Statistics don't lie, people do. But lying with statistics can certainly make for a heap of trouble. According to a study done for the purposes of litigation against Wal-Mart:
These statistics have been cited repeatedly in support of an historic class action certification, earlier this year, of 1.6 million current and former female employees of Wal-Mart who claim sex discrimination in promotions and pay at Wal-Mart stores around the country. Dukes, et al. v. Wal-Mart Stores, Inc., 222 F.R.D. 137 (N.D. Cal. 2004). In Dukes, the court found as sufficient, for this stage of the litigation, plaintiffs' expert's opinions that gender is a statistically significant variable in accounting for salary differentials between men and women, and that there was a shortfall of women promoted to in-store management compared to statistically "expected" promotion rates. Id. at 160-61. That certification is on appeal to the Ninth Circuit.
Do the cited statistics suggest that Wal-Mart is guilty of discrimination? No. The truth is that use of bottom-line analysis, the kind that is employed in the statistics cited above, can imply discrimination where none exists.1 Statistical disparity is not synonymous with discrimination and it is doubtful that the statistics as presented ought to support even the class certification. Comparative statistical analysis, properly done, might tend to show that a specific relationship - like pay and gender - is not random. But even statistically significant deviations from "expected" rates do not justify the assumption that the deviations were caused by the factor in question, in this case gender or gender discrimination.
When statistics are used as a test for discrimination, gender cases may be particularly sensitive to "false positives." That's because gender would seldom be completely independent of factors relevant to pay and job assignment; examples that apply most particularly in this case include seniority, work-related experience, managerial experience, work schedule preferences, willingness to work in a "bad" neighborhood and self-selection (application rate).
The effects of benign gender correlation are not trivial. Men and women are, on average, different in many physical, psychological and cultural respects. On average, men and women differ in their preferences and life choices. Average ought not to matter when making employment decisions; individuals ought to be promoted based on individual merit. But ironically it is the plaintiffs' case in Dukes, not Wal-Mart, that makes how men and women behave on average both relevant and important. Despite the fact that the overlap between the two groups may usually exceed differences in virtually any parameter, it should be no surprise that statistical computations reveal differences, on average, between the sexes. When gender independently correlates to factors relevant to pay or promotion, such correlation would give the appearance of discrimination where none in fact existed.2
Let's take job preference, as an example. It is not difficult to imagine in at least some cases that men, as a group, and women, as a group, may differ in their perceptions of attractive job characteristics. On average, women may prefer jobs that require less overtime, fewer nighttime hours and fewer forced transfers. The resulting self-selection for certain jobs may have implications for pay and promotion, wholly separate from gender.
To show how bottom-line statistics can be used to lie, let's explore a simple example. Let's assume, hypothetically, that promotion to store management requires a willingness to work full time. Assume further that part-time workers are disproportionately younger and less experienced, with a greater turnover rate. Part-timers are offered full time as openings occur, but many decline such opportunities for reasons of family, school or inclination. In a hypothetical store, there are 75 full-time employees, including 50 men and 25 women. There are also 225 part-time employees, including 50 men and 175 women. In a two-year time period, 20 men and 12 women are offered promotions and all but two women accept. A bottom-line analysis shows that fewer women than "expected" were promoted and that being male is a clear predictor for promotion to management. At bottom line, men are promoted twice as often as women and their rate of promotion in the two-year period (20 out of 100) is four times as high as the rate of female promotion (10 out of 200). Discrimination? Not quite. Further analysis shows that the disparity from the "expected" simply reflects differences between the genders in part-time versus full-time employment. When that fact is considered, along with the information about promotions declined, it turns out that qualified women were actually offered promotions at a rate slightly higher than their male counterparts. It can easily be seen, therefore, that defining the relevant labor pool is a pivotal issue in any statistical analysis for employment discrimination.3
Where the plaintiff is effectively excused from identifying the specific employment practice that caused the offending statistical disparity (by relying on a claim of amorphous "subjectivity," for example), employers may be taking the rap for differences that they in no way caused. The danger that statistics will be misused in employment discrimination and other litigation supports the establishment of more demanding judicially imposed thresholds to the use of statistics in "proving" cases.4 In some benign situations, employers may nonetheless fear that they are all but convicted of discrimination based on statistics aloneÑeffectively reversing the standard of proof, and leaving it up to the employer to "prove" its innocence. Alarm at the possibility of such impact might tempt an employer to take gender-conscious steps to "fix" workplace disparities. We must draw the line between using a grossly statistically skewed workforce as evidence of discrimination without implicitly mandating (or at least encouraging) employers to hire in a statistically conscious manner Ñ that is, employ quotas.5
Implications Of The Case For Employers
The Dukes court seems to have viewed the largely subjective decision-making procedure at Wal-Mart with much skepticism and perhaps as a convenient pretext for discriminatory practices.6 In fact, the main lesson of Dukes may come from the court's focus on the lack of objective procedures in Wal-Mart's decision-making processes. The plaintiffs presented expert opinion that free-wheeling managerial decision-making, unfettered by objective standards, allows managers to perpetuate stereotypes. The expert concluded that promotion and pay decisions by a predominantly male management "are likely to be biased 'unless they are assessed in a systematic and valid manner, with clear criteria and careful attention to the integrity of the decision-making process.'" Dukes, 222 F.R.D. at 153. Pointing to the lack of job posting and other self-selection procedures, the court concluded that management training at Wal-Mart is largely the product of a "tap on the shoulder" process. Id. at 148. Similarly, the court concluded that, in making salary decisions, store managers were not constrained by written criteria or oversight. Id. at 146-47.
If success breeds copycats (and in litigation, it surely does), we can expect to see more gender- and pay-based class action lawsuits following the formula used in Dukes.7 What should an employer do? Dukes illustrates the value of job postings and accurately recording applicant flow data, especially in circumstances where the interest or skill level of any protected group may be different than "expected." But Dukes also instructs us that applicant flow data may not be enough to insulate an employer from class action discrimination litigation and the associated costs. Procedures designed to promote even-handed selection procedures, rein in subjectivity and check supervisory decision-making are needed. Employers should:
1 The danger that statistics will be misused in employment discrimination and other litigation supports the establishment of demanding judicially imposed thresholds to the use of statistics in "proving" cases. While it is beyond the scope of this article, a strong argument exists that statistical presentations in employment and other litigation ought to be subject to Daubert- level review before such evidence is deemed admissible . See generally, Daubert v. Merrell Dow Pharm., Inc., 509 U.S. 579 (1993).
2 Absent job-posting policies, foresight and meticulous record-keeping, "preference" can be very difficult to measure. See, e.g., Gay v. Waiters' & Dairy Lunchmen's Local 30, 489 F. Supp. 282, 304-05 (N.D. Cal. 1980) (rejecting the use of income to define the relevant interested market pool because of the multitude of other factors that influence job preference), aff'd, 694 F.2d 531 (9th Cir. 1982). The best evidence of preference is valid application flow data, but where not available, some employers have sought to prove preference "after the fact." Courts are divided over the use of "job interest surveys" as a substitute for an applicant flow analysis in an attempt to identify the "interested" workforce. In one case, the defendant successfully relied upon a job interest survey in order to rebut the plaintiff's statistical evidence. EEOC v. Sears, Roebuck & Co., 839 F.2d 302, 320-21 (7th Cir. 1988). The court emphasized that the EEOC's failure to present the testimony of any alleged victims of discrimination "confirmed the weaknesses of the EEOC's statistical evidence." Id. at 310. In another case, the court rejected an employer's use of a job interest survey in a disparate treatment case; allowing the survey results to insulate the employer from liability, the court decided, would be "to accept that an employer is permitted to discriminate against the minority of women who are interested in seeking non-traditional employment as long as the majority of women are not interested in such work." Stender v. Lucky Stores, Inc., 803 F. Supp. 259, 326 (N.D. Cal. 1992).
3 In fact, courts have repeatedly commented that "(t)here is perhaps no determination more critical" to the resolution of a pattern or practice suit than the definition of the relevant labor pool. Rivera v. City of Wichita Falls, 665 F.2d 531, 540 (5th Cir. 1982). See also Anderson v. Douglas & Lomason Co. Inc., 26 F.3d 1277 (5th Cir. 1994); Olson's Dairy Queens, 989 F.2d 165, 168 (5th Cir. 1993) (noting that the evidence must be "finely tuned" to reflect the proportion of qualified individuals in the relevant labor market.).
4 See footnote 2, supra.
5 The same law that allows plaintiffs to use statistics to "prove" discrimination also prohibits the use of quotas in most cases. Title VII protections are gender neutral, applying to men as well as women. Furthermore, section 703(j) of Title VII provides:
Nothing contained in this subchapter shall be interpreted to require any employer ... to grant preferential treatment to any individual or to any group because of the race, color, religion, sex, or national origin of such individual or group on account of an imbalance which may exist with respect to the total number or percentage of persons ... [of such groups] employed by any employer ... in comparison with the total number or percentage of persons of such ... [protected group] in any community, State, section, or other area, or in the available work force in any community, State, section, or other area." 42 U.S.C. § 2000e-2(j).
6 Other courts have viewed "excessive subjectivity" with disfavor. Cook v. Billington, 1992 WL 276936, 59 Fair Empl. Prac. Cas. (BNA) 1010, 1013 (D.D.C. 1992).
7 In cases like this, class certification itself counts as "success." Rather than a threshold issue, class certification is, for many purposes, the main prize. The merits of the dispute matter far less if, as many commentators suggest, the sheer size of the potential liability will force Wal-Mart to settle before trial.
Laura M. Franze leads the labor and employment section for the Dallas office of Akin Gump Strauss Hauer & Feld LLP.