Government Rhetoric and the Representation of Public Opinion in International Negotiations

The role of domestic public opinion is an important topic in research on international negotiations, yet we know little about how exactly it manifests itself. We focus on government rhetoric during negotiations and develop a conceptual distinction between implicit and explicit manifestations of public opinion. Drawing on a database of video recordings of negotiations of the Council of the European Union and a quantitative text analysis of government speeches, we find that public opinion matters implicitly, with the exact pattern depending on governments’ stance toward the EU. Pro-EU governments are responsive to public opinion in their support for compromises and attempts to stall negotiations, whereas Euroskeptic governments tend to remain silent when confronted with a public positively disposed toward the EU. Our results show that although governments implicitly represent public opinion, they do not systematically invoke their voters explicitly, suggesting the public matters but in different ways than often assumed.


INTRODUCTION
A growing body of work has looked at the role of public opinion in international negotiations. One strand of the literature focuses on bargaining outcomes and assesses whether a constraining public opinion at home can serve governments to improve their bargaining success at the international level (e.g., Baerg and Hallerberg 2016;Bailer 2006;Lundgren et al. 2019;Wratil 2019), especially when national elections are close (Chaudoin 2014;Rickard and Caraway 2014;Schneider 2018). In contrast, another strand-especially in the literature on European Union politics-investigates governments' negotiation positions and tests to what extent they are reflective of, or supported by, domestic public opinion (e.g., Hobolt and Wratil 2020;Schneider 2018;Târlea et al. 2019;Thomson 2011;Wratil 2018). Although this rich literature has provided us with important insights into whether outcomes and government positions during intergovernmental negotiations reflect public opinion, we know much less about the exact mechanisms and ways in which public opinion matters.
In this study, we draw on bargaining rhetoric to understand how public opinion manifests itself during negotiations as well as to provide an original test-in a domain of government behavior rarely studied systematically to date-whether public opinion has effects on negotiations at all, given conflicting findings in the literature (for null findings see, e.g., Bailer 2006;Lundgren et al. 2019;Târlea et al. 2019;Thomson 2011). 1 For this purpose, we develop a conceptual distinction between explicit and implicit ways of how public opinion can play a role in rhetoric, building on the existing literature. Do governments represent public opinion in implicit ways by adjusting their negotiation rhetoric to the public's support for international cooperation? And alternatively or in addition, do they raise voters and their views as an explicit topic of discussion?
Whether public opinion manifests itself in implicit and/or explicit ways in the rhetoric of international negotiations is important from various perspectives. Theoretically, the literature on domestic constraints and two-level bargaining games (Putnam 1988;Schelling 1960) expects that governments will sometimes explicitly speak about their public and its views to restrict the bargaining zone. Similarly, constructivist and sociological institutionalist approaches in international relations (cf. Krebs and Jackson 2007;Risse 2000;Rittberger 2005;Schimmelfennig 2001) suggest that some governments may talk about domestic public opinion, as it may justify their own positions and persuade others to change theirs. In contrast, from the perspective of democratic responsiveness to public opinion, governments may just implicitly follow the wishes of the public when taking positions, agreeing to compromises, or blocking negotiations but do not necessarily have to highlight this explicitly in their rhetoric.
Empirically, the relative prevalence of implicit and explicit manifestations of public opinion may illuminate bargaining dynamics. For instance, implicit references may be associated with misunderstandings and miscalculations about the importance of public opinion among partners, whereas explicit ones may be reflective of red lines hampering compromise and agreement. Normatively, the simultaneous explicit and implicit representation of public opinion represents the assumed standard of democratic representation in national parliaments. Yet, many are concerned that representation at the international level may be technocratic, lacking explicit references to citizens, or unresponsive, not tacitly accounting for citizens' views (e.g., Føllesdal and Hix 2006).
We study the explicit and implicit manifestations of public opinion in speeches given by governmental representatives in the Council of the European Union (henceforth, "the Council") in which delegates meet to negotiate over EU legislation. Although the closed nature of intergovernmental negotiations usually precludes researchers from studying the actual bargaining rhetoric of governments, we exploit a transparency change in the Council that demands a significant part of intergovernmental discussions to be recorded on video, allowing us to investigate what governmental delegates actually say during negotiations. Drawing on a corpus of more than 3,600 speeches made by national representatives in the Council during the period between 2010 and 2016, we apply a text-as-data approach to measure governments' engagement with public opinion in their bargaining rhetoric. Using structural topic models (Roberts et al. 2014), we investigate how government ministers' speeches in Brussels differ depending on the public image of the EU in their member state. As a direct consequence of this modeling strategy, we are able to test whether governments engage with public opinion implicitly and/or explicitly during negotiations.
Our premise is that Council negotiations are nowadays characterized by a divide between pro-EU and Euroskeptic governments and that coalition-building in the Council primarily revolves around the core of pro-EU governments, who form policy coalitions to pass proposals. On the basis of this distinction, we derive testable hypotheses about explicit and implicit manifestations of public opinion in governments' negotiation rhetoric from different theoretical perspectives in international relations and EU politics. To clarify from the outset, throughout the article we use "pro-EU" versus "Euroskeptic" to denote the ideology of governments and "positively disposed" versus "negatively disposed" to talk about domestic publics' view on the EU.
Our empirical results yield clear evidence for implicit manifestations of public opinion: pro-EU governments express more support for compromises if public opinion at home is more positive toward the EU and become more cautious and call for postponing negotiations if the public's image toward the EU is more negative. However, for Euroskeptic governments we do not find much evidence that public opinion influences their rhetoric; instead, we show that these governments tend to participate less in debates when domestic public opinion is more positively disposed toward the EU. In contrast to implicit manifestations, we find little evidence for explicit manifestations of public opinion: none of the topics prominent in negotiations in the Council are dedicated to concerns about public opinion, citizens, or voters. Importantly, the predominance of implicit over explicit representation of the public suggests that the Council is not fully resembling a representative institution like a national parliament, where citizens and their concerns are addressed directly. This finding sheds a critical light on the EU's expressed ambition over the last two decades to turn the intergovernmental branch of its legislature, the Council, into a transparent and politicized decision-making body that would bring Europe closer to its citizens.

PUBLIC OPINION AND INTERGOVERNMENTAL NEGOTIATIONS IN THE EU
Intergovernmental negotiations typically take place behind closed doors, rendering their study inherently difficult. We investigate the Council of the European Union, arguably the most transparent intergovernmental bargaining forum of any international organization around the world. As the Council is the EU's primary legislative chamber, alongside the European Parliament, governments and their representatives negotiate over EU legislation on a daily basis. What makes the Council particularly relevant for our study is that certain deliberations have become public over time, even though the public is not invited to the deliberations. Even though lower levels of the Council structure, such as working groups and committees, are still conducted behind closed doors, the Council's ministerial meetings have been gradually opened up to the public since (European Council 2006. Since the adoption of the Lisbon Treaty in December 2009, the ministerial Council must meet in public when negotiating or voting on legislative proposals. Sessions are public insofar as the speeches of ministers or member states' permanent representatives to the EU are recorded on video and made available on the Council's website. 2 Although it is always possible to move discussions to secretive informal forums (e.g., lunch meetings) or to lower, bureaucratic levels that are not public (e.g., working groups), ministerial Council meetings make up for a significant part of the negotiations in the Council structure (Häge 2008;. In particular, the meetings take up a large portion of the time national ministers spend in Brussels. In many instances, ministers switch between closed and open sessions within the same meeting: the presidency usually informs the attending ministers that the public session has started, meaning the installed cameras are recording, and that it has ended. But ministers remain in the same seat and room. During the public sessions, ministers discuss technical details of legislative proposals, set out and justify their policy positions, or ask colleagues for compromises (Wratil and Hobolt 2019). The transparency of negotiations in the Council should make national representatives more prone to defend proposals that are important to the domestic public, possibly engaging in public posturing, even though this may come at reputational costs and risk the delay or breakdown of negotiations (Stasavage 2004). However, there is no regular media reporting on the deliberations themselves and only the videos of the public sessions are made available on the Council's website (i.e., no verbatim protocols). This renders the Council an ideal case to investigate international negotiations. In contrast, many public forums in other international organizations were designed as public-facing forums from the outset, leading to the development of "theatrical practices" for public consumption while actual negotiations are moved to informal, secretive forums (e.g., the UN Security Council). Moreover, systematic collections of speech data are generally rare in international relations. The UN General Assembly corpus is an exception, but the statements in the General Assembly resemble annual legislative state-of-theunion addresses in domestic politics and do not reflect intergovernmental negotiations (Baturo, Dasandi, and Mikhaylov 2017).
A large body of literature in international relations has addressed the question of whether public opinion influences negotiations. Although one strand assesses whether observable bargaining outcomes suggest that negotiations were influenced by public opinion (for a recent summary, see Frieden and Walter 2019), another focuses on whether government positions during negotiations are reflective of, or responsive to, public opinion. However, in both strands the influence of public opinion on negotiations is not a settled issue. Some scholars have found that constraining or salient public opinion at home is associated with more bargaining success in international negotiations (Baerg and Hallerberg 2016;Mariano and Schneider 2022;Wratil 2019), but others could not detect any effect (Bailer 2006;Lundgren et al. 2019). Similarly, some studies show that governments adjust their substantive bargaining positions to domestic (policy-specific) public opinion (Hobolt and Wratil 2020;Schneider 2018;Wratil 2018), yet others find no such relationship Thomson 2011). Moreover, some evidence that is regularly thought to show that public opinion matters in international negotiations does not measure opinion directly. Instead, it shows that governments' success in the international arena and their ability to prevent negative outcomes (e.g., conditions of IMF loans or WTO disputes) increase as elections draw close, assuming that governments can use the domestic public as a persuasive argument vis-à-vis international counterparts in such situations (Chaudoin 2014;Rickard and Caraway 2014;Schneider 2018). Analyzing bargaining rhetoric as opposed to bargaining outcomes or government positions provides an original test of whether public opinion influences negotiations.
Moreover, the existing literature, especially that on bargaining outcomes, provides little evidence of the exact mechanisms through which public opinion matters. Instead, studies make varied assumptions about how public opinion could manifest during negotiations. The influential two-level games perspective (Putnam 1988;Schelling 1960) argues that negotiators could benefit from explicitly drawing attention to domestic constraints-such as public opinion-if it credibly binds their hands, asking for special concessions from negotiation partners to accommodate their situation. In the words of Putnam (1988, 440), "[l]amenting the domestic constraints under which one must operate is […] 'the natural thing to say at the beginning of a tough negotiation.'" 3 Similarly, Lundgren et al. (2019, 70) argue for public opinion being explicitly mentioned at the bargaining table: "Constrained governments can make it known to their counterparts that they are under domestic pressure and threaten a collapse of negotiations unless others accept their demands." In contrast, others view public opinion as a force in the background that tacitly influences governments' position-taking and logrolling behavior, remaining hidden to observers, or take no side on whether public opinion will be actively discussed among governments (Baerg and Hallerberg 2016;Wratil 2018;2019). Thus, it remains an open question whether public opinion makes governments speak explicitly about it, formulate additional demands, refuse to compromise, or remain silent.
In summary, we still have a very rudimentary understanding of whether and how public opinion matters in international negotiations. Existing work on bargaining outcomes and position taking reports inconclusive results on the influence of public opinion and offers varied and rather undetermined expectations for rhetoric (see Appendix G for a full discussion). A wide range of possible images of the ways through which public opinion affects international negotiations remain-from being a topic for discussion, being explicitly used to back one's argument, to only implicitly influencing strategies and positions or even having no effect. No work has opened the blackbox of negotiation talk to look at the microfoundations of the public's role. We seek to ascertain how public opinion is reflected in the arguments governments make when it matters-at the negotiation table.

NEGOTIATION RHETORIC, PUBLIC OPINION, AND THE PRO-ANTI EU DIVIDE IN THE COUNCIL
To study the role of public opinion in bargaining rhetoric, we introduce a conceptual distinction between explicit and implicit manifestations of public opinion. As an explicit manifestation of public opinion, we understand negotiation rhetoric or even discussions that use words directly mentioning the public and its views or raising it through synonyms (e.g., "taxpayers," "voters"). In turn, an implicit manifestation is a negotiation-relevant, rhetorical element that does not mention the public but is motivated by unexpressed concerns about the public (such as governments' willingness to compromise or their negotiation demands). Note that government rhetoric could contain only explicit, only implicit, explicit and implicit, or neither manifestation of public opinion. 4 We develop expectations about the conditions under which public opinion becomes manifest in Council deliberations based on the idea that governments use public deliberations to find a policy coalition of member states that can pass a legislative proposal over the decision threshold. Importantly, although we conceive of coalition building as issuespecific (e.g., the composition of coalitions may vary on different legislative proposals), we argue that government ideology on European integration increasingly structures bargaining and can serve as a proxy for how governments form coalitions. Euroskeptic governments that emerged during the last two decades (De Vries and Hobolt 2020;Hooghe and Marks 2018) and portray themselves as defenders of national authority and identity (Hodson and Puetter 2019) oppose rather than support wide-ranging European regulations and programs. In contrast, the core of about 20þ governments made up of pro-EU parties support integrationist legislation in general, despite existing policy differences among them.
The salient difference between these two types is their baseline probability of belonging to the policy coalition that will pass a legislative proposal. The pro-EU core will try to form a coalition among themselves that meets the qualified majority threshold (in most cases), whereas Euroskeptic governments will often be excluded (or exclude themselves) from these coalitions. There is ample evidence for this assumption. In Appendix H, we show that stronger Euroskepticism among the government parties is associated with less bargaining success for the government when comparing negotiation outcomes to the government's initial positions as reported by experts as well as with more opposition votes cast by the government against legislative proposals. Both effects have increased in magnitude over time. This suggests that the pro-EU core increasingly forms the centre of gravity in coalition building during the actual negotiations, whereas Euroskeptic governments are often excluded. This coalitionbuilding structure shapes each government type's incentives of how to engage with public opinion at home in their bargaining rhetoric.

Explicit Manifestations of Public Opinion in Negotiation Rhetoric
Governments may talk explicitly about voters, the public, citizens, or taxpayers and their views. We present two theoretically competing reasons why governments may want to do so. First, if public opinion acts as a domestic constraint, actively and explicitly reminding negotiation partners of this constraint may serve to extract concessions from them Putnam 1988;Schelling 1960). This should primarily be a strategy of pro-EU governments cross-pressured by negative public opinion at home. They can point to this negative opinion as a factor constraining their ability to enter an agreement at the EU level. In line with this argument, Mariano and Schneider (2022) show that pro-EU governments are more likely to achieve successful outcomes in negotiations when they face a public negatively disposed toward the EU. In contrast, governments facing public opinion in line with their own preferences (i.e., pro-EU with positive or Euroskeptic with negative opinion) are not credibly constrained by this opinion, and Euroskeptic governments cross-pressured by positive opinion cannot use this as a constraint vis-à-vis the pro-EU core, for the public as a domestic player is more aligned with the partners' than the government's position. This suggests: H1a Public as Domestic Constraint: Pro-EU governments will raise the topic of public opinion mostly if the public is negatively disposed toward the EU, whereas Euroskeptic governments are unresponsive to the public's stance in this respect.
Second, governments may talk about domestic public opinion to back an argument about the legitimacy of their negotiation positions. From sociological institutionalist perspectives (similar arguments could be built from a constructivist viewpoint, see Risse 2000), governments may outline their position and, to support it and convince others of it, point to public opinion at home that is in line with their stance. For such an appeal to have an effect, the other partners must accept the representation of public opinion as part of a standard of legitimacy-a set of shared values and norms that defines what kinds of behaviors and actions are considered rightful and proper (see Schimmelfennig 2001). Given the member states' long-standing concerns about the democratic deficit of EU politics (see Rittberger 2005), we would expect that referring to citizens' views is seen by governments as a particularly persuasive argument for one's position in light of shared democratic values (e.g., "government for the people"). However, one impediment to the power of this form of "rhetorical action" is the credibility of the argument. If pro-EU (Euroskeptic) governments point to positive (negative) public opinion, this appears consistent and credible. However, in the sociological institutionalist view, the argument will be much weaker if the sender's reputation is inconsistent with the content of the message (e.g., if Euroskeptic governments refer to pro-EU opinion to back their position; Schimmelfennig 2001). Thus, we would not expect cross-pressured governments to refer to voters. Instead, we expect: H1b Public Legitimacy: Pro-EU governments will raise the topic of public opinion mostly if the public is positively disposed toward the EU, whereas Euroskeptic governments do so when public opinion is negatively disposed.

Implicit Manifestations of Public Opinion in Negotiation Rhetoric
Beyond explicit mentions of public opinion, there are more subtle ways in which the public's stance on the EU may matter. First, governments may respond in their substantive policy positions to public opinion to forestall the chance that their behavior in the Council may be made a matter of public discussion by media or interest groups (e.g., Wratil 2018). Governments may not need to make their representation of public opinion explicit but can just implicitly adjust positions. In their general negotiation rhetoric, this may figure as signaling a higher or lower willingness to compromise. To build a policy coalition in the course of the deliberations, policy controversies must be overcome (e.g., Thomson 2011). Much work finds that these compromises are enabled by concessions from many or all sides, including from large countries, rather than only from small or seemingly weaker partners (e.g., Arregui and Thomson 2009;Cross 2013;Golub 2012). We expect pro-EU governments to be likely to belong to the policy coalition and endorse a compromise. Yet, if faced with negative opinion at home, they may be less willing to embrace a compromise to avoid sending a signal of concessions. In turn, Euroskeptic governments are generally unlikely to belong to the policy coalition, as they tend to oppose new EU-level legislation. But if they face a public that is positively disposed toward the EU, they may consider embracing a compromise, as voters may value their responsiveness to them, especially if governments can sell the supported EU-level policy compromise as a bargaining success. Thus, both types of governments should marginally reflect public opinion in their willingness to compromise (although, on average, pro-EU governments may be more willing to compromise): H2 Compromise: Governments will express more willingness to compromise in their speeches the more the public is positively disposed toward the EU.
A second implicit way in which public opinion may matter concerns pushes for delay. Legislative proposals by the European Commission regularly get stuck in the Council, as there is no time limit to adopt a common position on the Commission's proposal. Some proposals are only delayed for a few months or a year, whereas others are delayed for years or are eventually withdrawn by the Commission (Boranbay-Akan, König, and Osnabrügge 2017; Kleine and Minaudier 2019). We argue that pro-EU governments who are cross-pressured by negative opinion vis-à-vis the EU at home have an incentive to actively delay decision making to shift the potential electoral risks of the opposing domestic public into the future. Governments may not want to reveal explicitly that their reason for delay is public opinion but justify it with other factors. Delaying may be sufficient to pass a period of particularly strong hostile public opinion (potentially aggravated by electoral contests) before news of any agreement the government has supported at the EU level could reach anti-EU voters through interest groups or media (cf. Kleine and Minaudier 2019;Schneider 2018). For governments in the pro-EU core, delay is arguably preferable to outright opposition, as they themselves may be ideologically aligned with the proposal advanced by the pro-EU core. Asking for a delay keeps them in the policy coalition and sustains their influence. Conversely, if public opinion is favorable toward the EU, pro-EU governments should speed up and quickly conclude negotiations to minimize any risks of future negative swings in public opinion. In contrast, the incentives to delay or accelerate agreement are less clear for Euroskeptic governments, who do not regularly belong to the policy coalition. If they are not part of the coalition, they may invariably try to delay EU legislation, irrespective of public opinion. If they aspire to become part of the policy coalition-for example, due to pro-EU opinion at home-they may still benefit from delaying tactics to increase their bargaining leverage. We expect: H3 Delay: Pro-EU governments will argue for more negotiation delays in their speeches when the public is negatively disposed toward the EU.
Third, the formulation of negotiation demands may be influenced by public opinion. A very general expectation about intergovernmental bargaining is that those governments that are indifferent between being inside or outside a policy coalition can formulate additional demands to the pro-EU core to remain in the coalition or join it (Moravcsik 1998), especially if they are needed to pass a deal. Public opinion can make governments indifferent about their membership in the policy coalition if they are cross-pressured. This is most obvious for pro-EU governments facing negative opinion on the EU at home. They should formulate additional demands to the remaining pro-EU core to keep them on board in the policy coalition, without necessarily mentioning public opinion. 5 Crosspressured pro-EU governments represent the first set of potentially pivotal players that have to be accommodated for the pro-EU core to pass its compromise (see also Mariano and Schneider 2022). But in some situations, Euroskeptic governments facing positive opinion on the EU at home may formulate demands to be brought on board of the policy coalition. In circumstances when the pro-EU core struggles to gather a sufficient majority among itself, it may rely on crosspressured Euroskeptic governments. In turn, governments that are not cross-pressured should be most likely to make concessions: H4 Demands: Pro-EU governments will formulate more demands in their speeches when the public is negatively disposed toward the EU, whereas Euroskeptic governments will do so if public opinion is positively disposed. Table 1 summarizes our hypotheses. It demonstrates that our expectations relate to critical aspects of intergovernmental negotiations-namely coalition formation, the speed of negotiations, when and who raises demands versus makes concessions, and whether public opinion is an explicit topic of discussion. For each aspect, we expect governments' rhetoric to reflect public opinion.

DATA
To study governments' rhetoric during public deliberations, we follow the "Debates in the Council of the European Union" data collection approach (Wratil and Hobolt 2019) and make use of transcriptions of videos of the Council's public deliberations. 6 This data source has been demonstrated to have high face, convergent, and predictive validity (see Appendix F.1 for a summary). Although videos of Council deliberations are made public on a designated video streaming website, including audio tracks of the simultaneous interpretations into the union's 23 official languages, the Council does not provide written records or transcriptions of its deliberations. To overcome this challenge, we rely on automatic speech recognition systems (see Proksch, Wratil, and Wäckerle 2019) and build the first comprehensive database of public Council deliberations. Our data contain transcriptions, obtained using automatic captioning of Google's YouTube video portal, of virtually all speeches by national representatives (mostly ministers and permanent representatives) in five Council configurations (the Council meets in 10 different "configurations" depending on the subject matter) over a period of up to seven years between 2010 and 2016. 7 We study speeches delivered in the following Council configurations: (a) Competitiveness; (b) Employment, Social Policy, Health and Consumer Affairs; (c) Economic and Financial Affairs; (d) Environment, and (e) Justice and Home Affairs. We deliberately exclude the Agriculture and Fisheries configuration, as EU competence in this policy area is very well established and prior work has shown that the influence of pro/anti-EU opinion mainly applies to areas where legislation extends EU authority (Hagemann, Hobolt, and Wratil 2017). Moreover, we selected areas in which legislative proposals are connected to typical societal left-right/libertarian-authoritarian conflicts at the domestic level in European democracies (Bakker, Jolly,  and Polk 2012), as we expect these areas to be particularly salient for voters and governments. Thus, our results may not apply to negotiations about legislation in less politicized policy domains. Our data cover debates on prominent legislative files such as the EU's 80 billion euro research funding programme "Horizon 2020," the posted workers directive, the banking union, the greenhouse gas emissions trading system, or the general data protection regulation (see Appendix F.2 for a full list of topics covered).
In total, 397 debates with at least two speech participations by national governments (excluding the Council presidency) were held across the five configurations during the period covered by our data. We only retain speeches by national governments (e.g., we discard the Commission) as well as remove all speeches by the Council presidencies, as they mainly perform the role of moderator and their speech style considerably differs from that of the other delegations. Moreover, as we are interested in actual legislative negotiations and not in debates on "work programmes" or some other issues (which are likewise public, see Appendix A.1), we subset our dataset to debates that relate to at least one completed legislative procedure on which a final vote had been reported. This leaves us with a corpus of 224 debates comprising 3,631 speeches of government delegates with a mean length of 267 words. On average, about 15 national governments participated in any given debate. Participation levels vary between governments, with limited inequality. Out of the 224 debates, the maximum number of debates (75) were held in the Justice and Home Affairs configuration and the minimum number (22) in the Employment, Social Policy, Health and Consumer Affairs configuration (see Appendices A and F for details on the dataset).

Independent Variables
To measure public opinion on the EU in each member state, we use the "EU image" question from the Standard Eurobarometer survey series conducted by the European Commission twice a year. This question asks respondents whether they have a positive or negative image of the EU: "In general, does the EU conjure up for you a very positive, fairly positive, neutral, fairly negative or very negative image?" It has been used in previous work to measure public support for EU integration (e.g., Hix 2018; Rauh, Bes, and Schoonvelde 2020) and is available on a semiannual basis for our time frame. 8 Specifically, we use the mean response to this question in each country using poststratification weights, with higher values indicating a more positive image of the EU ("5" for "very positive" and "1" for "very negative"). To cover any day during our period, we linearly interpolate this measure between survey fieldwork start dates. Moreover, in line with common practice (e.g., Hagemann, Hobolt, and Wratil 2017; Wratil 2018), we use a six-month lag of this measure to reflect that governments first observe public opinion and then adjust their behavior to it.
To classify governments into pro-EU versus Euroskeptic governments, we rely on the Manifesto Project's (Volkens et al. 2019) coding of the government parties' election manifestos, focusing on the difference between the percentages of positive and negative quasi-sentences on EU integration in the manifestos. 9 Specifically, we use a dummy variable for Euroskeptic governments that is "1" if the seat-weighted position of government parties on the EU is negative or zero (i.e., per108 -per110 ≤ 0), and "0" otherwise. This classification of governments' pro-anti EU ideology has high face validity (see Appendix A.2). For instance, the British Cameron and May cabinets and the Hungarian Orban III cabinet are coded as Euroskeptic according to the manifesto data. 10 Figure 1 presents an overview of the data. Panel (a) plots the monthly number of Council meetings between 2010 and 2016 that we study in our analysis, whereas panel (b) shows the average government position on the EU in the Council from January 2004 to December 2017. It reveals a long-run trend since 2004 of the average Council member government becoming more Euroskeptic. This trend reversed shortly around 2013, but then continued again. Finally, panel (c) shows the average public image of the EU across member states, but separately for countries with pro-EU and Euroskeptic governments. Although the average public image remained relatively stable in countries with pro-EU governments, it experienced a significant dip in countries with Euroskeptic governments between the end of 2011 and 2014. But note that the standard deviation of the public image was quite similar during our observation period (0.243) compared with the periods before (0.228) and after (0.217; see also the similarity of distributions in Figure A2 in Appendix A.2). Nevertheless, one concern with our sample is that the period covered largely coincides with the sovereign debt crisis in the eurozone, providing for a particular political context in which decisions were taken in the EU. We address this concern in the conclusion.

Control Variables
We control for various factors that account for alternative explanations of governments' rhetoric during negotiations derived from the Council literature. First, a significant part of the literature on Council politics has argued that negotiations between governments are 8 Note that the alternative question about support for EU membership has not been frequently measured during our observation period. 9 Our information on cabinet composition is primarily taken from the ParlGov dataset (Döring and Manow 2019). 10 In Appendix J.2 we demonstrate the robustness of our main findings to an alternative definition of Euroskeptic governments, incorporating information from the Chapel Hill Expert Survey on parties (Jolly et al. 2022). We dichotomize the measure of governments' Euroskepticism for two reasons: existing research points to the robustness of a binary measure (Proksch and Lo 2012) and a binary measure facilitates the interaction analysis within the structural topic model. strongly shaped by economic factors. This strand of work identifies a redistribution cleavage between member states (e.g., Bailer, Mattila, and Schneider 2015;Zimmer, Schneider, and Dobbins 2005). We therefore include unemployment and inflation rates (from Eurostat) in all our analysis models. This takes account of the fact that public opinion and government rhetoric may both be moved by macroeconomic conditions. We also operationalize budgetary positions using countries' annual net receipts from the EU budget in percentage of national GDP.
Beyond economic explanations, some work has found that Council politics is characterized by a longstanding north-south divide (e.g., Thomson 2009). Given the spread of Euroskeptic public opinion in several member states in the south during the eurozone crisis, which falls squarely into our observation period, we control for whether a government is from the south, north, or center. Another strand of the literature has argued that coalition building in the Council increasingly unfolds along partisan lines (Hagemann and Hoyland 2008), which could lead to different rhetoric by left-wing governments as compared with right-wing governments, especially with respect to policy-specific rhetoric (e.g., talking about jobs versus growth). We therefore also control for government parties' leftright ideology, relying on the Manifesto Project's RILE left-right score. Specifically, we include the seatweighted RILE of the cabinet parties. For ease of interpretation, in the models below we z-score standardize all continuous variables. Last, we model structural characteristics of our data. First, we include fixed effects for our five Council configurations using dummy variables, as we expect word usage to vary systematically with the topics discussed in a configuration. Second, we include a dummy variable for whether a debate is on the EU budget (e.g., annual amending budgets), as such debates are known as particularly conflictual and their rhetoric may vary systematically. Third, we control for whether the debate is on an issue that requires unanimity voting in the Council, which may influence the rhetoric during negotiations (e.g., alienating partners is more costly). Details on the sources and definitions of all variables are in Appendix A.2. 11

METHODS AND RESULTS
We analyze the rhetoric of governments using structural topic models (STMs; Roberts et al. 2014). Structural topic models are Bayesian mixture models that assume a corpus of documents stems from a number of k latent topics, with k being a parameter set by the researcher. In the STM, the prevalence of topics can be influenced by covariates. This feature allows us to investigate whether certain topics are mentioned more often if public opinion toward the EU is positive or negative. Like standard regression models, STMs identify statistical associations. Any causal interpretation hinges on (further) assumptions, and we do not seek to make causal claims. Our main interest is in whether specific topics are mentioned at the bargaining table, given a particular state of public opinion. We grant that in one situation the government may have cued public opinion before the negotiations to use it vis-à-vis partners, whereas it may feel pressured by it in another. A primary advantage of unsupervised topic models for our application is that they address the issue that we have little a priori knowledge about our text form. Although we have formulated hypotheses about what topics ministers should emphasize under different states of public opinion, we do not know which specific aspects of the relevant constructs (e.g., delaying negotiations or formulating demands) ministers will stress and what words they will use to do so. The specific categories for text classification are to some extent unknown, which speaks against the use of dictionary or supervised learning methods (Grimmer and Stewart 2013). 12 The main challenge is to adequately identify and validate topics from the STM that relate to our hypotheses. We carefully engage in this task below.

Structural Topic Model: Estimation and Descriptives
We present results from an STM with k = 40 topics. We also demonstrate the robustness of our main findings with a series of further STMs in Appendix J (e.g., using 35 or 45 topics as well as alternative measures for the main variables and additional fixed effects for years). Before running the topic models, we divide each speech into a beginning, mid, and end part, representing 20, 60, and 20% of the words, respectively. 13 This allows us to capture diplomatic habits (e.g., congratulations) exchanged at the beginning or end of a speech. Our total number of "speech parts," representing our documents in the structural topic model, is N = 10,214. 14 We include not only single words (unigrams) but also all combinations of two consecutive words (bigrams) in our document-feature matrix. This ensures that we can capture frequent bigram expressions-like "European Parliament" or "member state"-in conjunction and not only through their constitutive words. In preprocessing, we lemmatize all words (i.e., group inflected forms of a word and variants) as well as remove stop words, numbers, punctuation, symbols, seperators, and words with less than three letters. Moreover, we exclude unigrams and bigrams that occur fewer than 10 times or in fewer than five documents.
We model topic prevalence as a function of our independent variables (public opinion and government position on the EU), their interaction term, and all control variables. In addition, we also control for the part of the speech (beginning, mid, end) as well as for a categorical variable that indicates the type of the debate (initial presentation, policy debate, debate on political agreement, mixed), which was coded by research assistants. We report estimates (including confidence intervals and standard errors) that have been obtained using a nonparametric bootstrapping procedure, which resamples random draws from the STM's variational posterior of the document-topic proportions at the country level to correct for the clustering of speeches within countries (see Appendix I.5 for details of this procedure).
We broadly classify the estimated topics into those that relate to a specific policy area and those that are overarching and address procedural aspects of negotiations. Policy-specific topics are defined by specific terms associated with particular policy areas or even legislative proposals-for example, "criminal," "tax," "biofuel," and "research" or even acronyms like "epp" (European Public Prosecutor), "ecb" (European Central Bank), or "dgs" (Deposit Guarantee Schemes). 15 In turn, procedural topics address the conduct of negotiations and do not pertain to specific policy areas. This includes diplomatic habits such as thanking the presidency or chair, talking about the legislative process 11 Note that we could also control for the size of a member state, assuming that more powerful member states speak differently. But most literature has found no effects of size on bargaining success in EU legislative negotiations (e.g., Arregui and Thomson 2009;Cross 2013;Golub 2012). Thus, we do not expect it to affect rhetoric significantly (also see the robustness check in Appendix K.6). 12 Also note that the STM will typically be more conservative than a dictionary approach in identifying certain constructs in a corpus, as a latent topic will only be captured if it is sufficiently frequent to belong to the k most prevalent topics. 13 Previous work has shown that speeches are often structured by these three functional parts (Wratil and Hobolt 2019). 14 For information on missing data see Appendix A.3. 15 For a discussion of transcription errors (e.g., "epp_oh" or "v80"), see Appendix I.
(e.g., negotiations with the European Parliament), negotiation tactics such as formulating a demand or expressing support for a compromise, and rhetorical styles (e.g., affirming a point, using cautious language). We label all topics and classify them as policy specific or procedural based on an investigation of the highestprobability words, the most frequent and exclusive (FREX) words, and the most likely texts for each topic. We find that 22 of the 40 topics are policy specific, whereas the remaining 18 are procedural. In Figures 2  and 3 we respectively display procedural and policyspecific topics with their grand mean proportions. We also group the policy-specific topics by the Council configuration in which they made up the largest share of speeches. In Appendices I.1 through I.3, we demonstrate the validity of the STM, including topic labels and classifications, in various ways. In particular, we show how the prevalence of policy-specific topics tracks the EU's legislative agenda and how procedural topics occur across policy domains.

Hypotheses Tests
Our hypotheses relate to expectations regarding the procedural topics. The identified procedural topics in Figure 2 correspond to primary aspects of how we envision intergovernmental bargaining. For instance, governments do talk about compromise ("Talking about reaching compromise," "Supporting the compromise"), address legal details of proposals ("Talking about legal text"), or exchange courtesy (e.g., three different topics on thanking and congratulating). In terms of our hypotheses, we find a topic that relates to "Supporting the compromise" (H2); three topics on "Delaying agreement," "More technical-level discussion needed," and "Cautious language" (H3); and two topics targeting "Formulating a demand" or "Raising a concern" (H4). This confirms that-as expected-these themes make up for a significant part of governments' rhetoric in deliberations. In contrast, we find no procedural topic that relates directly to public opinion, people, citizens, or voters (H1a/b). 16 Governments do not seem to explicitly frame a topic around public opinion during negotiations, as we would expect them to do to highlight a domestic constraint or appeal to shared democratic norms. The lack of such a topic in the STM reveals that governments rarely consistently talk about public opinion during negotiations, which is evidence against H1a/b. In the following, we therefore focus on testing H2, H3, and H4, and we revisit H1a/b in the next section. We report the main results here; details of the STM regression results and hypotheses tests are in Appendix B.1, and scenarios with expected topic proportions are included in Appendix I.6.
To test hypothesis H2, we focus on the topic "Supporting the compromise." Fittingly, "support" is the highest-probability word for this topic and the FREX words include "general_approach," "partial_general," and "text." In fact, "(partial) general approach" is the term used in the Council to refer to a (partial) compromise proposal that lays out the essential contours for a political agreement. If a general approach is adopted, the Council has reached a compromise on the most important controversies with a coalition of sufficient  size supporting the text. The most strongly associated text for this topic is: 17 "On the proposed text, we fully support its goals and we fully endorse the general approach on the basis of this text without…" 18 In turn, the topic "Talking about compromise" relates to coalition building but not to the willingness to compromise; it often includes calls by governments on peers to "reach" an "agreement" "today" or more general talk about the compromise. We therefore do not consider it for H2. Figure 4 plots the estimated effect of public opinion on a topic's prevalence for all hypotheses. Specifically, as we standardize public opinion, the estimates show a change in topic proportion for a standarddeviation change in public opinion. A positive coefficient means that a more positive image of the EU in the member state increases the attention to this topic by the government in the Council, whereas a negative coefficient implies a decrease in attention. Given our interaction effect between government ideology and public opinion, we display estimates separately for pro-EU and Euroskeptic governments. We find partial support for H2. Both types of governments  express support for the compromise more often during public deliberations when public opinion at home is more positively disposed toward EU integration. However, we can only reject the null hypothesis for pro-EU governments (p < 0.01), whereas the confidence interval for Euroskeptic governments is very wide. When comparing two scenarios in which public opinion is at the 10th versus 90th percentiles, a pro-EU government would spend around 5.1% of its speech talking about support of the compromise if public opinion is skeptical, but this attention would increase to 6.0% if the public is strongly pro-EU, an increase in topic importance of almost 20%. This 0.9 percentage-point increase is sizable when considering that in a model with 40 topics the average topic will make up about 2.5%. We test H3 with three topics: calling for more technical-level discussions, delaying agreement, and cautious language. First, by emphasizing "More technical-level discussion needed" governments stress that a "discussion" has or has not gone "far" but that there are "still" unresolved "technical" "issue(s)." Although these are the top FREX words in this topic, the following also include "technical_level," "work_group," "expert_level," and "work_party" (see Appendix I.1), which all indicate that this topic is about shifting discussions to the technical level of working groups, resulting in a prolongation of negotiations. For instance, one of the top five associated texts is as follows: "But we have to be able to clarify the unclear questions in the future. The breakthrough so far is due to the work of the Polish presidency and therefore we propose that the remaining questions should be examined and studied at the expert group level. Thank you very much!" 19 Second, "Delaying agreement" represents the most explicit form of delaying tactic we observe in our corpus. For instance, an excerpt of one of the most highly associated texts with this topic is: "[Finland has] not yet officially finalized its national position on chapter four. As our president also said in his introduction, Effect of public image on expected topic proportion Effect of public image on expected topic proportion Note: Estimates reflect change in expected topic proportion for standard deviation change in public image; 95% confidence intervals as solid lines.
nothing is agreed until everything is agreed. Thank you." 20 Third, the "Cautious language" topic is implicitly related to delaying tactics as it captures the use of language that signals uncertainty, indecisiveness, or ambiguity. Governments say that they need to "think" more about issues or take another "look" at them. Among the top 20 FREX words are also terms like "perhaps," "feel," "carefully," "right_direction," or "idea," all of which signal (sometimes in polite ways) that a government does not want to take a decision at this moment (see Appendix I.1). The second-most associated text illustrates the thrust of the topic clearly: "…don't think we can work toward the one-stop shop along the lines that we have been working on so far and that's why I replied to the questions in the way I did. I think if we take the time to look at it carefully and look at the two legal advices next to each other, as suggested by the UK, I think that would be a very good idea and that would mean that we would be able to come back to this issue under the Greek presidency. Thank you!" 21 We did not identify any topic that clearly relates to the speeding up of negotiations. Figure 4 shows strong evidence that pro-EU governments make more use of delaying tactics if public opinion at home is negatively disposed toward the EU. For all three topics, the effect of public opinion on the topic's proportion is negative and statistically significant. In terms of substantive magnitude of these effects, we can combine the estimated proportions of all three topics. In a scenario, in which public opinion is positively disposed toward the EU (at the 90th percentile), pro-EU governments would spend about 10.8% of their speaking time on delaying tactics. In contrast, if public opinion is negatively disposed toward the EU (at the 10th percentile), this figure would increase to about 12.7%. This suggests that pro-EU governments try to delay agreement at the international level to cope with anti-EU opinion at home. In contrast, our results for Euroskeptic governments are inconclusive, as all effects are far from statistical significance. As expected, Euroskeptic governments call for a delay of agreement irrespective of public opinion at home-the effect is essentially zero. On the other two delaying tactics, the effects of public opinion point in opposite directions: Euroskeptic governments demand more technical-level work when public opinion is anti-EU but are also less cautious in their language. However, none of these effects are statistically significant.
For H4, we identified the topics "Raising a concern" and "Formulating a demand" as relevant. In both of these topics, governments point to certain aspects of negotiation substance (e.g., details of a proposal) that are particularly salient to them. However, in the "Formulating a demand" topic they often say that the Council or the presidency has to "take account" of a situation or an argument, whereas in the "Raising a concern" topic they explicitly label their point as a "concern" and sometimes also highlight that this concern is "important." From reading the top texts associated with both topics, the main difference is that the "Raising a concern" topic is more negative, implicit, and passive in tone, whereas the "Formulating a demand" topic is less negative but also more explicit and demanding in tone. Both topics clearly relate to the focus of H4 on governments raising demands vis-à-vis their partners. In contrast, we find no topic relating to governments making a concession-a move that speakers might find less pleasant to make explicit in their rhetoric. Figure 4 reveals little support for H4. In line with the theoretical expectation, pro-EU governments crosspressured by anti-EU opinion use the "Raising a concern" topic more often (p = 0.013), potentially using increased bargaining power due to their domestic constraint. However, in stark contrast, pro-EU governments actually decrease their usage of the "Formulating a demand" topic if public opinion is more anti-EU, even though this effect is not statistically significant (p = 0.252). When summing the proportions for both topics, the formulation of negotiation demands by pro-EU governments is independent of the state of public opinion at home (around 4.2% at the 90th and 10th percentiles). Similarly, we find no significant adjustment of negotiation demands by Euroskeptic governments, with the marginal effects of public opinion on both topic proportions being indistinguishable from zero. Governments do not seem to consistently formulate additional negotiation demands if they are cross-pressured by public opinion at home.
In Appendix K, we also investigate various other substantive expectations about how government rhetoric could be influenced by public opinion. 22 One important nuance we discover is that the public opinion effects on the "Supporting the compromise" and "Delaying agreement" topics are a bit stronger during final debates on political agreement as compared with earlier negotiation stages (see Appendix K.4). Both topics are relatively strong signals about governments' willingness to join the policy coalition when compared with topics such as "Cautious language." This could suggest that public opinion may matter more for clear commitments at the end rather than the beginning of negotiations. 20 Finland on October 10, 2014, in the 3336th Council meeting (Ecofin) on the general data protection regulation (2012/0011/COD). 21 Netherlands on December 6, 2013, in the 3279th Council meeting (JHA) on the general data protection regulation (2012/0011/COD). 22 One prominent argument pertains to whether pending elections influence the effect of public opinion. In Appendix D we demonstrate that governments facing elections at home are selecting themselves out of debates and speak less often in the Council. In Appendix K.2 we also show that pending elections consequently do not influence the topics governments talk about-they mainly silence them. Appendix L demonstrates that also the sentiment with which governments speak does not change before elections. Moreover, we investigate nonlinear effects of public opinion and the influence of negotiation stages, debate length, and government bond yields (see Appendix K). All these analyses support our main results.

DISCUSSION
Two of our results warrant deeper discussion and investigation. First, with respect to speaking about public opinion (H1a and H1b), we do not find an estimated topic that directly relates to public opinion as a major theme during negotiations. The absence of this topic is also replicated with a higher number of topics (see Appendix J.4). However, structural topic models do not allow us to rule out that governments raise the issue of public opinion parenthetically or on rare occasions. To assess whether governments mention public opinion at all, we code all mentions of the public in our corpus and analyze in Appendix C whether they occur as stipulated by H1a and H1b. This analysis yields limited evidence that when Euroskeptic governments mention the public, they rather do so when public opinion on the EU is negative, which is closer to H1b than to H1a. But we find no indication that pro-EU governments mention the public more often when it has a positive image of the EU to appeal to shared democratic norms.
Second, our analysis does not provide evidence that Euroskeptic governments adjust their negotiation rhetoric implicitly to domestic public opinion. Although this may partially be due to the small number of Euroskeptic governments, it might also be that Euroskeptic governments' representatives in the Council refrain from making statements in the first place when public opinion in their own country is positively disposed toward the EU. We examine this possibility by investigating whether member states make use of their speaking opportunities in Council meetings. This variable is coded "1" when a government delivers a speech in a given meeting and "0" otherwise. As independent variables we include the government's EU ideology, the domestic public image of the EU, their interaction, and a number of covariates. Figure 5a shows the predicted probability of speaking in a Council debate based on the results from a mixed effects logistic regression model with random intercepts for member states and debates (details are in Appendix D). Pro-EU governments have a high probability of speaking (around 0.64) that is independent of whether the public image is positive or negative. In contrast, Euroskeptic governments are significantly more likely to participate in debates when the public image of the EU is negative. The probability drops by half when the public image is positive. Thus, rather than changing the content of their speeches, Euroskeptic governments simply speak less frequently in the Council when they face a public with a positive image of the EU. Publics approving of the EU appear to silence Euroskeptic governments in Council debates.
Euroskeptic governments may be more likely to signal responsiveness in other ways, most notably in the final act of voting, which has been found to be used to signal to domestic audiences (Hagemann, Hobolt, and Wratil 2017;Schneider 2018). Figure 5b shows that Euroskeptic governments are clearly responsive to public opinion in final passage voting on the legislative acts covered in our data, casting more opposition votes when the public has a more negative image of the EU and fewer opposition votes when the public has a positive image. Albeit the small number of Euroskeptic governments, this effect is statistically significant in a mixed effects model with random intercepts for member states and legislative acts as well as our standard set of controls (see Appendix E for full results). As Euroskeptic governments usually do not belong to the policy coalition, they can use their votes as a signalling tool. In contrast, pro-EU governments in the policy coalition are expected to follow through with their support in voting and cast a "yes" vote, which diminishes their possibilities of responding to public opinion. Fittingly, we do not find a significant effect of public opinion on their voting behavior in our sample.
In summary, our findings suggest that the representation of public opinion in government rhetoric mainly occurs for pro-EU governments. Euroskeptic governments mostly participate in debates when supported by a public with a negative image of the EU. In such situations, these governments also use the opportunity to cast opposition votes against EU laws.

CONCLUSION
Our study opened the blackbox of international negotiations by focusing on governments' bargaining rhetoric. This allowed us to investigate different ways and mechanisms through which public opinion influences international negotiations-how it manifests itself in procedural moves such as offering compromises, formulating demands, or calling for a delay of negotiations. We thereby complement existing studies focusing on outcomes or bargaining positions that could not investigate how exactly public opinion matters during the negotiation process.
Our results show that public opinion matters for negotiation rhetoric, but largely in implicit ways. We find little explicit discussion about public opinion, justification of positions by reference to citizens' needs, or invocation of democratic norms of responsiveness. Public opinion remains a force in the background influencing, in particular, when governments express more willingness to compromise as opposed to remaining cautious and stalling the negotiations. Thereby, we demonstrate how coalition building in the Council is constrained and propelled by domestic public opinion.
In addition, we also uncover striking differences in the engagement with public opinion between Euroskeptic and pro-EU governments. Whereas pro-EU governments implicitly bring their rhetoric in line with public opinion, Euroskeptic governments select themselves out of Council debates if the public is positively disposed toward the EU. They simply remain silent. In contrast, if the public's image of the EU is negative, Euroskeptic governments raise their voice in the Council and become the most likely governments to explicitly talk about the public-but this only happens in a small fraction of speeches. Nevertheless, it suggests that Euroskeptic governments are most likely to explicitly represent their public. Christopher Wratil, Jens Wäckerle, and Sven-Oliver Proksch Our distinction between explicit and implicit manifestations of public opinion and our findings in this respect have important implications and provide impulses for future research. First, the distinction draws attention to the fact that when empirically analyzing negotiation rhetoric not everything may be explicit, but we have to look behind the rhetorical elements employed by negotiators and examine in what context and why they are used. Second, the distinction can serve future work to study more bargaining-related aspects empirically. The following are some open questions: Can governments' use of explicit references to public opinion ever increase their bargaining leverage? Does a dominance of implicit over explicit manifestations lead to misunderstandings and miscalculations between negotiation partners, potentially resulting in negotiation delays or breakdowns? Addressing such questions demands linking rhetoric back to outcomes and bargaining positions.
Third, the distinction also denotes two ways of how decision makers in representative institutions can try to create democratic legitimacy. Our results demonstrate that the Council (and potentially many intergovernmental forums in world politics) mostly engages in the implicit way of legitimizing itself vis-à-vis citizens. Despite the EU member states' declared goal of  Table D1 in Appendix D. The estimates and confidence intervals in panel (b) are based on the model results "Baseline Model" presented in Table E2 in Appendix E.
bringing the EU closer to its citizens during several treaty reforms throughout the last two decades that eventually led to the establishment of a de jure standard, bicameral legislature, Council politics still lacks the rhetorical "grandstanding" on behalf of and argumentative references to citizens' interests that is characteristic of national parliaments. Therefore, our study confirms a critical part of the argument that the EU suffers from a "democratic deficit" (Føllesdal and Hix 2006)-namely, EU politics' technocratic nature and detachment from citizens, who are rarely mentioned. Future work could investigate whether the predominance of implicit over explicit representation has consequences for citizens' legitimacy beliefs in the EU.
Fourth, the fact that we find little evidence for explicit manifestations has important implications for how theories of international negotiations should view the role of public opinion. Foremost, it challenges arguments based on sociological institutionalist or constructivist thought in international relations (e.g., Krebs and Jackson 2007;Risse 2000;Rittberger 2005; Schimmelfennig 2001) that necessarily expect governments to make public opinion a matter of explicit discussion to appeal to shared democratic norms or persuade others. These accounts focus on what is actually said between negotiators, and if public opinion plays no explicit role in rhetoric, its role is overall of less interest. Moreover, although this result does not fundamentally challenge theories of public opinion as a domestic constraint (e.g., Frieden and Walter 2019;Mariano and Schneider 2022;Putnam 1988;Schelling 1960), it suggests that the effect of such constraints on bargaining power seems to occur either tacitly with little need to emphasize constraintsfor instance, because everyone knows about and accepts them-or because constrained governments simply invest more bargaining resources on issues where they face a constraint. In both cases, public opinion will affect bargaining power but will not become an issue of explicit discussion.
Our study has a number of limitations. First, we focus on the ministerial Council as the highest layer in the Council's structure and we cannot exclude that public opinion plays a different role at lower levels. If governments may fear audience costs from not being able to deliver their public commitments (Fearon 1994), there may be an incentive to shift explicit discussions about the public to more secluded levels, as one could lose out in negotiations at the end. But recent contributions suggest that inconsistency between an initial position and a final outcome is not punished by voters in the EU and trade policy contexts, providing governments with an unreserved incentive to always openly commit to fight for the public (Chaudoin 2014;Schneider 2018). Future work should address this question-for instance, through expert interviews. A second limitation of our study is that we cannot observe (but only try to infer) intentions, plans of actions, or unwritten codes governments use. This simply represents a general limitation of political speech data. Third, although we provide the largest corpus of Council speeches ever collected, our observation period falls squarely with the EU's eurozone sovereign debt and migration crises. This may potentially limit the generalizability of our findings to other periods with less turmoil. However, research suggests the effects of public opinion may be muted in times of economic crisis (e.g., Ezrow, Hellwig, and Fenzl 2020), indicating our time frame should not make us less confident about the implicit manifestations of public opinion. More generally, future studies should investigate whether the effects of public opinion on rhetoric persist when studying a longer period with more government alternations, and especially more Euroskeptic governments, than during the period investigated here. Finally, our text analysis approach does not allow us to track how earlier statements on a legislative proposal influence later statements on the same issue. This could illuminate, for instance, how demands made in response to a public constraint early on in negotiations influence the willingness to compromise later. Advances in dynamic topic modeling could help to answer such questions.
Ultimately, our study draws a nuanced picture of the representation of public opinion in international negotiations, suggesting public opinion matters but in different ways than often assumed.

SUPPLEMENTARY MATERIALS
To view supplementary material for this article, please visit http://doi.org/10.1017/S0003055422001198.

DATA AVAILABILITY STATEMENT
Research documentation and data that support the findings of this study are openly available at the American Political Science Review Dataverse: https://doi. org/10.7910/DVN/JCT3F7. reviewers for their excellent and very constructive suggestions. Previous versions of this study were pre-