November 21, 2011 Articles

Statistical Analysis and Wal-Mart v. Dukes

The decision sharpens the focus on the use of economic and statistical expert analysis and testimony in class certification.

By Chip Hunter

To a practitioner in the field of economic analysis and econometrics, the most interesting part of the Supreme Court’s decision, and the part with the most impact on expert testimony, in Wal-Mart v. Dukes is the recognition of the plaintiffs’ failure to satisfy Rule 23(a)(2)—the requirement that states the party seeking class certification must demonstrate that the members of the class have common questions of law or fact. Although the statistical and economic approach proffered by the plaintiffs’ experts was similar to that in other cases that had been certified, the Dukes decision highlighted a fundamental disconnect between the plaintiffs’ theory of the case and the expert evidence they used to support it. The gap is the difference between presenting an analysis with conclusions that are consistent with a common question of law or fact and presenting an analysis that is evidence of a common question. The Court’s statement regarding the requirement for satisfying Rule 23(a) is as follows:

Rule 23 does not set forth a mere pleading standard. A party seeking class certification must affirmatively demonstrate his compliance with the Rule—that is, he must be prepared to prove that there are in fact sufficiently numerous parties, common questions of law or fact, etc.

Wal-Mart Stores, Inc. v. Dukes, 131 S. Ct. 2541 (2011) (emphasis added)

Thus, demonstrating the existence of a common question may require an examination of evidence commonly associated with the merits phase of the case. Addressing preliminary matters, the Court says, is a common part of litigation, and, in this case, “necessarily overlaps with respondents’ merits contention that Wal-Mart engages in a pattern or practice of discrimination.”

While the complete impact of the Dukes decision remains to be seen, it has clear implications for the nature of expert testimony in class-certification proceedings. The decision affects not only labor cases, but also other types of class-action cases in which there is a question about a defendant’s action and its impact on potential class members. Examples include both antitrust and consumer-fraud actions.

For experts and attorneys working with experts addressing questions of commonality, there are two primary points to note from the Dukes decision. First, for an analysis to be persuasive, it must be designed so that it can provide statistical insight into whether evidence of common questions exists for the class as a whole, not simply a finding that may be characterized as consistent with the existence of common questions. Second, the analyses must also be able to demonstrate that commonality of impact is due to the challenged behavior of the defendant and not other factors.

The first point primarily deals with the power of an econometric or statistical test to discern, at a fine enough level, whether the “impact” of the alleged bad act is sufficiently common across the class. The second point is focused on whether the results of the statistical analysis are sufficiently supportive of the theory of harm. These two points are complementary, and the questions they ask can usually be answered within the same analysis, but the Dukes decision divided them when it ruled that neither hurdle was surmounted.

Designing Analyses to Address the Existence of a Common Question
In Dukes, the plaintiffs claimed a company-wide practice led to a disparity in pay between women and men. The question common to class members was whether that practice impacted each of them. Accepting that an individual analysis of 1.5 million class members was impractical, the experts were left with the task of designing an analysis that, from a statistical point of view, would sufficiently demonstrate commonality within the proposed class. The Court concluded that the plaintiffs’ experts did not meet that burden.

Attorneys and experts are advised to remember that most statistical or econometric analyses designed to measure differential impact are measuring average effects, not individual effects. For example, a regression analysis designed to test for a wage disparity between two groups (class members and non-class members) measures the average disparity between the groups and whether the difference between the groups is statistically significant. That is, it answers the question “How different are class members from non-class members on average?” The commonality test, however, is concerned with answering the question “Is the impact of the alleged practice sufficiently common among class members?”

A simple example illustrates this point: Suppose there is a firm with 200 employees, half that work at Factory A and half that work at Factory B. Half of the employees at each factory are left-handed, and half are right-handed; each right-handed employee has to be matched with a left-handed employee to run a specialized machine. Left-handed employees at Factory A make $10 an hour, and right-handed employees make $15 an hour, and all Factory B employees make $15 an hour. A regression of hourly wage on “handedness” will result in a (very) statistically significant finding that left-handed employees make on average $2.50 an hour less than right-handed employees. However, by construction, we know that there is meaningful variation in wages within left-handed employees: Half of them make exactly what right-handed employees make. The regression coefficient on left-handedness has revealed nothing about the variation in wages within the left-handed employees, and it has not revealed that the difference in pay for left-handed workers is related to which factory they work in.

The plaintiffs’ statistical expert performed his econometric analysis at the level of a Wal-Mart “region,” each of which contains 80 to 85 stores. The Court found that, from a statistical point of view, the plaintiffs’ experts’ analyses were performed at too great a level of aggregation:

A regional pay disparity, for example, may be attributable to only a small set of Wal-Mart stores, and cannot by itself establish the uniform, store-by-store disparity upon which the plaintiffs’ theory of commonality depends.

In other words, any effect detected by the expert’s analysis is for such a large part of the proposed class that it would be reasonably possible that the results were driven by only a portion of the proposed class. A region consists of many stores, managers, and decision-makers. The regression analysis provided insufficient insight into whether decisions at a more granular level, such as the store level, were consistently impacting female employees. A moderate disparate impact occurring at all stores in a region and a strong disparate impact at only some stores could both deliver the same statistical outcome when examined at a disaggregate level.

The proper level of disaggregation is a function of the nature of the plaintiffs’ claim. For example, if the claim concerns a group of sellers allegedly colluding on prices in different geographic markets, the commonality inquiry should be performed at least at the level of each local market. If more disaggregate data are available, the plaintiffs’ case will be stronger if they can demonstrate commonality at even finer and finer levels of disaggregation. With the advent of increased data-processing power, often, such detailed data are available, as are the computational tools that allow for such an analysis.

For example, in the factory illustration above, regressions done at the factory level would reveal that left-handed people were paid less at only one location, which would weigh against certifying a company-wide class of left-handed employees.

Even among practitioners, the standards for assessing commonality are contested. In their amicus brief, a group of labor economists and statisticians advocated that “when there is a legitimate choice between two possible levels of analysis, an expert generally should use the broader level.” Brief of Amici Curiae, Labor Economists and Statisticians in Support of Respondents at page 15. The Court’s decision, however, has ruled out broad, aggregate analyses from the set of legitimate choices. The amici cite literature supports the idea that aggregate data are useful for studying “general patterns.” The relevant goal, however, for a class certification inquiry is to determine whether there is commonality within a group, not an average effect for the group. And, unlike many academic studies where only aggregate data are available, the parties in the Dukes case had access to very detailed data, which allowed for a finer examination of the alleged impact.

Connecting Statistical Evidence with the Theory of Harm
One likely change we may expect as a result of the Dukes decision is that successful class actions will have smaller proposed classes supported by expert analyses that rely on disaggregate data. A perhaps less obvious change comes from the Court’s highlighting the importance of tying together the theory of harm with the statistical analysis.

There is another, more fundamental, respect in which respondents’ statistical proof fails. Even if it established (as it does not) a pay or promotion pattern that differs from the nationwide figures or the regional figures in all of Wal-Mart’s 3,400 stores, that would still not demonstrate that commonality of issue exists.

Dukes, 131 S. Ct. 2541

Stated differently, the expert’s statistical analysis must not only demonstrate that the alleged impact is sufficiently common to the class, but it also must support an inference that the specific practice being challenged is the common cause of the impact. In the Dukes case, the Court found no specific employment practice alleged as the basis for disparate impact.

Return to the stylized example from above but now assume that every left-handed worker is paid only $10 per hour. A statistical analysis would show that left-handed workers were paid less at both locations. To satisfy their burden, however, plaintiffs would need to show that the wage disparities at each location are due to a common challenged practice. To the extent practicable, experts should include in their analyses the effects of local labor market conditions to help ensure that the differential impact is not due to factors other than the alleged practice. For example, a surplus of left-handed workers in the market near Factory A could lead to reduced wages for left-handed workers there.

The Dukes decision sharpens the focus on the use of economic and statistical expert analysis and testimony in class-certification matters, and the depth of inquiry will be more akin to what is done at the merits stage. Plaintiffs’ experts will likely need to perform analyses at a very disaggregate level to demonstrate that impact is widely distributed among the class and not simply present at some aggregate level. Simultaneously, the statistical analyses must be able to discern whether plaintiffs’ theory of harm is reasonably the cause of any measured disparity and not due to other factors. This will likely lead to fewer class-action cases being filed, while those that do get filed will likely have smaller class sizes and more focused claims, making them perhaps easier to certify.

Keywords: litigation, class actions, derivative suits, statistical analysis, aggregate analysis

Chip Hunter – November 21, 2011

Copyright © 2011, American Bar Association. All rights reserved. This information or any portion thereof may not be copied or disseminated in any form or by any means or downloaded or stored in an electronic database or retrieval system without the express written consent of the American Bar Association. The views expressed in this article are those of the author(s) and do not necessarily reflect the positions or policies of the American Bar Association, the Section of Litigation, this committee, or the employer(s) of the author(s).