 U.S. 385(1986):

While the omission of variables from a regression analysis mayrender the analysis less probative than it otherwise might be, it can hardlybe said, absent some other infirmity, that an analysis which accounts forthe major factors "must be considered unacceptable as evidence ofdiscrimination." Normally, failure to include variables will affect theanalysis' probativeness, not its admissibility.

Id. at 400 (citation omitted). The Court acknowledged, however, that "there may, ofcourse, be some regressions so incomplete as to be inadmissible as irrelevant." Id. at 400n.10.

Frequently, one party, generally the defendant, argues that the other party's experthas omitted one or more variables from an analysis so as to render the analysisinadmissible. As the D.C. Circuit has characterized the issue, the issue then becomeswhether the omitted variables are "clearly major variables." Coward v. ADTSec. Sys.,

Sprenger ~> Lang PLLCATTORNEYS AT LAW

3

140 F.3d 271,274 (D.C. Cir. 1998) (quoting Koger v. Reno, 98 F.3d 631,637 (D.C. Cir.1996)). In the same two opinions, the court classified education and prior workexperience as "clearly major variables." Identifying whether omitted factors are "major"and how to proceed ifthe omitted variables are "major," however, is not always easy.

For example, plaintiffs' statistical analysis survived defendant's summaryjudgment motion in McReynolds v. Sodexho Marriott Servs., 349 F. Supp. 2d 1 (D.D.C.2004), although it initially did not control for education and experience. The electronicdatabase did not contain such information. The defense expert, however, reconstructedsuch information for a sample of about 20% of employees from paper records, andconcluded that with that information included in the analysis, there were no statisticallysignificant disparities. !d. at 23-24. Despite this prodigious effort, the court did not grantsummary judgment. Plaintiffs attempted to poke various holes in the defense analysis,and by making various adjustments to it, were able to generate new analyses showingstatistically significant disparities. !d. at 24-27.

McReynolds is hardly alone in giving weight to analyses that fail to control forapparently "major" variables. See, e.g., Velez v. Novartis Pharmaceuticals Corp., 244F.R.D. 243, 260 (S.D.N.Y. 2007) (concluding that there was a common statisticalquestion as to disparities in performance ratings even though plaintiffs' analysis failed tocontrol for employees' job level, which would have eliminated the reported disparity).

On the other hand, courts in many recent class cases have rejected expert analysesperformed for plaintiffs based on omission of one or more factors. See Anderson v.Westinghouse Savannah River Co., 406 F.3d 248,262-63,266 (4th Cir. 2005) (affirmingexclusion of plaintiffs' expert's testimony concerning ranking system used to set payraises when expert used EEO job groupings instead ofjob titles to control for job andexpert failed to control for education and experience); Morgan v. UPS, 380 F.3d 459 (8thCir. 2004) (affirming grant of summary judgment when plaintiffs' expert failed to controlfor past pay and plaintiffs had not shown that past pay decisions were discriminatory andfailed to control for all years ofperformance appraisal data that existed and may havebeen relevant); Bennett v. Nucor Corp., 2006 u.s. Dist. LEXIS 61609, at *44 (E.D. Ark.Aug. 25, 2006) (holding that commonality requirement not established because plaintiffs'expert "failed to eliminate common non-discriminatory reasons for the disparity in ratesof disciplinary action and training between white and black applicants"); Jones v. GPO,Inc., 2005 U.S. Dist. LEXIS 18820, at *38-41 (E.D. Pa. Sep. 1,2005) (denying motionfor class certification in part because plaintiffs' expert omitted crucial variables such aseducation and work experience); Yapp v. Union Pac. R.R. Co., [FULL CITE}, 229 F.R.D.at 620 (denying motion for class certification in part because plaintiffs' expert did notcontrol for department or conduct a thorough inquiry into minimal qualifications); seealso Ellis v. Costco Wholesale Corp., 240 F.RD. 627, 647-48 (N.D. Cal. 2007) (excludingone ofplaintiffs' analyses for failure to include data from six months during liabilityperiod).

The issue becomes more difficult if the expert fails to control for a variablebecause the data to control for that variable have not been produced. Over the past fewdecades, statistical experts typically have relied primarily on parties' computer-readable

Sprenger <~ Lang PLLCATTORNEYS AT LAW

4

human resources and payroll databases rather than having to create databases from paperrecords. The databases produced to the experts, however, frequently are incomplete,because of changes in databases over the years, failure to record any information at allabout relevant information, especially about prior qualifications such as education andprior job experience, and gaps in otherwise-populated fields of data. 2

Although plaintiffs generally lose ifboth parties' statistical analyses areinadequate because plaintiffs have the burden ofproof on both class certification andliability, many courts have held that plaintiffs' claims should not be rejected when anexpert report does not control for an important factor because the data simply does notexist. "[P]laintiffs cannot legitimately be faulted for gaps in their statistical analysiswhen the information necessary to close those gaps was possessed only by defendantsand was not furnished either to plaintiffs or to the Court." Palmer v. Shultz, 815 F.2d 84,110 (D.C. Cir. 1987) (quoting Troutv. Lehman, 702 F.2d 1094,1102 (D.C. Cir. 1983),vacated on other grounds, 465 U.S. 1056 (1984)); see EEOC v. Radiator Specialty Co.,610 F.2d 178, 185 n.8 (4th Cir. 1979) (the allocation of the burden ofproof may fall onthe party "with the most ready access to the relevant information"). As the D.C. Circuitexplained:

The appropriate degree ofrefinement of the plaintiffs' statisticalanalysis, moreover, may depend on the quality and control of the availabledata. See Vuyanich [v. Republic National Bank}, 505 F. Supp. [224,] 356[(N.D. Tex. 1980]. Ifthe plaintiffs account for the effects of extraneousvariables to the extent reasonably permitted by the available data and theevidence presented strongly supports an inference of discriminatorytreatment, the District Court properly may conclude that the plaintiffs havemade out a prima facie case.... ["]Deficiencies in the data base 'may, ofcourse, detract from the value of [statistical evidence,' but ordinarilywould not obliterate its evidentiary value." [Detroit Police Officers' Ass 'nv. Young, 608 F.2d 671,] 687 (citations omitted) (quoting [Int'! Bhd. oj]Teamsters [v. United States,] 431 U.S. [324,] 34Q [(1977)].

2 Sometimes the data made available to the expert witnesses are incomplete because theemployer reads plaintiffs' document requests narrowly and refuses to produce data. Thisstrategy is, I believe, often flawed. Ifthe employer fails to produce data reasonablywithin the scope of a document production request, a court is likely to prohibit thedefense expert from using such data, and the defendant might find later in the litigationthat the data are important to its defense. Similarly, if otherwise populated fields of datahave missing records for a substantial number of employees, defense counsel's decisionto withhold from plaintiffs a category of information from which such data could havebeen filled in, whether because production would be overly burdensome or otherwise,may prove shortsighted because it may justify a judicial decision to excuse the failure ofplaintiffs' expert to fill in the gaps.


5

Trout v. Lehman, 702 F.2d at 1101-02. See also Brown v. Nucor Corp., 2009 U.S. App.LEXIS 17643, at *10-18 (4th Cir. Aug. 7, 2009) (directing district court to certify a plantwide promotion class based on statistics that used change-of-status forms to identifypromotions for two years and assumed that the racial composition of the bidding pool forthose jobs was the same as in other years, when defendant had destroyed the data for thetwo earliest years of the liability period); McClain v. Lufkin Indus., Inc., 519 F.3d 264,280 (5th Cir. 2008) (affirming district court decision that gave greater credence toplaintiffs' promotion analysis based on potential pools than defendant's analysis based onbid sheets when "the bid data were incomplete and unreliable, [with] more than half ofthe promotions ... missing from the bid database which Lufkin prepared in anticipationof this litigation); Wright v. Stern, 450 F. Supp. 2d 335 (S.D.N.Y. 2006) (rejecting attackon plaintiffs' report "when most of the alleged errors cited by defendants ... results fromthe nature of the raw data maintained and provided to plaintiffs by [defendant]"); but seePhillips v. Gates, 2009 U.S. App. LEXIS 9706, at *10-15 (6th Cir. May 5,2009)(affirming bench trial ruling for employer when experts' analyses to account for missingdata were found to be "equally problematic" and hence employees had not carried theirburden, even though some of that data was destroyed in violation of a retention order).Courts recognize the principle that plaintiffs generally should not be punished when dataare unavailable even when ruling against plaintiffs for other reasons. In Pottenger v.Potlatch Corp., 329 F.3d 740 (9th Cir. 2003), the court rejected plaintiffs' analyses forfailing to take certain variables into account but the court held that it would have ruleddifferently if the information were not available. Id. at 748.

Plaintiffs, however, generally cannot decline to include a variable because it isavailable only in paper records, not computer-readable information. See Anderson v.Boeing Co., 2006 U.S. Dist. LEXIS 76964, at *34 (N.D. Okla. Oct. 18,2006) (rejectingplaintiffs' analysis that used a proxy instead of the actual groups from which decisionswere made: "plaintiffs cannot base their statistical analysis on different measures solelybecause the more accurate data is more difficult to obtain"). Nor can plaintiffs rely oncomputer-readable data if it appears to be seriously flawed and alternatives are available.In Williams v. Boeing Co., 2006 U.S. Dist. LEXIS 3417 (W.D. Wash. Jan. 17,2006), forexample, plaintiffs' expert relied on an indicator in the employer's database to identifypromotions. The employer, however, was able to show that it had not assigned theindicator to a significant number ofpromotions. Id. at *10-11. The defense expert usedsalary increase as a proxy for promotions. The judge acknowledged that salary growthwas an imperfect proxy, but still found the defense expert's analysis to be "moreprobative." Id. at *11-12.

Proxies may be used in other contexts as well. For companies using "tap on theshoulder" selection systems, there are unlikely to be any applications to measure levels ofinterest. In Wards Cove Packing Co., Inc. v. Atonia, 490 U.S. 642 (1989), the SupremeCourt endorsed the use of either applicant flow or labor force analyses injob selectioncases. Id. at 650-51. For hiring claims, labor force analyses often will rely on census orother data not restricted to the employer to calculate the percentage ofminorities orwomen in categories ofjobs (often in the geographical area in which the employer isoperating), with the assumption that absent discrimination the employer probably couldhave hired roughly that percentage ofminorities or women. For promotion claims,

Sprenger ~> Lang pucATTORNEYS AT LAW

6

experts are likely to use the relevant characteristics ofthe selected employee to identifythe members ofthe relevant internal labor pool (e.g. all employees with the same jobgrade, same type ofjob, and at least the same number of years of experience as theselected employee). In the absence of reliable applicant data, courts often find such laborforce analyses acceptable. See, e.g., Maleve v. Potter, 320 F.3d 321, 326-27 (2d Cir.2003) (reversing a finding for the employer because the plaintiffs' expert properly usedlabor force data when applicant data was unavailable).

Instead ofproxies, scientifically selected samples may be used, as in McReynolds,in hiring and promotion cases when the universe of applications is voluminous and thoseapplications have not been entered into a database. In Rhodes v. Cracker Barrel OldCountry Store, Inc., 213 F.R.D. 619, 655-56 (N.D. Ga. 2003), for example, the court heldthat defendant had established differing levels of interest based on a sample showing that70% of non-African Americans applied for server jobs but only 44% of AfricanAmericans did.

Rather than rely on labor force statistics or samples, defendants or their expertsoccasionally have tried to use surveys to fill in gaps in the data, often concerning interestin certain positions at issue in the case. Surveys prepared for purposes of discriminationlitigation generally have failed because of attorney involvement that may taint the results,poor design, or other reasons. See Dukes v. Wal-Mart, Inc., 222 F.R.D. 189, 196-97(N.D. Cal. 2004) (striking testimony based on store manager survey designed andadministered by counsel that exhibits bias on its face);3 Yapp v. Union Pac. R.R. Co., 301F. Supp. 2d 1030, 1037 (E.D. Mo. 2004) (granting motion to strike defendant's expertreport because its conclusion that analysis should be performed on a departinent-bydepartment basis was based on a non-random survey, defense counsel was heavilyinvolved "in the design and conduct of [the] survey," and the survey otherwise did notuse "reliable scientific methods"). Nonetheless, there is no reason why a survey designedto fill in data gaps inherently is inadmissible if designed and administered in ascientifically reliable manner by experts without attorney involvement.

When a database contains no information concerning employees' workexperience before they commenced employment with the defendant, experts may attemptto estimate such experience through measures such as "time since highest degree" or"age minus 21 minus tenure" to approximate prior experience. Several courts haverejected use ofthis type ofmeasure. See Cooper v. Southern Co., 390 F.3d 695, 717(11 th Cir. 2004) (holding that period oftime since employee finished formal education"may have measured the passage of time, [but] did not in any way factor in the quality,type, or relevance of an employee's experience"); Trout v. Garrett, 1990 u.S. Dist.LEXIS 10060, at *15-24 (D.D.C. Mar. 30, 1991) (citing several other decisions andholding that an age proxy model for experience is substantially biased in favor ofwomen). Some courts have refused to invalidate plaintiffs' analyses that failed to employ

3 A Ninth Circuit panel has now issued two opinions concerning class certification inDukes, but because the Ninth Circuit has now vacated those opinions and heard argumenten bane, I do not cite to either panel decision.

Sprenger ~~ Lang PLLCATTORNEYS AT LAW

7

such a variable and made no attempt to control for prior experience. See Taylor v.District ofColumbia Water & Sewer Auth., 241 F.R.D. 33,43 (D.D.C. 2007).

The expert's alternative to not controlling for prior experience or using a proxy isto recreate from resumes, applications, or other records (if they exist) the prior experienceof a sample or the full universe of employees. This approach also faces an analyticalproblem: how to value the actual experience. Someone presumably must determine inan unbiased manner ifthe actual experience is relevant to the type ofjob for whichallegedly discriminatory selections have been made.

As the above survey of the cases indicates, there is no single rule to follow when adatabase is missing important data. Judges have dealt with data gap issues on a case-bycase basis, often seemingly based on their own predispositions. Perhaps the bestguidance is to determine how important the missing data is likely to be, how defensiblethe use of a proxy would be, whether underlying paper information is likely to bereasonably complete and, if not, is likely not to be systematically skewed, and the cost ofinputting the data from the paper records.

III. CLASS CERTIFICATION: STATISTICAL QUESTIONS COMMON TOTHE PROPOSED CLASS

A. The Commonality Requirement in Employment Cases

The requirements for certification of a class action are set out in Rules 23(a) and(b). In determining whether plaintiffs have satisfied those requirements, courts attempt tobalance two principles. They strive to avoid "conduct[ing] a preliminary inquiry into themerits of a suit," Eisen v. Carlisle & Jacquelin, 417 U.S. 156, 177 (1974), whileengaging in a "rigorous analysis" to ensure "actual, not presumed, conformance" with thecriteria for class certification. Gen. Tel. Co. ofthe Southwest v. Falcon, 457 U.S. 147,160-61 (1982).

In attempting to strike this balance, appellate courts increasingly have beendirecting district courts to resolve any factual or legal issues necessary to decide classcertification, but not to reach beyond the Rule 23 elements to resolve other factual orlegal issues. In the most comprehensive recent discussion, the Second Circuit explainedthat a district court judge must "resolve[] factual disputes relevant to each Rule 23requirement" even ifthose disputes "overlap with a merits issue." In re Initial PublicOffering Securities Litigation ("IPO"), 471 F.3d 24, 41 (2d Cir. 2006). But, the SecondCircuit added, a district court must not "assess any aspect of the merits unrelated to aRule 23 requirement" and must ensure "that a class certification motion does not becomea pretext for a partial trial of the merits." Id.

The party seeking certification under this level of review must show under Rule23(a) that: (i) the class is so numerous that joinder of all members would beimpracticable ("numerosity"); (ii) the claims "raise questions oflaw or fact common tothe class" ("commonality"), (iii) the claims of the proposed class representatives aretypical ofthose of other class members ("typicality"); and (iv) the proposed class

Sprenger ~~ Lang PLLCATTORNEYS AT lAW

8

representatives can fairly and adequately represent the class ("adequacy"). In addition,the party seeking certification must satisfy one of the three subsections of Rule 23(b). Inemployment discrimination cases, the two prongs ofRule 23(b) most often at issue are(b)(2), which requires that the party opposing the class have acted in a manner generallyapplicable to the class that justifies injunctive or declaratory relief, and (b)(3), whichrequires that common issues predominate over individual issues ("predominance") andthat a class action be superior to other procedural mechanisms for resolving the dispute.

At the class certification stage of the case, the statistical expert's testimony maytouch on numerosity by calculating the number of class members, but in almost all casesthe primary focus is on commonality, typicality and (if certification is sought under Rule23(b)(3)) predominance. Predominance is related to commonality in that the issue iswhether those common questions that permit plaintiffs to satisfy the commonalityrequirement predominate over questions unique to individual class members. For ease ofdiscussion, I will refer only to commonality in this paper.

Parsing the language of the commonality clause, it actually creates twoanalytically distinct sub-requirements: the claims must raise one or more "questions,"and those questions must be common to the class members. Practitioners often lose classcertification motions, and judges often write muddled opinions, by not focusingseparately on each sub-requirement.

Whether statistical reports create a question and whether that question is commonto class members requires also an understanding of the two types of employmentdiscrimination class action claims: disparate treatment "pattern or practice" claims anddisparate impact claims. In a "pattern or practice" claim, the plaintiffs must prove thatintentional discrimination is the employer's "regular rather than the unusual practice."Teamsters, 431 U.S. at 336. Although plaintiffs also should present policy and anecdotalevidence, statistical evidence often is the most important type of evidence to establish aprimafacie pattern or practice case. See Ottaviani v. State Univ. ofNY., 875 F.2d 365,371 (2d Cir. 1989); Allard v. Ind. Bell Tel. Co., 1 F. Supp. 2d 898, 909 (S.D. Ind. 1998).

By contrast, in a disparate impact case, plaintiffs need not prove that the employerhad a discriminatory motive, but only that an employment practice adversely affected theclass members. Griggs v. Duke Power Co., 401 U.S. 424,429-30 (1971). Statisticsprovide the primary means of establishing the prima facie elements of a disparate impactclaim is statistical. From the point of view of statistical evidence, there is little differencebetween a pattern or practice and a disparate impact claim other than in the specificity ofthe practice being analyzed.

B. Is There a Bona Fide Statistical Dispute?

Turning to the first "commonality" sub-requirement, a question exists when thereis a bona fide dispute. If plaintiffs have presented an analysis showing statisticallysignificant disparities between the effects of an employer practice on the members of twogroups, and defendants present a counter-analysis showing that there are no statistically


9

significant disparities, then a question exists, unless one analysis is so flawed that theapparent question does not really exist.

In the federal courts, the standards that an expert analysis must meet to beadmissible are set forth in Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 u.s. 579(1993). At class certification, there generally is no evidentiary hearing and hence,technically, admissibility is not an issue. However, if plaintiffs' expert's analysis doesnot meet the Daubert standards at class certification, then plaintiffs have notdemonstrated that there is a dispute, or question, and hence have not settled thecommonality requirement.

This does not mean that defendants must file a Daubert motion to disprove theexistence of a question. Instead, it could show in its opposition to class certification thatplaintiffs' report does not meet the Daubert standards. It should be required, however, topresent such an argument by motion or by opposition to plaintiffs' class certificationmotion in order to oppose commonality on the ground that there is no statistical question.

In recent years, some courts expressly or impliedly have concluded, incorrectly,that Daubert challenges are premature at the class certification stage. See Serrano v.Cintas Corp., slip op. at 4 (E.D. Mich. Mar. 31, 2009) ("A Daubert analysis at this stageoflitigation is unnecessary."); Taylor v. District ofColumbia Water & Sewer Auth., 241F.R.D. 33,43 (D.D.C. 2007) ('''statistical dueling' is irrelevant to the certificationdetermination"); Bennett v. Nucor Corp., 2006 U.S. Dist. LEXIS 61609, at *31 (E.D.Ark. Aug. 25, 2006) (explaining that expert "testimony should be judged on the basis ofwhether it supports class certification" and that whether it "is admissible can be raisedlater in the proceedings"); Williams v. Boeing Co., 225 F.R.D. 626,635 (W.D. Wash.2005) (declining to rule on defendants' challenge to plaintiffs' statistical analysis ofcompensation and promotions). Other courts have imposed a standard that may havebeen similar to Daubert, but have not referenced Daubert or otherwise sufficientlydefined their standards to be certain. See Gutierrez v. Johnson & Johnson, 2006 U.S.Dist. LEXIS 80834, at *7 (D.N.J. Nov~ 6,2006) (following Second Circuit in holding that"the appropriate level of review for the admissibility of expert testimony in classcertification hearings" is "to ensure that they are not so fatally flawed to be inadmissibleas a matter oflaw"); Hnot v. Willis Group Holdings Ltd., 228 F.R.D. 476, 483-84(S.D.N.Y. 2005) (certifying class despite defendant's challenge to plaintiffs statisticalevidence because "at the class certification stage, only a plausible position needs to be setforth,,).4 Other courts properly have evaluated expert reports submitted in support ofclass certification under Daubert. See, e.g., Dukes v. Wal-Mart, Inc., 222 F.R.D. 189(N.D. Cal. 2004).

4 The Second Circuit remanded Hnot to the district court after its decision in lPG, and thedistrict court concluded that it had reached the correct result under the "rigorous analysis"mandated by IPo. Hnot v. Willis Group Holdings Ltd., 241 F.R.D. 204, 210 (S.D.N.Y.2007) (explaining that, under lPG, the disagreement over "whose statistical findings andobservations are more credible ... is relevant only to the merits ofplaintiffs' claim andnot to whether plaintiffs have asserted common questions of law or fact")


10

While some courts have erred on the side of leniency, other courts have imposed alevel of scrutiny that, regardless of what they may say is their standard, appears to go farbeyond determining admissibility. For example, in Rhodes v. Cracker Barrel OldCountry Store, Inc., 213 F.R.D. 619 (N.D. Ga. 2002), the court analyzed the validity ofthe analysis of plaintiffs' expert at length in light ofthe criticisms leveled by defenseexperts. Id. at 654-69. The court concluded that plaintiffs' expert's hiring, promotionand pay analyses were flawed in various ways. For example, the promotion analysesfailed to control for location, for interest as reflected in a company-wide surveyconducted by the company before the lawsuit was filed, and for whether or not anemployee chose to take the required test for promotion. There was no attempt to considerthe report in light of the Daubert factors. See also Cooper v. Southern Co. COMPLETE

Of course, even when a court applies a Daubert-like standard, differences inquality of expert reports and among judicial perceptions of the minimum necessary tosatisfy the admissibility standard are likely to lead to varying outcomes. CompareGarcia v. Johanns, 2006 U.S. App. LEXIS 7892, at *22-24 (D.C. Cir. 2006) (affirmingdenial of class certification in lending discrimination case against U.S.D.A. whenplaintiffs' expert failed to use multiple regression analysis and thereby failed to accountfor relevant factors) and Remien v. EMC Corp., 2008 U.S. Dist. LEXIS 15940 (N.D. Ill.Mar. 3, 2008) (striking plaintiffs' expert declaration in support of class certificationbecause of many analytical flaws, including failure to use data available to measuredispute in litigation and imposing arbitrary limits on persons included in analysis) withArnold v. Cargill, Inc., 2006 U.S. Dist. LEXIS 41555, at *28 (D. Minn. June 20,2006)(concluding that defendant's assertion that plaintiffs' expert "should have controlled forcertain variables impacts the Court's determination of what weight to assign [his] reportwhen considering whether to grant class certification, but does not impact the issue ofadmissibility"). However, judicial use of a consistent standard tied to the language ofRule 23(a)(2) should lead to greater consistency of results as to whether opposing expertreports raise a "question."

C. Is the Dispute Common to the Proposed Class?

The language ofRule 23(a)(2) doesn't require just that there are "questions"; itrequires that the "questions" be common to the class. While the standard for showing theexistence of a "question" should be lenient - a disagreement in the conclusions betweentwo admissible expert reports creates a "question" - the standard for showing that thequestion is common to the class should be higher.

The mere fact that plaintiffs' and defendants' experts disagree about whether adisparity between all ofthe members of a proposed class and persons outside the class isstatistically significant does not mean that the question is common to the entire class.The disparity may exist entirely in certain business units, job grades, or job families, andthe results may be statistically neutral for the rest ofthe proposed class.

Unfortunately, no statistical test determines whether a statistical disparity iscommon to a class. Several defendants have argued that a sophisticated analysis called a


11

12

"Chow test" is such a test, but courts have virtually uniformly rejected the argument.s

See Taylor, 241 F.R.D. at 43 (rejecting criticism ofplaintiffs' analysis based on Chow testas "neither supported by authority nor the circumstances presented here"); Dukes, 222F.R.D. at 157-58 (declining to exclude report of plaintiffs' expert for failure at classcertification stage to perform a Chow test before aggregating data on a regional level asthe present factual record does not "compel the conclusion that a Chow test must beperformed"); Rossini v. Ggilvy & Mather, 615 F. Supp. 1520, 1522-23 (S.D.N.Y. 1985),vacated and remanded on other grounds, 798 F.2d 590 (2d Cir. 1986) (finding indecision on merits that "categorical rejection" of a single regression combining all databefore and after limitations date was "not warranted" despite results of Chow testshowing statistically significant differences in regressions run on data before and afterthat date); Vuyanich v. Republic Nat'l Bank, 505 F. Supp. 224,299,314 (N.D. Tex.1980), vacated on other grounds, 723 F.2d 1195 (5th Cir. 1984) ("Because thecontroversy here [regarding the Chow test] appears to center on an issue on the frontier ofeconometrics, and there seems, at least to a court unschooled in the intricacies ofeconometrics, to be genuine conflict between the experts as to the proper approach, we donot decide the issue"). But see Coates v. Johnson & Johnson, 756 F.2d at 541-42(affirming as not clearly erroneous trial court decision not to pool data across years in anon-employment case based on results of Chow test conducted by defendant's expert).

Without a single definitive test, courts have decided, and must continue to decide,in a more impressionistic manner whether a statistical dispute is common to the class.The very nature of employment discrimination class actions, in contrast to securities andantitrust class actions, should help plaintiffs.

In securities fraud cases, adverse reliance is an element ofliability. Accordingly,unless plaintiffs can establish the conditions for use of the fraud-on-the-marketpresumption of reliance, established in Basic Inc. v. Levinson, 485 U.S. 224, 245-47(1988), they must show adverse reliance by each member of the proposed class for it toprevail. This runs afoul of Rule 23(b)(3)'s "predominance" requirement. (Plaintiffs donot generally seek meaningful injunctive relief in securities class actions, foreclosingcertification under Rule 23(b)(2).) See, e.g., Teamsters Local 445 Freight Div. PensionFund v. Bombardier, Inc., 546 F.3d 196, 199-200 (2d Cir. 2008); lPG, 471 F.3d at 42-43;Gariety v. Grant Thornton, LLP, 368 F.3d 356, 367-68 (4th Cir. 2004). Antitrust classaction plaintiffs face similar difficulties because of the need to show that each classmember suffered antitrust injury as an element ofliability. See, e.g., In re HydrogenPeroxide Antitrust Litig., 552 F.2d 305, 311-12, 322-25; Blades v. Monsanto Co., 400F.3d 562, 572-74 (8th Cir. 2005).

S To use a Chow test, an expert first must identify the independent variables, includingthe protected class indicator, that the expert believes should be used as controls, e.g. job,grade, education level, tenure, and gender, in a multiple regression analysis for whichcompensation (or theoretically another type of employment decision) is the dependentvariable. The expert then divides an employer into a number of decision-making units,which could correspond to departments, locations, or other organizational units. TheChow test examines whether the values of the independent variables yielded by amultiple regression analysis for each unit differ to a statistically significant extent.

Sprenger <~ Lang PLLCATTORNEYS AT lAW

By contrast, in an employment discrimination class case, a class can prevail on a"pattern or practice" disparate treatment claim or a disparate impact claim, obtaininjunctive relief for the class as a whole, and shift the burden to the defendant to disprovethat it discriminated against any class member who seeks individual relief, withoutoffering any proof about any individual member. See generally Int 'I Bhd. a/Teamsters v.United States, 431 u.s. 324, 360-62 (1977). Evidence that tends to prove or disprove a"pattern or practice" or disparate impact is common to the affected class regardlesswhether every class member has suffered discrimination.

But an overall statistically significant disparity between the treatment of theproposed class and other employees plus the nature of employment discrimination classactions generally is insufficient for plaintiffs to show that a statistical dispute is commonto the class. It is important also to show that the policies or practices that supposedlyproduced the disparities are common across the class. See, e.g., Yapp v. Union Pac. R.R.Co., 229 F.R.D. 608, 619-22 (E.D. Mo. 2005) (declining to find commonality because,among other things, plaintiffs failed to identify a common practice across the variousdepartments and plaintiffs' expert failed to give any valid reason for aggregating hisanalyses across departments instead of analyzing disparities on a department-bydepartment basis); but see Brown v. Nucor Corp., 2009 u.s. App. LEXIS at *17-18(reversing denial of certification to a class of African-American employees in amanufacturing plant when statistics showed an overall statistically significant disparitywithout any discussion of whether that disparity was similar throughout the sixproduction departments in the plant).

Most employment discrimination class actions over the past twenty or more yearshave claimed that the employer's policies are excessively subjective.6 Courts generallyare more reluctant to approve a class action in these cases than in cases involving theapplication of a single test or a single non-discretionary policy to large numbers ofemployees. See NAACP v. North Hudson Regional Fire & Rescue, 2009 u.s. Dist.LEXIS 12155, at *45 (D.N.J. Feb. 18,2009) (holding that plaintiffs, largely throughstatistics, had established that a residency requirement gave rise to a prima facie case ofdiscrimination). They fear that these types of claims will give rise to numerousindividualized decisions by scores or (if the company is sufficiently large) hundreds orthousands of decision-makers. See, e.g., Cintas, supra at 8-9; Reap v. Cant 'I Cas. Co.,199 F.R.D. 536, 544 (D.N.J. 2001).

Plaintiffs should attempt to give the court reason to believe that the subjectivepolicies are applied consistently against the members of the proposed class. Possiblenon-statistical evidence includes expert testimony from stereotyping experts or proof thathigh-level executives expressed biases against class members that spread to decisionmakers.

6 These cases are distinct from claims that the employer engaged in "a general policy ofdiscrimination" manifested through an "entirely subjective decision-making process[]."The Supreme Court in Falcon explained that such a claim, if substantiated, would beappropriate for class treatment, 457 U.S. at 159 n.15, but very seldom is an employer'sprocess purely subjective.


13

Prudent plaintiffs' lawyers, however, generally also should have their expertpresent statistical evidence to show that the overall disparity is not just a product of a fewbusiness units or job categories. Otherwise, they risk a judge concluding that plaintiffshad failed to establish commonality. But see Gutierrez, 2006 U.S. Dist. LEXIS 80834, at*10 (failing to address defendant's argument that plaintiffs' expert report assumedaggregation was proper rather than demonstrating "that her reported aggregate averagedifferences stem from common employment practices"). Plaintiffs have several potentialapproaches to this analysis, although none is perfect. I briefly identify four approachesbelow which can be employed separately or together. For ease of expression, I willdiscuss these approaches in the context of a hypothetical gender discrimination classaction in which the plaintiffs claim that women are paid less than similarly situated male

. employees.

First and most simple-minded, plaintiffs' expert can present simple averagecompensation figures for men and women in each business unit, grade, and/or job family,to show that women are paid less than men in each grouping. But if the expert hasemployed a multiple regression analysis that controls for numerous other variables toshow that, overall, there is a statistically significant disparity in pay between men andwomen and explained the need for controls, simple averages, while suggestive, areunlikely to be convincing. They are best used in combination with one ofthe other typesof analysis discussed below.

Second, the expert could total the residuals (the difference between the actual andexpected values for each employee) associated with the employees in each business unit,grade, and/or job family and determine whether, for each unit, grade and/or family, theresiduals are positive or negative for women and the size of the average residual. This ispreferable to the first approach because it controls for all the same variables used in theoverall analysis. A potential line of attack, however, is that it will not show whether thedisparity in each unit, grade, and/or job family is statistically significant, and some judgesseem to think, erroneously, that plaintiffs must show that the disparities in each group aresimilarly statistically significant. See Puffer v. Allstate Ins. Co., 2009 U.S. Dist. LEXIS3533, at *40 (N.D. Ill. Jan. 15,2009) (holding that "although most of [plaintiffs'] results[by grade level] exceed the level oftwo standard deviations," plaintiffs had not showncommonality partly because of "the wide discrepancy in standard deviations betweengrades"). Plaintiffs must educate a judge that, even ifthe employer identicallydiscriminated against women throughout the company, the disparities would vary fromone unit, grade or family to the next. No analysis can explain 100% ofthe decisions,especially with subjective practices. In my experience, regression or other statisticalanalyses in employment discrimination cases, as evaluated by the R2 measure, typicallyvary from 20% to 90%. The unmeasured factors, which typically are associated withfrom 80% to 10% of the dependent variable, can cause huge variances from one unit,grade or family to the next even if the impact of a discriminatory practice is identical.See generally William T. Bielby & Pamela Coukos, "Statistical Dueling withUnconventional Weapons: What Courts Should Know About Experts in EmploymentDiscrimination Class Actions, 56 EMORY L.J. 1563, 1602, 1609-11 (2007).

Sprenger ~> Lang PLLCATTORNEYS AT lAW

14

Third, the expert could try to tum the tables on defendants and perform a Chowtest, or a Joint F test, to show that the values of the variables collectively, or the gendervariable in particular, in each business unit, grade and/or job family is sufficiently similarthat the results may be aggregated. Of course, judges who are not convinced by Chowtests or Joint F tests when used by defendants are no more likely to be convinced by themwhen used by plaintiffs. And these types of sophisticated tests require time to programand perform, adding to the expense ofthe statistical analysis.

Finally, the expert could perform separate multiple regression analyses for each ofthe business units, grades, and/or job families to show that the disparities, measured instandard deviations, are similar for each ofthem.7 This is probably the most expensiveapproach and is quite risky. It adopts defendants' typical "divide-and-conquer"methodology of dividing the workforce into a number of small units and analyzing eachone separately, a methodology that increases the likelihood that statistically significantdisparities will not be observed. See generally Bielby & Coukos, supra at 1599-1608(explaining how conduct of analyses of many small units decreases statistical power andthereby increases likelihood of yielding non-significant results). This type of approachalso may fall victim to judges who seem to want consistency ofresults across units. SeeCintas, supra at 10 ("even assuming Plaintiffs' own interpretation of the statistical data,although some Cintas locations under-hired women and racial minorities, other locationsover-hired women and racial minorities during the same time period"; moreover, "even atindividual locations, patterns ofunder-hiring and over-hiring fluctuated from year toyear"); Anderson v. Boeing Co., 2006 U.S. Dist. LEXIS 76964, at *12 (N.D. Okla. Oct.18, 2006 (finding "no evidence of class-wide discrimination" when in three business unitsmen were paid statistically significantly more than women, in many units there were nostatistically significant differences, and in some units, "women were paid slightly morethan men, though these numbers were not statistically significant"); Arnold v. Cargill,Inc., 2006 U.S. Dist. LEXIS 41555, at *50,52-53 (D. Minn. June 20,2006) (denyingclass certification because of variances in results among business units, with some unitsunfavorable to black employees to a statistically significant extent, some favorable toblack employees to a statistically significant extent, and most without statisticallysignificant differences). Some judges don't seem to realize that, even if there is uniformdiscrimination across the company, they should expect a variety of results across subgroups, shaped roughly like a bell curve (if there are a sufficient number of such subgroups). Some of the sub-groups may show positive compensation for women, even ifthere is a uniform practice of discrimination. See Bielby & Coukos, supra at 1606-08.However, if the judge understands that uniformity of results is not to be expected, a seriesof separate regressions showing a pronounced pattern adverse to women provides strongevidence in support of class certification.

7 Plaintiffs could limit this type of analysis to units large enough for meaningful statisticalanalyses, see Carlson v. CH. Robinson Worldwide, Inc., 2005 U.S. Dist. LEXIS 5674, at*27,30 (D. Minn. Mar. 31, 2005), or conduct sign tests on the units, i.e., show thatalthough most of the units are so small that any disparities are not statistically significant,there are statistically significant disparities between the number ofunits with negativeregression results and the number with positive results.


15

In opposition to this type of evidence, a defendant may be tempted to show that,using its expert's methodology, there is much greater variation in results among businessunits, grades, or job families than shown using plaintiffs' methodology. This type ofargument, however, should not logically suffice to defeat a showing that the statisticaldispute is common to the class. Ifplaintiffs' methodology, which passes Daubertscrutiny, shows that the dispute is common to class members, and defendant'smethodology shows that it is not, then the dispute over statistics remains common to theclass members: ifthe factfinder believes plaintiffs' statistical evidence, this evidence willbenefit all class members.

Defendants instead should attempt to show that, using plaintiffs' methodology,the statistical dispute is not common to the class. If, for example, plaintiffs attempt toshow a common pattern using the second type of analysis described above, defendants'expert may want to test the patterns using separate regression analyses. If plaintiffs haveevaluated the consistency of the disparities between men and women only from onebusiness unit to the next, it may be because they know or suspect that there would be alack of consistency in the disparities between men and women among grades or jobfamilies, and defendants might want their expert to examine the consistency of thedecisions along that vector.

If defendants believe that the results among business units, grades or job familiesare too inconsistent for commonality to be satisfied, they have two approaches. First, thedefendant can try to carve out pieces of the proposed class, showing that even if there aredisparities between the compensation ofmen and women in certain units, grades or jobs,there are no gender disparities in others. It can argue that those units, grades or jobsshould be excluded from the proposed class. Second, it can attempt to show that thevariations in the size of the disparities among all units, grades and jobs using plaintiffs'methodology are so great that the class should not be certified. Ifplaintiffs contestdefendants' characterization of the implications of their statistical report, some courtshave expressed, probably incorrectly, reluctance to resolve the dispute. See Velez v.Novartis Pharmaceuticals Corp., 244 F.R.D. 243,261 (S.D.N.Y. 2007) (concluding thatit is not appropriate to resolve at the class certification stage dispute over breadth ofpaydisparities, in which defendants claim that plaintiffs' statistics show statisticallysignificant disparities for only 3 of 12 job types, and plaintiffs respond that men werepaid more than women in 11 of 12 job types).

Not surprisingly, with judges often failing to analyze properly what type ofstatistical showing would suffice to support or undermine commonality, decisions nowvary greatly in result. Compare, e.g., Gaston v. Exelon Corp., 247 F.R.D. 75 (E.D. Pa.2007) (explaining that if a "decision by a company to give managers the discretion tomake employment decisions, and the subsequent exercise of that discretion by somemanagers in a discriminatory manner ... suffices to certify a class of all black employees,every employer that gives its managers discretion in employment matters will face asimilar suit") and Armstrong v. Powell, 230 F.R.D. 661,677 (W.D. Okla. Aug. 10,2005)(denying class certification in promotion case while giving short shrift to statisticalevidence because "[s]tatistical evidence ofunderrepresentation is, by itself, insufficient toestablish commonality") and Rhodes v. Cracker Barrel, 213 F.R.D. 619 (N.D. Ga. 2003)

Sprenger ~> Lang pucATTORNEYS AT LAW

16

(denying class certification where no nexus between subjective decision-making andstatistical disparities) and Reid v. Lockheed Martin Aeronautics Co., 205 F.R.D. 655(N.D. Ga. 2001) (same) and Abram v. UPS, 200 F.R.D. 424 (E.D. Wis. 2001) (same)with, e.g., Caridad v. Metro-North Commuter R.R., 191 F.3d 283,292 (2d Cir. 1999)(reversing denial of class certification based on statistical and anecdotal evidence ofcommonality and explaining that "[w]here the decision making process is difficult toreview because of subjective assessment, significant statistical disparities are relevant todetermining whether the challenged employment practice has a class-wide impact") andEllis v. Costco Wholesale Corp., 240 F.R.D. 627 (N.D. Cal. 2007) (certifying nation-wideclass based largely on evidence of overall statistical disparities despite defense evidencethat statistically significant disparities existed in only two of seven regions, whenplaintiffs introduced evidence that decisions in all regions were reviewed at corporateheadquarters) and Dukes v. Wal-Mart Stores, Inc., 222 F.R.D. 137 (N.D. Cal. 2004)(certifying class based largely on statistical analyses of disparities throughout company)and McReynolds v. Sodexho Marriott Servs., 208 F.R.D. 428,434-36,441 (D.D.C. 2002)(same).

Of course, even if courts adopted a consistent approach to the use of statisticalevidence in support of or opposition to class certification, the facts and judicialpredispositions would lead to variance in results. However, more rigor and consistencyin analysis should produce more predictability in results. This would benefit both parties,facilitate settlement discussions, make employment discrimination class litigation moreefficient, and at least somewhat reduce the costs associated with statistical analysis inthese cases.

IV. ATTRIBUTING DIFFERENCES IN RESULTS TO INHERENT ORSOCIALLY-BASED DIFFERENCES BETWEEN THE PROTECTEDGROUP AND OTHER EMPLOYEES NOT REFLECTED IN THE DATA

The Supreme Court's decision last term in Ricci highlights an issue thatfrequently is at play in employment discrimination cases, even if it is never mentioned inany expert report or filing. In Ricci, there were statistically significant differences,measured by the EEOC's four-fifths test, between the passing rates of white and racialminority candidates on a screening test used by the City ofNew Haven to evaluatefirefighters for promotions. The Supreme Court reversed the lower courts' rulingsupholding New Haven's decision not to use the test results.

For purposes of this paper, the important point is that the Supreme Court accepteda screening test in which racial minorities performed sufficiently worse than whitecandidates that the results could not be attributed to random chance. It raises the issue:was there something wrong with the test, were the minority candidates poorer test-takersthan the white candidates, or did the minority candidates actually have less of the skillsand attributes that the City of New Haven desired?

A similar question arises in any employment discrimination case in which thedefendant's expert cannot make the statistical disparities disappear through use ofadditional variables, disaggregating the population, or other means: does the disparity


17

reflect differences among the population group and not discrimination by the employer.If it is a race case, does the disparity reflect poorer quality education and social exposurefor African-Americans? If it is a gender case, does the disparity reflect women's greatercommitment to home and family and lesser commitment to work? If it is an age case,does the disparity reflect that older employees tend to be poorer performers than youngeremployees in the same job and grade because it took younger employees on average lesstime to reach that job and grade than it did older employees?

Although defendants' experts have asserted in cases in which I have beeninvolved that race disparities that cannot be explained away may be a product of societaldifferences, this type of issue has remained under the surface in race discrimination cases.Explanations of gender disparities based on supposed differences in interest levels havesurfaced in several gender discrimination cases over the past ten years, but courts so farhave avoided basing decisions on this type of argument. For example, in Carpenter v.Boeing Co., 456 F.3d 1183 (10th Cir. 2006), the employer sought to explain awaydisparities in overtime based in part on "potential differences in the rates at which menand women volunteer for overtime," id. at 1196, and in Anderson v. Boeing Co., 2006U.S. Dist. LEXIS 76964 (D. Okla. Oct. 18,2006), the employer argued in part thatplaintiffs had not accounted for "the amount of overtime offered to female as compared tomale employees," id. at *28 (emphasis in original). The court in each case did not haveto address these arguments because it instead ruled against plaintiffs on the ground thatplaintiffs' expert had not adequately taken into account the organizational rules underwhich overtime was awarded.

Similar arguments are made in class actions in which women claim discriminationin advancement or compensation. In Ellis v. Costeo Wholesale Corp., 240 F.R.D. 627(N.D. Cal. 2007), for example, the employer presented an expert report arguing that"gender disparities, if they exist, are based upon factors, such as women's lack of interestin jobs requiring early morning hours, which are unrelated to Costco' s culture andpromotion processes." Id. at 638. The court decided that it did not need to resolve thisissue at the class certification stage because the issue was "common to the class, and, ifanything, supports the commonality factor." Id. at 640. The court was correct that theissue supported a commonality determination (see Part III above), and should not bedecided on the merits at the class certification stage.

But because of defendants' persistence in raising the interest or commitment issuein gender discrimination cases - it has been raised in one way or another in almost everygender discrimination class action in which I have been involved - sooner or later a courtwill have to address it. Lawyers on both sides should be aware of the need to developevidence to advance or rebut expert reports that seek to attribute gender disparities todifferences in interest or commitment to the workplace.

In age discrimination cases, defendants' experts typically opine, and defendantstypically argue, that disparities in promotions, pay raises, terminations, or other personneldecisions between older and younger employees may not be the result of agediscrimination, but may instead be the result of a phenomenon similar to the "PeterPrinciple." They maintain that employees tend to reach a ceiling in their careers set by


18

their skills and abilities, and as employees reach that ceiling, their rate of advancementand their pay increases tend to decline. If younger employees catch up to olderemployees, it is because their skills and abilities better meet their employer's needs.Thus, younger employees will tend to have a better skill set than older employees in thesame job and grade. And when experts perform analyses of employment decisions at agiven point in time or over a period such as a year controlling for job and grade, youngeremployees will tend to have more favorable outcomes than older employees even in theabsence of age discrimination because of their superior skill sets for the job. As a result,the types of cross-sectional analyses typically used in employment discrimination cases,defendants argue, either cannot be used in age discrimination cases or can be used only ifadequate performance-related controls are employed that plaintiffs typically characterizeas "tainted variables." This defense interpretation of statistical disparities in agediscrimination cases is best advocated in an article published two years ago, PaulGrossman, Paul W. Cane Jr., & Ali Saad, "Lies, Damned Lies, and Statistics": How thePeter Principle Warps Statistical Analysis ofAge Discrimination Claims, 22 LAB. LAW.

251 (2007).

It is unclear at this time how much traction this type of argument will have withthe courts. At least as of the time of drafting this paper, the article had not been cited inany decision reported in LEXIS. Although the article cited several decisions that seemedto adopt similar views to those espoused, there are probably an equal number of decisionsin which courts employed the same types of cross-sectional analyses typically used inrace or gender discrimination claims.

It is also unclear how much traction this type of argument should gain with thecourts. It is based primarily on theory, not empirical studies. There have been fewpublished studies of the relationship between age or tenure and advancement orproductivity within a single employer. Those studies that have been conducted haveproduced conflicting results, with some seeming to support this defense theory, andothers seemingly at odds.

I do not attempt in this paper to analyze the strengths and weaknesses ofthisdefense argument. Rather, I want to emphasize the need to think through the issues itposes during the pretrial process as early as possible. Defense counsel need to thinkabout the information that their experts will need in support of the argument. In one casethat we are handling, after defendant had resisted our efforts to obtain discovery prior tothe liability period under the continuing violations doctrine, its experts decided at the lastminute that they would like employee history data for 30 years to try to refute our claims.Defendant dumped that data on us on the last day of the discovery period. The judgedenied our efforts to declare the data off-limits or to impose other sanctions; other judgesmight not be so lenient.

Plaintiffs' counsel needs to think about how it will establish that older employees'performance or productivity has not been inferior to the performance or productivity ofyounger employees who are now in the same grade and job, without hopefully usingtainted data. This may involve some creative discovery, including possibly into recordsthat have not been computerized. By recognizing early that the "Peter Principle"


19

argument is very likely to arise in age discrimination class cases, counsel can be preparedto take positions during discovery that might be very different than positions that mightbe taken in race or gender discrimination class cases.


20

HOT ISSUES CONCERNING USE OF STATISTICAL … · HOT ISSUES CONCERNING USE OF STATISTICAL EVIDENCE IN EMPLOYMENT DISCRIMINATION ... ofstatistics in employment discrimination class

Documents