Guides · Antitrust

Overcharge Regressions in Price-Fixing Cases

In a price-fixing case, damages are usually the overcharge: the difference between what purchasers actually paid and what they would have paid absent the conspiracy. Nobody observes the but-for price, so the plaintiffs' economist estimates it, almost always with multiple regression, and the defense economist tries to show the estimate is an artifact of the model rather than the cartel. This guide covers how the models are built and the appellate decisions that test them. Class certification questions about common impact, averaging and pass-through are covered separately in Common Impact and Pass-Through at Class Certification.

Guide 8 of 10 · October 19, 2026 · 11 min read
On this page
  1. The methods
  2. Choosing the clean
  3. Control variables
  4. False positives
  5. Separating lawful
  6. A model that
  7. How the defense
  8. What to prepare

The methods

The standard reference is Daniel L. Rubinfeld's Reference Guide on Multiple Regression in the Reference Manual on Scientific Evidence (3d ed. 2011). As the Urethane court put it, multiple regression determines "the relationship between an unknown variable (the 'dependent' variable) and one or more 'independent' variables that are thought to impact the dependent variable." In re Urethane Antitrust Litig., 768 F.3d 1245, 1260 (10th Cir. 2014). Overcharge experts use a few familiar designs.

Before-during-after. The expert compares prices during the alleged conspiracy with prices in a clean period before it, after it, or both. Because costs and demand also change over time, the regression controls for those drivers so that the remaining difference can be attributed to the conduct.

Yardstick. Instead of a clean period, the expert uses a clean market: a comparable product, region or customer group the cartel did not reach. The defense attack usually focuses on how the comparison market differs.

Dummy-variable versus forecasting. A dummy-variable model pools the benchmark and conspiracy periods in one regression and adds an indicator for the conspiracy period; its coefficient is the estimated overcharge. A forecasting model estimates the price relationship on benchmark data only, then predicts but-for prices during the conspiracy; the overcharge is the gap between actual and predicted prices. The dummy-variable approach assumes the price relationship is the same in both periods; the forecasting approach avoids that assumption but rests entirely on a representative benchmark.

Urethane describes a forecasting model. The plaintiffs' statistical expert "chose a benchmark period and tested various independent variables to find the combination that would accurately predict prices during the benchmark period. That combination of variables was then applied to the conspiracy period to calculate the prices that would have existed but for the conspiracy." Id. at 1260. Comparing predicted and actual prices, he estimated overcharges of 15.6 percent for MDI, 14 percent for TDI and 14.9 percent for polyether polyols. Id. at 1251.

Choosing the clean period and the conspiracy period

Every before-during-after model depends on two dates: when the conduct started affecting prices and when it stopped. Move either one and the overcharge moves with it.

In Urethane, Dow argued that the expert engaged in "benchmark shopping" by moving 2004 from the conspiracy period into the benchmark period "in order to manufacture supra-competitive prices during the conspiracy period." Id. at 1262. The plaintiffs answered with the expert's explanation that he included 2004 in the benchmark based on test results showing that 2004 prices "were more consistent with competition than collusion." Id. at 1263. The Tenth Circuit called the dispute "a swearing match" that the district court had resolved for the plaintiffs, and found no abuse of discretion. Id. It went further, holding that "even if Dow could prove otherwise, its benchmark-shopping argument does not implicate the reliability of Dr. McClave's methodology." Id. at 1262-63.

The period dates can also fail on the facts. In the railroad fuel surcharge litigation, the plaintiffs suggested the conspiracy began earlier than the class period, which could explain overcharges the model found on pre-conspiracy contracts. The D.C. Circuit was unmoved because the plaintiffs had offered no specific evidence of that possibility, "say, by rerunning both models from an earlier start date," and because their own collusion evidence was a "structural break" at the start of the class period. In re Rail Freight Fuel Surcharge Antitrust Litig., 725 F.3d 244, 254 (D.C. Cir. 2013). A start date the expert chooses for liability purposes has to work for damages too.

For both sides, the benchmark choice should rest on a documented, testable reason, and the expert should show what happens to the estimate under the alternatives.

Control variables and omitted variables

An overcharge regression is only as good as its controls. If raw material costs rose during the conspiracy period and the model does not account for them, the conspiracy coefficient absorbs the cost increase and overstates damages. Typical controls include input costs, energy, capacity utilization and downstream demand.

The governing standard on omitted variables comes from an employment case. In Bazemore v. Friday, 478 U.S. 385 (1986), the Fourth Circuit had rejected regression analyses of salary disparities because they did not include every measurable variable. In an opinion by Justice Brennan, joined by all other Members of the Court, the Court held that "While the omission of variables from a regression analysis may render the analysis less probative than it otherwise might be, it can hardly be said, absent some other infirmity, that an analysis which accounts for the major factors 'must be considered unacceptable as evidence of discrimination.'" Id. at 400. "Normally, failure to include variables will affect the analysis' probativeness, not its admissibility." Id. The Court added a limit: "There may, of course, be some regressions so incomplete as to be inadmissible as irrelevant." Id. at 400 n.10.

Urethane applied Bazemore to an overcharge model. "The validity of a regression analysis depends on selection of the appropriate independent variables," but defects in that selection "do not generally preclude admissibility, and courts allow use of a regression model as long as it includes the variables accounting for the major factors." 768 F.3d at 1260-61. The court also quoted the D.C. Circuit's view that "a defendant cannot undermine a regression analysis simply by pointing to variables not taken into account that might conceivably have pulled the analysis's sting." Id. at 1261 (quoting Koger v. Reno, 98 F.3d 631, 637 (D.C. Cir. 1996)).

Urethane also shows how an expert defends a specification choice. Dow argued that the TDI model should have included domestic demand variables. The expert had tested them, found no statistically significant relationship to price, and saw price fall as demand rose, a "nonsensical negative sign." Id. at 1261-62. He also relied on a 2004 Bayer document identifying exports as a driver of TDI prices. Id. at 1262. The court called the treatment of domestic demand "open to debate" but left it to weight. Id. A variable choice survives when it has a documented economic reason, not only a better fit; the district court had found the expert had "a basis, beyond statistical fit, rooted in general economic theory and particular documents." Id. at 1252.

False positives: when the model finds injury that cannot exist

The most damaging critique of an overcharge model is not that it omits a variable but that it produces results that are impossible on the facts. If the model finds overcharges where the conspiracy could not have reached, there is little reason to trust what it finds elsewhere.

That was the defect in Rail Freight. The defendants showed that the plaintiffs' damages model, applied to shippers bound during the class period by rates negotiated before any alleged conspiracy, produced overcharges similar to those it found for class members. "But the same methodology also detects injury where none could exist." 725 F.3d at 252. The expert conceded the model measured overcharges to those legacy shippers and class members alike. Id. at 253. The court rejected the plaintiffs' answer that the critique did not logically disprove classwide overcharges: "It is not enough to submit a questionable model whose unsubstantiated claims cannot be refuted through a priori analysis." Id. at 254. "As things stand, we have no way of knowing the overcharges the damages model calculates for class members is any more accurate than the obviously false estimates it produces for legacy shippers." Id. The court vacated certification and remanded, stating that Rule 23 "not only authorizes a hard look at the soundness of statistical models that purport to show predominance," but "commands it." Id. at 255. What happened on remand, where the model's own results showed uninjured class members, is covered in the common impact guide.

Rail Freight arose at class certification, but the defense can run its test in any posture: apply the model where the conduct could not have reached and see whether it still reports an overcharge.

Separating lawful from unlawful effects

A related failure is a model that attributes every price effect to the challenged conduct, including effects with innocent explanations. Concord Boat Corp. v. Brunswick Corp., 207 F.3d 1039 (8th Cir. 2000), involved market share discounts rather than a cartel, but it is a frequently cited example of a damages model that cost the plaintiffs a jury verdict for this reason.

The plaintiffs' expert assumed that in a competitive market the defendant and one rival would each hold 50 percent, and assessed an overcharge whenever the defendant's share exceeded that level. The Eighth Circuit held that the model "was not grounded in the economic reality of the stern drive engine market, for it ignored inconvenient evidence." Id. at 1056. The defendant had held a 75 percent share before the challenged discounts and acquisitions began, and the model "also failed to account for market events that both sides agreed were not related to any anticompetitive conduct," including a rival's engine recall and the problems around a competitor merger. Id. The court held that the opinion "should not have been admitted because it did not incorporate all aspects of the economic reality of the stern drive engine market and because it did not separate lawful from unlawful conduct." Id. at 1057. Because the jury's damages matched the expert's figures, the error affected the defendant's substantial rights, and the court held the defendant's motion for judgment as a matter of law should have been granted. Id.

Concord Boat and Bazemore mark the two ends. A model that accounts for the major factors goes to the jury. A model that ignores known, significant innocent causes, especially ones the expert concedes, can be excluded and take a verdict down with it.

A model that survived to verdict

Urethane is a useful example of an overcharge regression that held up through trial and appeal. The jury found that Dow participated in a price-fixing conspiracy, that the conspiracy caused class members to pay more, and that the injury did not precede November 24, 2000. It awarded $400,049,039 in damages; after trebling and deducting settlement amounts, judgment was entered for $1,060,847,117. 768 F.3d at 1252. The expert had calculated $496,680,486. Id. at 1249.

Dow attacked the model on several fronts. On variable shopping and benchmark shopping, discussed above, the court held the disputes went to weight. Dow also invoked Rail Freight, arguing that the models detected injury where none could exist. Id. at 1267. The court held Rail Freight did not apply: there, the methodology "yielded damages for a time period in which prices had been freely set." By contrast, "Dow has not identified a single class member for whom injury was impossible." Id. Finally, Dow argued the jury's shorter injury period meant it had rejected the models. The court agreed with the district court that "the jury may have fully credited Dr. McClave's models, but found the evidence insufficient to find an injury before November 24, 2000," and that a jury may reduce an expert's figure without the means to rerun the model. Id. at 1267-68.

The lesson is that a false-positive attack needs a concrete category of purchasers or periods for which injury is impossible; an inference from a jury's partial findings was not enough.

How the defense attacks the model

Defense econometric work usually runs along these lines:

  • Benchmark sensitivity. Re-estimate with alternative clean periods and conspiracy dates. An overcharge that appears under only one choice needs a justification beyond the result.
  • Omitted and mismeasured drivers. Under Bazemore, the defense has to show an omitted factor is major and changes the estimate, not just that it exists.
  • Impossible results. Apply the model to purchases the conduct could not have affected, such as pre-conspiracy contracts or unaffected products, as in Rail Freight.
  • Lawful causes. Show that recalls, mergers, capacity changes or demand shocks unrelated to the conduct explain the price movement, as in Concord Boat.
  • Structural stability. Test whether the price relationship holds across periods, and whether the benchmark is long and representative enough to forecast from.
  • Pooling and averaging. Test whether one overcharge can be applied across products, customers or channels; see the common impact guide.

The Rule 702 standard matters here. Urethane relied on the principle that reliability is "primarily a question of the validity of the methodology employed by an expert, not the quality of the data used in applying the methodology or the conclusions produced." 768 F.3d at 1263 (quoting Manpower, Inc. v. Ins. Co. of Pa., 732 F.3d 796, 806 (7th Cir. 2013)). The December 2023 amendment to Rule 702 requires the proponent to show the court that it is more likely than not that the opinion is based on sufficient facts or data and reflects a reliable application of the methods to the facts of the case, and the committee note criticized decisions treating those questions as going only to weight. Defense counsel can frame benchmark and variable choices as application questions under Rule 702(b) and (d); plaintiffs' counsel can answer that Bazemore and Urethane still describe the line between a defective model and a debatable one. How courts have actually ruled on these motions is summarized in How Often Courts Exclude Antitrust Economists.

What to prepare

  • Ground the benchmark period and conspiracy dates in documents and structural break tests, and report results under the reasonable alternatives.
  • Give every control variable an economic reason as well as a statistical one, and record the variables tested and dropped, with the reason.
  • Account for the major cost and demand drivers; be ready to explain why any omitted factor is minor.
  • Run the model on purchases the conduct could not have reached before the other side does, and investigate any overcharge it finds there.
  • Identify recalls, mergers, capacity changes and other lawful events in the period, and show how the model treats each one.
  • Prepare the record for amended Rule 702: show reliable application to the facts, not only an accepted method.

The library's expert class pages for the cartel-overcharge damages economist and the defense cartel econometrician list the rulings behind each side's typical opinions.

Supreme Mind builds Expert Witness AI Agents for litigation, for any expert on any matter, from case intake to settlement. A lawyer points an AI Agent at a matter and picks a class of expert, such as a cartel-overcharge damages economist, and one run returns a brief in four sections: the expert's likely opinion, ranked cross-examination weaknesses, the Daubert / Rule 702 attack surface, and a directional settlement range. Experts are de-identified composites of a class, never a named individual; supporting quotations are checked verbatim against the public record, authorities carry their check result, and nothing it produces is filed or testifies. A complete sample brief is free, and the method explains how it works.

Sources

This guide is general information, not legal advice.

How each court of appeals applies amended Rule 702: see the Rule 702 tracker.

Run this analysis against the expert on your own matter.Start freeRead a sample brief →

Experts in this guide's practice

Cartel-overcharge damages economistPlaintiff common-impact economistIndirect-purchaser pass-through economistDefense class-certification antitrust economistMarket-definition economistExclusionary-conduct damages economistLabor-market antitrust economistDefense cartel econometrician

The series

10Mass tort · October 20, 2026General Causation in Mass Torts: Bradford Hill Under Rule 70209Securities · October 19, 2026Section 11 Damages and Negative Causation08Antitrust · October 19, 2026Overcharge Regressions in Price-Fixing Cases07Antitrust · October 19, 2026How Often Courts Exclude Antitrust Economists, and Why06Securities · October 13, 2026Loss Causation, Damages and the Inflation Ribbon05Securities · October 6, 2026Market Efficiency at Class Certification: Cammer and Krogman04Any expert · September 29, 2026Amended Rule 702 After Nearly Three Years: What Courts Exclude03Antitrust · September 22, 2026Common Impact and Pass-Through at Class Certification02Securities · September 15, 2026Price Impact at Class Certification After Goldman01Securities · September 8, 2026Daubert Challenges to Event Studies in Securities Cases

Related: the expert agent library, a complete sample brief, and how it works.