Old tools for the new economy? Counterfactual causation in foreclosure assessment and choice of remedies on data-driven markets

Abstract

Under the Digital Markets Act, gatekeeper platforms abide by numerous ex-ante rules, including narrow data-sharing obligations that address data-driven market tipping. Gatekeepers also remain subject to ex-post competition interventions. While interventions under both ex-ante and ex-post (antitrust) mandates are not novel, they must be approached with extra care in data-driven digital markets, where uncertainty and error costs are high. Therefore, we argue that post-DMA competition enforcement against gatekeepers requires rigorous ex-post economic effects analysis at two levels: foreclosure assessment and devising appropriate remedies. Regarding foreclosure assessment, we propose a ‘multiple-counterfactual-causation-test’ for establishing anticompetitive effects, which should ensure a nuanced analysis of the competitive landscape on data-driven digital markets in the post-DMA world. Regarding remedies, a fine-grained, proportionality-driven assessment of the need for ex-post data-related obligations is vital.

1. INTRODUCTION¹

In light of ongoing enforcement against digital giants (gatekeepers) under the European Union’s (EU) Digital Markets Act (DMA),² this article underscores the need to adapt ex-post competition enforcement towards (even more)³ careful and economically sophisticated analysis. This analysis should accommodate dynamic changes in digital markets using tools like counterfactual thinking, which is available but currently underutilized in EU antitrust enforcement. Our argument is illustrated with data-specific harms and their remedies under the DMA (data-sharing mandates).⁴ However, it could also apply to other practices sanctioned under Articles 5 and 6 of the DMA,⁵ specifically the so-called ‘Type B cases’ identified by Cremer and co-authors, ‘for which only some of the gatekeeper’s conduct falls with the DMA’.⁶ In such cases, narrow data-sharing mandates may cause unintended downstream effects, such as higher prices for advertisers and consumers due to decreased gatekeeper monetization. These issues can be addressed through ex-post competition enforcement under Article 102 TFEU, using multiple counterfactual causation analysis based on econometric models that we propose in Section 4 of this article.

While a possible alternative to our approach—Article 8(3) DMA—allows for regulatory dialogue between the Commission and gatekeepers, it has been criticized for raising concerns of regulatory capture.⁷ It was also avoided by the Commission in April 2024 when the enforcer instead chose to escalate against gatekeepers using non-compliance proceedings under Articles 20 and 29 DMA.⁸ Moreover, other alternatives under the DMA, which allow the Commission to specify or update obligations, aim to improve contestability and fairness, but do not directly curb consumer harm.⁹ These provisions have also been criticized for their lack of detail and ‘teeth’ in the final legislative text.¹⁰

Given this status quo, ex-post competition enforcement, despite known delays,¹¹ is crucial as a corrective mechanism for the unintended consequences of narrow data-sharing mandates under the DMA. Our proposed analytical strategy urges authorities to employ rigorous counterfactual causation assessment under ex-post, dominance-based (Article 102 TFEU) effects analysis of data-driven harms (Section 4).¹² Rigorous counterfactual methods in ex-post antitrust foreclosure analysis enhance decision-making certainty¹³ and reduce error costs.¹⁴ These considerations, in our view, outweigh possible enforcement delays, now a concern under the DMA as well.¹⁵ Counterfactual thinking, a mainstay of US antitrust enforcement since the 1970s,¹⁶ though slower, is practicable and has also been employed in New Zealand (albeit with more limited success).¹⁷ Counterfactual thinking can also be used at the ex-post remedial stage (a method that has both its proponents and detractors),¹⁸ which we discuss in ‘The counterfactual link and (data-driven) remedies’ section.¹⁹ Before elaborating on our proposals in Section 4, we outline the possible reasons for the narrow nature of DMA’s data-sharing obligations (Section 2) and critically examine the final version of the DMA in this respect (Section 3).

2. DATA-DRIVEN MARKET TIPPING: A CASE FOR EX-ANTE MANDATED DATA SHARING?

In order to understand the interaction of ex-ante enforcement under the DMA and ex-post competition scrutiny, it is important to first become aware of the debates preceding the DMA that likely also explain the ‘narrowness’ of its current data-sharing mandates. In the ‘Data-sharing mandates and their alternatives’ section, we claim that exactly this narrow nature of DMA’s data-sharing mandates makes them difficult to enforce and could create enforcement gaps that, in turn, could be ‘closed’ by employing competition law.

Data-driven market tipping: the theory

The advent of digital markets, and in particular the faster pace at which market power manifests and entrenches itself, has brought to the forefront the question of how this phenomenon can be addressed through policy. In this context, some suggest enhanced competition/antitrust²⁰ scrutiny (the current approach in the USA), while others go further and propose the introduction of ex-ante economic regulation (regulating big tech as public utilities, an approach that is exemplified by the EU’s DMA).²¹

The motivation for some of the latter proposals lies in the now-established workings of data-driven indirect network effects for big data companies (providers).²² These effects are a relatively recent phenomenon related to the advent of big data, whereby a ‘user’s utility is not directly affected by other users but the amount of other users increases the provider’s stock of user information, which in turn decreases the cost of innovation for the provider.’²³ Data-driven indirect network effects are a source of invariable market skewing towards monopoly (ie, tipping) that, according to economic thought,²⁴ is very difficult to reverse and should be prevented through ex-ante intervention. Market tipping is a particularly ‘sticky’ phenomenon, as it decreases the incentive to innovate ‘sharply for all market players’.²⁵ Thus, because incentives to innovate on tipped data-driven markets are low, it is difficult—though not impossible²⁶—for them to ‘untip’.

Hence, one can see market power and consequent tipping in data-driven markets as close to, but just short of, a permanent market failure.²⁷ This in-between position warrants debates regarding how one should approach data-driven markets from a policy perspective. Traditionally, as both legal and economic experts point out, when a market is characterized by a failure other than market power—for example, information asymmetry—ex-ante regulation can be seen ‘as a prerequisite for there to be effective competition’.²⁸ However, when the issue is market power (a type of non-permanent market failure), the base slogan is ‘competition where possible, regulation where necessary.’²⁹ Hence, one-off ex-post competition interventions are the preferred default, with ex-ante regulation applied only exceptionally. Given this model, data-driven market tipping—being close to permanent market failure but based on market power (itself a non-permanent market failure)³⁰—has sparked vigorous debate about the appropriate regulatory approach to address it.³¹

While many voices, including the European Commission itself,³² expressed support for an invariable ex-ante solution to prevent tipping through employment of data-sharing mandates, management and business strategy literature remained more skeptical. From the Commission’s perspective, scholars such as Prüfer and Graef suggested that the adequate policy response was to introduce ex-ante regulation, including broad data-sharing obligations that would enable competitors of companies with large market shares to get access to progressively larger chunks of the latter’s data.³³ From the business and management perspective, fewer voices advocated for intervention,³⁴ with many calling for moderation in regulation or even for avoiding interference with digital markets altogether.³⁵ According to Evans, several peculiar features of digital markets make them work competitively, amongst which are low entry costs, low costs associated with adding or changing programme features through code, enhanced user experiences due to direct and indirect network effects, and the possibility of users switching between different providers (or ‘multi-homing’).³⁶

While the latter contribution is older and thus superseded by more recent accounts that demonstrate the specific potential of data-driven indirect network effects to tip markets,³⁷ a more recent account by the legal scholar Nicolas Petit has put forward a theory of ‘moligopolistic’ competition³⁸ that also discovers competition dynamics in these contexts. As Petit points out,³⁹ one should not look for competition on the narrower, tipped markets where digital giants hold their respective monopolies. Instead, enforcers should look at the big picture and examine whether there might be competition in the broader context of digital ecosystems. Hence, data-driven indirect network effects operating in a core market could be superseded as competition begins to emerge in adjacent markets. According to Petit, such competition is currently happening in the markets for payments, messaging, video streaming, productivity applications, hardware, cloud computing, mobile phones, among others.⁴⁰ This creates conditions for oligopolistic competition in these broader markets and, if not competitive markets, at least contestable markets.⁴¹ The oligopolistic competition in broader markets, in turn, keeps (near-) monopolists on their toes in the narrower markets which they dominate, as they are concerned that users might simply move on to the next ‘hot’ technology, abandoning the monopolists’ dominant offerings. Hence, the threat of the emergence of superior, differentiated substitutes is the engine behind the above-described ‘moligopolistic’ competitive dynamics. For example, Facebook was clearly concerned that users might leave its social network for the messaging app Snapchat, prompting its management to introduce new features that mimicked those of Snapchat. About 15 years prior, Google’s new Android mobile operating system (OS) posed such a serious competitive threat to Microsoft’s Windows desktop OS—due to the rise in mobile phone usage—that Microsoft began investing heavily in developing its own mobile OS.

The upshot for regulating moligopolistic structures is not to intervene ex-ante based on economic thinking that views market power as a market failure (this issue can be addressed through ex-post competition interventions), but rather to use tools such as consumer and data protection law that can better address the non-economic harms that moligopolists often bring with them.⁴² Under this view, there would have been no reason to intervene prospectively (ex-ante through the DMA) to address cases of data-driven market tipping.

The policy response following the above-described expert exchange was a curious amalgam of the considerations mentioned so far. In line with the Commission’s views, and contrary to the advice of Evans and Petit, an ex-ante regulatory approach was endorsed through the adoption of the DMA. The DMA is justified on grounds that are very different from market power—namely, fairness and contestability. In that sense, its rhetoric resembles that of ‘non-economic’ harms (failures different from market power) as outlined by Petit. The choice of ex-ante regulation makes sense when viewed in this context. At the same time, its core provisions⁴³—the positive and negative obligations imposed on gatekeepers—heavily borrow from the market-power-based logic of competition law. From this perspective, and if the logic is market-power-based, the justification for an ex-ante approach is less strong. This dichotomy was likely known to the Commission and might also potentially be the reason for the very carefully drafted, specific, and narrow scope of the DMA prohibitions. However, this approach also leads to a piecemeal (as opposed to holistic) tackling of the underlying market failures described above. This observation is all the more relevant for data-sharing obligations under the DMA, none of which prescribes true, large-scale data sharing; as such, the DMA data-related obligations can be defined as ‘fragmented’ or piecemeal. This fact creates gaps in the DMA’s protective scope, which we believe could be filled through competition law. In Section 4, we develop this point further and suggest careful ex-post enforcement, based on counterfactual thinking, should the narrow data-sharing mandates under the DMA trigger classical competition problems. Such a situation is highly likely, given their narrow basis that is not supported by theory, as explained in the following section.

Data-sharing mandates and their alternatives

Data is a crucial asset without which many 21^st century businesses cannot function properly. It has also come to feature prominently on the antitrust agenda on both sides of the Atlantic, with the EU⁴⁴ and its Member States⁴⁵ adopting respective legislation, and the US Congress probing data-driven giants through formal hearings.⁴⁶

The algorithms underlying digital services and apps, connected devices, and industry 4.0⁴⁷ production facilities require data to work effectively. Thus, data are at the core of digital companies’ business models⁴⁸ and are essential for their competitiveness in the market. More data⁴⁹—at least up to a point—allow for continuous improvements to the algorithms underlying apps, digital services, and smart devices.⁵⁰ This, in turn, enables companies to match the preferences of their customers ever more precisely.⁵¹ For example, search engines that collect more user data provide more relevant search results.⁵² Social networks with large data repositories show more interesting content and provide better friend suggestions to their users. Data-rich online retailers suggest the most desirable products to their users.

Additionally, better preference-matching abilities drive the acquisition of ever-increasing amounts of data.⁵³ Users who see highly relevant, personalized content are more likely to engage with it. This user engagement results in the creation of more data, which enables better algorithms and, ever more precise preference-matching abilities. A virtuous cycle (from the perspective of the companies that collect the data) ensues, allowing firms that own large data repositories to collect increasingly more data, become increasingly indispensable to their users, and consequently acquire greater market power.⁵⁴ This dynamic is driven by data network effects as mentioned above: these network effects create a winner-take-all (or winner-take-most) dynamic in which data-rich companies become increasingly more powerful. Data network effects are indirect and differ from direct network effects, or people network effects,⁵⁵ which occur when an additional user makes a product or service more valuable for existing users. Products or services exhibiting data network effects do not depend on the addition of new users to increase in value: while the addition of new users is sufficient to generate additional data points and enhance the value of the product or service, it is not necessary to add new users for this to happen. The engagement of existing users with the product or service is enough for new data to be generated and the value of the product or service to be increased. The more diverse services a digital platform offers (such as email, messaging, video, music, and telephony), the higher the engagement of users with the platform, and the more new data the platform can collect.

Several consequences follow from this. First, products and services that display both data and people network effects do not need to grow their user base in order to become more valuable for existing users; user engagement with the product or service is enough to increase the product’s or service’s value. Secondly, if the user base of a product or service exhibiting both people and data network effects does grow, the value of the product or service to existing users increases at an even faster rate. In that case, people and data network effects combine to make the product or service ever more valuable to existing users through improvements in preference matching and through growth in the number of people to whom existing users can connect. Thirdly, because a product or service exhibiting data network effects increases in value every time existing users engage with it, it is exceedingly hard for new entrants in the same market to steal away users from the product or service in question.⁵⁶

In other words, products or services displaying data network effects can build data moats that make it very difficult for new entrants in the market to compete.⁵⁷ In such a scenario, the product or service that first collects enough data to become very efficient at preference matching⁵⁸ develops an ever-widening lead over competing products or services. This means that competition in the market becomes less and less dynamic until only very few firms—or even just one firm—are left. These dynamics have been argued to constitute market failures (albeit not of a permanent nature), leading some legal scholars and economists to suggest interventionist measures to restore the balance of market forces in data-driven contexts as we saw earlier.⁵⁹ The remedy, according to these scholars, is to mandate data-driven companies that have captured a certain market share to make (a chunk of) their data available to competitors. The reasoning goes that if everybody has access to the same data, then everybody has the same chance of winning in the market.

While data-sharing mandates may seem like an ‘optimal remedy’⁶⁰ at first glance, the central question is whether the potential market failures introduced by data network effects are truly irreparable and whether other means of market correction is impossible. After all, data-sharing mandates are heavy-handed interventions that might have negative consequences for innovation and dynamism in digital markets. For one, they might undermine incentives for companies to try and collect large data sets on their own in order to develop sophisticated preference matching algorithms and get ahead of their competitors. This problem, although not related to data itself but to access to incumbents’ telecommunication networks, has already come to the fore in competition cases within regulated industries.⁶¹ In this sense, Geradin maintains that what can discredit mandatory access to a network is evidence showing efficient own rollout by newcomers on telecommunication markets: ‘[i]t is well known that facilities-based operators are much better able to compete than service-based competitors relying on the incumbent’s network’.⁶² At its core, this suggestion involves a counterfactual demonstration that discredits the initially alleged essentiality of network access. A similar reasoning, as shown in Section 4, can also be applied to ex-post competition assessment of data-driven harm and the decision for or against subsequent imposition of ex-post data-sharing mandates.

Data-sharing mandates might also punish particularly well managed, successful companies that had the foresight, innovation prowess, talent, and technologies to start accumulating, processing, and analysing data early on and thereby achieve a lead over the competition. Thus, ex-ante regulation through data-sharing mandates should only be applied when there is no other way to stop or counteract the mechanism leading to the market failure in question, ie, the observed market failure is nearly irreparable. To the extent that the state of ‘no turning back’ is not unanimously supported in the literature regarding data-sharing obligations, the ones currently existing under the DMA should be enforced with care. Additionally, should it, in the future, be needed to impose further data-sharing obligations for novel data-driven harms on the same or new infringers, these could be dealt with under a flexible, counterfactual ex-post competition assessment detailed in Section 4.⁶³

Also, one needs to keep in mind that there are emerging political and business initiatives that might reduce or eliminate the need for mandated data sharing in the first place. For one, France, Germany, and other European partners have invested in the GAIA-X federated cloud infrastructure,⁶⁴ which allows smaller companies to gain access to the data of larger firms that these firms share (albeit on a voluntary basis). In addition, the EU made it mandatory in the General Data Protection Regulation (GDPR) that companies make user data portable, not only allowing data subjects to obtain the data that a data controller holds on them but also enabling the transfer of that data to another data controller. This empowerment of data subjects fosters competition between different data controllers as a welcome side effect.⁶⁵ Moreover, there are voluntary data-sharing initiatives driven by industry itself such as the ‘Open Data Initiative’⁶⁶ and the ‘Data Transfer Project.’⁶⁷ However, as numerous regulatory governance scholars have pointed out,⁶⁸ it is not uncommon for business actors to come up with their own rules in order to prevent regulatory advances by the government (which often go further than private initiatives). Thus, voluntary initiatives might not be as well designed and inclusive as public initiatives. For example, the companies involved in the ‘Open Data Initiative’ are mostly large and medium-sized companies, not small startups, limiting the benefits of the initiative for smaller firms. Despite these caveats, private data-sharing initiatives of this kind can help counteract the winner-take-all dynamics created by data network effects and produce at least some of the same results as mandated data-sharing agreements.

However, data sharing might not even be necessary to achieve welfare-enhancing effects in data-driven markets. Thus, industry experts have identified various business and technical strategies that enable small companies to bootstrap the ‘minimum viable corpus’⁶⁹ of data that is needed to begin developing sophisticated algorithms (eg, internalising early users’ externalities by offering them something in return for their data; using mechanisms such as web crawling to automate data capture from available sources; relying on transfer learning to repurpose data from other domains; and programmatically creating data to train against).⁷⁰ The last point refers to the possibility of generating large data sets with synthetic data, ie, data artificially created on the basis of a generalised statistical approach that removes personally identifiable information from datasets. As one expert points out, ‘[t]his lowers the barrier to deploy data science by removing the need for vast volumes or real data,’⁷¹ in effect allowing small companies to amplify their datasets and compete better with large, data-rich companies.

Proponents of data-sharing mandates might counter that none of the aforementioned methods, regardless of their merits, ensure that smaller firms have access to similarly large-scale data sets as established data-driven firms. However, such access might not be needed (as might be implicitly acknowledged by those advocating the sharing of only a slice of established companies’ data with smaller competitors). Thus, it has long been known that while more data are always strictly welfare enhancing,⁷² there are scenarios where the incremental cost of more data goes up after a certain point and further data collection or analysis does not lead to more accurate algorithms.⁷³ This means that, after that point has been reached, gathering more data is not economical and therefore makes very little sense. Thus, while larger datasets are always at least as valuable as smaller datasets (ignoring their higher costs),⁷⁴ a smaller data set will often perform just as well in terms of generating an accurate algorithm. In many contexts, smaller datasets can therefore be as helpful as large data sets and reach the right conclusions faster, more reliably, and at lower cost.⁷⁵ As Bryson pithily puts it: ‘[M]uch data is redundant except for surveillance purposes (or even for those).’ ⁷⁶ For example, researchers at Stanford University used a ‘small data’ model based on high-dimensional statistics, which maximises the joint predictive power of all data in a data set, to significantly reduce the number of tests needed to identify a range of common diseases.⁷⁷ Industry developments might decrease the need for big data even further. Thus, Renda points out that the move to edge computing and to local data collection, processing and analytics will bring about a number of changes, including that ‘small data become more important than big data; data storage occurs mostly at the edge and in devices; and [core] distributed architectures are possible, with possible consequences also for competition.’⁷⁸

Thus, as laid out above, while data network effects lead to the accumulation of market power, the potential market failures introduced by data network effects are not irreparable and correction of the market is theoretically possible. Finally, there are several practical difficulties regarding the implementation of data-sharing mandates. For one, it is unclear how exactly large digital platforms would share data with smaller competitors. ‘Fresh’ data are often (though not always) more valuable than older data, so one key question is whether data would be shared in real time. Another issue is how smaller companies with fewer server capacities would handle the immense server storage space and processing requirements associated with huge data sets. It is also worth mentioning that large data sets will only be useful if companies actually have the talent and technology to analyse these data, draw the right conclusions from it and write sophisticated algorithms that result in performance improvements (ie, in better preference matching abilities). Given that large data-driven platforms pay among the best salaries in the industry, smaller competitors might struggle to attract the right employees to actually do something useful with the data shared by dominant digital platforms.

The above discussion puts in perspective the Commission’s regulation of digital markets and acknowledges that—although ex-ante intervention is certainly one tool for going after data-driven market power and concomitant tipping—less intrusive alternatives to data sharing remain available. However, even without siding with a particular ‘camp’, if we go back to the theory of data-driven network effects described in ‘Data-driven market tipping: the theory’ section, we see that the prescription is either for no data-sharing mandates whatsoever (Petit)⁷⁹ or for all-out broad data sharing (Prüfer and Graef).⁸⁰ Hence, it becomes clear that the DMA, with its narrow data-sharing mandates, follows neither of the two approaches. As Prüfer and Graef put it regarding data-sharing duties under the DMA, ‘the scope of the duty is restricted to search data and to a few large online platforms acting as gatekeepers’.⁸¹ Hence, although the above-reviewed literature suggests that — theoretically — either broad or alternatively no ex-ante data-sharing mandates are the two plausible regulatory responses, this is not where policy went in practice. Instead of a theory-inspired approach (either broad or no ex-ante data-sharing mandates), the DMA opted for a piecemeal approach of narrow ex-ante data-sharing mandates as expressed in—but not limited to—Article 6(11) DMA.⁸²

We maintain that this approach opens the possibility for greater antitrust harm downstream than would occur under a broad or no such mandate. This is because a narrow data-sharing mandate is like granting only partial access to infrastructure in classical utility industries. For instance, one could imagine a hypothetical situation where only copper wires, under telecom regulation, would be subject to access and price controls (excluding coax cables and other new generation infrastructure). This will create opportunities for the owner of the infrastructure to cross-subsidize its lower revenues on the regulated part of the network with higher pricing on the non-regulated parts. The same effect is also likely to be observed under narrow data-sharing mandates in digital markets—the lost value of data due to mandated sharing in search services can be recouped on other markets where no such sharing mandate exists.⁸³ This can be done through usage of unfair terms or exploitative data practices that are prohibited under Article 102(a) TFEU. Hence, both businesses and end users in these (connected) markets will be worse off. Under a broad data-sharing mandate, no such cross-subsidization would have been possible (at least in the short run).⁸⁴ Additionally, narrow mandates are also worse that no mandates, as they impose enforcement costs while failing to solve the problem of data-driven market tipping. As Sauter and Akker aptly put it, certain regulatory rules—such as the DMA’s data-sharing rules in our case—can create ‘imbalances’ in the applicable ex-ante framework. Importantly, they argue that ‘[…] in cases of imbalance, the regulation creates room for competition abuses. Excessive pricing cases can be seen as a sign of imbalance of the underlying regulation.’⁸⁵ Although this statement was made in the context of the pharmaceutical industry, we maintain that this is precisely what we are likely to see under the DMA’s data-sharing regime as well. Hence, while co-enforcement against gatekeepers under both the DMA and EU competition law is inevitable, caution is advised when antitrust is applied to fix the imbalances of the DMA’s data-sharing regime (Section 4).

To summarize, given that anti-competitive harm can materialize as unintended consequence of the DMA’s narrow data-sharing rules, careful foreclosure assessment under antitrust will be pivotal. Therein, the existence of ex-ante narrow data-sharing mandates should inform the building of an appropriate counterfactual by being considered as part of the ‘conditions on the relevant market’.⁸⁶ Additionally, appropriate remedies (which could include imposing a different type of data-sharing mandate) should directly stem from the findings of the counterfactual foreclosure assessment (through applying a proportionality test).⁸⁷

In order to develop these points, we first discuss, in more detail, the concrete narrow ex-ante data-sharing mandates laid out in the DMA (Section 3). Section 4 proposes that the latter be employed in the building of counterfactuals for the purposes of effects/foreclosure analysis that then feeds into the remedial stage and informs the decision regarding the need for ex-post data-sharing mandates. While the latter point (remedies need to logically/proportionally flow from the infringement) is established in literature,⁸⁸ counterfactual thinking at the level of anticompetitive foreclosure in ex-post competition analysis under Article 102 TFEU remains a rather unexplored topic. Although suggested in a Commission Guidance Paper,⁸⁹ courts have not yet explicitly engaged with this standard under Article 102 TFEU, and scholars have not been vocal about it either. Hence, in Section 4 this contribution shows how picking up the old proposal for counterfactual reasoning to establish foreclosure might work in the context of data-driven markets.

3. DATA-SHARING MANDATES UNDER THE DMA: A NARROW APPROACH

The core of the current DMA comprises Articles 5 and 6, which list practices that are outlawed should they be performed by a gatekeeper, a provider of a core platform service. Several of these provisions have bearing on market tipping caused primarily by data-driven indirect network effects,⁹⁰ but—as explained above—all of them are rather piecemeal and largely borrow from existing specific competition cases.⁹¹ The provisions that deal concretely with data-related obligations comprise 5⁹² out of 18 obligations in total and, although most of them are not inspired by past or ongoing competition cases, they do seem to track individual business complaints against data practices of digital gatekeepers.⁹³ Hence, a point for improvement for the DMA might be the introduction of a more integrated approach that relies less on competition precedents or complaints and more on theory. Thus, the DMA could have introduced a broad data-sharing mandate,⁹⁴ or none at all,⁹⁵ as explained above.

To summarize, the Commission seems to be basing itself in practice, which invariably leads to a piecemeal approach and to narrow data-sharing obligations without further delineation. Under the current Articles 6(9) and 6(10), the obligations imposed are for (i) the portability of end-user data and (ii) portability of business user data, respectively. These provisions, however, only refer to claiming back data already generated by certain business or private users and as such do not constitute broad data sharing. The only provision that allows for broad data sharing but is circumscribed in scope (applying only to search engines) is Article 6(11) DMA. Finally, Articles 5(2) and 6 (2) are inspired by specific narrow practices that formed the subject matter of the Facebook Germany case (Article 5(2))⁹⁶ and the Commission’s investigation against Amazon, respectively.⁹⁷ All in all, it seems that only Article 6(11) DMA comes close to the kind of broad data-sharing mandates envisioned, inter alia, by Prüfer and Graef, with the important caveat that this provision only applies to the specific case of search engine data, which again narrows it down significantly.

Mandates for sharing data with competitors are, moreover, not limited to the DMA. They have also been included in other EU regulations as well, such as the revised Payment Services Directive (PSD2) passed in 2015, which enables financial technology (fintech) firms to gain access to the customer account data of large banks with customers’ consent. In Articles 66 and 67, PSD2 does not mandate banks to share all their data (eg, risk assessment and credit scoring data or internal operations and security data). Instead, it requires them to share the subset of their data (customer account data) that is the most relevant and valuable for fintech firms. Having access to these data does not just allow fintech firms to understand customer behaviour and tailor their services to customers’ preferences but also allows them to see (at least some of) the fees that banks charge for their services. This, in turn, enables them to potentially offer these services at lower prices than banks, thereby creating real competition in the market. We regard the data-sharing mandate entailed by PSD2 as more sensible than the one entailed by the DMA, as it gives fintech firms access to the data that they need to compete effectively with banks. The DMA, in contrast, only gives gatekeepers’ competitors access to limited (self-generated or search) data, which is valuable for other search engines but by no means covers all of the data that gatekeepers hold that are relevant to smaller competitors and that, if shared, might create real market competition. Thus, it is likely that ex-post legal challenges will be brought forward by gatekeepers’ competitors if these competitors believe that broader access to data would safeguard fair market competition.

Additionally, there are enforcement hurdles as explained by Lundqvist—while ‘the combination of Articles 6(2), 6(9) and 6(10) creates a compulsory access and use regime […] stopping just short of a property right’, the actual ability of business customers of platforms to enforce in the current DMA setup ‘remains a bit unclear’.⁹⁸ In particular, the relationship with legislation potentially blocking the type of data sharing envisioned in the above DMA provisions (such as the GDPR) leaves a lot to speculation. Therefore, the concern with the above three provisions taken as a whole is that they do not ‘have teeth’ in practice. By contrast, the issue with Article 6(11) DMA is its overly narrow scope, as already explained. Hence, as argued above, based on theory, the mandated data-sharing obligations under the DMA should have been either broad or none at all. This is not currently the case, which prompts the need for adapting ex-post competition enforcement accordingly. Thus, our proposal for refining counterfactual causal thinking when establishing anticompetitive foreclosure and remedies under Article 102 TFEU. Ensuring careful counterfactual thinking ex-post would ideally mitigate (i) potential regulatory inconsistencies between Article 102 and DMA enforcement and (ii) concomitant error costs (false condemnations/‘false positives’ and false acquittals/‘false negatives’, respectively).

In summary, when it comes to the current data-sharing mandates under the DMA, it seems that their piecemeal nature leaves the door open for antitrust challenges and harms. In this situation, the question arises as to how to enforce competition law towards an already regulated gatekeeper in order to close the abovementioned gaps and without creating (i) regulatory inconsistencies and (ii) committing enforcement errors.

4. THE CASE FOR A STRENGTHENED COUNTERFACTUAL CAUSATION TEST IN EX-POST COMPETITION ENFORCEMENT AGAINST DIGITAL GATEKEEPERS

Setting the scene

Taking DMA data-sharing obligations that already apply to designated gatekeepers as an example,⁹⁹ we argue that all of the above objectives could be rather well achieved through the use of multiple counterfactual effects analysis in ex-post abuse of dominance cases against gatekeepers. This is because the establishment of antitrust harm under Article 102 TFEU through the use of multiple counterfactuals could become a carefully crafted exercise, taking due account of pre-existing regulatory obligations (and their limitations/narrowness), as will be demonstrated below. Therefore, this section proceeds to make a case for the use of a counterfactual causation method in ex-post competition enforcement against gatekeepers as follows: first, in ‘Type I, Type II errors and overlapping regulatory regimes’ section, the problem of committing false-positive (and false negative) errors by enforcers when two regulatory regimes overlap will be charted out, in order to then argue that the proposal we make could mitigate both error costs. In this section, we also address the additional and related concern of ‘regulatory inconsistency’ between the two overlapping regimes (competition law and the DMA). Thereafter, in ‘Counterfactual logic in antitrust’ section, we introduce counterfactual thinking in antitrust more generally. ‘Causation and counterfactual causation in Article 102 TFEU’ section demonstrates how it can concretely be deployed in the realm of ex-post competition enforcement under Article 102 TFEU. Finally, 'The counterfactual link and (data-driven) remedies’ reflects on how such thinking can be used not only to assess potentially anticompetitive behaviour of designated gatekeepers under Article 102 TFEU but also to inform the design of appropriate remedies under the same provision. The last section concludes this article.

Type I, Type II errors and overlapping regulatory regimes

Exactly 40 years ago, Frank Easterbrook introduced his influential framework on antitrust error cost analysis, concluding famously that—given the self-correcting power of markets—relevant enforcers are better off underenforcing (committing false negatives) than over-enforcing (committing false positives).¹⁰⁰ Since then, this framework has been questioned most prominently by industrial organization¹⁰¹ and behavioural economists¹⁰² who showed that many markets are far from ‘perfectly competitive’ and ‘self-correcting’, thus mitigating the fear of false positives in antitrust analysis.¹⁰³

However, the issue of over-enforcing still looms large specifically when firms are subject to overlapping economic regulatory regimes, as in the case of digital markets in the EU. For instance, in its Android judgment of 2022, the General Court showed a preoccupation with false positives by stating that ‘close examination of the actual effects’ of a practice are needed before it can be concluded that said practice is harmful to competition. The court then elaborated that such an examination ‘serves to reduce the risk that penalties may be imposed for conduct which is not actually detrimental to competition on the merits […]’,¹⁰⁴ in other words—the GC was clearly watchful for potential false positives.¹⁰⁵ This preoccupation is also shared in scholarship. In this regard, Unver points out the impossibility for gatekeepers to provide defenses for their behaviours under Articles 5 and 6 DMA—an option that is available in other regimes, such as the UK, for instance.¹⁰⁶ A similar sentiment regarding the DMA prohibitions is voiced by Ntemuse who states that ‘the imposition of blanket per se rules is likely to significantly heighten the risk of Type 1 errors (false positives)’.¹⁰⁷ In the context of Article 5(2) DMA concretely, Plantinga also reasons that ‘certain practices that could be procompetitive are sanctioned by default’.¹⁰⁸ In the upcoming sections, we argue that the above concerns could be mitigated by counterfactual effects-based analysis of anticompetitive foreclosure by gatekeepers under Article 102 TFEU because—as the GC itself maintains in Android discussed above—examination of the actual effects of a practice serves to reduce false positives. Unfortunately, in the judgment, the court does not engage with counterfactual effects assessment in sufficient depth, as will be explained in ‘Current practice under Article 102 TFEU’ section. This fact will hopefully be corrected by the Court of Justice, where the case is currently awaiting a final judgment on appeal from the GC.¹⁰⁹

It is also important to note that overlapping (ex-post and ex-ante) enforcement does not only raise issues of potential false positives. A risk of false negatives, which is equally problematic,¹¹⁰ also exists. For instance, if a gatekeeper firm is treated too leniently in a national antitrust investigation, under an (erroneous) assumption of a national antitrust enforcer that the DMA prohibitions have already debilitated the gatekeepers’ anti-competitive potential.¹¹¹ We argue that—through its potential for careful calibration of effects—multiple counterfactual assessment under Article 102 TFEU could address both issues. ¹¹² This consideration is all the more relevant on digital markets because of the following concern expressed by Crocioni: ‘not only may emerging markets change the overall probability of incurring type I or type II errors, they may also, perhaps more importantly, increase the costs of making such errors compared with an “average” case’.¹¹³

Lastly, and related to the error costs concerns mentioned above, a potential ‘regulatory inconsistency’ between the DMA and competition law is being observed in scholarship.¹¹⁴ More specifically, the calibration function of counterfactual assessment in ex-post competition enforcement is of paramount significance now that the DMA is being enforced, given concerning developments¹¹⁵ around the interaction between competition enforcement and sectoral regulation. To illustrate, in the Slovak Telekom case, the first instance court [the General Court (GC)] made and won the argument that—because the defendant is subject to an ex-ante duty to deal under telecommunications regulation—there is no need to apply the ‘indispensability’ criterion to it under an Article 102 TFEU refusal to supply analysis. While ‘indispensability’ is the most crucial prong of the classical abuse test for refusal to supply,¹¹⁶ according to the logic of the GC, if an undertaking is subject to an ex-ante, regulatory duty to deal, its services are by definition ‘indispensable’. This is so because the finding of ‘indispensability’ of a certain service or infrastructure is a prerequisite for the latter being subjected to ex-ante regulation. However, as Geradin and O’Donoghue observe, the real issue is ‘[…] whether the substantive conditions under which ex ante regulation was imposed were the same as the conditions for ex post abuse’.¹¹⁷ Given that the two regimes have different objectives and enforcement designs, the assessment of the question of whether a service or infrastructure is ‘indispensable’ is bound to vary. Hence, a finding of ‘indispensability’ under one regime cannot be automatically used to inform the other. Otherwise, one risks a leap of logic that abridges the antitrust test for refusal to supply in a dangerous way, leading to a higher likelihood of faster (or even false) convictions. As concluded by Berqvist, in such cases: ‘sector regulation not only creates but also forms an antitrust infringement, allowing plaintiffs to present their claims in the manner most beneficial for this and not as dictated by fact and the law’. It is also notable this is not the first case where competition law and telecom regulation interact counter-productively and hence produce regulatory inconsistencies.¹¹⁸

It is not unthinkable that similar dissonances could emerge in the interaction between the DMA’s (data-sharing) obligations and ex-post competition enforcement on digital markets. It is in this context that Ganesh remarks: ‘Considering this, it might be ideal to have competition authorities and regulatory authorities play a joint role for the betterment of consumers and competition.’¹¹⁹ While this is a very apt solution at an institutional design level, in this article we argue for the need of a solution at the level of substantive assessment of gatekeeper abuse of dominance cases under Article 102 TFEU as well. To prevent the above-charted dissonances and the concomitant decision-making uncertainty and potential regulatory errors, a flexible counterfactual approach to assessment of anti-competitive effects (and remedies) is needed under Article 102 TFEU. Specifically, the counterfactual analysis needs to—among others—carefully consider the presence of regulatory obligations under the DMA and their effects on the market at the time of ex-post inquiry.¹²⁰

Counterfactual logic in antitrust

Counterfactual analysis is a type of causal test that law borrows from the social sciences and humanities.¹²¹ The test requires the following logical connection between two events: ‘Event E counterfactually depends on C iff,¹²² if C were not to occur, E would not occur.’ This logical postulate is derived from the realm of philosophy.¹²³ Translated to the field of EU competition law and the abuse of dominance doctrine that is relevant for digital gatekeepers, it would require a causal demonstration as follows. Once the law has identified a potential of a certain behaviour to exert negative effects on competition, the enforcers abstract themselves from the alleged behaviour and construe the economic picture that would have prevailed had that behaviour not occurred. As explained by Scott-Hemphill,¹²⁴ a second way for constructing a counterfactual would be to devise a model simulating the market where a different possible action than the suspicious conduct was taken instead of the actual conduct. These are also the two varieties of counterfactual thinking prescribed in the 2009 Guidance Paper on Article 102 TFEU (the Guidance Paper), which is the first instrument containing the word ‘counterfactual’ in the abuse of dominance domain.¹²⁵ We will turn to 2009 Guidance Paper in more detail under ‘Causation and counterfactual causation in Article 102 TFEU’ section, while we now proceed to outlining counterfactual logic in antitrust assessment more generally.

Counterfactual thinking is associated with the so-called ‘more economic’ approach in EU competition law,¹²⁶ which ensued exactly 20 years ago, with the entry into force of Regulation 1/2003.¹²⁷ Since then, even though the approach has been fully embraced in the enforcement of anticompetitive agreements (Article 101 TFEU), it has not been widely or consistently endorsed by courts in abuse of dominance cases.¹²⁸ This is also true for counterfactuals which—as instruments introducing economic analysis in the law—have only been mentioned in six Article 102 cases of the CJEU so far but abundantly used under Article 101 TFEU.¹²⁹ As rightly observed by Colomo, this situation de facto means that economic ‘effects would have a different meaning under Articles 101 and 102 TFEU’.¹³⁰ Acknowledging this situation, scholars have called for an equally ‘economic approach’ under the doctrine of abuse of dominance.¹³¹ In this context, and zooming in on data-driven digital markets, this work puts forward the potential of counterfactual assessment as a method for nuanced antitrust analysis in complex regulatory settings, such as the one at hand.

Nuance can be achieved through counterfactual thinking because it is based on logic and—in the context of competition law—economic analysis, while also allowing for versatility. Here, we draw on the conclusions of Scott-Hemphill from the US antitrust context.¹³² The author shows that there are different types of counterfactuals, all of which are logically coherent and could be economics-laden, but—depending on their type—afford varied scope for intervention to courts. Thus, we first examine the logic of counterfactuals as per Scott-Hemphill and then show how economics can help further ‘build’ them.

Scott-Hemphill charts out the following counterfactual types: ‘net effects test’ and ‘least restrictive alternative (LRA)’ test, with sub-types ‘dominant LRA’ and ‘balanced LRA’. The former test is a ‘cost-benefit analysis that explicitly balances, or trades off, the incremental harms and incremental benefits of the conduct compared to a world without the challenged conduct.’¹³³ This test is more difficult for courts as it requires them to engage in balancing—‘courts have only a limited capacity to assess and compare the pros and cons. Quantification is difficult given that parties deny the existence of a tradeoff rather than providing guidance about making the tradeoff’.¹³⁴ Instead, the latter (LRA) test aims ‘to compare the conduct to a hypothesized alternative and ask whether the alternative action is less harmful in the particular sense that it is “less restrictive”’.¹³⁵ In the latter scenario, two sub-options are available—a ‘dominant LRA’ is one whereby the alternative action is both less restrictive to competition and as effective as the impugned conduct. As Scott-Hemphill puts it, the inquiry here is ‘could the good have been achieved equally with less bad’?¹³⁶ However, such dominant LRAs are difficult to find in antitrust inquiry.¹³⁷ Most available alternatives are only better ‘on the balance’ (balanced LRAs), in the sense they are less restrictive but also less effective than the impugned conduct.¹³⁸ What are the implications of this categorization for our inquiry? First, from this discussion it transpires that it is possible to rank counterfactuals, with dominant counterfactuals taking precedence over balanced counterfactuals, while both of these are superior to a net effects test. ranking might allow Additionally, there is one counterfactual option that is often not considered in antitrust inquiry, but is superior to all of the above—namely, a counterfactual that is more effective/beneficial than the impugned conduct while being less restrictive.¹³⁹ Hence, if enforcers are to adopt counterfactual thinking under Article 102 TFEU, they could take into account the abovementioned ranking,¹⁴⁰ which could potentially help them economize resources by exploring only the most promising counterfactuals (the lowest threshold could be drawn at balanced LRAs, for instance). In this sense, they are also aided by economic analysis that is used to ‘build’ counterfactuals. Economic analysis—depending on the sophistication of the causal inquiry—could use either regression analysis (lower level of ability to make causal predictions)¹⁴¹ or structural models (higher ability to make causal predictions).¹⁴² In this sense, courts could distinguish between superior and inferior counterfactuals as well. Finally, as suggested by Marsden and Colley, ranking of counterfactuals could also happen on the basis of probabilistic analysis—‘ie identifying the range of plausible scenarios and attaching a probability weighting to each instead of having to identify one above all others’.¹⁴³ Given these considerations at the level of theory, we now proceed to examining how counterfactual thinking under Article 102 TFEU actually works.

Causation and counterfactual causation in Article 102 TFEU

As noted in ‘Counterfactual logic in antitrust’ section, counterfactual thinking under Article 102 TFEU was endorsed by the European Commission in the 2009 Article 102 Guidance Paper,¹⁴⁴ currently under review.¹⁴⁵ Paragraph 21 suggests assessing anticompetitive foreclosure by comparing the actual or likely future market situation with a counterfactual, such as the absence of the conduct (net effects analysis under Scott-Hemphill’s definition) or another realistic scenario (LRA under Scott-Hemphill’s definition). These options, combined with Scott-Hemphill’s insights, provide a basis for devising and ranking multiple counterfactual scenarios. However, the 2009 Guidance Paper and its counterfactual logic did not gain traction,¹⁴⁶ meaning that the problem with counterfactual causation in Article 102 cases still remains the non-application or—at best—incoherent application of the test by the CJEU.¹⁴⁷

Given counterfactual thinking’s importance for error reduction and regulatory consistency, we recommend taking seriously the European Commission’s proposal to use counterfactuals in ex-post abuse of dominance analysis (especially on data-driven digital markets). This analysis can also inform appropriate remedies. For example, if a dominant platform raised prices due to data-sharing obligations, a suitable remedy might be temporary exemption from these mandates.¹⁴⁸ Alternatively, if price increases result from a vertically integrated monopoly in AdTech,¹⁴⁹ structural separation might be appropriate.¹⁵⁰ Hence, our argument emphasizes careful counterfactual consideration at both the foreclosure analysis and remedial stages.

Current practice under Article 102 TFEU

Digital giants and specifically their market-distorting practices on the territory of the EU¹⁵¹ are covered by Article 102 TFEU—a provision with the primary purpose of prohibiting abuses of dominant position on the Internal Market. The structure of the general prohibition, which borrows elements from both tort and criminal law, requires that several elements are fulfilled before it obtains. For the purposes of the current argument, the following conditions are of importance. The first two are (i) the establishment of a dominant position on a specifically defined relevant market and (ii) the abuse of that position, defined by its actual or potential anti-competitive effects on that market.¹⁵² Crucially, the establishment of a causal link between the abuse and its effects on the market (foreclosure assessment) is also a core part of the test under Article 102 TFEU, the law requires it, and we argue it should be established by means of counterfactual reasoning only, at least when it comes to digital markets for reasons explained in ‘Counterfactual logic in antitrust’ section.

It should also be noted that while a causal link between abuse and effects on the market has to be established, there is no hard obligation as to what kind of causal test is used in this respect.¹⁵³ As Veljanovski explains, the current case law on dominance mostly uses ‘benchmarks’ that serve as shortcuts to establish causation.¹⁵⁴ In other words, if the defendant manages to put forward a credible alternative story to an anti-competitive effects allegation by the Commission, this suffices. This rather lax standard also explains problems at the remedial stage of Article 102 TFEU cases, for which practice shows remedies that do not work as intended—namely, they miss the mark of restoring effective competition.¹⁵⁵ Hellstöm¹⁵⁶ attributes this fact to the phenomenon of ‘mirroring’, reasoning as follows ‘[…] mirroring the abuse may be the intuitive remedy of choice. However, in the event that this intuitive remedy is not likely to be successful or effective, one must further explore what instruments are at the Commission’s disposal—taking the effect of the infringement as a starting point.’¹⁵⁷ Hence, establishing the actual effects of the infringement at the foreclosure stage—an exercise for which counterfactual analysis is singularly suitable—is of pivotal importance for being able to come up with working remedies at the next stage of the assessment of an impugned practice. While we will examine the precise mechanism that we suggest for establishment of remedies in ‘The counterfactual link and (data-driven) remedies’ section, we first offer a practical demonstration of the centrality of the causation concept and the importance of using a counterfactual to establish a causal link between abuse and its effects.

Let us take the Post Danmark I case of the CJEU as an example.¹⁵⁸ Therein, the court did acknowledge the centrality of the so-called ‘as-efficient-competitor’ (AEC) test, itself a counterfactual construction,¹⁵⁹ in order to gauge the anti-competitive effects of discriminatory pricing by the defendant dominant undertaking (Post Danmark). However, the counterfactual AEC test itself was performed by means of an abridged, or ‘benchmark’, analysis, as Veljanovski puts it. Namely, the established case law test for abusive below-cost pricing was applied,¹⁶⁰ using the cost data of the dominant undertaking. Therefore, a full-fledged analysis of the market was not performed—specifically, the actual costs of competitors could also have been looked at, with the aim to see whether they could truly survive competition in the long run. Instead, the court signaled that it was convinced by the argument of the defendant—namely, that during the period of alleged abuse, its competitor ‘managed to maintain its distribution network despite losing the volume of mail related to the three customers involved and managed, in 2007, to win back the Coop group’s custom and, since then, that of the Spar group’.¹⁶¹ It is not difficult to see that arguing that your competitor managed to survive does not exclude the possibility of market power abuse on your side, nor of further negative effects on the market beyond the setbacks suffered by your competitor. In this particular case, a true counterfactual approach through the AEC test could have extended to examining actual competitor costs, but also to modelling market conditions without the alleged abusive behavior. This is all the more important, given that a market connected to the relevant market in the case was subject to regulation—a fact that, as discussed in the case itself, had an impact on the dominant undertaking’s costing structure (but likely also on the access conditions to the in casu relevant market). An explicit counterfactual assessment of this latter point did not figure in the decision while it could have shed more light on the effects of the alleged anti-competitive behaviour.

Let us also examine a more recent case where counterfactuals play a role—Google Android.¹⁶² The case is under appeal at the Court of Justice at the time of writing and includes several counterfactual examinations, put forward by both the Commission and the claimant Google. For the purposes of illustration, we only examine the first claim of the case, regarding the so-called ‘MADA’ agreements, where the theory of harm was anticompetitive tying due to Google obliging mobile phone manufacturers to pre-install Google Search and the Chrome browser in order to obtain access to the Play store. In its criticism to the Commission foreclosure analysis, Google argued, by means of a counterfactual, that its Android system would not have been able to survive on the relevant mobile operating systems market had Google not used tying practices in the mid-2000s when that market was dominated by its competitors.¹⁶³ The argument—though supported by academics¹⁶⁴—was dismissed by the General Court as it did not ‘reflect the content of the contested decision’¹⁶⁵ due to being too broad-brush. According to the CJEU, the Commission’s reasoning and the submitted counterfactuals were going against specific counterfactual practices of Google (the so-called MADA’s) and not against its decision to offer Android as an open and free licensing system.¹⁶⁶ Still, given that the two italicized elements are inter-related, the reasoning of the CJEU does not convincingly explain why it could not examine the ‘broader’ counterfactual argued by Google vis-à-vis the ‘narrower’ one submitted by the Commission. This is not a novel approach either—testing the potentially abusive conduct against alternatives is called ‘scoping the counterfactual’ by Marsden and Colley,¹⁶⁷ and has been extensively used in competition decisions by common law courts.¹⁶⁸ It is therefore of pivotal importance that the CJEU clarifies what the law is in this regard when it delivers its final decision on the case.¹⁶⁹

All in all, we believe that the ‘benchmark’ or any other approach¹⁷⁰ to establishing a causal link short of a fully-fledged multiple (at least ‘balanced LRA’) counterfactuals poses serious risks for over-simplification in the context of fast-changing digital markets. As exemplified by the case law discussed above, this concern is not merely hypothetical and will play a significant role in future antitrust administrative and judicial decision-making in the (data-driven) digital realm.

Proposal for novel counterfactual causation approach on data-driven markets under Article 102 TFEU

Although counterfactuals are not widely accepted as seen above, scholars have argued that applying counterfactual causation under Article 102 TFEU can—at least methodologically—align dominance case law with the type of counterfactual reasoning used under Article 101 TFEU for decades.¹⁷¹ As observed above, using counterfactual analysis for anticompetitive assessment can also better align remedies with impugned conduct through a shared focus on tackling anticompetitive effects.

Geradin argues that there are multiple ways in which counterfactuals could be constructed, but the common feature of Article 102 TFEU counterfactuals is that they are invariably based on the reconstitution of the past.¹⁷² However, there are multiple ways/counterfactuals through which one can reimagine the past, and all of them merit attention on dynamically changing data-driven markets. Hence, this article’s call for a multiple counterfactual causation analysis, whereby one can use econometric simulations to compare the current state of affairs with (i) a world without the infringement in question or (ii) with an ‘alternative realistic scenario’. Such a Guidance-paper inspired approach, entailing the creation of multiple counterfactuals, is a practice not unheard of in jurisdictions beyond the EU.¹⁷³

One way in which the ‘realistic alternative scenario’ counterfactuals can be constructed is by using the parameters of anti-competitive foreclosure as identified under paragraph 20 of the Article 102 Guidance Paper. Most of these parameters echo Porter’s ‘five forces analysis’¹⁷⁴ (conditions of entry and expansion; position of customers and suppliers; position of competitors) and can be varied, ceteris paribus.¹⁷⁵ For example, if there are indications that market exit could happen (eg, due to a failing firm), one can create an ‘alternative scenario’ counterfactual with that exit factored in. If it is shown that competition in the alternative scenario will deteriorate further, then the current scenario is to be preferred.¹⁷⁶ Indeed, this was one of the counterfactual claims made by Google in the Android case as seen above. As suggested by Scott-Hemphill above, if any of the above scenarios is either a dominant LRA or—even better—more beneficial LRA, it has to be preferred over the rest. This conclusion can even be further strengthened if the probability of materialization of the said counterfactual is high, as per Marsden/Colley.¹⁷⁷

By the same token, if there are indications that regulation prevents or somehow distorts competition, or the defendant puts forward such a theory as a defense, a counterfactual without regulation (or one that varies some regulatory obligations) can be created.¹⁷⁸ Given that our focus in this article is on data-driven markets, and ex-ante data-sharing mandates in particular, these specific regulatory obligations can be varied as they constitute ‘conditions of entry and expansion’ in the framework of the Article 102 Guidance Paper. Hence, several ‘alternative scenarios’ can be counterfactually constructed—with (i) no mandates, (ii) less, or (iii) more intrusive ex-ante data-sharing mandates. Such an exercise will enable the assessor to isolate the influence of data-sharing mandates on the impugned conduct, thus providing information about the relationship between mandated data sharing and anti-competitive effects. This information can then be fed into the remedial stage that, as will be seen in ‘The counterfactual link and (data-driven) remedies’ section, can show the need for additional ex-post data-sharing mandates or for no data-sharing mandates at all. Of course, multiple other counterfactuals can be constructed, including the classical ‘without infringement’ counterfactual. However, the above-described ‘alternative scenarios (LRA)’ counterfactual should be considered first, given the ranking introduced by Scott-Hemphill in ‘Counterfactual logic in antitrust’ section.

Once generated and ranked as per the framework in ‘Counterfactual logic in antitrust’ section, if there is still doubt regarding the choice of counterfactual, the outcomes can also be normatively assessed on the basis of the reasonableness standard. The so-called ‘reasonableness’ theory in tort and criminal law is applied in cases where several events could have caused a single particular outcome (ie, an anti-competitive effect).¹⁷⁹ In order to decide whether a causal connection can be established, the criterion of ‘reasonable attribution’ of a cause to an effect is used. This approach evidently introduces a subjective judgment regarding what is ‘reasonable’ and therefore inserts a nuance that helps courts decide so-called ‘hard cases’. Given the complexity of digital markets described above, it is almost certain that competition cases involving data-driven dominance will be ‘hard cases’ whereby multiple possible causal counterfactual scenarios explaining the observed anti-competitive outcome could be envisioned. When faced with a number of different causal counterfactual scenarios, enforcers might find it difficult to decide which counterfactual is the most convincing, especially if the econometric models underlying the various plausible counterfactual scenarios present differences in terms of their design and data required and used. In such cases, it is important to probe these structural models thoroughly in order to understand the accuracy of the data and the assumptions underlying the models. As Cragg and co-authors point out, ‘[c]ritically, the validity of [a] structural model is determined in the first instance by its ability to reflect accurately what happened in the actual world.’¹⁸⁰ Thus, econometric models need to be consistent with the relevant institutional factors and facts in the setting being modeled and produce ‘results that can be evaluated in the context of other evidence’.¹⁸¹ Of course, thoroughly assessing the validity of all the different structural models and attendant counterfactuals presented in a given case increases the work of the Commission. However, it also—importantly—helps to avoid both false convictions and false acquittals and therefore aids in safeguarding efficiency in dynamic data-driven markets.

This test—composed of multiple counterfactuals, a ranking per Scott-Hemphill and/or a reasonableness criterion for their outcomes’ causal attribution to a given effect—gives more room for manoeuvre both to the enforcer and to the undertakings involved in an antitrust dispute, as it introduces more and more rigorous steps in the establishment of causation. It also introduces a normatively inspired criterion of reasonableness for causal attribution. As observed by some commentators,¹⁸² this subjective criterion can cushion the effect of overzealously applied counterfactuals and can result in giving the ‘benefit of the doubt’ to some defendants. Hence, this approach would also contribute to a more equitable weighing of the interests of the parties to an antitrust dispute—a development much needed in fast-moving, dynamic digital markets.

The counterfactual link and (data-driven) remedies

As argued by Lianos, the link between effects and remedies should be logically coherent, established by a proportionality test.¹⁸³ This means the counterfactual outcome from foreclosure assessment should be linked to necessary, suitable, and proportionate remedies.¹⁸⁴ Turner maintains that remedies, like theories of harm, should be based on economic analysis.¹⁸⁵ We agree but emphasize that, contrary to some views,¹⁸⁶ remedies do not need a separate counterfactual assessment at their determination stage. Instead, the proportionality principle should guide the selection of remedies that can address anti-competitive effects identified through counterfactual analysis at the foreclosure assessment stage.¹⁸⁷

For example, if counterfactual analysis shows that ex-ante data-sharing mandates hinder competition by reducing smaller firms’ innovation incentives, ex-post remedies should avoid imposing additional data-sharing obligations on the dominant firm. This could include temporarily suspending the firm’s ex-ante data-sharing obligations under Article 9 DMA. Rolling back data-related obligations can also occur under Article 12(2)(e) DMA pending a market inquiry.¹⁸⁸ However, the competition authority cannot directly roll back ex-ante remedies; it can recommend this to the ex-ante regulator, who can then act under Article 9 DMA (temporary) or Article 12(2) (e) mandates.¹⁸⁹

Another option would be that counterfactual assessment at the foreclosure stage points to the need for even more stringent data-sharing mandates than the ones imposed ex-ante under the DMA. At that point, from all possible proposed mandates, the one that best fits the proportionality requirement should be chosen. Whatever the concrete case may be, a more general point merits attention: literature suggests that data-sharing mandates might work better as an ex-post remedy than an ex-ante one. As Gal and Petit argue: ‘the role of data as a barrier to entry depends on the attributes and context of each market.’¹⁹⁰ Since the DMA’s ex-ante regulation is not market-assessment-based, ex-post competition enforcement is better suited to address data-related concerns at the remedy stage. Thus, our ultimate conclusion is that data-sharing mandates, if imposed, should be within the ex-post enforcement framework.

5. CONCLUDING REMARKS

In the face of the DMA, which imposes a number of ex-ante obligations, including limited data-sharing mandates on digital gatekeeper platforms, this article emphasized the importance of enforcement adaptation of ex-post competition enforcement. Underlying the article’s focus is the recognition that gaps in the DMA’s data-sharing mandates are likely to trigger ex-post enforcement given that both regimes (DMA and antitrust) can address the issue of data-driven market tipping. The extent to which one enforcement domain impacts the other can, for instance, be seen in competition policy’s neighbour domain of telecommunications, where, as illustrated by the Slovak Telekom case, interaction led to an abridged test for abuse of dominance under Article 102 TFEU.

We argued that such regulatory inconsistency between the two domains needs to be approached with care, not the least because of the risk of error costs that it entails. Digital markets are subject to fast-evolving competitive dynamics, which can make the ideas enshrined in ex-ante requirements superfluous or ineffective by the time a gatekeeper platform is under ex-post scrutiny for novel types of harm. For example, as we described in detail in the ‘Data-driven market tipping: a case for ex-ante mandated data sharing?’ section, emerging policy and business initiatives as well as new technologies and techniques for creating and analysing data are bound to make data-sharing mandates ever less effective at restoring competition in digital markets. We therefore made the case for measured intervention based on the recognition that dissonance between the ex-ante DMA and ex-post competition enforcement might stifle dynamic digital markets.

Specifically, we argued that authorities need to proceed carefully with regard to two aspects of ex-post interventions in dynamic digital markets already governed by ex-ante obligations. First, authorities need to ensure that there is a high threshold for such interventions, requiring rigorous analysis to determine whether companies have abused their dominant position. Second, authorities need to carefully choose appropriate remedial actions if abuse of dominance is found.

With regard to establishing that a company has abused its dominant position, we proposed a multiple counterfactual causation test. We maintained that such a test should be based on careful examination of anticompetitive foreclosure, taking into account all of the factors mentioned in paragraph 20 of the European Commission’s Guidance Paper, and ranking the viability of each counterfactual through reasonableness analysis as understood in the neighboring domain of tort law. We also suggested alternatives for ranking based on US antitrust thinking, recommending the ‘least restrictive alternative’ (LRA) test in particular.

Finally, with regard to selecting appropriate remedial actions when abuse of dominance has been established, we maintained that it is crucial for authorities to determine the actual effects of the infringement at the foreclosure stage—an exercise for which counterfactual analysis is particularly well-suited—in order to then come up with working remedies through proportionality analysis. Doing so ensures that authorities choose remedies that are effective at actually reversing the anticompetitive harm caused by the dominant company’s abusive behaviour. Through the steps proposed above, ex-post competition enforcement becomes sufficiently flexible to respond to narrow regulatory obligations under the DMA, while preserving its current logic of effects-based assessment.

Footnotes

Acknowledgements

The authors would like to thank the researchers and staff of the International Relations Chair (IR Chair) of the Munich School of Politics and Public Policy (Hochschule für Politik München) for the various comments they provided on this project. A note of thanks is also due to the Economic Public Law (EPL) Department of Utrecht University for the support and intellectual exchange on the paper, as well as to the RegulAite project team within the Political Science Department of the University of Amsterdam (UvA). We are also indebted to Helena Malikova (European Commission) and Joachim Henkel (TUM) for generously offering to comment on this work on several occasions. The authors have nothing to disclose; all errors remain their own.