Here to stay: regression analysis in follow-on cartel damages

: Private damage claims that follow after a competition authority ’ s infringement decision require an accurate estimation of the harm caused, in order to avoid under- or over-compensation. The right method for valuation of damage will depend on the specifics of a particular case, and will need to balance the need to allow for a sufficient level of detail, while remaining tractable and practical for the case overall. Regression analysis is often the method that best balances these competing objectives. This article discusses the increasing use of regression analysis in fol-low-on damage claims in Europe. It outlines possible reasons why this widespread application of regression analysis is not yet extensively reflected in final judgments by national courts, and considers how this may change in the future. It concludes that the regression analysis is here to stay.


Introduction
Cartel investigations carried out by the European Commission or national competition authorities aim to establish the characteristics of the infringement of competition law committed by undertakings and the liability of the companies involved. While the authorities' investigations also typically result in substantial fines being imposed on the cartelists, these do not compensate the losses suffered by direct or indirect victims of the infringement. Private enforcement (follow-on damages claims) that follows competition authorities' decisions aims to compensate these victims for the harm suffered by them.
In the EU, those claims should 'place a person who has suffered harm in the position in which that person would have been had the infringement of competition law not been committed', while over-compensation should be avoided. 4 At the heart of private enforcement, therefore, lies a high standard for the quantification of damages, which aims to ensure that claimants are fully compensated without making the defendant cartelists pay more in damages than the harm caused (although the combined pillars of public and private enforcement should generally cause the cartelists to pay, through fines and damages, more than the harm caused).
National court proceedings for the award of damages require courts to establish some essential elements, based also on the findings of the competition authority in its infringement decision: 'when' the conduct harmed the claimant, 'what' practices caused harm (such as higher prices, lower quality products, smaller rebates, etc.) and 'how much' that harm was, 'which' purchases were affected and 'whether' the claimant managed to pass on some of the harm to its own customers. Since cartel damages represent a monetary loss that was suffered in the past, the compensation should also include interest to ensure full compensation. To answer these questions, economics can play an important role in helping assess the way competitive markets operate and the effects of cartels on market outcomes.
As part of the 'how much' question, the claimant or claimants seek to quantify the overcharge, or the level by which prices during the cartel were higher than their competitive level. This will often be a key point of contention with the defendant or defendants. The main reason for this is that the overcharge is based on the 'price that would otherwise have prevailed' absent the cartel (often referred to as the counterfactual price). 5 That counterfactual is unobserved by definition and therefore needs to be estimated through analysis.
As the Commission accurately states, 'the success of an antitrust damages action [depends also] on the quantification of the harm suffered by the victim [which] can be a quite technical and complex exercise'. 6 In order to assist national courts, in 2013 the Commission published a Practical Guide that discussed in detail the different tools that are available to undertake such an analysis. 7 The Commission's efforts to promote private enforcement in Europe did not stop then; further guidance on calculating the pass-on of overcharges was published in 2019. 8 As noted by the Commission in its 2013 Practical Guide, 'econometric techniques can increase the degree of accuracy of a damages estimate and may thus help in meeting a higher standard of proof'. 9 Econometric techniques (such as regression analysis) are essentially 'statistical techniques which help to investigate patterns in the relationship between economic variables and to measure to what extent a certain variable of interest [such as the price of the cartelised product] is influenced by the infringement as well as by other variables that are not affected by [it] (e.g. raw material costs)'. 10 Regression analysis is not only used by economistsit has been used for decades by different professions (e.g. in medical and social studies). It is a well-established and recognized way of identifying and quantifying different effects in complex analyses.
The remainder of this article discusses the usefulness and application of regression analysis in the quantification of damages, based on our experience from over 50 cartel cases in Europe and discussions with our colleagues at Oxera. 11

Why use regression analysis?
It has been argued that regression analysis adds complexity to the assessment of follow-on damages and might appear as a 'black box' to legal professionals, making it difficult for judges to rely on. 12 Nevertheless, it remains in common use by economists in the valuation of damages arising from competition law infringements. There are a number of factors that explain the continued prominence of the method.
In this section, we highlight the usefulness and advantages of regression analysis more generally before discussing, in section 3, its use in the different stages of cartel cases.

The complexity of price determination warrants richer approaches
The final price of a product (or service) is typically influenced by a large number of factors which are related to both the demand for and the supply of that product. These include the characteristics of the product (e.g. size and quality), the variable costs of production and distribution, and macroeconomic factors such as disposable income and truly exogenous factors, such as temperature or rainfall. 13 Some of these variables can change substantially over time and across countries, preventing a simple 'comparator approach' from yielding informative results or an accurate assessment of damages. 14 For example, imagine that one would like to compare the prices of a hypothetical ice-cream cartel with those from another, non-cartelised ice-cream market in order to assess the counterfactual prices. If the differences in the key inputs of the price (such as milk and labour) and weather conditions (influencing consumption with higher temperatures) are different between the two markets, the comparison could be misleading and under-or over-estimate the losses caused by the cartel. This is where the power of regression analysis comes in. As explained in the Commission's Practical Guide: Regression analysis … makes it possible to assess whether, and by how much, observable factors other than the infringement have contributed to the difference between the value of the variable of interest observed on the infringement market during the infringement period and the value observed in a comparator market or during a comparator time period. Regression analysis Commission Practical Guide (fn 3), para 92. 10 Commission Practical Guide (fn 3), para 69. 11 The authors have a collective experience of 30 years in follow-on damages assessment. 12 The term 'black box' is often used as a critique concerning the perceived opaqueness of regression analysis: see, for instance, J. Paha, 'Empirical methods in the analysis of collusion' (2011) 38(3) Empirica 389-415. 13 Others include the way in which goods and services are procured, changes in the capacity installed in a market (e.g. a new plant starting production), allowing imports from a certain date and the appearance of competing new technologies. 14 The comparator approach is based on the idea that, absent the infringement, the prices of the affected product and of the comparator product (a different product, time and/or geography) would follow the same pattern. When, during the infringement, there is a divergence between the affected product and the comparator, that difference can be attributed to the infringement.
is thus a way to account for alternative causes for the difference between the compared data sets. 15 Of course, this does not mean that other, simpler methods (e.g. arithmetic margin comparisons) should never be used. For example, these might be employed when there are few differences (aside from the impact of the cartel) between the product of interest and the comparator. 16

Regression analysis is fit for purpose
Occasionally, sceptics of regression analysis argue that it can be obscure and complex and should not be preferred over simpler methods, such as a comparison of average prices. We disagree. Regression analysis has a number of features which makes it fit for purpose in the damages estimation. We discuss these below.
First, regression analysis generates model results that can be tested against the factual background. The estimated effect of each factor on the prices can be identified and compared with the original expectation. For example, if a regression analysis predicted that organic ice cream with whole fruits was less expensive than lowquality ice cream, this would indicate that there is something wrong with the model. In addition, the robustness of the overall model and its ability to 'fit' the data can (and should) be assessed by performing sensitivity analyses.
Second, regression analysis can produce robust estimates under conditions of uncertainty and variation. It provides a rigorous statistical framework to assess the uncertainty in any estimate. Standard statistical tests can be performed to check whether observations belonging to two different groups are significantly different (in the statistical sense) from each other and whether the chosen specification is well-suited to the observed data (i.e. explains the data sufficiently well). Using this statistical framework, an analyst can quantify the uncertainty associated with any estimate. A result based on an arithmetic comparison or economic theory will also come with a degree of uncertainty around it, but this may be less visible than the ranges (confidence intervals) produced through regression estimates.
Third, regression analysis is readily replicable. Replicability means that results can be scrutinized and challenged by the other parties following sufficient disclosure. 17 As is often the case, experts can experiment with variants of the analysis to test if the evidence leads to robust conclusions.

Economist practitioners as expert witnesses
In civil proceedings, judges may seek the opinion of experts on technical matters in the form of reports, on which, depending on the jurisdiction, they might also be cross-examined. For example, physicians or construction engineers might be called in to explain technical issues and provide their professional assessment to the court concerning specific facts. Experts in proceedings often contribute with the methods and insights from some scientific discipline, and this applies to ballistic experts or construction engineers as much as to economic experts, who are increasingly being called in to assist in such matters.
The role of the expert economist is to convincingly, concisely and correctly explain the assumptions and results of his or her analysis. This approach should help the court reach a balanced judgment based on the facts of the case and the evidence (including expert evidence). 18 2.4. The use of regression analysis in competition, regulation and more widely is well-established Finally, the use of regression analysis is well-established in public enforcement such as merger control or the regulation of monopolies. For example, the European Commission uses regression analysis to assess the impact of market entry and exit on prices in order to make inferences about the post-merger situation and a merger's impact on competition. 19  into account regional differences that are beyond the company's control (e.g. population density). 20 On a broader basis, the use of regression analysis to inform decision-making is very common in many areas of the economy. For example, such analysis is used as part of COVID-19 research. 21 It has also been common for many years to analyse the impact of policy changes, such as investment in schooling, health programmes or the revenue impact of tax changes with the help of such techniques. 22

Use of regression analysis over the lifetime of a case
In our experience, regression analysis has been playing an increasing role in follow-on damages, particularly in the last five years and since the transposition of the EU Antitrust Damages Directive by the Member States. Even though, according to a recent paper discussed below, there are currently few court decisions in Europe in which the judge has considered regression analysis in the determination of the overcharge, regression analysis has been submitted in a number of court cases. Further, evidence given to a court only represents the tip of the iceberg. That is because the vast majority of cases are settled out of court between the parties, and it is not uncommon for the parties to already share their experts' econometric analyses with each other before settlement is reached.
In this section, we set out briefly how regression analysis is used by both claimants and defendants in the different stages of private enforcement.

Before a competition authority starts an investigation
Before formal proceedings are initiated, regression analysis can help enforcers to 'screen' suspicious markets before a cartel member has 'blown the whistle'. 23 Those screens can be used to search for collusive patterns, often called 'markers', in market data, such as prices, quantities, market shares or costs. These screens can take various forms. One way to screen is to test for structural breaks in prices, price-cost margins or other variables over time. If markers are identified, authorities can look further into the market (e.g. interviewing participants) before launching dawn raids and starting an official investigation. For example, the UK's CMA has recently announced that it has implemented a price monitoring tool to detect resale price maintenance. 24 We note, moreover, that leniency applicants may also be interested in knowing the likely scale of effects, among other factors, in assessing the consequences of blowing the whistle, and regression analysis on their data or retail data would be an obvious path to that damage estimation.

The competition authority investigation
Public enforcement in Europe under Article 101 TFEU (and equivalent Member State competition legislation) considers cartel infringements as 'object' rather than 'effect' cases. For these cases, there is no requirement on the enforcing authority to establish any anti-competitive effects of the cartel. However, many authorities will nevertheless decide, or be required, to consider effects alongside an object assessment of the conduct in question when setting fines, and the parties involved may follow suit.
For example, the French Competition Authority has explicitly considered overcharges during cartel investigations. 25 There is a specific provision under the French Commercial Code that the so-called 'damage to the economy' arising from anticompetitive practices (i.e. the effect of such practices) can be considered when determining the appropriate level of fines. provision is also included in the national competition laws of Spain, Poland and Hungary. 26 In a similar way, the Italian Competition Authority links the scale of fines to the presence of actual effects. In the Italian fines guidelines, there is an explicit 'claw back' factor that is linked to the extra profits that were generated through the competition infringement. 27 In such a setting, regression analysis can be used with an explicit link to the legal framework by the competition authorities and/or submitted by the companies under investigation for the assessment of effects (and fines).

Preparing a follow-on damages claim
The preparation of a potential claim is relevant for both claimants and defendants. Potential claimants (or litigation funders) aim to understand the potential value of such a claim before starting proceedings. Defendants might seek similar insights in order to determine the financial risk that these claims represent to their business.
Regression analysis often features quite prominently at this stage, even though (at least in the case of potential claimants or litigation funders) data availability might be limited prior to disclosure from the other party (or parties). In such a situation, economists may use public data or theoretical techniques in order to gauge a range for the damages that might potentially be recoverable. In our experience, competition lawyers are also increasingly adept at evaluating the results of the regression analysis prepared by economists, and use it to assist their client in making strategic decisions.

Pre-trial stage
In most follow-on damage claims, the claimant and the defendant(s) are in contact long before the beginning of the trial. They might disclose information, exchange witness evidence (including expert evidence and analyses), and they often engage in negotiations regarding out-ofcourt settlements.
Regression analysis is often employed to assist in these early negotiations, sometimes before an economist's expert report is finalised. The main reason for using regression analysis at this stage is to add credibility to the claim and to start negotiations with a well-defined estimate of loss.
When it comes to the exchange of expert reports, regression analysis can (and often does) feature prominently, and the discussions between experts can focus on the way the analyses have been performed. In our experience, when both parties are using regression analysis, any divergence between them is typically due to the data and time horizon chosen for the analysis and the interpretation of the statistical evidence. In the vast majority of cases, there is no debate concerning the usefulness of regression analysis as a tool.
In our experience, the established nature of regression analysis in this context depends neither on the country where the case is litigated nor on whether the cartel was EU-wide or national. From Spain to Germany, and from Italy to Finland, regression analysis is widely used at this stage.
It is widely recognized that the vast majority of cases are settled out of court long before the trial. 28 There are a number of factors that can contribute to the likelihood of settlement. For example, settlement may be more likely if an on-going supplier/buyer relationship exists. Also, reputational effects can be best mitigated through settlement or where legal matters such as the issue of joint and several liability with the settling parties are agreed. The presence of an economically well-founded valuation of losses from both sides using robust analysis can also assist, and in our experience this is often an important contributory factor to the settling of cases after the exchange of expert reports.

Trial stage
If settlement discussions fail to resolve a dispute, the case in question will usually proceed to trial. The reliance on experts in different jurisdictions, and the way they participate in the court proceedings, varies across the EU and also within Member States. This has implications for the way in which regression analysis is considered at this stage. For example, in some courts, economic experts will be cross-examined on the results of their regression analysis, while elsewhere, this evidence might not be considered in detail during the trial. In any case, the court (or a court-appointed expert) will have to assess all submitted evidence, including the conclusions drawn from regression analysis contained in experts' reports. Therefore, even if the decision of the court might rely explicitly on simpler analyses, this does not necessarily mean that the judges have not considered the regression analysis findings when reaching their conclusions. 29

Increase in the use of regression analysis
Sections 2 and 3 above have summarized how, in our experience, regression analysis is playing a prevalent and important role in the assessment of cartel effects. In this section, we discuss a number of reasons why we expect the use of regression analysis to persist and grow in the future. We also outline possible reasons why this widespread application of regression analysis is not yet extensively reflected in final judgments in Europe, and consider how this may change in the future.

Decisions of national courts in Europe to date
There are a limited number of studies which have looked thoroughly into the way national courts in Europe have ruled on cartel overcharges. Amongst these are the studies undertaken by Jean-François Laborde. The latest one (from 2019) was prepared with the assistance of a large number of practitioners across Europe. 30 Although these studies do not cover out-of-court settlements, which, in our experience, are likely to account for the majority of cases, 31 they do provide interesting insights. For example, it appears that claimants are increasingly able to secure redress for the harm suffered as a result of a cartel: in the last couple of years, well over two-thirds of the court judgments have either awarded damages to the claimants or established liability; before 2015, the majority of claims were dismissed.
However, according to Laborde, none of these 59 positive judgments were based on regression analysis. This finding might be surprising given the discussion in the previous sections, and is contrary to our own experience as practitioners that indicates that regression analysis has been used increasingly in the last few years.
Why might it be that national European courts appear to buck the trend toward the use of regression analysis in their final judgments? A definitive answer would require a careful consideration of the casework that underlies the judgments on which the Laborde study is based, which is beyond the scope of this article. Nevertheless, we are able to offer a number of possible explanations based on our own experience of these cases.
First, by their nature, cases that make it to court and do not settle beforehand tend to be cases in which the parties have a very different view on the damages and the way in which those damages should be calculated. Therefore, the finding that regression analysis has not yet been explicitly relied on by courts in awarding might be intrinsically linked to the fact that in the past, such analyses hadn't been submitted to the court in the first place by at least one party (and this contributed to a failure to settle). This appears to be the case for many cases in Laborde's sample. For example, of the 23 judgments considered by Laborde before 2017, for those cases that likely originated long before the implementation of the EU Antitrust Damages Directive, overcharge estimates based on regression analysis had only been submitted twice. 32 Second, many of those cases were judged long before the EU Antitrust Damages Directive was implemented by the Member State in which the court was located. For example, around half of the French judgments (which make up nearly 40% of the total sample) were handed down before France implemented the EU Antitrust Damages Directive in March 2017. In the future, the EU Antitrust Damages Directive will apply to more cases, and this will make the use of regression analysis more frequent than in the past, given, inter alia, the compensatory standard already discussed, the availability of data (Article 5 on the disclosure of evidence will ensure greater availability and disclosure of data) and more assistance to courts from national competition authorities in quantifying loss (according to Article 17(3), courts can ask competition authorities for assistance with quantification).

What does the future hold?
Looking forward, we see a number of reasons to expect greater use of regression analysis in future cases, not only 29 This is particularly likely to be the case where technical content, such as econometric estimations of overcharges, is pre-digested for the judge by a court-appointed expert in an expert's proceedings phase supporting the courte.g. in the Italian 'CTU' stage of damages cases, following a dialogue with the parties' experts. 31 Laborde (fn 27) notes, at para 23, that many UK cases 'settled before any judgment on the merits' (para 18) and that the 70 cases in which German courts established liability but did not address the question of damages tended to be followed by settlements. in cases that are resolved through settlements, but also for cases that end up in court. First, in most EU Member States, the EU Antitrust Damages Directive has made substantial changes to the ability of claimants to achieve redress for the harm caused by cartels. We are only just starting to see a difference in the way courts are dealing with such cases, as the EU Antitrust Damages Directive was only transposed into national law a few years ago by most Member States. Also, even cases heard by courts following the respective transposition dates might not be subject to the Directive's provision if the claim was not started beforehand.
Second, judges might not yet be comfortable with engaging with complex economic analysis. This is likely to change. As the number of cases grows continuously, judges will become more familiar with such cases, and gain experience of understanding and challenging regression analysis. To support this process, the European Commission is actively supporting the training of national judges. 33 Such a development can already be observed in the US, where regression analysis is more widely used and applied in cartel cases 34 and associated with the virtue of prudence: '[the] prudent economist must account for differences and would perform minimum regression analysis when comparing price before relevant period to prices during damage period'. 35 Third, as the number of cases grows, specialist courts that can deal effectively with damages might be founded across Europe. In the UK, the Competition Appeal Tribunal combines expertise in law, economics, business and accountancy and is increasingly being chosen as a forum for follow-on damage claims. If this model is replicated in other countries, it is likely to boost further the use of regression analysis and its consideration in court judgments. As an alternative, in other jurisdictions, court-appointed experts may play a role in attaining similar outcomes.
Last but not least, the potential of regression analysis is growing with the complexity and size of available data. Since competition infringements have happened in datarich areas like digital platforms or financial services, these generate large amounts of data that need to be addressed with suitable analytical techniques. Making correct inferences from such large and rich databases, in our view, requires statistical analysis of some form to process the information correctly (and in an expedited manner).

Conclusions
The use of regression analysis is well-established in various disciplines and is considered at various stages of follow-on damages claims. Even though it has not yet featured prominently in judgments of national courts in Europe, this is likely to change in the near future as more cases that deal with quantum go to trial and judges become more and more familiar with economic evidence.
We expect that regression analysis in follow-on claims is here to stay and has an important role to play in assisting courts to accurately estimate damages caused in order to fully compensate victims without unduly punishing the companies involved in the infringement.
33 See e.g. European Commission, Training of national judges and judicial cooperation in the field of EU competition law, available at: https://ec. europa.eu/competition/court/training.html. 34 See e.g. Clark and Sander (fn 25).