Introduction

The role of interest groups in EU policy formulation remains a significant gap in interest group research on EU policy (Bernhagen et al. 2015). Most of the literature on EU lobbying is based on theories of resource exchange and access. Interest groups offer so-called access goods, such as expertise or legitimacy, in exchange for access to political decision-making (Bouwen 2002, 2004, 2009). However, access to EU institutions does not automatically translate into “influence” on EU policy-making (Bouwen 2004, 337–38; Klüver 2013, 10). The concept of influence is used by many scholars, but is often defined vaguely and in some cases very differently in the interest group literature. For example, while Chalmers (2011) “conceived of influence as a function of an interest group’s information processing capacity”, Dür and De Bièvre (2008) define it as “control over political outcomes”: the closer interest groups manage to align political outcomes with their ideal outcomes, the more influence they are said to have (Dür and De Bièvre 2008, 28; Dür 2008; Schneider and Baltz 2003, 2005). Some researchers attempt to analyze interest groups’ so-called ideal positions on the basis of their commentary on EU policy issues (Yackee and Yackee 2006; Yackee 2006) and to measure their distance from the actual policy outcome (Klüver 2012, 2013). Methodologically similar to that, (Yackee 2020) measures interest group success in lobbying during the US agency guidance document development process, through a content analysis.

These approaches are very promising for measuring preference attainment and seeing who has influence on policy outcomes and to what extent, but they leave the underlying mechanisms in a blackbox. However, I am interested in how lobbyists achieve this influence within the EU policy formation process. Another very difficult aspect of conceptualizing influence is its implicit causality. To assure causal influence within the decision-making process, careful process tracing is urgently advised. To make matters worse, this causality, whether lobbyists’ actions affect the actual political outcome, is almost impossible to assess in such a concealed and opaque relational structure like lobbying. Maybe the think tank analyst Weidenbaum (2010, 136) cited was right, when he stated “You can’t measure influence”. But what we can do is find a proxy to infer the mechanisms within the blackbox of EU lobbying. This is where Discourse Network Analysis comes into play. It allows us to link all the actors participating in the whole policy debate to their specific claims and arguments. Thereby we are able to see shifts in their argumentation and in the content they produce. For example, when key policy-makers, such as members of the European Parliament, Member States, or institutional actors, such as committees, abandon their previous reasoning and adopt the claims of interest groups before influencing the framing of the final regulation.

The huge emissions of passenger cars, the possibility to shape their future eco-design, and the great potential in this field to reduce the global emissions of greenhouse gases were reasons for the EU to start regulatory efforts with the Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles. In retrospect, the huge fraud by German car manufacturers, the huge increase in electric vehicles, and the still growing threat of climate change underscore the importance of this case—not only for the future of private transportation and mobility in the EU, but in the whole world. If manufacturers have to adapt to a regulatory framework, for example to meet binding \(\hbox {CO}_{2}\) ceilings in such a central market, they would also be forced to redesign their products, which in turn would shape the premises of car design worldwide and lay the foundation for all future regulations.

Against the backdrop of an increasingly public discourse on the climate crisis on the one hand and a severe economic recession in the wake of the financial crisis of 2007 on the other, the EU initiated its Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles. These prerequisites alone promise a heated debate with a comparatively high level of public attention between the automotive industry, environmentalist NGOs, as well as decision-makers with opposing political convictions, within the EU and the individual Member States. The latter are divided between member states where the automotive industry plays a key economic role, on the one hand, and on the other Member States where the fight against climate change is a higher priority. They also make the political discourse around this regulatory effort an ideal case for the analysis of lobbying in the EU in a societally relevant field, which allows many insights into its process due to its wide range of interests.

There is also a methodological reason for choosing this particular case. It has been the showcase for the great approach of Klüver (2013), who calculates similarity scores of interest groups’ comments on the European Commission’s first draft and the policy output at different stages of the legislative process in order to draw conclusions about the winners and losers of the policy-making process (Klüver 2013, 205). But I want to know more: At what stage within the interinstitutional policy process are interest groups able to shape the bill? Which institutions, parties, committees, and individual decision-makers are crucial for influencing the final legislation? To answer these questions, it is necessary to delve into the case, to know all the texts, drafts and claims. Most importantly, I do not want to consider the EU institutions as monolithic actors, but to evaluate the behavior of all the entities they contain. Therefore, I had to manually code all documents related to the case, such as comments from interest groups, debates in the Parliament, committee drafts or amendments. In addition, the research design had to be geared towards detail in order to take into account the institutional and policy specific rules of the Brussels business that affect the lobbying success of interest groups on this specific issue. Only through this approach was I able to identify points of interconnection between stakeholder involvement and the policy debate, and to assess which of these are crucial for interest group influence on the final legislation. Thus, I had to sacrifice generalization for in-depth knowledge of the underlying mechanisms.

While a growing literature on European Union lobbyism attempted to grasp influence on EU decision-making, surprisingly few authors analyze political discourse or political debates on specific legislative proposals, even though those verbal interactions between political actors about a given policy are extremely useful for studying actors’ ideal outcomes, their learning, and their influence on the final legislation. Finalized EU legislation evolves from this political discourse, containing a selection of claims and contributions proposed by individual actors. It is thus fairly straightforward to determine which of these were successful and which were not. This enables researchers to gain insights into which lobbyists are able to place their favored claims into the final legislation and which EU institutions did in fact adopt them during different phases of the policy formation process. The actors’ statements are highly relational both temporally and cross-sectionally, which characterizes them as a network phenomenon. Discourse network analysis takes this relational dimension of political actors and their statements into account while analyzing the structure of the political discourse and inferring its generative processes (Leifeld 2017).

Discourse network analysis combined with inferential network analysis is a well-established methodological combination, although it has only been used sporadically in interest group research and, to my knowledge, not yet as a proxy for approximating lobbying influence, especially when applied to lobbyists’ comments and relevant documents that emerge throughout the EU policy process on the respective issue. In addition, the methodological and data collection approach is geared towards detail, close to process-tracing, and allows hints on causal mechanisms, worthy to be investigated in further research and broader scope. In order to avoid the fuzzy and sometimes misleading concept of influence, I analyze the reproduction of interest groups’ ideal concepts by political decision-makers to account for lobbying success. Influence, like all causal mechanisms, cannot be measured directly, but we can infer it. While we cannot rule out the possibility that actors are successful simply because they are lucky and policymakers adopt their framing without prior knowledge, this should not be statistically significantly related to their type or characteristics (Dür et al. 2019, 43).

In a final step, we use Exponential Random Graph Models (ERGMs) to model which argumentative adoptions by political decision-makers are really statistically significant. Most debates evolve endogenously. Participants may simply reproduce claims because they are popular, are repeated to some extent, or sound plausible. The use of ERGMs allows me to account for these effects. By explicitly linking claims to their originators through discourse network analysis, I can then model which stakeholder comments are significantly entangled in the argumentation of political decision-makers. But before elaborating further on the methodology, let’s dive into the case study and hypothesize about the mechanisms of lobbying influence within the inter-institutional policy debate on the issue.

The discourse around the \(\hbox {CO}_{2}\) reduction strategy

The discourse on the EU Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles, as well as the discourse network mapping it, is based on four variables: actors, concepts, agreement, and time (see Leifeld (2017) for an in-depth introduction). The actors are in this case institutional actors (like the EU Commission, the EU Parliament, its Committees, etc.), lobbyists (like the automotive industry or environmentalist NGOs), as well as political parties and EU Member States. During the debate, they express claims that are subsumed under so-called concepts. Concepts are abstract representations of the contents that are discussed. In a broader sense, concept “is a neutral word for policy beliefs, preferences, justifications and so forth at the content level” (Leifeld 2017, 304). In this specific case, concepts subsume claims that actors make during the policy debate. For example, if they state that “the aim of this regulation should be a reduction target of 130 g/km,” we would then allocate this specific claim with others containing the same target or content into the concept “Reduction target 130 g/km” (as seen in Table 3).

To extract the concepts of all participating actors, all publicly available documents on the Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles that arose during the policy process were coded manually and inductively by the author via the Discourse Network Analyzers (DNA) (Leifeld 2010) and the R-Package rDNA (Leifeld and Gruber 2018). The 129 documents contained inter alia drafts, proposals and communications from various EU institutions, comments from interest groups, oral contributions and amendments by members of the European Parliament, as well as the opinions of individual EU member states and, last but not least, the final legislation. The data is time-continuous. This allows me to examine the changes throughout the debate and across the different EU institutions.

Since the discourse is highly dependent on EU procedures, the discourse and the corresponding network have been divided into time slices (as shown in Table 1). These represent the individual phases of the bill and the subsequent debates, which are determined by the publication of the EU Commission’s draft, the EU Commission’s proposal, which can already be influenced by lobby interests, and the final regulation.

Table 1 The individual phases of the policy discourse within the EU institutions
Full size table

Before presenting its proposal, the European Commission publishes a draft on which interest groups, lobbyists and basically anyone interested can comment within a certain time window. These comments are made public by the Commission, so the interest groups’ concepts can be captured, analyzed and their ideal positions can be extracted. These first discourse coalitions are then augmented with all political decision-makers, committees, institutions, Member States, etc. that participate in the following discourse—as well as their expressed concepts. Two major interest group coalitions intervened in the debate, representing the automotive industry on the one side and environmentalist NGOs on the other. Their composition is shown in table 2. The individual actors were assigned by the author to these two coalitions, according to the interests they represented.

Table 2 The interest group coalitions
Full size table

Several other interest groups participated in the discourse. Consumer organizations, automotive suppliers, business associations, fuel associations and trade unions were part of the discourse, but only played an insignificant role and did not align themselves with the two coalitions. Therefore, they are not the focus of this article, although they are included in the discourse networks. The 127 actors vividly discussed ten issues that represented the main political discussion around the proposed regulation with the participation of interest groups. These issues consisted of 37 concepts in total and two to seven conflicting concepts per issue (see Table 3).

Table 3 Included frames and concepts
Full size table

By treating the bill and the final regulation as a specific entity (or like a non-human actor, following Latour (2005)) and linking them to the concepts they contain, I can show how the draft, the proposal and the final legislation move between the two interest coalitions. It is thus possible to see which interest groups share concepts with the bill in its individual phases. This reveals which interest group coalition shaped the regulation in its favor during the particular phases of the policy process. Furthermore, this allows me us to identify possible belief changes of EU institutions. This is very important, because actors constantly reproduce their beliefs, which subjects them to procedural change (Schneider and Leifeld 2009, 140). This is the crux of our proxy for inferring influence. But before specifying this, we want to derive some hypotheses from the existing theories on the lobbying success of interest groups lobbying the EU.

Hypothesizing the mechanisms of interest groups’ influence on the EU policy debate and its decision-making

Theories on the mechanisms of lobbying success in regard to EU policy debates are well developed, so I rely on some expectations from the aforementioned literature, which is very well summarized and extended by Dür et al. (2019). Many of the legislative proposals on the EU's agenda represent a departure from the status quo of minimal regulation or no regulation at all. This applies for this case, too. As mentioned in the introduction, the EU’s Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles is the EU’s first regulatory effort in this direction. According to Dür et al. (2019), most business interests are close to this status quo, while citizen groups, as well as the Commission and the EP, tend to advocate policy change. As a result, in most of these debates, business interests tend to take an almost unified position. Thus, the structures of conflict in EU legislative politics put business interest and the position of civil society actors in opposition (Dür et al. 2019, 39–40). Although so-called strange bedfellow coalitions exist (Beyers and De Bruycker 2018; Abel and Mertens 2023), many authors assume typical NGO-business conflicts, especially in the field of climate policy (Abel and Mertens 2023; Ingold 2011; Ingold et al. 2017).

Hypothesis 1:

Business interest groups and environmentalist NGOs should form two opposing coalitions within the debate.

To assess the possible pathways for interest groups to influence the policy debate, we have to account for the multitude of institutional venues and the effect of contextual variables. “Interest group scholars have not yet fully analyzed the complex interplay between individual interest group behavior and the overall institutional or policy context in which interest groups operate” (Klüver et al. 2015, 449). We cannot trace every venue of influence, but we can analyze where interest groups do find political decision-makers who reproduce their claims. Berkhout et al. (2021) analyzed the relationship between political ideology and political contact, based on survey data, and found: “For instance, Green parties, in their ideologies, have a clear conception of the nature of citizenship activism, and this belief is likely to reverberate in their contacts with citizen groups” (Berkhout et al. 2021, 427). Authors working primarily with the Advocacy Coalition Framework on environmental and energy policy tend to make a similar assumption: that right-wing and pro-business actors on the one hand and left-liberal and environmental actors on the other are more likely to form coalitions (Ingold et al. 2017; Ingold 2011).

Hypothesis 2a:

Environmentalist NGOs will side with green and progressive MEPs, while the automotive industry might find an ally in conservative and liberal MEPs.

Dür et al. (2019) derive from interviews with EU Commission officials that business groups lack an institutional ally in the EU, while citizen groups have two—the European Commission and the European Parliament. That’s why the latter often end up being more successful in EU legislative politics than business interests (Dür et al. 2019, 133–34). Thus, the EC and EP are much closer aligned with citizen groups than with business actors, and not even the Council is a steady ally for business interests (Dür et al. 2019: p. 157). Business interests start with a disadvantage, but they can compensate for this disadvantage by using technical information to reduce the distance between them and these two policy-making institutions (Dür et al. 2019, 157).

But when business interest groups provide high quality information, the Commission and the Parliament are closer to them (Dür et al. 2019: p. 157). Similarly, Rasmussen (2015) states that business interests are more likely to influence the European Parliament, when they are united, when they lobby on highly technical issues, and when mainstream committees examine proposals. Regarding the issue and recalling the concepts (listed in Table 3), the information provided by interest groups is of a highly technical nature. Klüver et al. (2015), as well as Marshall (2015), who draws on several surveys of MEPs, also emphasize that committees are attractive and important targets for interest groups.

Hypothesis 2b:

While environmentalist NGOs may find an ally in the EU Commission, business groups, if they are united, should be able to influence the debate in the European Parliament as well as the outcome in its committees.

As described earlier, for this small, limited case, I want to use a holistic approach that covers the entire legislative process and leaves no blind spots. Before testing the hypotheses, I will outline our methodological approach.

Conceptualizing the inference of influence

In contrast to most of the studies cited above (with the exception of Klüver’s (2013) study), this methodological approach relies exclusively on primary sources. All official contributions to the policy debate are included in the discourse network analysis. Institutions are not considered as monolithic actors. Every voice is included.

Therefore, we will also gain depth of field regarding the content, which is broken down to every single concept. All the concepts are linked to their originator and their supporters within the debate, even if that support changes at a later point in time.

The methodological approaches of the abovementioned studies, which analyze the influence of interest groups on EU policy-making, measure the distances between interest groups’ comments or ideal positions and the policy outcome—sometimes a joint decision of an EU institution (Dür et al. 2019), sometimes the different stages of the draft bill, or the final legislation (Klüver 2013). But mechanisms remain in a black-box, though. How did interest groups get there? Which MEPs, for instance, did they convince? Discourse network analysis is very useful for this kind of process tracing by incorporating the whole range of actors and linking them to the contents of the debate. Hereby, I can reconstruct the debate from the first draft of the commission and the initial comments by interest groups, throughout the discussions within the EU institutions, up to the wording of the final legislative text.

But by simply counting the links between interest groups, concepts, and decision-makers, I would highly overestimate their status, as for example “Environmentalist NGO” or “Automotive Industry,” as well as the attributes of decision-makers—for instance as an German MEP or member of the Green Party. Every discourse and every debate has strong endogenous effects. Verbal interaction builds on previous statements and content. Perhaps the reason why a concept is reproduced by a decision-maker is merely because of its popularity and visibility within the debate. Or maybe it just sounds plausible, regardless of the entity backing it. With the Exponential Random Graph Models (ERGMs), I am able to account for these endogenous effects and can then calculate which interactions between different interest groups and specific decision-makers are really significant. This allows me to infer whether, for example, green MEPs reproduce the content of environmentalist NGOs significantly more often than by chance or by a random actor without this interest group type. I will introduce discourse network analysis and ERGMs briefly in the following subsections.

Discourse networks

By combining the four variables (actors, concepts, agreement, and time), a bipartite affiliation network is generated (as seen in Fig. 1).

Fig. 1
figure 1

Discourse network illustration. Bipartite affiliation network based on Janning et al. (2009): The edges between the actor and the concept levels represent the affiliation network, while the dashed edges on the individual levels represent the specific projections

Full size image

Discourse network analysis operationalizes connections between network actors based on their agreement on concepts. The agreement relation captures the statements’ sentiment. In our case, only positive connotations were coded. If actors rejected a concept, their alternative proposal was coded instead of a negative connotation. Date stamps on the statements allow us to analyze changes in the policy debate. Thereby, an actor’s statement is an edge to the corresponding concept at a specific point in time (Leifeld 2017). Network analysis is less a theory than a toolbox for measuring relational configurations and their structural characteristics (Kenis and Schneider 1991, 44). Linkages can be conceptualized as resource exchange (Gulati et al. 2005; Kenis and Schneider 1991). Interest group coalitions that exchange their resources and/or information (Heaney and Leifeld 2018) then try to include key actors within the EU institutions in their discourse coalition (Hajer 1993) to mold decisions for their benefit. This discourse coalition formation can be easily examined with this approach to infer the influence on the policy formation and the mechanisms enabling that influence. Furthermore, EU institutions are not seen as monolithic entities, which provides insight into potential conflicts within institutions, for example between parties or committees within the European Parliament. It is thus possible to capture problem definitions of single actors as well as their interplay with potential policy changes. The aim is to grasp changes and to analyze them as determinants of or in dependence on other data, such as lobby influence and resource exchange, veto player constellations, or external events (Leifeld 2009, 391).

To analyze which actors are able to place their concepts in the final legislation and which political decision-makers adopted the concepts that lobbyists supported, the discourse network was analyzed using Exponential Random Graph Models (ERGMs). The following section briefly introduces ERGMs and the model terms used in this paper to operationalize the success of interest group concepts and to capture the replication of their concepts by key decision-makers, which ultimately serves as a proxy for lobbyists’ influence on the legislation.

Exponential Random Graph Models

ERGMs model the probability of a network and the measures of network features, like clustering, homophily or the parameters we will specify later in this section (B. A. Desmarais and Cranmer 2012; see also Koskinen and Daraganova 2012; Robins and Lusher 2012; Robins and Daraganova 2012). They are also able to model the effects of covariates on the status of relationships and at the same time the prominence and significance of structural dependencies. The latter distinguishes them from conventional regression (Bruce A. Desmarais and Cranmer 2017). They model a network by describing its composition of endogenous local structures and how its structure is additionally codetermined by exogenous covariates. The exogenous covariates are for example nodal attributes that increase or decrease the tie probability of a connected dyad (Heaney and Leifeld 2018). Before specifying the endogenous and exogenous model terms (see Table 4), the basic function of ERGM will be described very briefly. The ERGM formula (see for example Goodreau et al. 2009; B. A. Desmarais and Cranmer 2012) can be expressed by:

$$\begin{aligned} \Pr \left( \varvec{Y} = \varvec{y}\ |\ \varvec{X} \right)= & {} \ \frac{\exp \left[\varvec{{\theta }}' \text {g}\left( \varvec{y},\ \varvec{X} \right) \right]}{\kappa \left( \varvec{\theta },\varvec{X},\varvec{\mathcal {Y}} \right) } \end{aligned}$$

\(\varvec{Y} \subseteq {[1,\ \ldots ,\ n]}^{2}\) shall be the set of potential dyads among \(n\) nodes. Nodes are in this case, as introduced above, actors and concepts. The network \(y\) is represented by a set of ties. Here, the statements of the actors link them to the concepts. The set of possible sets of ties, \(\varvec{Y} \subseteq 2^{\varvec{Y}}\), is the sample space, then \(\varvec{y} \in \ \varvec{Y}\). \(\varvec{X}\) is an array of covariates that contains attributes of nodes and/or dyads. These are, for example, the actors’ status-group, such as “Environmentalist NGO,” a political party or an EU institution, or the attribute of a concept, e.g. whether it has been dismissed or is in the final legislation. \(\textbf{g}(\varvec{y},\ \varvec{X})\) represents a vector of network statistics with a vector of coefficients, \(\varvec{\theta }\), for those statistics and the normalizing constant \(\kappa \left( \varvec{\theta },\varvec{X},\varvec{\mathcal {Y}} \right) = \ \sum _{\varvec{y} \in \ \varvec{\mathcal {Y}}}{exp\left[\mathbf { \varvec{\theta }' \text {g}}\left( \varvec{y},\varvec{X} \right) \right]}\) is the sum over the space of possible networks \((\varvec{Y})\) on n nodes.

As described above, the discourse network has been cut into time slices or repeated network cross-sections over time, representing the phases of the EU policy process. This enables the observation of the complete network at these points in time and to model the dynamic network processes on the same node set. To do so, the individual time-slices are compared to their predecessor via conditional maximum likelihood estimation (Krivitsky and Handcock 2019). This is important not only because it is interesting to know what happened in these phases of policy-making, but also to find clues to causal mechanisms. It would be impossible to analyze lobbying influence in a single, cross-sectional network, because linkages of actors over concepts would not reveal anything about the direction of the influence. However, if a political decision-maker significantly supports a concept that a lobbyist stated earlier on, the lobbyists’ efforts can be considered successful. Since the individual concepts are so fine-grained, it is very unlikely that the same claim is made by chance. For example, it is highly unlikely that two actors claim that the reduction should only apply to “vehicles not exceeding 2,610 kg” by chance, without knowing about the claim beforehand. Even if a lobbyist and a policy-maker were to repeat a concept already expressed by a third entity unknown to us, the lobbyist would still be successful: after all, her ideal position has been reproduced. Merely our discourse coalition would then be incomplete. That means, knowing that lobbyists’ statements are in the very beginning of the policy debate, a linkage over a concept between a lobbyist and a political decision-maker that partakes in the debate at a later point in time is an indicator of lobbying success.

To analyze these dynamic processes, we estimate the effect of endogenous and exogenous model terms on the formation of ties between each of the time slices described in Table 1. Basically, we model the transition from \(t_1\) to \(t_2\), from \(t_2\) to \(t_3\), and so on. For our case, only the formation dynamics are important since actors only added concepts to the discourse and did not “withdraw” them.

The formation form, conditional on only adding ties, can be expressed as follows:

$$\begin{aligned} \Pr \left( \varvec{Y}^{+} = \varvec{y}^{+}\ |\ \varvec{Y}^{t};\ \varvec{\theta }^{+} \right) = \ \frac{\exp \left[{\varvec{\theta }'}^{+}\textbf{g}^{+}\left( \varvec{y}^{+},\ \varvec{X} \right) \right]}{\kappa \left( {\varvec{\theta }}^{+},\varvec{X},\varvec{\mathcal {Y}}^{+}\left( \varvec{Y}^{t} \right) \right) } \end{aligned}$$

Given \(\varvec{y}^{t}\), a formation network \(\varvec{Y}^{t}\) is generated from an ERGM with formation parameters \(\varvec{\theta }^{+}\), the formation statistics \(\textbf{g}^{+}\left( \varvec{y}^{+},\ \varvec{X} \right)\) and \(\varvec{y}^{+}{\in \varvec{Y}}^{+}\left( \varvec{y}^{t} \right)\) (Krivitsky and Handcock 2014; Krivitsky and Handcock 2019).

Readers who are already aware of ERGMs might be already familiar with this terminology from the Separable Temporal Exponential Random Graph (STERGM) function of the statnet package. Unfortunately we had to create the transition matrices between the specific time slices manually to model the formation of ties from a network \(\varvec{Y}^{t}\) at time \(t\) to a network \(\varvec{Y}^{t + 1}\) at time \(t + 1\), because statnet requires the same set of nodes in every time-slice and therefore works with isolates, nodes with no edges at the specific point in time. The latter however are theoretically forbidden in our case, since the isolates are not only absent from the discourse at the time of the analyzed transition, but are also forbidden to engage by the strict EU policy procedure. Including these actors in the model, like for example a European Parliament Committee, even though it is not allowed to participate in this phase of the discourse, would bias the estimation. To avoid this, and to avoid the STERGM-function’s automatic usage of dissolution models, which are not relevant for me either, I constructed the transition matrices manually, using a logical disjunction of the two specific network matrices observed at the abovementioned points in time to obtain the transition matrix. Then I ran ERGMs on these transition matrices to model the formation of ties between the individual phases of the policy formation.

The following table introduces and explains the endogenous and exogenous model terms that will be used to model the ERGMs for our bipartite discourse networks.

Table 4 ERGM terms
Full size table

As described above, the endogenous model terms simulate the structural effects of the network while the exogenous model terms analyze the actor-relation effects we want to measure. To model the endogenous structure of the discourse network and its evolution, several structural effects are included. The edge-term accounts for the network density and the two curved, geometrically weighted terms on both levels model the centralization of actors and concepts (Hunter 2007). In this case, they model the general behavior of the actors (if they tend to jointly support other concepts, suppositionally they already share a concept) and the concept popularity (how many actors do support them and the tendency to support an already popular concept).

This study uses two exogenous configurations. The bipartite node-match terms (Bomiriya et al. 2014) for the two-mode ERGM enable insights, which interest group coalitions are more successful in “bringing” their concepts into the final legislation. All concepts that are represented in the final legislation, or are in other words successful (the highlighted concepts in Table 3), were tagged with the attribute “Legislation” while all others were labeled “Dismissed.” By doing so and by using the node-match term for bipartite networks, it is possible to measure if concepts of the automotive industry actors or the concepts that environmentalist NGOs support are significantly more likely to be represented in the final legislation. We are thus able to estimate their influence on the conceptual composition of the regulation.

The 2-star-configurations on the concept level for the ERGMs on the transition matrices enable us to measure whether political decision-makers reproduce concepts previously introduced or expressed by lobbyists. If an actor expresses a concept at an earlier time point in the network and other actors support that concept at a later time point in the observed network, this hints on influence and success for the first actor. This is the central measure to understand which political decision-makers and EU institutions shared lobbyists’ concepts and how lobbyists’ favored concepts were brought into the final legislation—in other words, to infer the mechanisms of their influence and who enabled it. This is basically the operationalization of hypotheses 2a and 2b and can also be used to assess actor homophily with respect to their status as environmentalist NGO or automotive industry actor.

Glimpsing the lobbying influence on the discourse

To assess the endogenous goodness-of-fit of the individual models, various diagnostics have been conducted to assure the endogenous network properties of the models match the ones of the observed discourse network. This is important to rule out any possible bias of the estimates. All models are well fitted, as the graphical representations of the goodness-of-fit diagnostics, that are contained in the appendix, show. For the estimation of the statistical models included in the cross-sectional ERGM, the ergm package (Handcock et al. 2008; Handcock et al. 2018) from statnet (Handcock et al. 2008) in the R environment (R Core Team 2018) was used.

One first step to grasp a glimpse on lobbying influence on the legislation, and to determine which interest group coalition was more successful in placing its concepts in the final legislation, is to include various node-match statistics across the two levels of the cross-sectional bipartite network in an ERGM (Bomiriya et al. 2014). This allows me to estimate which interest groups express concepts that were dismissed later on or those that are represented in the final legislation, and is my proxy for measuring the success of actors, as well as their discursive influence on the debate and the final conceptualization of the regulation. The concepts (as seen in Table 3) represented in the final legislation were marked as ‘Legislation’ while those that were dismissed during the policy debate have the attribute ‘Dismissed.’ The ERGM in Table 5 includes several purely structural effects (in the upper part) that control for the overall network structure. They account for the geometrically weighted degree on the concept level of the bipartite network (Hunter 2007) to model the clustering of the network. The strong negative effect of the geometrically weighted degree distribution on the concept layer means that we encounter a more centralized network than by chance and a tendency towards nodes with higher degrees.

Table 5 Bipartite ERGM analyzing lobbyists’ success in placing their concepts into the regulation
Full size table

The interaction effects in the second section show that the automotive industry, although supporting a statistically significant number of dismissed concepts, too, is significantly more likely to “bring” their favored concepts into the final legislation. The environmentalist NGOs, on the other hand, have statistically significant estimates on the node-match configurations that include dismissed concepts. Obviously, the concepts of the automotive industry’s interest groups were more likely to appear in the final EU legislation: therefore, they were the more successful lobby coalition by comparison and seem to have had a stronger influence on the legislation’s concept composition.

To assess why the automotive industry was more successful and which political decision-makers enabled it to do so by reproducing its claims and concepts, we use the time-dependent approach. As outlined above, other network terms and the dimension of time are necessary to operationalize the reproduction of concepts by key political players in order to provide indications of the influence of lobbyists on the legislation. Therefore, the individual transition matrices representing the discourse network formation between the single cross-sectional discourse networks, shown in Table 1, were analyzed with ERGMs. Pursuant to the above described time-slices of the discourse network, Model 2 reveals the formation patterns of ties from the first phase of the debate (first section of Table 1) to the second (second section of Table 1). Therefore, the discourse network slice in the second section of Table 1 is compared with the small network only consisting of the EU Commission’s draft, the concepts it contains, and the accompanying comments by the EU Commission. This formation model analyzes the contributions by the interest groups as well as the first comments of members of the European Parliament, its committees and the EU Member States.

In accordance with the ERGM in Table 5, the endogenous network formation was controlled by purely structural terms (see the section above the actor-relation effects). The structural effects are robust and significant in all models and show a strong centralization tendency on popular concepts.

The interaction effects are the 2-star-configurations on the concept level, described above. They estimate which actor-types share concepts and, bearing the time-dimension in mind, which actor types replicate other actor types’ concepts. Several 2-star-effects had to be dropped, because they were unstable. Some actor-concept-actor constellations simply developed no new 2-stars in the specific time-frame with actor attributes of interest. The interaction of the two interest group coalitions were examined in Model 2. As expected in Hypothesis 1, environmentalist NGOs are significantly more likely to share concepts with other environmentalist NGOs. The same applies for the actors representing the automotive industry, while the two coalitions do not tend to agree on the same concepts. This means, each of the two coalitions show tendency to within-coalition homophily and a polarization of the discourse along the coalition membership—they do not only share their denotation, but also concepts in the debate.

Table 6 Bipartite STERGMs analyzing the influence of lobbyists on political decision makers
Full size table

The time-dependent models in Table 6 and the other models show the categorized actors in the chronological order of their participation in the discourse—this applies to the vertical order of the actors in each model as well as to the examined actor pairs on the concepts. This means, for example, that in the first actor pair “EC Draft & Environmentalist NGO” the draft of the EU Commission was published first and then the interest groups reacted to it with their comments.

Model 2 shows strong positive and significant results for this actor pair, indicating that 2-stars between these two actors are way more likely to be seen than by chance—in contrast to the 2-star configuration between the EC Draft and the automotive industry. Translated to our research question, this means that in their comments environmentalist NGOs strongly supported the concepts proposed by the European Commission in its draft regulation. Some Member States submitted written contributions before the time window for interest groups’ comments was opened. Only actors representing the automotive industry show a statistically significant but slightly negative agreement with Member States’ concepts. The probability that the latter would reproduce a concept introduced by Member States is less than by chance. The same applies for the environmentalist NGOs and the EPP-ED, the conservatives in the European Parliament, as well as for the automotive industry and members of the PES, the socialist party. We must note, however, that only a few MEPs commented on the regulation efforts at this stage of the debate.

Following the first comments of some Member States, the comments of the interest groups, and the first statements of MEPs, the European Commission published its proposal for the regulation. Not surprisingly, the proposal contains a lot of concepts from the draft, which explains the high positive and significant 2-star effect between these two documents. It is striking, however, that the automotive industry shows a strong and significant 2-star effect with the proposal. While the draft has received approval by the environmentalist NGOs, the proposal seems to replicate much more concepts favored by the automotive industry than by chance. Translated to our general research question, this hints on influence by the automotive industry on the concept composition of the proposal. Influencing the Commission's proposal was the first step for the automotive industries’ achievement in changing the wording of the bill in their favor.

The last phase of the discourse is analyzed in model 4. A first look at the endogenous processes reveals that the strong tendency towards centralization at the concept level decreases. It is interesting to see that the negative estimates of the geometrically weighted degree distribution parameters decrease in the course of the debate, as more alternative concepts enter the stage and/or gain popularity. This goes hand in hand with the results of the previous models and with the estimates of the ERGM 1, shown in Table 5. In Model 1 we already saw hints on the success of the automotive industry, which introduced many concepts that ultimately ended up in the final legislation, even though the concepts included in the first draft of the European Commission were hailed by the environmentalist NGOs. So there had to be a shift in popularity towards alternatives advocated by the automotive industry. In this phase, the Members of the European Parliament and the Member States in the Council introduced their amendments, too.

Thus, for the successful implementation of these concepts in the legislative framework, they had to be replicated by political decision-makers. Yet, the estimates of Model 4 show only significant negative estimates when it comes to party affiliation of MEPs and the support for the interest coalitions’ concepts. On the one hand, ALDE and EPP-ED members are less likely to share concepts with environmentalist NGOs than by chance, while Greens-EFA and PES members, on the other hand, are less likely to support the automotive industry’s concepts. I cannot support Hypothesis 2a, though, since the results only confirms its negation. We neither see significant support by conservative MEPs for concepts of the automotive industry nor by Green MEPs for environmentalists’ concepts. Contrary to this, the origin of the individual MEPs mainly shows positive effects on the 2-star configurations. French, German, Italian and Spanish MEPs, as well as their colleagues from the UK, show a strong support for the concepts of the automotive industry, which is not surprising though, since they are home to Europe’s largest car manufacturers. Only MEPs from the Netherlands and Sweden show a stronger support for environmentalist NGOs’ concepts than by chance—just as German MEPs, which may surprise at first, but goes along with the fact that a large part of the Greens-EFA is German. The automotive industry’s concepts are significantly reproduced by the European Parliament Committees. This provides support for Hypothesis 2b, indicating that business groups remained united and exerted influence over both the European Parliament's debate and the outcomes in its committees. Environmentalist NGOs had an ally in the Commission—at least in the very beginning of the policy-making process. Many concepts in the first draft were supported by them. The automotive industry, though, was able to mold the subsequent Commission proposal as well as convince the majority of MEPs of their point of view. The Member States (in this model active after the interest groups expressed their comments) are, in total, slightly less likely to reproduce automotive industry’s concepts.

Conclusion and discussion

For both interest coalitions, it was crucial to shape the conceptual composition and focus of the regulation, as this also meant shaping the basis of the EU's efforts to regulate \(\hbox {CO}_{2}\) emissions from cars in the long term. The regulation presented here was the first of its kind and will influence further renewals, too, which could be analyzed in further studies to increase the in-case insights.

Using discourse network analysis in combination with ERGMs, we were able to find a suitable proxy for lobby influence and thus assess which interest coalition was more successful. This allowed us to assess not only which lobby actors were able to shape the composition of the regulatory concepts, but also how the mechanisms of their influence on the debate and EU policy formation worked in this case. Claiming concepts is only the first step. The nexus between the lobbyists’ concepts and the determinant of their success or failure are the key decision-makers within the EU institutions who reproduce them and enable the lobbyists’ influence.

The automotive industry claimed pretty successful concepts in their commentary. Although the first concepts by the EU Commission’s draft were positively received by the environmentalist NGOs, the environmentalist coalition’s concepts were rather dismissed. The automotive industry, on the other hand, was able to gain broader support and thus was able to place their concepts in the European Commission’s proposal. Building on that, the concepts they supported were significantly more likely to appear in the final legislation. Shaping the Commission’s proposal in combination with the strong support by MEPs from car-manufacturing Member States in the later phase of the debate, translated to this success. Despite its actors claiming several concepts that were dismissed during the debate, the automotive industry coalition was significantly successful in placing concepts into the final legislation and influencing the outcome of the EU Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles—in contrast to their opponents, environmentalist NGOs.

Admittedly, this single-case study provides insights only for readers interested in the lobbying processes surrounding the EU Community Strategy to reduce \({CO}_{2}\) emissions from passenger cars and light-commercial vehicles and who are curious about an assessment of lobbyists’ influence on this discourse. The substantive findings of the study are by no means generalizable to lobbying in the EU, but the methodological part is. First, the operationalization of influence and, more generally, of individual actors on discourses and debates is transferable to other cases and research settings. It stands out for its ability to add more cases in a comparative way and for its expandability, both theoretically and methodologically. This allows for more case studies of lobbying on EU regulations in different policy fields that could be compared later, and thus for more sophisticated and generalizable findings on the matter of influencing EU policy-making. Moreover, as it relies on network analysis, which is itself a fairly agnostic methodological toolkit, it is furthermore modularly extendable. Not only could more case studies based on the methodology provide a broader understanding of EU lobbying, but the modular expandability of the approach itself could provide an additional advantage over conventional approaches research on lobbying. The individual levels of the bipartite discourse network could be extended by other relational data, embedding other theories on interest groups and/or discourse formation, into a multidimensional network.

There are a variety of other hypothesized conditions that might determine the ability of actors to influence political decisions. In conclusion, and as a brief outlook, I would like to highlight a few of the most popular research strands that could be incorporated into further network based analysis of lobbyism and its influence estimation. The actor level, for example, could be enriched with ties that contain information about whether actors are sitting together in EU expert groups or if they are entangled in some other way. Does this translate to influence in the EU policy formation? Population ecology (Baum and Dutton 1996; Berkhout et al. 2015) and comitology (Chalmers 2014; Gornitzka and Sverdrup 2008) could inspire such an integration of actor interconnections. The concept level could contain further information about the combinations of concepts to integrate framing (Daviter 2007); thus, the analysis of the way policy concepts are associated with a subjective interpretive framework or pattern. Interest groups, for instance, attempt to steer the debate in a specific direction by highlighting certain aspects of a legislative initiative—and ignoring others—to influence how other players, like political decision-makers or the public, see these issues collectively. Baumgartner and Mahoney (2008) call this individual level framing. In this case, frames would be combinations of different concepts from different issues. Framing possibilities for interest groups and their framing strategy choices are, on the other hand, highly dependent on the context (Baumgartner and Mahoney 2008; Klüver et al. 2015). Political interference by framing will never stay unchallenged (Daviter 2012; Jones and Baumgartner 2002). The specific framing strategies are linked to various factors, like interest group population (Jones and McBeth 2010, 346), involved institutions and interest groups’ form of organization (Klüver et al. 2015, 495). Last but not least, it refers to the collective dynamics of the individual level frames. All these multifaceted approaches navigate around the interaction and relation that are the core mechanisms of lobbying, but they are not able to fully encompass them—the network approach, on the other hand, does and could enrich and combine them. The network approach to study interest groups does not reinvent the wheel, but has the potential to gather these loose spokes and align them into a better one. Consequently, extending the bipartite discourse network to a multidimensional network might provide a promising starting point to unite the tessellated but interdependent strands of research on EU lobbying, to explain it in a more multifaceted way that might ultimately allow a comprehensive analysis of lobbying influence and detailed insights into its causal mechanisms.