chevron-down Created with Sketch Beta.

ARTICLE

The Sea of Tuna Decisions: Key Takeaways

Kelly Lear Nordby

The Sea of Tuna Decisions: Key Takeaways
Giordano Cipriani via Getty Images

Introduction

On November 14, 2022, the U.S. Supreme Court declined StarKist Company’s petition to review the Court of Appeals for the Ninth Circuit’s en banc opinion upholding certification of three subclasses of tuna purchasers in Olean Wholesale Grocery Cooperative, Inc. v. Bumble Bee Foods LLC, 31 F.4th 651, No. 19-56514 (9th Cir. Apr. 8, 2022) (“Olean”). These decisions received much attention because defendants argued that almost a third (28%) of the direct purchaser plaintiffs (DPPs) class was not injured by the alleged price-fixing., The main issue addressed on appeal was whether the plaintiffs’ regression model, along with other expert evidence, was capable of showing classwide antitrust impact.

Using a regression model that pooled data from all three defendants, the DPPs’ expert found that DPPs were overcharged, on average, by 10.28 percent. This article discusses the Ninth Circuit’s analysis of the DPPs’ and defendants’ experts’ regression models, focusing on the approaches used to estimate the number of potentially uninjured class members. I identify three findings and discuss key takeaways of the decision for economic analyses of the predominance requirement under Federal Rule of Civil Procedure 23(b)(3) for damages classes within the Ninth Circuit and possibly beyond.

First, the Ninth Circuit rejected any categorical argument that pooled regression models are inherently unreliable. Second, citing the Supreme Court’s decision in Tyson Foods, Inc. v. Bouaphakeo (“Tyson”), the court found that a regression model can be relied upon to establish classwide liability if each class member could rely on the model “to show antitrust impact of any amount,” if he or she had brought an individual action. Third, the court found that the district court did not err or abuse its discretion in concluding DPPs’ expert’s pooled regression model, along with other evidence, was “capable” of showing antitrust impact on a classwide basis. In reaching this conclusion, the district court accepted DPPs’ expert’s argument that putative class members with insufficient data can rely on the pooled regression model’s results for other “similarly situated plaintiffs.”

In view of these findings, one can expect plaintiffs’ experts to continue to proffer pooled models as evidence of classwide impact. In such cases, it is important to test econometrically whether the evidence suggests the pooled model’s common overcharge estimate holds for all or nearly all putative class members. Additionally, when confronted with data limitations, experts may consider whether putative class members with sufficient data are representative of those with insufficient data. If the evidence suggests that the putative class members with limited data differ from those used to estimate the model in ways that would likely have affected prices, then one cannot simply assume the pooled model’s results are unbiased and hold for all putative class members.

Uninjured Plaintiffs and the “Fear” of “Monstrously Oversized Classes”

In its petition, StarKist asked the Supreme Court to address two important questions, specifically:

“[w]hether, and in what circumstances, the presence of uninjured class members precludes the certification of a class under Federal Rule of Civil Procedure 23(b)(3)”

and

“[w]hether, and in what circumstances, a plaintiff may rely on representative evidence such as averaging assumptions to establish classwide proof of injury to satisfy Rule 23’s requirements.”

Several organizations also filed Amicus Curiae briefs urging the Supreme Court to review the en banc court’s decision and clarify Rule 23’s requirements. The Supreme Court declined.

The prospect of allowing certification of classes that potentially contain large numbers of uninjured members is concerning because, as the dissent noted, class action cases rarely proceed to trial. In Olean, dissenting Judge Lee, joined by Judge Kleinfeld, “fear[ed] that [the] decision will unleash a tidal wave of monstrously oversized classes designed to pressure and extract settlements,” with implications that extend to class actions beyond price-fixing.

“Pooled” Regression Models and Potentially Uninjured Plaintiffs

When experts rely on regression models that “pool” (i.e., gather together) sets of data for putative class members to examine the common question of whether there is evidence of antitrust impact in the form of higher prices paid by all or nearly all class members, criticisms related to the use of “averaging assumptions” and the presence of potentially uninjured plaintiffs are intertwined.

More specifically, in antitrust class actions that allege price-fixing, plaintiffs’ experts often employ multiple regression analysis where price is modeled as a function of supply and demand factors, as well as other factors that economic theory suggests influence prices, such as product, customer and market characteristics. A “before-and-after” approach may be used to assess whether there is evidence of classwide impact by estimating the model using data from the period when the anticompetitive conduct allegedly occurred (the “impact period” or “class period”), as well as before and/or after the anticompetitive conduct. The period before and/or after the impact period serves as a benchmark for how prices would have been set without (“but-for”) the alleged misconduct.

An indicator variable (also known as a “dummy variable”) that is set equal to one during the impact period and zero otherwise can be included in the model to isolate and estimate the effect of the alleged conduct on prices over the alleged impact period. In the simple model where a single coefficient is estimated on the impact period dummy variable using sales transaction data that is “pooled” across all putative class members, products, and/or defendants, the single coefficient represents the average effect of the anticompetitive conduct on prices during the impact period across all putative class members, products, and/or defendants, respectively. If the estimated coefficient on the impact period dummy variable is statistically significant and the sign is consistent with the plaintiffs’ theory of harm, i.e., positive in a price fixing case, plaintiffs interpret the results as evidence of common impact (i.e., overcharges) for all putative class members.

Pooled regression models are often criticized by defendants and practitioners for employing “averaging assumptions” that “mask individual differences” and for presuming, rather than proving, common impact. Thus, before assuming the single dummy variable pooled model is appropriate, experts should test econometrically whether the common overcharge estimate holds for all or nearly all putative class members.

Three approaches used commonly by experts to test estimates of average impact from pooled regression models include the following:

  1. Sub-regression approach: Estimate the pooled regression model using subsets of the data, such as by putative class member or by partitioning the proposed class into subclasses based on factors that economic theory predicts ex ante account for differences in prices. Regressions estimated on subsets of the data are also referred to as “sub-regressions.” The “sub-regression” approach results in a different overcharge estimate for each subset. In addition, this approach allows the coefficients on all other explanatory variables (“covariates”) in the regression model to vary by subset. The expert can then test whether the coefficient on the impact period dummy variable for each subset is positive and statistically significant. Additionally, if the coefficients on the other covariates in the sub-regressions vary widely across putative class members, then it is unlikely that a single model provides a common method of proof.
  2. Interaction approach: Estimate the regression model using pooled data but allow the overcharge estimate to vary for each putative class member (or group of class members) by interacting the impact period dummy variable with dummy variables that identify each putative class member (or group). Like the sub-regression approach, this approach results in a different overcharge estimate for each class member/group but it estimates a single regression using all the data, rather than running separate sub-regressions using subsets of the data. The expert can then test whether the coefficients on the impact-interaction dummy variables are consistent with each class member/group being injured (i.e., a statistically significant and positive). Without additional interaction terms, this regression model imposes the restriction that the effects of the supply and demand factors on price were common across class members and were not affected by the alleged conduct.
  3. Prediction approach: Use the estimated coefficients from the proffered regression model and the underlying data to obtain predicted prices without (“but-for”) the alleged anticompetitive conduct for each putative class member and then determine if the actual price charged exceeds the predicted but-for price., An actual price that is above the predicted but-for price is consistent with the plaintiff being overcharged. The expert may then report statistics such as the percentage of class members with at least one positive overcharge or the percentage with positive mean overcharges.

With the first two approaches, estimated coefficients on the impact period dummy variables that are not statistically significant or that have a sign that contradicts the expected effect of the alleged conduct (e.g., a negative overcharge) are typically interpreted as evidence that not all class members were harmed. With the third approach, for any given putative class member, if the predicted but-for prices always exceed the actual prices paid, then the model would not support a conclusion of injury for that putative class member.

The Supreme Court’s Decision in Tyson and the Use of Statistical Evidence

The Ninth Circuit’s en banc decision relies heavily on the Supreme Court’s decision in Tyson, a wage-and-hour class action where the Court held that representative evidence (specifically averages estimated from a sample) could be used to show predominance and affirmed certification. Tyson focused on the use of representative proof “to fill an evidentiary gap” created by the defendant, specifically due to defendants’ failure to keep adequate records. In Tyson, the Supreme Court stated:

[W]hile petitioner, respondents, or their respective amici may urge adoption of broad and categorical rules governing the use of representative and statistical evidence in class actions, this case provides no occasion to do so. Whether a representative sample may be used to establish classwide liability will depend on the purpose for which the sample is being introduced and on the underlying cause of action.

The Supreme Court further held that one way a sample could be relied upon for proving classwide liability is “by showing that each class member could have relied on that sample to establish liability if he or she had brought an individual action.” As discussed below, the Ninth Circuit extended these concepts to statistical evidence (again averages) derived from regression analysis.

Key Takeaways for Economic Analyses of the Predominance Requirement

The Ninth Circuit’s en banc majority’s decision in Olean has at least three key takeaways for regression models used to assess classwide antitrust impact.

1. “[A]ny categorical argument that a pooled regression model cannot control for…individualized differences among class members must be rejected.”

The en banc court rejected any argument that pooled regression models “involve improper ‘averaging assumptions’ and therefore are inherently unreliable,” noting that Tyson rejected “any categorical exclusion of representative or statistical evidence.” Additionally, the court stated that, “any categorical argument that a pooled regression model cannot control for variables relating to the individualized differences among class members must be rejected,” and that, in antitrust cases, “regression models have been widely accepted as a generally reliable econometric technique to control for the effects of the differences among class members and isolate the impact of the alleged antitrust violations on the prices paid by class members.”

In the wake of this decision, one can expect plaintiffs’ experts to continue to proffer pooled models as evidence of classwide impact. As always, it will be important to examine whether the model correctly measures and controls for individualized differences and other factors that affect prices. For instance, it may be that certain explanatory variables in the model capture observable differences believed to influence prices that vary by putative class member (or groups of putative class members), such as customer type or geographic location. Using prices that account for promotional credits, discounts and rebates would also enable one to account for individual customers’ bargaining power.

However, a pooled model that specifies a single overcharge estimate does not control for possible differences in the effects of the alleged price-fixing across putative class members over the impact period. As discussed above, the sub-regression and interaction approaches allow one to test econometrically whether the single coefficient dummy variable model applies to all putative class members.

In Olean, both sides’ experts proffered alternative models to test whether the data suggest the single overcharge estimate is valid for all or nearly all putative class members. Defendants’ expert revised the DPPs’ expert’s regression model to allow the overcharge estimate to vary by putative class member and reported that, of the 604 direct purchasers, 169 direct purchasers (28 percent) could not rely on the model to show a positive, statistically significant overcharge. In addition, the DPPs’ expert allowed the overcharge estimate to vary by customer type and found large, statistically significant overcharges for every customer type. The description of the evidence suggests both experts used the interaction approach (#2) described above to examine whether overcharges differ by putative class member or class member type, respectively, although the sub-regression approach (#1) could achieve the same objective.

Economists recognize that the distribution of such individual- or group-specific overcharge estimates can inform the question of classwide impact. If the distribution of the plaintiff-specific overcharge estimates has a large variance, such that the estimates differ in magnitude and/or sign, then there may still be a substantial number of uninjured plaintiffs.

When certifying the class, the district court credited the DPPs’ expert, who asserted that the defendants’ expert’s model could not provide any result for 61 direct purchasers (approximately 10 percent of the putative class) because they did not make any purchases during the benchmark period, and that many other purchasers had too few transactions to generate statistically significant results using any regression model. Thus, the Ninth Circuit concluded that, “[c]ontrary to the dissent’s claim… [defendants’ expert] did not show that 28% of the class potentially suffered no injury.”

2. A regression model can be relied upon for proving classwide liability if each class member could have relied on the model to establish liability in an individual action.

The Ninth Circuit noted that Tyson established “the rule” that plaintiffs could satisfy their Rule 23(b)(3) requirement if “‘each class member could have relied on [the plaintiffs’ evidence] to establish liability if he or she had brought an individual action,’ and the evidence ‘could have sustained a reasonable jury finding’ on the merits of a common question.” In Olean, the en banc court extended the Tyson “rule” to expert evidence derived from regression models.

More specifically, in Olean, the defendants argued that the pooled model could not sustain liability in individual proceedings because individual plaintiffs pursuing their own claims showed overcharges both above and below the average (10.28 percent). The en banc court rejected the defendants’ argument, noting that, “[f]or purposes of determining whether each member of the DPP class can rely on the model to prove antitrust impact,” the “question is whether each member of the class can rely on [plaintiffs’ expert’s] model to show antitrust impact of any amount.” The court recognized that,

“[w]hile individualized differences among the overcharges imposed on each purchaser may require a court to determine damages on an individualized basis…, such a task would not undermine the regression model’s ability to provide evidence of common impact. Accordingly, we reject the Tuna Suppliers’ argument that the regression model could not sustain liability in individual proceedings.”

When rebutting defendants’ expert’s criticisms, DPPs’ expert examined the pooled regression model’s predictions for each member of the class (i.e., approach #3 above) and found that 94.5 percent of purchasers made at least one purchase at an inflated price. The en banc court stated that this “again provided further evidence that the conspiracy had a common impact on all or nearly all the members of the DPP class.” However, defendants’ expert claimed the calculation was “misleading” because it was premised on the 10.28 percent average overcharge estimate. In addition, DPPs’ expert examined the predicted prices using the defendants’ expert’s revised model and found positive overcharges for 94 percent of customers with any result. Although these results could be interpreted as suggesting that as much as 6 percent of the DPP class was uninjured, the court did not address this issue because defendants did not argue that this result precludes certification. Other courts have found that 5 or 6 percent “constitutes the outer limits of a de minimis number” of uninjured plaintiffs who may be included in a certified class.

Thus, when determining whether a proposed model is capable of showing common impact for all class members, going forward, courts may place greater weight on analyses of but-for prices and overcharges predicted by regression models for individual class members (approach #3) over analytical approaches that provide individual-specific overcharge estimates (approaches #1 or #2 above) when the latter regression models are vulnerable to small sample size criticisms.

3. Putative class members with limited data can rely on the regression model’s results for “similarly situated” plaintiffs.

In Olean, the en banc court disagreed with the defendants’ and the dissent’s contention that the district court failed to resolve a dispute among the parties as to whether 28 percent of the putative class were not injured. The en banc majority held that the defendants and the dissent “mischaracterize[d] the import” of defendants’ expert’s findings, and concluded that defendants’ expert “did not make a factual finding that 28 percent” of the DPP class were uninjured. Rather, the en banc majority found that the district court “resolved [the] methodological dispute…in favor of [plaintiffs’ expert] by crediting [plaintiffs’ expert’s] rebuttal that even class members with limited transactions during the class period can rely on the pooled regression model as evidence of impact on similarly situated class members.” The en banc court concluded that “the district court determined that [plaintiffs’ expert’s] pooled regression model was capable of showing that the DPP class members suffered antitrust impact on a class-wide basis.”

This finding that putative class members with limited data (i.e., few transactions) can rely on the pooled model’s results as evidence of impact is noteworthy because it raises some of the same questions as the presence of potentially uninjured plaintiffs. For example, what is the maximum percentage of putative class members for which the proposed model cannot yield any result (or any statistically significant result) and still be deemed capable of showing classwide impact? Here, the court permitted at least 10 percent (i.e., those with no benchmark period transaction) and perhaps as high as 28 percent (i.e., those who could not rely on the model to show a positive, statistically significant overcharge).

Additionally, the en banc majority remarked that the defendants’ reference to DPPs’ expert’s regression model as “representative evidence” was “imprecise” because

“representative evidence generally refers to a sample that represents the class as a whole…. By contrast, a regression model analyzes available data to determine the degree to which a known variable, such as collusion, affected an unknown variable, such as price, while eliminating the effect of other variables.”

While it is true regression analysis controls for the effects of the explanatory variables in the model, this statement overlooks that regression coefficients are also estimates of population parameters that measure the linear relationships between the dependent variable and the covariates in the model. Thus, like sample estimates, regression coefficients can be affected by the underlying data if the data are incomplete in a way that is not random.

More specifically, a non-random sample will only yield unbiased coefficients if the model is (i) the same for all subsets of the population (i.e., those included and those excluded from the data) and (ii) the regression data exhibit variation in the covariates that is comparable to what is observed in the population, including encompassing observations for which there is insufficient data to run the regression. However, the first assumption is exactly what plaintiffs’ analysis aims to prove.

Experts therefore may seek to examine whether class members with insufficient data differ from those with sufficient data in order to assess whether the evidence indicates they are, in fact, “similarly situated.” In other words, does the evidence suggest the putative class members with sufficient data are representative of those with insufficient data? If the evidence suggests this assumption does not hold, then it may not be appropriate to assume the pooled model’s results for a subset of plaintiffs hold for all putative class members when evaluating the question of common impact. Defendants may therefore seek to show that exclusion of certain putative class members from the regression biases the results.

Conclusion

In sum, Olean has at least three key takeaways for expert regression analysis of the predominance requirement in antitrust class actions within the Ninth Circuit and possibly beyond. First, it suggests pooled models will continue to be proffered as common evidence of impact and that plaintiffs’ experts may assert that class members with limited data can rely on the pooled model’s results as evidence of impact. Before relying on a pooled model’s single overcharge estimate, experts should test econometrically whether the estimate holds for all or nearly all putative class members. Second, going forward, courts may place greater weight on analyses that examine a regression model’s predictions of the impact on individual putative class members, rather than estimates of impact from regression analyses based on limited data for individual plaintiffs (or subsets of plaintiffs). Third, when determining whether individual plaintiffs with limited transaction data can rely on the regression model to prove impact, courts may consider whether the facts and data indicate these class members differ systematically from those with sufficient data, in order to assess whether the evidence suggests they are “similarly situated.” If the evidence suggests plaintiffs with limited data differ from those used to estimate the regression in ways that may have insulated them from the alleged conduct, one cannot merely assume the regression model is capable of reliably demonstrating antitrust impact for all putative class members.

The views expressed herein are those of the author and not necessarily the views of Ankura Consulting Group, LLC., its management, its subsidiaries, its affiliates, or its other professionals. Ankura is not a law firm and cannot provide legal advice. 

    Authors