Two important issues in the case related to (i) the question of uninjured class members (and whether the use of average overcharge models masked their presence); and (ii) whether it can be shown that all or nearly all class members were impacted by the alleged conduct. On the question of uninjured class members, defendants’ expert put forward an analysis relying on DPPs’ and CIIPPs’ experts’ direct purchaser overcharge models. To overcome this challenge, plaintiffs relied on analyses put forward by DPPs’ and CIIPPs’ experts that plaintiffs claimed established impact to all or nearly all class members resulting from the alleged conduct. We unpack these analyses in the next two sections.
Uninjured Class Members
Defendants challenged certification based on the presence of uninjured class members and the question of whether the use of average overcharge models would mask the presence of such class members. In Packaged Seafoods and Lamictal, discussed in more detail below, the courts grappled with the question of uninjured class members and to what extent their presence might be masked by averaging. The presence of uninjured class members was also key in some previous decisions that denied class certification.
There is no consensus on whether a showing of some uninjured class members results in a denial of class certification. In Packaged Seafoods, the district court initially certified the three classes at issue, finding that it was sufficient for plaintiffs’ experts to have put forward a method that is capable of showing impact on a class-wide basis, but that the question of whether there was class-wide impact could be left for a later stage. A 9th Circuit panel overturned the district court opinion, finding that the district court needed to resolve the factual dispute over the number of uninjured class numbers before certifying the classes. A 9th Circuit en banc decision ultimately reaffirmed the district court’s initial opinion.
In Lamictal, the district court initially certified the class of direct purchasers, finding no issue with defendants’ criticism that plaintiffs’ expert’s averaging of discounts masked the presence of some uninjured class members, stating “a class can be certified with de minimis uninjured parties.” A 3rd Circuit panel disagreed, noting that defendants’ expert found that 25 of 33 purchasers likely were not injured from the alleged conduct given the but-for world posited by defendants’ expert. Crucially, the circuit court panel noted that the district court needed to resolve several factual disputes at class certification to determine whether plaintiffs’ expert’s use of averages masked the presence (or absence) of individualized injury. The circuit court panel vacated class certification and remanded the case for the district court to “analyze the evidence and arguments submitted as part of class certification.” On remand, the district court ultimately denied class certification twice following the circuit court decision.
In Turkey, defendants claimed that DPPs’ expert’s model was not capable of showing impact to all or nearly all class members because his own analysis showed that some class members were not injured. Specifically, defendants offered an analysis by their expert, who applied DPPs’ expert’s overcharge model individually on the 532 DPPs with sufficient data. Defendants’ expert found that 186 of those 532 putative class members did not sustain a positive and statistically significant overcharge, constituting 35% of the 532 customers analyzed and 11% of the entire putative class. In reply, DPPs’ expert claimed that for each of the 186 customer-specific regressions for which defendants’ expert found no overcharge, defendants’ expert discarded at least 99.6% of transactions used in DPPs’ expert’s aggregate regression models. DPPs’ expert argued that when regression models are applied to only subsets of the data, estimated coefficients become less precise and “may make little economic sense.” DPPs’ expert further claimed that defendants’ expert’s application found more customers without overcharges among smaller customers, which DPPs argued could not be correct. DPPs’ expert also claimed that all top 50 customers and 97 of the top 100 still have an overcharge, and that customers with a positive overcharge represent 95% of all transactions. Finally, DPPs’ expert relied on what is referred to in the Turkey cases as an “in-sample prediction method” to show that all or nearly all class members were impacted on at least one transaction. We discuss this method in the next section.
Ultimately, Judge Harjani ruled that DPPs’ expert’s argument, that defendants’ expert’s application of his overcharge regression model to small samples is a flawed methodology, is a classic battle of the experts about the proper approach to the regression analysis that must be left for the fact finder to resolve. Judge Harjani also referred to Broilers in terming this a factual dispute that could be “resolved at summary judgment or trial.” Judge Harjani further noted that defendants’ expert offered no other affirmative approach to show the number of unimpacted class members beyond her application of DPPs’ expert’s model, which DPPs’ expert disagrees with.
Similarly, in the CIIPP case, defendants’ expert applied CIIPPs’ expert’s overcharge models to different subsets made up of specific direct purchasers, products, and defendant-product combinations, and she claimed that her regressions based on CIIPPs’ expert’s models show no overcharges “for 38.5% of all Class Product sales to resellers who sold to CIIPPs and thus, there is no overcharge to pass through to the CIIPPs from these resellers.” Specifically, defendants’ expert found no statistically significant overcharges on sales to Costco, whole bird turkey products sold to Gordon Food Service, or breast meat sold to Sam’s Club; and sales of Foster Farms’ Frozen Ground Turkeys, Jennie-O Grand Champion Raw Boneless Roast with Foil, Skin-On, and all ground turkey sales by Cooper Farms, Farbest, and Prestage. In reply, CIIPPs’ expert made a similar argument to DPPs’ expert and claimed that defendants’ expert’s econometric results are flawed and unreliable because “her customer-specific, product-specific, and Defendant-specific regressions are all plagued by small-sample size issues,” stating that defendants’ expert’s “overcharge sub-regressions did not use 99 percent of the data [CIIPPs’ expert] used in his analysis.” As discussed in the next section, CIIPPs’ expert also relied on the so-called in-sample prediction methodology for a showing of common impact.
Ultimately, Judge Harjani ruled that “the structure of the turkey market, the documentary record, and basic principles of economics,” among other things, fell in favor of granting class certification. Judge Harjani further opined that defendants’ arguments went to the probative value rather than admissibility of the overcharge regression results.
Plaintiffs’ Experts’ So-Called In-Sample Prediction Methodology
In order to demonstrate that the “anticompetitive effects of the alleged conspiracy were widespread across members of the proposed [DPP] Class, causing harm to all or virtually all Class Members,” DPPs’ expert performed a procedure that involves predicting transaction-level prices based on his average overcharge regression model. Similarly, CIIPPs’ expert applied “his direct overcharge model to the third-party distributor sales data and [. . .] found that over 99.9 percent of CIIPPs (234,006 of 234,171 customers), have at least one impacted transaction.”
This methodology, as applied by DPPs’ expert, involves the following steps:
- Estimate the overcharge regression model. This model provides an average overcharge estimated by comparing prices paid by consumers during the alleged conduct period with prices in a benchmark period, controlling for other observable factors that affect pricing but are not related to the alleged conduct. DPPs’ expert estimated an average overcharge of 11.3%.
- Use the estimates from the overcharge regression model to generate predicted prices but for the alleged conduct. The predicted but-for prices are the prices the model predicts for each transaction (known as “fitted prices”) minus the estimated average overcharge of 11.3%.
- Compare these predicted but-for transaction-level prices to actual transaction-level prices. DPPs’ expert labeled this difference between actual prices and predicted but-for prices as the “total overcharge” and flagged any transaction where the actual price was higher than the predicted but-for price as overcharged.
- Flag as harmed any customer with at least one overcharged transaction among all its purchases in the alleged conduct period. This methodology led DPPs’ expert to conclude that 99.8% of customers in the data were harmed by the alleged conduct.
As defendants’ expert explained, because class members are deemed harmed if even a single transaction is flagged as overcharged, the likelihood that any given class member is deemed harmed by the application of this methodology increases when the class member makes multiple purchases. To illustrate, in a hypothetical model with no average overcharge and residuals that are independent of one another, defendants’ expert pointed out that a customer has about a 50% chance that the residual for any given transaction will be positive (i.e., will be flagged as harmed by the methodology). If a customer has two transactions (and assuming the residuals are independent), the chance that at least one is flagged as injured is 75%, and if the customer has 10 transactions, the chance that at least one transaction is flagged as overcharged is over 99.9%. Given this, defendants’ expert argued that “to say my common impact model found 99 percent impact to me is not at all meaningful in a market setting where customers make multiple purchases.”
Judge Harjani noted that this “in-sample prediction approach” has been used and accepted as proof of common impact in prior litigations. However, defendants in prior cases have argued this methodology is flawed and is not capable of showing common impact. For example, in Pork, defendants’ expert argued that the methodology yields false-positive results. In Capacitors, defendants’ expert argued that the methodology is “non-standard compared to what is found in the academic econometric literature, [and] it also leads to biased and incorrect estimates since it is based on a single overcharge estimate . . . .” He also stated that he had never seen the methodology used in scientific or academic literature, or in any textbook. In Peanut Farmers, defendants argued that the methodology “manufactures common impact by applying the same average underpayment rate to all growers,” and also that the methodology “produces false positives.” In Dealer Management Systems, defendants argued that “the methodology is circular: it cannot prove “common impact” because it assumes it (i.e., by applying an aggregate overcharge at step one).”
Defendants and defendants’ expert made similar critiques of the methodology here. First, defendants argued that the basis of DPPs’ expert’s approach is a regression model that “shows average impact, not individual impact.” Defendants also argued that the methodology “would show injury on half the transactions in the benchmark period even when there was no conspiracy.”
Defendants’ first argument referenced by Judge Harjani is that this method is flawed for the purpose of showing common impact because it starts with a model that assumes a single average overcharge. In other words, the basis of the methodology is a regression model that does not estimate or test for different overcharges for different customers (or groups of customers) in the class. Instead, the apparent variation in overcharges at the transaction level generated by the methodology comes from the fact that the regression residuals—the differences between actual prices and the prices predicted by the model—are included as part of the “total overcharge.” DPPs’ expert explains that the transaction-level overcharge he estimates using this methodology has two parts, a common component (which is the 11.3% average overcharge estimate from his overcharge regression) and an individual component—the regression residual. He further claims that the regression residual is an unbiased estimate of the individual component of overcharges.
As defendants’ expert noted in Capacitors, this interpretation of residuals as “individual overcharges” does not appear in scientific or academic literature. Even a well-specified regression model has residuals because no model perfectly predicts every individual price. By design, a regression fits the underlying data by predicting prices that are too high about half of the time and predicting prices that are too low about half of the time. It does not follow that a given residual represents an “individual overcharge,” as defendants’ second critique described below illustrates.
The second critique by defendants that Judge Harjani references is that the so-called in-sample prediction methodology would show injury even where none exists. To show this, defendants’ expert appears to have applied the methodology to transactions in the benchmark period, where no overcharges would exist. She found the methodology indicates “injury” for half of the transactions in the benchmark period. This finding by defendants’ expert is expected, because as explained above, the regression model fits the data by overpredicting prices for some transactions and underpredicting for others. This means that even in the benchmark period where there is no overcharge, the model generates positive residuals that would be flagged as overcharges using this methodology. In response to this critique, DPPs’ expert argued that “the residual is interpreted differently in the damages period than the benchmark period because there is no impact in the benchmark period.” Given the presence of pricing variation that is not explained by the regression model (i.e., the regression residuals) even in the benchmark period where no conduct is alleged and there are no overcharges, it is unclear what the basis is for DPPs’ expert’s claim that any unexplained pricing variation for transactions in the conduct period must represent individual overcharges.
Conclusion
The question of whether class-wide impact can be shown using average overcharge models (as plaintiffs’ experts might argue) or whether the use of such models masks the presence of uninjured class members (as defendants’ experts might argue) is frequently argued at class certification, with courts arriving at different conclusions in different cases. In Turkey, Judge Harjani ultimately found plaintiffs’ experts’ explanation—that finding no statistically significant overcharge when applying plaintiffs’ experts’ own models to a subset of the data could be because small samples make it difficult to arrive at reliable results—sufficient to conclude that this particular issue is a battle of the experts. Judge Harjani also noted that with an affirmative analysis of uninjured class members by defendants’ expert, he would have been inclined to give more weight to defendants’ arguments on this particular issue. The so-called in-sample prediction methodology used by DPPs’ and CIIPPs’ experts to demonstrate common impact has appeared in several previous cases and will likely continue to be put forward by plaintiffs’ experts. Given the competing econometric arguments by plaintiffs’ and defendants’ experts related to this methodology, an independent evaluation of this methodology by a third-party expert would be beneficial for future cases.
