A. Untangling the relationship between the algorithm and the findings

Fastgreedy,Footnote ¹⁷ Louvain,Footnote ¹⁸ Label propagation,Footnote ¹⁹ Walktrap,Footnote ²⁰ and Infomap, Footnote ²¹ all widely used algorithms, define the concept of community differently, implementing complementary strategies to map and organize the network. Recall that a network is a structure of nodes (judgements) and edges (citations between judgements) – with citing judgements pointing to the cited judgements. Networks are represented with graphs; in the network of the Court’s case-law, the graph is directed.

Two algorithms – Fastgreedy and Louvain – define communities as tightly connected groups of nodes (judgements) that are loosely connected with the rest of the network. Practically, they typically find bigger communities. To illustrate, Figure 1 below represents the most relevant communities found by Fastgreedy. The horizontal view shows the relative growth of a community over time (darker shades of the rows imply faster growth). The vertical view illustrates the structure of the case-law at any given time (the horizontal axis marks the year). Darker shading indicates a larger share of the community in the network.

Figure 1.

Communities of case-law (clusters). Fastgreedy identifies 19 cohesive communities (clusters); of those four are large (rows 1, 4, 8, and 5).

By contrast, Walktrap defines a community as a group of nodes (judgements) in which a random walk, a process of visiting the judgements by moving randomly through citations, is likely to be trapped. Infomap employs the concept of an infinite random walk to approximate the information flow in the network. It defines a community as a set of nodes in which such an infinite random walk lingers longer. Finally, Label Propagation does not a priori define the concept of community but applies an iterative process to generate labels in the network. Initially, each node has its own label; subsequently, each node updates its label according to the most frequent label amongst its neighbouring nodes, until no further update is possible. The same labels cluster into one community.

The six squares in Figure 2 present a large network organized in smaller groups of judgements based on citations, regardless of citation content and thus the theme of individual communities. The first impression clockwise is that some algorithms find more groups than others (comparing the top and the bottom square in the middle of the two rows in Figure 2), and that the clusters found by each algorithm vary in size.

Figure 2.

Visualizes the effect of community detection on the Court’s citation network. The initial unordered adjacency matrix does not show any recognisable pattern. By applying community detection algorithms and discovering communities in data, the patterns become evident, and each algorithm shows some specific aspect of the data. Fastgreedy divides the networks into four big communities that account for ∼10000 nodes (∼85 per cent of the network) and a bunch of smaller ones. Louvain finds 18 communities plus some other small ones. Louvain mainly breaks down the four biggest communities from Fastgreedy into smaller ones. Label propagation finds a very large community which clusters together the first and the third largest communities found by Fastgreedy, another community which corresponds to the second largest community from Fastgreedy and a third smaller community that is a subset of the fourth community found by Fastgreedy. Walktrap highlights a core-periphery organisation of the network with one large community serving as a core and other 12 smaller communities followed by 550+ micro-communities. Infomap clusters the network in a large number (700+) of small communities. For all algorithms, it is clear how the detected clusters have higher internal density of connections than they have externally. Furthermore, the smaller a community, the higher its internal density.

The number and the size of the communities vary depending on the algorithm. The below comparison illustrates a hierarchical relation between the communities: Algorithm A groups all judgements into one community and algorithm B divides the judgements in two sub-communities. Both produce similar findings, on a different level of detail/generality. For example, work-related benefits judgements including unemployment, sick leave or family benefits always remain in the same sub-group but group into a larger legally coherent community of free movement of persons with the judgements dealing with the entry and residence of family members, like cases on residence permits or reunification. The community labelled free movement of persons, which contains both sets of judgements, is hierarchically superior in network science terms.

Figure 2 shows the findings of Fastgreedy, Louvain, Label Propagation, Walktrap, and Infomap algorithms, employed with adjacency matrices. An adjacency matrix is a square used to represent a graph or a network, indicating whether pairs of nodes, meaning the points on the graph, or judgements are close together, that is, adjacent. A black dot is a connection from a node i (row) to a node j (column) – from judgement i to judgement j. Figure 2 first visualises the Court’s citation network with its unordered adjacency matrix (first square to the left in the first row). Next, the rows and columns of the adjacency matrix are reordered according to the communities found by each algorithm and in decreasing order of community size/granularity. Reordering highlights the patterns in the network, distinguishing between densely connected areas on the main diagonal that are sparsely connected with the remaining network.

Figure 2 represents clear differences in community structures. A closer analysis reveals that Fastgreedy finds 103 communities, four of which are rather large and account for 85 per cent of the total nodes in the network (Figure 1 represents only 19 largest communities). Louvain detects 74 communities and exhibits the broadest distribution of community size amongst the algorithms. Its largest 18 communities cover 98 per cent of the judgements in the network. Label propagation detects 127 communities, of which the largest three account for 86 per cent of the judgements. Walktrap detects 589 communities, from which it is still possible to recognise a core-periphery structure of the network with the largest community as a dense internally connected network core and the periphery of smaller communities interacting with the core. Finally, Infomap detects 713 small and dense communities (the largest one contains 277 nodes as opposed to over 3000 nodes detected by Fastgreedy as presented in Figure 1, second adjacency matrix in the middle of the top row).

The adjacency matrices imply that algorithms could produce vastly dissimilar configurations, owing to the choice of the algorithm. The measures of similarity and hierarchical relationship, developed in network science, address this issue. They are discussed in the following Section.

B. Similarity as hierarchy

A legal analysis of the community structures (the reading of judgements in the same community) reveals whether the communities detected by the algorithms have similar content and by which legally relevant parameters, for example, a policy field or legal basis. The network science, by contrast, measures similarity as a hierarchical relation between the communities, meaning the extent to which different algorithms partition the network at different levels of granularity (thus creating communities of different sizes) and the hierarchical structure between them. This means that one algorithm groups together several communities discovered by another.

Algorithm performance is measured in terms of approximating the so-called ground truth. The appraisal of similarity and hierarchy implies a preliminary measure of ground truth – a labelling of elements that describes the objective or true assignment of elements to communities.Footnote ²² Labels offer verifiable information about each judgement, serving as a reference point. While the subject matter, which the Court assigns to each judgement, requires further reckoning before the findings of community detection can be used for legal analysis (a question addressed in Section 4), it is a solid basis for the assessment of similarity and hierarchy in network science terms.

The available measures can appraise the similarity between two cluster structures, or the overlap between a cluster structure and ground truth. Each has its own assumptions, reference models, and biases;Footnote ²³ none is superior.Footnote ²⁴ The first measures of association used in this analysis combine three commonly used complementary association measures in data mining: adjusted rand index (ARI),Footnote ²⁵ adjusted mutual information (AMI),Footnote ²⁶ and Fowlkes-Mallows index (FM).Footnote ²⁷ The second measure combines informedness (BI), markedness (MK) and their geometric mean, the Matthew’s correlation coefficient (MCC).Footnote ²⁸

Regarding the association measures, the ARI penalises algorithms wrongly separating a true cluster in too many clusters,Footnote ²⁹ while AMI is biased towards cluster structures with more clustersFootnote ³⁰ and FM is biased towards skewed cluster sizes.Footnote ³¹ The measures yield a score of 0 if two variables are independent and 1 if they are the same.

When comparing clustering with ground truth or the real structure of the network, informedness compares algorithm’s findings against a random guess. If the guess is random, informedness is zero. Markedness indicates how consistently the algorithm clusters a pair of elements together (or apart). A random guess yields a zero score. Informedness and markedness are the two components of the Matthew’s correlation coefficient. To illustrate, informedness of 0.5 means that the algorithm captures 50 per cent of ground truth, while markedness of 0.5 means that the algorithm correctly clusters every second pair of elements.

When comparing communities found by two algorithms (indicated with A and B) with informedness and markedness, informedness (A, B) measures the portion of A correctly captured by B. Markedness (A, B) measures the portion of B captured by A. The correlation coefficient MCC (A, B) is a symmetrical measure of similarity between A and B. It is possible to have instances of high informedness and low markedness. This means that the algorithm captures a large portion of ground truth (high informedness) but wrongly clusters together pairs that belong to different clusters or vice versa (low markedness). High informedness and low markedness usually indicate that the algorithm discovers overly big communities, clustering together dissimilar pairs. For example, Label Propagation clusters together cases on harmonization and liberalization. Conversely, low informedness and high markedness signals that the algorithm fails to account for a large share of ground truth but is very reliable for the small part of ground truth that it captures. This signals that the algorithm identifies overly small communities, separating pairs that should be clustered together; for example, the algorithm groups cases covering different aspects of social policy into different communities or clusters. Again, these measures take the perspective of the network science; the legal analysis, interpreting the findings, might arrive at different conclusions about the accuracy of a given algorithm. Whatever the algorithm chosen, the legal analysis must disentangle the connections and processes behind the clustering. Network clustering and the legal analysis are interdependent. The researcher selects the algorithms but does not determine the results.

Table 1 compares the communities discovered by five algorithms. Some differences are apparent. For instance, according to some measures (the AMI), the most similar communities are those found by Louvain (LM) and Walktrap (WT) (the score 0.53 in Table 1.A.). According to others (ARI, FM, and MCC), the most similar communities are those found by Fastgreedy (FG) and Label propagation (LP) (ARI score 0.35 in Table 1.B; FM score 0.54 in Table 1.C., and MCC score 0.39 in Table 1.D.). Overall, the co-occurrence matrix displayed in Figure 2 clearly shows that the communities discovered by the five algorithms are nested within each other.

Table 1.

Table 1 compares the communities discovered by the five different algorithms: Fastgreedy (FG), Louvain (LM), Label propagation (LP), Walktrap (WT), Infomap (IM). A) Adjusted Mutual Information, B) Adjusted Rand Index, C) Fowlkes-Mallows score, D) Matthews Correlation Coefficient, E) Informedness, F) Markedness.

Figure 3 shows the hierarchical organization between the communities. Concretely, the darker colour in the left square of Figure 3 indicates that fewer algorithms clustered a pair of nodes together. Black areas indicate that no algorithm clustered these nodes together. The central square shows that the largest community found by Label propagation (6629 judgements) joins two large communities discovered by Fastgreedy, which further contain communities discovered by Louvain, Walktrap and Infomap. This means that the communities are hierarchically structured. The first square from the right shows that the communities discovered by Louvain are nested into the communities discovered by Fastgreedy.

Figure 3.

Shows the hierarchical organization of the Court’s citation network. The co-occurrence matrix shows how the clusters found by the algorithms are nested into each other (left square). Hierarchy-ALL shows the adjacency matrix of the Court’s case-law reordered according to the hierarchical order suggested by all the algorithms (middle square). Hierarchy F-L, shows the adjacency matrix reordered according to the hierarchical order suggested by Fastgreedy and Louvain (right square).

Table 2 illustrates the relationship between the communities found by Fastgreedy and Louvain in a table format, looking at the labels assigned to the judgement. As stated above, it shows that the communities discovered by Louvain are nested into the communities discovered by Fastgreedy measured as the relationship and overlap between the labels.

Table 2

The frequent subject matter of judgements in selected significant communities found by Fastgreedy (left column) and Louvain (right column). Some labels are identical, meaning that the communities overlap entirely. For instance, Fastgreedy community C15 includes the label ‘common customs tariff’ which corresponds to an identical Louvain community C23. Most communities, however, discovered by Fastgreedy and Louvain are similar with respect to the labels they include, such as the community dealing with taxation (third row).

Concretely, Fastgreedy groups judgements labelled as approximation of laws in diverse policy areas in one community. Louvain separates these judgements into six communities, all dealing with harmonization: (1) environmental policy, pollution, and waste, (2) social policy and consumer protection, (3) asylum immigration, and cooperation in criminal matters, (4) judicial cooperation in civil matter, (5) intellectual, industrial and commercial property, and (6) freedom to provide services and freedom of establishment.

A. Communities of case-law

The description of each individual community includes a Figure, demonstrating the composition of the community by subject matter – the primary point of reference in Section 2 and the starting point of the legal analysis. The horizontal view shows the relative growth of each community over time with darker shades implying faster growth. The vertical axis shows the content. Some communities are more heterogeneous (contain judgements with many different labels and several rows). Darker shading indicates a larger share of the subject matter/label at a given time (the horizontal axis displays the year).

Community 0/European Coal and Steel Community (ECSC)

The ECSC community covers 66 years of European integration (1954–2020), nicely illustrating the potential of community detection to reveal institutional change. The judgements in the community are labelled by the Court as ECSC and approximation of laws. Until 1992, the community contained only judgements labelled ‘ECSC’, while post 1992, the label ‘approximation of laws’ became dominant. This shift extends to the type of proceedings: Annulment actions dominate until 1992 but are replaced by preliminary references then. Upon the expiration of the Merger Treaty in 2002, the community practically disappeared.Footnote ³³

Figure 4.

Displays the most frequent labels in the community ‘ECSC’.

The turning point is the judgement in Pierrel,Footnote ³⁴ where the Court declared that the suspension or revocation of a market authorization could rely exclusively on the grounds laid down in Community legislation.

The judgements issued after 1990 mostly deal with market authorizations: for foodstuffs, or most commonly for medicinal products, frequently questioning the possibility of parallel imports. They raise questions of the application of secondary law, usually in relationship with the Treaty. A good example is the judgement in Paranova,Footnote ³⁵ addressing whether the withdrawal of market authorisation for medicinal products complies with the free movement of goods (Articles 28 and 30 EEC) when it simultaneously withdraws the authorisation for parallel imports. The Court held that such withdrawal was compatible with the Treaty only if the Member State could justify it with the need to safeguard public health.

Community 5/Taxation

Community ‘Taxation’ is one of the larger communities with 12 per cent of all judgements in the network. The most common subject matter is ‘taxation’, followed by ‘external relations’, ‘free movement of goods’ and ‘agriculture and fisheries’. Despite the varied subject matters the judgements are closely linked to tax: The cases under ‘taxation’ are generally VAT cases, whereas the cases under the subject matters ‘free movement of goods’ and ‘agriculture and fisheries’ relate to fiscal charges and customs. ‘External relations’ cases refer to dumping and customs.

Figure 5.

Displays the most frequent labels in the community ‘taxation’.

The community highlights the changes in litigation in response to legislative changes.Footnote ³⁶ The community has expanded since the eighties, coinciding with the harmonization of the VAT in 1977.Footnote ³⁷ The Court facilitated/enhanced the enforcement of the new rules by allowing individuals to invoke the unimplemented VAT Directive,Footnote ³⁸ and declaring that national courts could apply European Union law ex officio, irrespective of and contrary to national procedural law.Footnote ³⁹ The fastest growing period in the community spans from 2007 to 2019, following the recasting of the Sixth VAT Directive in 2006.Footnote ⁴⁰ The development of the community can only be understood against the enforcement effort of the Commission, in particular inerasing cross-border hurdles to trade which could impair the internal market.Footnote ⁴¹ Around half of the infringement proceedings judgements in the community fall in this period of growth.Footnote ⁴² Furthermore, the incomplete harmonization of VAT has kept the litigation high in the area via preliminary references.Footnote ⁴³

Community 15/Customs Union

Community 15 comprises 2.28 per cent of cases. This relatively small community includes only four labels. ‘Free movement of goods/customs tariff’ is the main label. Most cases refer to the classification of products under the Common Customs Tariff Code (CCT). For instance, in the first case in the community, the Court declared that Member States could not issue binding interpretations relating to the headings of the CCT.Footnote ⁴⁴ The second most common group of cases in the community is labelled ‘agriculture and fisheries’ with judgements referring to the classification of agricultural products and foodstuffs under the CCT. The labels, ‘external relations’ and ‘approximation of laws’, appear only in the late eighties, remaining sparse. The judgements with those subject matters also concern the classification of agricultural products.

Figure 6.

Displays the most relevant labels in community ‘Customs Union’.

The judgements were delivered between 1970 and 2020. The fastest growing periods in the community arose late in the integration process.Footnote ⁴⁵ Figure 6 highlights three such periods. The first and largest occurred between 1993 and 1997. The community also expanded in the period of 2006-2009 and, to a lesser extent, between 2013 and 2015. The three periods of expansion relate to the years following the three most relevant amendments to the Common Customs Tariff Regulation.Footnote ⁴⁶

Community 23/Social policy

Community 23 counts 61 judgements labelled social policy, representing only 0.5 per cent of the cases in the network. All were preliminary references about very concrete issues, like the calculation of pensions, subsidies and benefits, and the rules in case of overlapping.Footnote ⁴⁷ The first case in the community is Keller, where the Court declared that when the insured periods completed under the legislation of one Member State were sufficient to derive a right to a pension, that Member State did not need to consider other periods completed under the legislation of another Member State.Footnote ⁴⁸ The most recent judgement is Blanco Marqués, which discusses whether Spain could detract a supplement from the plaintiff who also received a pension from Switzerland (it could not).Footnote ⁴⁹

Figure 7.

Displays the most frequent labels in the community ‘social policy’.

Figure 7 indicates punctuated growth with periods of fast expansion (late 70s to the early 90s) and decline (since 2000).

Community 24/Free movement of agricultural products

Community ‘Free movement of agricultural products’ illustrates the potential of community detection to identify growth and stagnation. The community stalled at the turn of the century, reflecting broader changes. Three main subject areas in the community are ‘agriculture and fisheries’, ‘external relations’ and ‘free movement of goods’, all related to agricultural products. Judgements labelled ‘external relations’ deal primarily with agricultural products,Footnote ⁵⁰ and those with the label ‘free movement of goods’ with Article 34 TFEU and free movement of agricultural products.Footnote ⁵¹ This is intuitively sound, given the proximity of free movement of goods to agriculture and fisheries in the early years of the European Communities.Footnote ⁵²

Figure 8.

Displays the most frequent labels in the community ‘free movement of agricultural products’.

The changes in the community structure reflect the policy shifts in the internal market. The community grew fastest from the late seventies to the early nineties, coinciding with the implementation of the ambitious Commission’s Single Market program,Footnote ⁵³ and legislation.Footnote ⁵⁴ Many legal instruments were adopted in the areas of agriculture and fisheries and the free movement of goods, as regulatory adjustments were required for the correct functioning of the incipient common market for agricultural products.Footnote ⁵⁵ From 2000, no judgements are labelled as ‘external relations’ and ‘free movement of goods’ in the community, which is consistent with the general disappearance of cases on goods from the docket of the Court.Footnote ⁵⁶ Judgements with the label ‘agriculture and fisheries’ become sparse, reflecting the sparsity of new legislation, due to the difficulties in passing legislative reforms in this area, where the interests of the Members States diverge greatly.Footnote ⁵⁷

Community 30/The regulation affecting intellectual property

Community ‘The regulation affecting intellectual property’ extends from 1971 to 2020. It is a good illustration of the relationship between the meaning of citations and labels (subject matter). The judgements in the community carry a priori unrelated labels like ‘intellectual, industrial and commercial policy’, ‘competition’, ‘free movement of goods’ and ‘agriculture and fisheries.’ That said, all cases touch on intellectual property matters from slightly different angles. The ‘competition’ judgements focus on the role of patents in the competition law provisions. The ‘free movement of goods’ judgements address general matters of intellectual property, like parallel imports of pharmaceutical products, where the patent is key. Finally, (fewer) ‘agriculture and fisheries’ judgements refer to market authorizations, keeping a looser link to intellectual property. In the mid-nineties, the ‘approximation of laws’ judgements appear in the community. Those cases refer to intellectual property but in regulated areas (secondary legislation).Footnote ⁵⁸

Figure 9.

Displays the most frequent labels in the community ‘the regulation affecting intellectual property’.

Community 32/Politically significant policies

Community labelled ‘Politically significant policies’ illustrates the changing density of the communities, reflecting the strength of the connections between the judgements and allowing for the examination of special relationships. The community forms in 1973 but becomes denser around 2000, when litigation becomes more frequent. Today, it is one of the densest and fastest growing communities.

Figure 10.

Displays the most frequent labels in the community ‘politically significant policies’.

The composition of the community is varied and eclectic. Many judgements, particularly from 2000, concern the area of freedom, security, and justice (AFSJ), and equally many cases deal with electronic communications and data protection. The judgements do not share the subject matter but are bound by increased regulation of more politically significant areas without a direct and close relationship to the regulation of the four freedoms, notably asylum, criminal and police cooperation and data protection.Footnote ⁵⁹ Moreover, the structure of the community reflects broader institutional changes: The Treaty of Lisbon abandoned the pillar structure,Footnote ⁶⁰ which included the AFSJ among the shared competence under the jurisdiction of the Court,Footnote ⁶¹ and increased regulation.Footnote ⁶²

Community 63/Legal basis/Article 115 TFEUFootnote ⁶³

The ‘Legal basis’ community includes judgements labelled mostly as ‘social policy’, ‘approximation of laws’, ‘freedom to provide services’ and ‘freedom of establishment’. All ‘social policy’ judgements deal with the protection of workers in the event of transfers of undertakings.Footnote ⁶⁴ ‘Approximation of laws’ judgements are mostly related to civil liability linked to the use of vehicles,Footnote ⁶⁵ and fair price comparisons.Footnote ⁶⁶ Judgements labelled ‘freedom of services’ and ‘establishment’ relate to commercial agents,Footnote ⁶⁷ so they share a connection to companies and frequently groups of companies.Footnote ⁶⁸ A closer analysis reveals that all Directives were adopted on the same legal basis (Article 115 TFEU, ex Article 100 TEC). Moreover, legislative changes account for the fast growth of the community (1990, 1995 and 2017–2019), with litigation dealing with those legislative instruments. At times, legislation did not change, but the Court expanded its previous case-law. This has been the case for the Acquired Rights Directive.Footnote ⁶⁹ Dating back to the golden period of social policy, the Directive was reformed in 2001, but the label expanded in recent years, as the Court built on its case-law to apply the principles developed on its transfer of undertakings’ jurisprudence to ‘new commercial situations and matrices (albeit without pushing the boundaries too far)’.Footnote ⁷⁰

Figure 11.

Displays the most frequent labels in the community ‘Legal basis/Article 115 TFEU’.

B. Summary

As illustrated in Sections 2 and 3, network analysis and legal analysis mutually inform and cross-validate each other in three ways. First, community detection unravels relations among judgements, which more routine methods of identifying legal communities of cases would miss or underplay. Thus, the findings force a closer investigation of overlooked connections and processes that might prove significant. For instance, Community 30 grouped judgements that at first appeared dissimilar, as the labels were seemingly disconnected. However, a closer look articulated the connection, as all cases referred to different aspects of intellectual property. Sometimes the connections would have been even harder to grasp using other methods, as shown by Communities 63 and 32. For the former, the cases are connected by the same legal basis, for the latter, diverse and eclectic subject matters come together because the judgements relate to new secondary legislation in areas with no apparent connection to the market. However, only legally informed interpretation of the findings can support convincing conclusions. Second, community detection shows which institutional, structural, political, and legal developments are reflected in litigation, but also that litigation is not only a carnival mirror of social reality. Third, community detection supports the investigation of micro- (where the network is partitioned in smaller communities) and macro-processes (where the network is partitioned in fewer bigger communities), adding knowledge to the ongoing debates about the state, the priorities, the direction, and the appeal of integration.

Institutional change primarily explains the structural shift of some communities. In this sense, Community 0/ECSC (Figure 4) evolves from ECSC to the approximation of laws around 1992, when the latter subject matter becomes dominant,Footnote ⁷¹ disappearing after 2002 with the expiration of the Merger Treaty.Footnote ⁷²

Legislative changes, frequently followed by the enforcement effort of the Commission, contextualize the shifts in litigation patterns in Community 5/taxation (Figure 5). The latter took off in the eighties, following the harmonization of the VAT in 1977.Footnote ⁷³ The Court facilitated the enforcement of the new rules by allowing individuals to invoke the unimplemented VAT Directive,Footnote ⁷⁴ and declaring that national courts could apply European law ex officio, irrespective of and contrary to national procedural law.Footnote ⁷⁵ This is reflected in Figure 5 (row four from the bottom), which shows an expansion of the label ‘taxation’ during those years. The fastest growing period between 2007 and 2019 follows the recasting of the Sixth VAT Directive in 2006,Footnote ⁷⁶ and the Commission’s subsequent enforcement effort. Half of the infringement proceedings in the community, approximately 100 cases, corresponds to this period of growth, which again translates into an expansion of the label ‘taxation’ in the community (Figure 5, row four). This coincides with doctrinal analysis linking the ‘incomplete’ harmonization process in VAT with a never-ending litigation.Footnote ⁷⁷ Similarly, the periods of faster growth in Community 15/Customs Union follow the three most relevant amendments to the Common Customs Tariffs Regulation:Footnote ⁷⁸ 1993–1997, 2006–2009 and, to a lesser extent, 2013-2015.

By contrast, litigation dwindles where legislation and case-law settle. This translates into communities that grow slowly or altogether stop growing. For instance, Community 23/social policy (Figure 7) grew particularly fast from the late seventies to the early nineties, and stagnated afterwards. This is unsurprising when we consider the cases in the community, which share the label ‘social security’ and relate to very concrete aspects of the calculation of pensions, subsidies and benefits, and the rules in case of overlapping.Footnote ⁷⁹ The legal questions dominating the early years of integration are largely solved,Footnote ⁸⁰ making litigation redundant.

Conversely, litigation intensified after 2004 in the areas where the Union has traditionally had little to no competence to regulate but rather to support and coordinate the policies of the Member States. Community 5/taxation reflects this. As shown in Figure 5, in the last two decades the community has incorporated new labels, like trans-European networks (row eight from the top) or data protection (row six from the top), where the Union has taken a more active role only recently, often crystalized in new legislation.Footnote ⁸¹ Community 32/politically significant policies similarly shows that in the last two decades, the judgements with the labels ‘AFSJ’ (Figure 10, first row from the top), ‘data protection’ (third row) and ‘telecommunications’ (fourth row) increased.Footnote ⁸²

A. A progressively harmonized market

Community 4 developed in the mid-1970s and expanded in the mid-2000s. Its slow start is unsurprising given the unsuccessful early unification attempts, which would have required a political will inexistent at the time.Footnote ⁸⁴

In the early 1980s, the Commission took several Member States to Luxembourg for not living up to their commitments in the environmental domain, notably pollution.Footnote ⁸⁵ It filed an action against the Netherlands for not transposing the Council Directive on the quality of bathing water,Footnote ⁸⁶ and against BelgiumFootnote ⁸⁷ for the non-implementation of a directive on waste from the titanium dioxide industry, and the Directive on waste.Footnote ⁸⁸ This is reflected in the expansion of the label ‘environment’ in Community 4 (Figure 12, row 19 from the bottom). Around the same time, the Commission launched an ambitious internal market program, aiming to complete the internal market and remove all physical, technical and fiscal barriers.Footnote ⁸⁹ Several legal instruments were enacted as a result, especially oriented to ensuring the free circulation of safe products across the Union.Footnote ⁹⁰ Concerned with non-compliance (either for lack of transposition or deficient implementation), it also pursued an aggressive enforcement strategy,Footnote ⁹¹ which is visible in Community 4 in the expansion of the label ‘free movement of goods’ between the mid-eighties and the mid-nineties (Figure 12, row 4 from the bottom).

Figure 12.

Displays the most frequent labels in the community ‘harmonized market’.

Social policy was supposed to correct the market. Its growth, particularly from the mid-eighties, corresponds to the enactment of key legislative instruments,Footnote ⁹² of which three especially stand out. First, the Equal Treatment DirectiveFootnote ⁹³ gave rise to intense litigation during the eighties,Footnote ⁹⁴ leading to milestone judgements like Johnston, Von Colson, Kamann, and Marshall,Footnote ⁹⁵ which forge the central principles of European Union law: judicial review, conform interpretation/indirect effect. Those judgements shared the label ‘social policy’, which as shown in Figure 12 expanded in the community during the eighties and nineties (row 11 from the bottom).Footnote ⁹⁶ Second, the first Working Time Directive, enacted in 1993,Footnote ⁹⁷ spurred a wave of litigation including an annulment proceeding brought by the United Kingdom.Footnote ⁹⁸ The Court’s judgements substantially expanded on aspects that the Directive did not address. For instance, the Court ruled that on-call duty was to be considered working time.Footnote ⁹⁹ The Member States did not welcome the expansion, and a proposal for amending the legislation, essentially to contain the interpretation of the Court, was pursuedFootnote ¹⁰⁰ but ultimately rejected. Litigation continued, including several infringement actions brought by the Commission for absent or deficient implementation.Footnote ¹⁰¹ The Court confirmed its previous interpretations in new rulings, overwhelmingly deciding in favour of employees and against national or employers’ practices,Footnote ¹⁰² thus probably prompting more litigation. Finally, the Framework Directive on Fixed Term Work in 1999 led to many preliminary questions, mainly from Spain and Italy.Footnote ¹⁰³ The Court has also taken a rather protective stance, at least in the less controversial aspects of the Directive,Footnote ¹⁰⁴ without fully assimilating the Directive in other anti-discrimination instruments.Footnote ¹⁰⁵ In other words, all instruments spurred intensive litigation in the area of social policy, as captured by Fastgreedy in Community 4. As shown in Figure 12 (row 11), the label ‘social policy’ becomes more relevant in the community in the years following the enactment of those instruments: the mid-eighties, after the coming into force of the Equal Treatment Directive; the nineties, following the first Working Time and the Framework Directives, and again in the mid-2000s, coinciding with key developments in the Court’s case-law on the Working Time DirectiveFootnote ¹⁰⁶ and the recast of the Equal Treatment Directive.Footnote ¹⁰⁷

Taxation became vibrant in the eighties with the harmonization of VAT in 1977,Footnote ¹⁰⁸ and has continued to grow,Footnote ¹⁰⁹ as reflected in the expansion of the label ‘taxation’ in Community 4 (Figure 12, row eight from the bottom). In Becker, the Court declared that an individual could rely on the unimplemented VAT Directive.Footnote ¹¹⁰ As shown in Figure 12, this coincided with the first expansion of the label ‘taxation’. A decade later, the Court held that national courts could apply a provision of European Union law ex officio even when national procedural law excluded this possibility.Footnote ¹¹¹ This is again reflected in Figure 12, with another growth of the label ‘taxation’ around those years.

Consumer protection judgements surged in the nineties after the adoption of the Maastricht Treaty, which established the Union’s shared competence in that policy field,Footnote ¹¹² and liberated consumer protection from the constraints of linking consumer rights to the protection of the internal market.Footnote ¹¹³ Many preliminary references touched on the Unfair Terms Directive,Footnote ¹¹⁴ particularly mortgages and the banking sector.Footnote ¹¹⁵ Figure 12 shows (row 23 from the bottom) that the label first appears in Community 4 in the mid-nineties, but drastically expands after the economic crisis in the 2000s, where litigation intensified and led to many judgements regarding mortgages and with the label ‘consumer protection’.Footnote ¹¹⁶ As put by Reich and Micklitz, the economic crisis in the 2010s revived the Unfair Contract Terms Directive.Footnote ¹¹⁷

Environmental cases date to the eighties (in line with the strengthening of the Commission’s take on infringement),Footnote ¹¹⁸ with litigation growing exponentially after 2000,Footnote ¹¹⁹ as reflected in Figure 12 (row 19, label environment). The timing corresponds to the combined effect of the Commission’s legislative and enforcement strategies. The first Barroso Commission prioritised the energy and climate package, featuring a handful of directives;Footnote ¹²⁰ the Commission has traditionally taken a strict stance towards infringements,Footnote ¹²¹ at least until being able to rely on private parties and decentralized enforcement.Footnote ¹²²

The newly minted area of justice and home affairs took off after 2005, particularly after its full inclusion in the Treaty of Lisbon. In 2005, the Court held in Pupino Footnote ¹²³ that national criminal courts could apply a Council Framework Decision in the context of police and judicial cooperation in criminal matters. As shown in Figure 12 (row 12 from the top), the label first enters the community in the mid-2000s, and has grown steadily ever since.

The most recent judgements in the community increasingly deal with more contested matters like asylum, immigration, services of general interest, protection of fundamental rights, and taxation (VAT).Footnote ¹²⁴ Those are also increasingly subject to approximation and high on the list of legislative priorities,Footnote ¹²⁵ although political preferences of the Commission might change following political debate and contestation.Footnote ¹²⁶

B. European priorities, legislative change, and institutional action

Figure 1 shows a faster growth of Communities 4, 5, and 32 and a slower growth of Communities 8, 23, and 24. The shifts in litigation patterns in the communities respond to a large extent to institutional and legislative changes, notably the expansion of European Union’s competences to regulate novel policy areas prone to politicization and contestation, as well as the Commission’s enforcement priorities. However, the most interesting development pertains to the multiplication and the spread of the label ‘approximation of laws’.

Around 15 per cent (1985) of judgements are labelled the approximation of laws. Most concern services (866), environment (282), consumer protection (276), goods (265), agriculture (167), taxation (154), social policy (55), and various aspects of intellectual property like patents or trademarks (478).

The label ‘approximation of laws’ is notably present in ten of twelve largest communities; most judgements, 83 per cent, are spread among the four largest communities, and 41 per cent are in a single large community – thus labelled ‘the progressively harmonized market community’ (Figure 12).

Apart from the proliferation of secondary legislation, the label ‘approximation of laws’ highlights a more frequent interaction between the communities. Regulation triggers disputes that fall between policy areas. European citizenship is a classic example of the spillover from the free movement of workers and social security to the area of political and family rights.Footnote ¹²⁷

Individual communities have become more heterogeneous, as reflected in the diversification of the subject matter. In Community 4/harmonized market, more classic European Union law fields like freedom to provide services coexist with newer labels like consumer protection or AFSJ. In Community 32/politically significant policies (Figure 10), the expansion of regulation is indicated by the emergence of new labels. The label ‘consumer protection’ first appears in the community in the mid-nineties but grows after 2010. The increase coincides with stark litigation about mortgage contracts following the economic crisis,Footnote ¹²⁸ mostly referred to the Unfair Terms Directive. Interestingly, the Directive gave rise to very little litigation in its first years, but became central with the surge of over-indebted consumers (mostly property owners) after the economic crisis, particularly in some Member States.Footnote ¹²⁹ The Court responded in a ‘perhaps even activist way’,Footnote ¹³⁰ encouraging more references from national courts, though the interest of the Court in these cases seems to have decreased recently.Footnote ¹³¹ Similarly, the label ‘AFSJ’ appears after 2010, following the Treaty of Lisbon, the inclusion in the shared competence, and the growing litigation in the area of asylum and migration.Footnote ¹³² The multiplication of the label ‘environment’, present since the late eighties, expands in the mid-2000s due to 1) institutional change (the Lisbon Strategy on ‘sustainable development’), 2) legislative process (the push of the Barroso Commission) and 3) the Commission’s role in enforcement/the enforcement strategy.Footnote ¹³³ Notwithstanding numerous pushback and crises, the environmental acquis has proved resilient.Footnote ¹³⁴

The example also illustrates that the changes in litigation frequently respond to the Commission’s priorities and enforcement strategies.Footnote ¹³⁵ Environmental protection case-law after 1980 is the result of numerous infringement proceedings (Figure 12, label environment),Footnote ¹³⁶ and its decline of the reprioritization of environment with energy and climate package by the first Barroso Commission.Footnote ¹³⁷ The label persists due to the decentralized enforcement via preliminary references through national courts (Figure 12, row 19 from the bottom).Footnote ¹³⁸ Similarly, a shift in the Commission’s enforcement policy in consumer protection can be observed as a decline of infringement actions and the rise of preliminary references on the Unfair Terms Directive,Footnote ¹³⁹ particularly with respect to mortgages and the banking sector.Footnote ¹⁴⁰ The latter explains the persistence of litigation since 2000, and the prevalence of consumer protection within community 4 (Figure 12).

Yet, decentralized litigation via preliminary references and the centralized push of infringement action alone cannot explain the expansion and development of the communities. Both piggybacked on the Court’s commitment to the strengthening of the European Union legal order. Generally, the expansion of the harmonized market community in the eighties is parallel to the Commission’s launch of an ambitious internal market program aiming to erase all physical, technical, and fiscal barriers to trade (leading to SEA).Footnote ¹⁴¹

The Court’s case-law has historically opened an alternative channel for legal gap-filling, legislative amendment, and novel legislative agendas. As aptly put by Scharpf, the Court has ‘strategic value as an instrument of European legislation’.Footnote ¹⁴² In its long shadow, the litigants seized the opportunity to initiate, accelerate, or terminate the process of policy making,Footnote ¹⁴³ with litigation driving integration.Footnote ¹⁴⁴ The communities largely reflect this, for instance, the expansion of consumer protection in Community 4.Footnote ¹⁴⁵ Similarly, the expansion of the label ‘AFSJ’ in Communities 32 (Figure 10, first row from the top) and 4 (Figure 12, twelfth row from the top) reflects the mobilization of legal actors across different Member States, Footnote ¹⁴⁶ not always with the explicit opposition of national governments.Footnote ¹⁴⁷ The Court interpreted rights generously (which explains the expansion of Communities 4 and 32),Footnote ¹⁴⁸ particularly in the context of arrest warrants and expulsion.Footnote ¹⁴⁹ Strikingly, the emphasis on fundamental rights at times allows for substantial differentiation among Member States.Footnote ¹⁵⁰

Finally, litigation remains high in areas heavily reliant on judge-made law, given the scarce legislative output. A good example is intellectual property, in which undertakings have engaged in long-term litigation strategies to shape policy.Footnote ¹⁵¹ As shown in Figure 12 (row 12 from the bottom), the label ‘intellectual property’ keeps a high and steady number of cases, which the creation of a specialized court dealing with matters related to patents is unlikely to alter.Footnote ¹⁵²