Beyond expertise: the public construction of legitimacy for EU agencies

ABSTRACT How is the power of independent agencies legitimized? This is a central question in modern democratic societies. Earlier research has privileged technical expertise as the predominant source of legitimacy for such agencies. While recent contributions have challenged this assumption, we have seen few attempts to systematically analyze the conditions under which different sources of legitimacy are established in public discourse. We address this gap by proposing a conceptual framework of four legitimation arguments and test their prevalence through an empirical analysis of the public legitimation of EU agencies. We hypothesize that the prevalence of each argument depends on characteristics of the agency, especially its scientific ‘hardness’ and its public salience. We test our hypotheses in three steps. We first combine automated text classification and qualitative content analysis to analyze Swedish news media coverage of three EU agencies, 2005–2019. In a third step, we quantitatively analyze aggregated data on the Swedish news coverage of all EU agencies 2005–2019. We find more technical-expertise discourse in coverage of hard-science agencies, and more political-control discourse where agencies are ‘softer’ or more salient. Our findings are therefore relevant for ongoing normative and empirical discussions on the legitimacy of independent agencies.


Introduction
A steadily growing array of policy tasks in the European Union (EU) has been delegated to EU agencies. These are EU-level non-majoritarian institutions set up at an arm's length from elected officials to perform a specified task. They are created by secondary law, staffed by specialists or experts, and are not led by, nor under the direct command of, democratically elected representatives (Wonka and Rittberger 2010). A central issue is how such institutions can be legitimized, given their removal from majoritarian electoral politics.
A prominent argument in academic literature and EU official discourse is that independent expertise can legitimize EU agencies. Expertise has been described as the 'be-all-andend-all criterion for legitimacy in the EU regulatory state' (Busuioc and Rimkutė 2020b, 5). This perspective, however, overlooks the potential for domination that arises when unelected experts in EU agencies gain power at the expense of democratically elected politicians (Holst and Molander 2019). This is a pronounced risk particularly seen from normative approaches that regard democracy 'as aggregative, participatory, and intrinsically justified' (Christensen and Holst 2017). In contrast, a 'deliberative, elite, and epistemically justified' approach to democracy would accept expertization more easily, if not outright recommend it (Christensen and Holst 2017, 821). These contrasting ideas of democracy, and in particular the tension between epistemic and (majoritarian) democratic concerns (Krick 2021), may give rise to different arguments about a non-majoritarian institution's legitimacy.
In this paper, we explore the idea that different agencies may face systematically different legitimacy demands. In order to get a better understanding of the conditions under which different arguments about agencies' legitimacy come into play, we empirically examine legitimation arguments about EU agencies in the public sphere. While we have a fine-grained understanding of EU agencies' establishment, accountability, self-presentation, and reputation management strategies (Busuioc and Lodge 2016;Busuioc and Rimkutė 2020b;Rimkutė 2019;Wood 2018;Thatcher 2011;Majone 1997;Wonka and Rittberger 2010;Scholten 2014), we know less about how the legitimacy of EU agencies is negotiated, promoted and contested in the media. We treat legitimacy as a multidimensional concept, and ask: Which legitimation arguments about EU expert agencies are present in the public sphere, and under what conditions do we find each argument?
While there are several channels for debate in the public sphere, we take news media to be one of the most central. Our empirical strategy is a mixed-methods approach comprising three analyses of EU agencies' news coverage in Swedish news media from 2005 to 2019. In the first and second analyses, we focus on full-text news coverage of three EU agencies that give us high variance along our variables of interest-the European Border and Coast Guard Agency (Frontex), the European Environmental Agency (EEA) and the European Banking Authority (EBA). A quantitative text analysis is combined with a qualitative content analysis of news articles selected from the same corpus. The third stage is a keyword-based analysis of all 36 EU agencies 2005-2019.
This paper contributes to existing literature by taking a novel approach to the study of EU agencies' legitimacy. We present and operationalize a conceptual scheme that empirically examines four normative legitimation arguments. We find that technical expertise underpins (de-)legitimation claims only under certain circumstances. Where hard science evidence is lacking or public salience is high, political control is a more prominent category. Fundamental rights discourse is much more prominent in the coverage of Migration and Home Affairs agencies. The findings offer a starting point for a new direction for the study of EU legitimacy. The determinants of different legitimacy arguments warrant closer attention from several angles, both through empirical studies and normative theory.
Our paper proceeds as follows: We first discuss four conceptualizations of legitimacy (the evidence-based, legislators' command, public participation and fundamental rights arguments) that can serve as the basis for the public (de-)legitimation of independent expert bodies. We then formulate hypotheses towards their prevalence in the media of input, throughput, and output legitimacy (Scharpf 1999;Schmidt 2013). Input-oriented approaches focus on the popular-sovereignty component of democratic legitimacy, in essence that decisions should be responsive to the concerns of the constituency (Bellamy 2010, 5). Output-oriented approaches focus on the content of policies and whether they effectively provide for the common good (Schmidt 2020). Throughput is sometimes added as a procedural criterion, presented as an opening of the 'black box' between input and output (Schmidt 2013), concerned with the 'quality of governance processes as judged by (…) accountability (…), transparency, inclusiveness and openness' (Schmidt and Wood 2019, 728). The systems-theoretical framework offers a useful heuristic to systematize normative debates over legitimacy. Our concern is more specific: the discursive legitimation of non-majoritarian agencies. This requires us to go beyond the categories of in-/through-/output and focus on how the power of institutions are justified. We maintain, however, that such justifications should be understood as appeals to different aspects of democratic legitimacy.
As Jens Steffek (2015) points out, the systems-theoretical scheme is sometimes taken to imply an understanding of democracy where democratic legitimacy is equated with input, while output legitimacy simply requires that an institution provides efficient problemsolving as something entirely unrelated to democracy. We agree with Steffek that this is a reductive understanding. Instead, we take democracy to entail a commitment both to popular sovereignty and to the common good (Pettit 2004). Output legitimacy demands that an institution must be evaluated by its contribution to the common good. This is, by implication, an argument about its democratic legitimacy. For this argument to be persuasive, however, we need it to be grounded in something more than pure efficiency-it must also credibly provide a mechanism that ensures the common good. For instance, an independent central bank is legitimized by reference to its institutional independence and its economic expertise, which together will secure price stability-an outcome in the long-term public interest (Steffek 2015, 271). The same goes for independence: It seems unlikely that citizens would find a powerful institution without any clear expertise basis, accountability mechanisms, or participatory procedures legitimate simply because it is independent from electoral politics-even if it has generated efficient or 'good' (however defined) outcomes so far. In this scenario, we have neither reason to trust that such an institution will continue generating efficient outcomes in the future, nor do we have means to constrain it if it begins to perform poorly. In addition to independence, then, the power of a non-majoritarian institution must be justified with reference to some legitimizing argument (E. O. Eriksen 2021a, chaps. 1-2).
The systems-theoretical framework remains underspecified about what such legitimizing arguments might be. We aim to identify which dimensions of democratic legitimacy an agent might appeal to when justifying the power of a non-majoritarian institution. Some arguments are more in line with input-oriented approaches, some are more in line with output-oriented ones, but they are all geared towards the specific context of non-majoritarian institutions. They therefore presuppose some degree of institutional independence and insulation from electoral politics.

Four types of discursive legitimation
Building on central contributions in the normative and empirical literature (E. O. Eriksen 2021b;Bellamy 2010;Maggetti 2010;Majone 1997;Scharpf 2009;Buchanan 2011), we identify four main types of justificatory arguments for the legitimacy of non-majoritarian institutions: The evidence-based, the legislators' command, the public participation and the fundamental rights arguments. The first three categories are based on, and get their names from, the typology developed in E.O. Eriksen (2021b).
These arguments may be treated as empirically observable analytical categories derived from dominant normative ideas of legitimacy. They are not mutually exclusive nor necessarily competing. It is conceivable that different institutions have different legitimacy bases, and one institution may draw on different legitimacy arguments at different points in time. We hypothesize that the prevalence of any of these four types of justifications depends on two main variables: An agency's scientific 'hardness' and its public salience. The general idea of a 'hierarchy of the sciences' has been with us since the days of Auguste Comte. While there is some academic debate on the empirical tractability of different sciences' hardness (see Fanelli and Glänzel 2013;Hedges 1987), we are here concerned with perceived scientific hardness in the public eye, and not the metaphysical status of a given discipline. As Smith et al. (2000) show, public perceptions of the hardness of different sciences follow Comte's scheme quite well: Respondents on average ranked physics and chemistry at the top, biology, medicine and psychology a little lower, economics in the middle and sociology at the bottom. We define salience as the amount of attention given to an agency. This can for instance be measured by the number of media articles about an agency or the number of parliamentary questions mentioning it (Koop 2011;Wood 2018). We use both measures in turn and return to the exact specifications below.
As we have seen, there are several other ways to systematize EU agencies, for instance by their formal independence or their regulation, information, or decision-making powers. We do not suggest that such traits are less important than the salience of an agency or the hardness of its expertise. An analysis of relative importance would require a different research design, which, although interesting, is outside the scope of this study.

The evidence-based argument
In the literature on independent agencies, much attention has been given to agencies' role as independent expert bodies. The standard account is that the agencies' technical-scientific conduct-not their place in a 'chain of command' from electorates to legislators and governments-gives them their legitimacy (see e.g. Majone 1997;2001;Ossege 2015;Radaelli 1999). This argument is most clearly developed in the regulatory state literature (Majone 1994; and overlaps with notions of 'output legitimacy' and 'technocratic justification' (Boswell 2009;Scharpf 1999). According to the evidence-based approach, non-majoritarian institutions need not be responsive to politicians or the public, because their conduct is based on a validation procedure that ensures that objective truth (and ultimately a commitment to the common good) guides the agency's conduct. This validation procedure would typically be the scientific method, but other validation procedures-for instance risk assessment-would also fall under this category (Rothstein, Borraz, and Huber 2013).
The European Commission's own rhetoric encapsulates this idea. In a 2002 communication, the Commission writes: 'The main advantage of using the agencies is that their decisions are based on purely technical evaluations of very high quality and are not influenced by political or contingent considerations' (Commission of the European Communities 2002, 5). As Eriksen (2021c) points out, however, 'hardly any analyst agrees with this claim.' There is widespread agreement in the academic literature that an agency's technical assessments are often intertwined with political values.
Drawing on the above, we theorize that an appeal to science or technical expertise can be a powerful legitimizing device under certain circumstances. We hypothesize that two main factors determine the argument's prevalence. First, the agency's policy field should matter. It is reasonable to assume that appeals to objective evidence are more readily accepted in some fields than others. For instance, the regulation of medicines is predominantly based on randomized controlled trials-so-called 'gold standard' scientific evidence (Maor 2007). But border control, for instance, has no 'gold standard' to appeal to; the appeal to science might therefore be more difficult to accept in such a field (Fjørtoft 2020). Note that there is a difference between an agency being based on hard science and being discussed in evidence-based terms. We expect there to be a correlation, but we will test this expectation, not assume it. Second, the salience of an agency should matter. If a matter is highly salient or politicized, then-all else being equal-we expect the evidence-based argument to be less accepted. When stakes are high, a 'limited deference' to agencies' judgment and technical expertise might arise (Gormley 1986, 614).
If the argument is challenged, we expect it to happen along two main lines. One is a radical anti-technocratic line: The experts do not represent the people or the common good, and are therefore illegitimate (E. O. Eriksen 2021b, 20). This line would be skeptical to the authority of scientific objectivity altogether. Another is a less radical neutrality-questioning line: Scientific objectivity is a valid source of authority, but the agencies do not live up to it. They may be too close to industry or political interests, or have too limited an evidence base. In summary, this gives the following hypothesis: H1: The evidence-based argument is prevalent in fields with high access to 'hard' science procedures and/or low public salience.

The legislators' command argument
The legislators' command argument maintains that the route to legitimacy is answerability to, and control by, elected politicians. EU agencies may be non-majoritarian, but they cannot escape one core attribute of democratic rule: that they should be accountable (Lord and Beetham 2001, 446). This argument emphasizes the need for agencies to be responsive to their constituencies, i.e. the public as mediated through elected politicians (Flinders 2004). Non-majoritarian institutions, including EU agencies, have often been criticized for their accountability deficit (Schillemans 2011;Landwehr and Wood 2018). Institutionally, the legislators' command argument would call for, for instance, accountability to the European Parliament, safeguards that limit expert discretion, political control on the hiring of executive directors, or member states' representation on agencies' management boards.
We can expect the legislators' command argument to take hold under two conditions. First, contrary to the evidence-based argument, we expect that the public more readily calls for legislators' command in fields that are less technical (less 'hard science') and more value laden (Gormley 1986). If there is no readily acceptable technical or scientific foundation, we expect that expertise and independence are valued less, and political control and accountability valued more. Second, the degree of public salience could matter. When an issue is salient, it by definition receives more attention from politicians. The dynamic might extend to the public debate. We theorize that public calls for elected politicians to exercise control over an agency are more rampant if the issue area is contested or highly salient (Gormley 1986).
We expect the main challenge to the argument to fall along the lines of undue politicization. Political parties and political elites are sometimes seen as corrupt or illegitimate. We might therefore imagine that under certain conditions, political attempts to control an agency will be seen as an undue act of influence and power instead of an exercise of democratic accountability and a strengthening of legitimacy. This counter-justification is most developed in the economic domain, where credible commitments and the time-inconsistency problem are widely accepted grounds for delegation to independent central banks (Kydland and Prescott 1977;cf. van't Klooster 2020). We see that there is a certain symmetry between the legislators' command and evidence-based arguments. The conditions under which we expect the evidence-based argument to succeed are the ones where we expect the legislators' command argument to run into challenges-and vice versa. This gives the following hypothesis: H2: The legislators' command argument is more prevalent in fields with less access to 'hard science' procedures or with high public salience.

The public participation argument
We call the third argument the public participation argument. The academic debate on a 'participatory turn' describes the pressure to open up policymaking to the public as a major trend in modern governance (Krick, Christensen, and Holst 2019). The argument is that non-majoritarian institutions need the direct participation of lay citizens to be legitimate, or that participation has epistemic merits of its own (Fischer 2009;Brown 2009;Jasanoff 2011). The idea builds on the direct model of democracy: it sees direct participation as a better source of legitimacy than indirect representation through elected representatives. Its normative potential is evident: Participation offers a clear link back to the public. It can both compensate for excessive technocratic developments and offer protection against partisan politicization, and may as such be well-placed to resolve the 'epistemic-democratic challenge' (Krick 2021, 2). Institutionally, participation is often achieved through the involvement of civil society, for instance through so-called partnership agreements on the EU level. We also treat transparency and public access to information as central components of participation. These values are prerequisites for meaningful involvement and have 'arguably attained 'quasi-religious significance' in contemporary debates about legitimacy and good governance' (Hood 2006, 3; cited in Krick, Christensen, and Holst 2019).
We expect that public participation will be requested or accepted in issue areas that are highly salient in the public debate. Attention is a limited resource. It is reasonable to assume that people accept or request more participation in issues where they have a lot at stake (Røed andHansen 2018, 1447). This can in principle happen in any type of policy area. We therefore do not expect the scientific 'hardness' of the issue to be consequential-only its public salience. We furthermore expect that if public participation is challenged, it will be along one of two lines. Its potential lack of representation can be questioned, given that only a select few are realistically able to participate in the workings of an agency. Another critique is epistemic: Some issues are complex and demand specialized expertise-perhaps these are best left to the experts (E. O. Eriksen 2021b, 27). In summary, the hypothesis is: H3: The public participation argument is more prevalent in fields with high public salience.

The fundamental rights argument
The literature on EU legitimacy has been overwhelmingly concerned with legitimacy as a debate between effectiveness and quality on the one hand, and representation or participation on the other. A typical example is the debate between input-oriented and outputoriented legitimacy. This debate, however, overlooks one fundamental component of legitimacy: Human rights. 1 Recent analyses of the normative legitimacy of states and international organizations have argued that the respect for human rights is a minimal precondition of legitimacy (Buchanan 2011;Schmidt 2020). Nevertheless, this dimension has been largely lacking in normative treatments of the EU's legitimacy.
The reason for this blind spot may be exactly the fundamental nature of human rights commitments. The respect for human rights is so basic that it is taken for granted. If we understand EU as a limited organization dealing mostly with 'cross-border economic activity' (Føllesdal and Hix 2006, 551), this backgrounding of human rights may make sense. But the EU and its agencies are continually expanding in competences and scope (Genschel and Jachtenfuchs 2016). When the EU's power expands to core state powers, previously taken-for-granted minimal preconditions might come to the foreground. The EU is becoming an actor that not only regulates markets but engages in direct 'factual' or operational conduct towards citizens and third state nationalsconduct that does not involve the adoption of legally binding acts. One example is the physical act of preventing persons from entering a territory (Fink 2020, 533). With such conduct, the EU has the capacity to violate fundamental rights in a direct way. Moreover, human rights are expected to be especially salient in external contexts. A relatively minimal view of human rights says that 'human rights set standards for states' foreign policy, and particularly for the legitimacy of foreign intervention' (Valentini 2012, 182). Whatever else we take human rights to do, they at least do this. Therefore, when EU agencies take on state-like tasks with consequences for people and states external to the union, we expect fundamental rights arguments to be especially prevalent.
In summary, we expect fundamental rights arguments to be prevalent in (1) policy areas that deal with people and states outside of the EU itself, or (2) agencies whose conduct is 'factual' (operational) rather than regulatory (Fink 2020). In order to test this expectation, we leverage the fact that most agencies that satisfy one or both of these two conditions are under the responsibility of the European Commission's DG Migration and Home Affairs (DG Home 2021). There are two EU agencies dealing with fundamental rights as a policy area, the Fundamental Rights Agency (FRA) and the European Institute for Gender Equality (EIGE). We expect coverage of these agencies to feature many fundamental rights terms, but this is not the kind of coverage we are focusing on here. These agencies, however, are not under DG Migration and Home Affairs. Our results will therefore not be driven by these agencies. Since the respect for human rights is a widely accepted norm, we expect explicit critiques of the norm itself to be rare. Instead, we expect that the human rights discourse will be relatively absent in the coverage of agencies that do not satisfy the conditions outlined above. In hypothesis form: H4: Fundamental rights arguments will be more prevalent in coverage of Migration and Home Affairs agencies than other agencies.

Method and data
This paper follows a methodological triangulation strategy (see e.g. Neuendorf 2017), comprising three analyses (two quantitative and one qualitative). All analyses are based on a dictionary approach (Busuioc and Rimkutė 2020a; Rooduijn and Pauwels 2011). We define a priori four sets of keywords that we consider to be indicators of the four legitimacy arguments, giving four legitimacy dictionaries. These terms are derived from academic texts and policy documents. For a full list of terms and a detailed explanation of how these were selected, see the Appendix.
In a first step, we quantitatively analyze full-text media coverage of three EU agencies from 2005 through 2019 in Swedish media. The second, qualitative step allows for a more fine-grained understanding of how and under which conditions the legitimacy discourse about the three agencies differ and gives a validation of the quantitative findings. In the third step, we analyze aggregate data on the Swedish news coverage of all EU agencies 2005-2019. This iteration between small, medium, and large N allows us to leverage the relative strengths of each approach while mitigating their relative weaknesses.
As we are interested in variation over time and between agencies, we choose to keep the geographic dimension constant and focus our analysis on one country: Sweden. The country's absence of strong distorting factors in its political and media systems-it is not extreme in any relevant sense-makes Sweden well suited as a domain for this paper's theory building and testing. The country's high level of trust in science and expertise makes it a most-likely case for the evidence-based strategy and a least-likely case for deviations from it. Deviations from the evidence-based strategy found here might therefore plausibly generalize to national contexts with lower trust (Levy 2008). Although Sweden is a relative late-comer in the EU and a rather reluctant member-state with a number of opt-outs (most notably, it is not part of the Eurozone), it has one of the lowest rates of infringement of EU legislation and is among the most compliant member-states in terms of transposition of EU directives into national legislation (European Commission 2019). Crucially, Swedish citizens have a positive outlook of the EU and its future at a rate that is well above the EU average, while their trust in EU institutions has also been above EU average and stable over a long period of time (Eurobarometer 2019). Two additional factors render the Swedish case especially suitable for a study of the public legitimacy of EU agencies. The first one is its political tradition of consensus politics (Christiansen et al. 2010). The second is its democratic corporatist media system (Hallin and Mancini 2004), whereby political parallelism has coexisted and developed alongside a strong mass-circulation press, a high level of journalistic professionalization and a strong tradition of limits on state power, including an early development of freedom of the press. The combination of a high baseline level of social trust and favorable conditions for public debate makes Sweden a suitable environment to test our theoretical expectations. All research designs entail trade-offs between generality, feasibility and parsimony (Wiedemann 2013). Our restriction to a single country is warranted for a study like this, which aims to develop theory as well as to test it (Flyvbjerg 2011). A cross-national comparative perspective would introduce more complexity with only marginal benefits for theory development.
Our corpus of news articles is collected from the Swedish news media database Mediearkivet. The database allows us to access articles from almost all news media sources in Sweden, not just elite broadsheets or selected tabloids. This broad approach therefore gives a complete view of Sweden's media discourse. It is also preferable to an analysis of EU-level elite newspapers like Politico. Such papers might offer a 'bird's eye' comparative perspective by focusing on the EU as such and not a single country, but their specialization and elite readership make them less well suited to pick up (de-)legitimation debates as they play out in a broader public sphere.

Analysis 1 and 2: full-text analysis of three EU agencies
Design For the full-text analysis, we select three European Union agencies that display variation along the 'scientific hardness' and salience axes, and that operate in areas that have all, at different times, been high on the agenda and endured crises (Immigration, climate change and financial regulation): The European Border and Coast Guard Agency (Frontex), the European Environment Agency (EEA), and the European Banking Authority (EBA). Frontex is tasked with the management of the Schengen area's external borders, and was established in 2004. EEA was established in 1990 and is mandated with providing environmental information to policymakers and the public. EBA was established in 2011, in the aftermath of the 2008 financial crisis. It is part of the European System of Financial Supervision. Its main objective is 'to maintain financial stability in the EU and to safeguard the integrity, efficiency and orderly functioning of the banking sector' (European Banking Authority 2020).
For this analysis, we can draw up a 'hard-to-soft' axis from the EEA, based on the natural sciences, via the EBA, based predominantly on economics (for the 'hardness' of economics, see Akerlof 2020), to Frontex, which relies mainly on intelligence, policing, demographics and related fields. We use the number of news articles in Swedish media as our salience indicator. While there is large variation in salience between the three agencies analyzed here, they are all relatively salient compared to most EU agencies. This is necessary to have enough news articles for the qualitative analysis (most EU agencies receive very little media attention). See the total number of articles mentioning an agency in Figure 1.
Our analysis is based on full-text media articles. We downloaded all articles mentioning an agency and at least one term in one of our legitimacy dictionaries over the period 2005-2019 (N = 8302). 2 Details on the data collection are given in the Appendix. 3 For the quantitative stage, we calculate term frequencies for each term and article by counting the number of times a term in a legitimacy dictionary appears in an article. We finally calculate each term's TF-IDF by dividing its term frequency (how many times it appears in a document) by its inverse document frequency, i.e. how often the term appears across all documents in the corpus. These measures are summarized within each legitimacy dictionary, giving each article one total term frequency and one total TF-IDF for each legitimacy dictionary.
For the qualitative analysis, we selected the top 40 articles by TF-IDF in each dimension for each agency for qualitative content analysis. This gives, after accounting for ties and some categories with less than 40 articles, a total of 344 articles for the qualitative stage of analysis. One author read each article and coded any passage matching any of the four legitimacy arguments. Articles are coded on a sentence level, so each article may contain multiple passages with different codes. The coder also coded, where applicable, whether the passage is negative (delegitimizing) or positive (legitimizing), the role of the speaker (e.g. politician, agency representative, NGO, etc.), as well as other categories inductively emerging during the analysis process. For this stage of our analysis we followed a qualitative, interpretative approach (Hijmans 1996). To ensure dependability of the coding, text extracts selected by author 1 were discussed among the authors and collaboratively interpreted, until consensus was reached in the formulation of codes and their use to interpret the selected texts. The codes are designed to triangulate the findings of the quantitative analysis from Step 1. They are applied to the most relevant articles in our corpus, not a representative (random) sample of them. We therefore interpret their content rather than their frequencies.

Results
Deploying our dictionaries on our corpus of Swedish news articles, the first quantitative stage aims to map out (1) how the three agencies differ along the legitimacy dimensions, and (2) whether there is any variation over time. Panel A of Figure 2 shows the mean number of words belonging to any of the four dimensions per article, by dimension and agency. Panel B of Figure 2 shows how the measure has developed over time. While there are some spikes, the legitimacy dimensions are relatively stable over time. A summary of the main qualitative coding frequencies is reported in the Appendix's Table 9. The claims analyzed are put forth by a range of different actors. We report the role and affiliation of speakers where relevant, and a full overview is found in the Appendix's Table 8.
Next, we move on to evaluate our hypotheses. In this section, we integrate the quantitative and qualitative evidence from analyses 1 and 2.
Hypothesis H1: evidence-based arguments depend on hard science and low salience. The first part of hypothesis H1 says that we would find more emphasis on the evidencebased dimension in fields characterized by hard science. We find quantitative support for this hypothesis (Figure 2). For the agencies relying on 'harder' expertise, EEA and EBA, we see that the evidence-based dimension is clearly the most common. We also find support for this hypothesis in qualitative terms. The EEA and EBA are very often referred to as neutral arbiters of facts and numbers. While we rarely find explicit legitimation (of the form: 'agency X is staffed by highly qualified experts and therefore deserves our support'), we count this taken-for-granted position as a source of data, evaluations and facts as an implicit form of evidence-based legitimation.
Notably, the EEA sometimes explicitly criticizes EU policies. Its numbers are also used by media sources critical towards the EUs environmental policies. In a manner we do not find for the other two agencies, the EEA is portrayed in the media as a 'watchdog': more like an external institution than an EU agency. The EEA acts as a kind of an ombudsman for the environment: A horizontal accountability institution that offers an internal check on other branches of the EU system. Its main source of credibility, and hence legitimacy, nonetheless lies in its provision of accurate numbers and facts.
The EBA is most often mentioned in the media for its stress tests. In articles about the agency, its evidence-based dimension is often linked to the need for trust in markets and institutions. This quote from two Swedish MEPs for the Social Democrats is an example: 'Every country must show respect for [the EBA's] work and make sure that the domestic banks that receive a bad test score immediately plug the black holes' (Göteborgs-Posten 2011). 4 This emphasis on credibility is a classic case of evidence-based legitimacy. Positive and negative evaluations of the agency are often linked to the perceived accuracy, performance and 'toughness' of these tests. For instance: Anders Kvist, finance director at the SEB, states that this year's test was 'much tougher' than earlier, which he welcomes. 'It's about creating trust among investors, analysts and the general public about the risks in the banking sector as a whole,' he says. (Gefle Dagblad 2011) The EBA is not often criticized in evidence-based terms. But we find some evidence-based delegitimation when it fails to use its expertise in satisfactory ways. For example, the EBA stress tests received critique for failing to foresee, at various points, bank collapses in Cyprus, Ireland and Belgium (see e.g. Svenska Dagbladet 2013).
One important prerequisite for evidence-based legitimacy is independence. When the independence of the EBA is criticized, this should therefore be taken as an instance of delegitimation. What we find is that the EBA has at various points been criticized for being either too close to the banks or too close to politics-or both. For example, in a story about new regulations on financial technology, a startup entrepreneur complains that the EBA is 'fully on the side of the banks' (Ny Teknik 2017). In another story, the EBA's stress tests are criticized for being too vulnerable to member states' political pressures, making it impossible to include certain realistic but politically sensitive scenarios in the tests (Svenska Dagbladet 2011a). Finally, another article highlights the need for independence both from financial lobbies and member states to make the EBA 'more robust' and to build up its 'authority and weight' against pressures from the financial lobby and national interests (Svenska Dagbladet 2011b).
Frontex is also sometimes discussed in the evidence-based discourse, but here the picture is not as clear-cut. We find positive legitimation of Frontex in stories about the agency sending experts to border areas, for instance in Greece. The type of expertise highlighted is not hard science-based: One story run by 15 outlets reports that the EU sends 175 specialists with expertise in 'counterfeit documents, stolen cars, dog handlers, illegal border crossings etc' to the external borders (Barometern 2010). On the whole, however, the dominant story about Frontex is more critical and less concerned with evidence or expertise. We explore this further in the next section.
In summary, we find that both critique and praise using evidence-based arguments is more common for the EEA and EBA than Frontex. Independence, 'toughness' and the provision of facts are the most important components. This supports hypothesis H1.
Frontex's status as a neutral facts provider is more disputed. A typical line of critique is represented by this op-ed writer: It is exactly this cold political consciousness that is presented as functional economic policy for the Swedish nation and the European Union, when it in fact has strong streaks of inhumanity. Frontex and the EU's member states carry detailed statistics on police interventions against migrants and deportations, but do not count the number of deaths. (Värmlands Folkblad 2019) In this reading, even the statistics themselves-the hallmark of technical expertise-have humanitarian repercussions. Frontex is here delegitimized in evidence-based terms and portrayed instead as an 'inhuman' actor, showing how the different legitimacy arguments are sometimes presented in opposition to each other.
Hypothesis H2: legislators' command argument more prominent in value-laden fields or where salience is high In fields with lower access to hard science or where salience is high, we expect higher demands on oversight and control by elected officials-be it member states or the European Parliament. Frontex is a more salient agency than the two others in our corpus. If we additionally accept that Frontex's expertise is of a 'softer' kind than the two others, this hypothesis finds quantitative support. For Frontex, the legislators' command dimension is nearly three times as common as the evidence-based (Figure 2).
Demands for parliamentary control of Frontex in our corpus are often coupled with humanitarian arguments, as this excerpt illustrates: We want Frontex employees to be educated in order to ensure compliance with asylum law. We also want the European Parliament to scrutinize the agency to a larger extent, says Philip Amaral from the Jesuit Refugee Service in Brussels. (Miljömagasinet 2010) This quote represents rather clear delegitimation along the legislators' command axis. Frontex is described as lacking in legitimacy because it lacks accountability both to member states and the European Parliament. We do not find this kind of critique against the other agencies.
In the rare occasion when Frontex makes a claim in an article, we see that the agency uses its mandate to deflect criticism: A Frontex spokesperson says, however, that it is not the agency's task to control whether or not human rights are respected, and points out that the agency's task is 'purely coordination.' (Miljömagasinet 2010) Restricting the agency's mandate to 'purely coordination' is an act of depoliticizing the agency, shifting responsibility to the member states-an attempt at legislatorscommand style legitimation.
We qualitatively find that Frontex in some periods is a top political issue. Before the 2019 EP elections, Frontex features frequently in articles discussing different parties' general policy positions. When the EU at large is discussed, Frontex is often part of the discussion. In contrast, the other two agencies feature in narrower, more specialized debates.
We expected that the legislators' command dimension would be less prominent-or even negatively discussed-in articles about the EBA and EEA. The theoretical argument goes that there can, in some instances, be too much 'legislators' command.' This is what we have called undue politicization. In the literature, such an argument is often found in relation to independent central banks. As we also saw under Hypothesis H1 above, we find traces of this rationale in our corpus when the politicization of the banking sector is problematized. In articles such as this one, the emphasis is on the ECB, but the EBA's stress tests are mentioned as part of a system that might restore popular trust in European banks: In several important countries, the banking sector is heavily politicized.
[…] Especially in the heavily indebted countries, the question arises about how the ECB assesses their risks. What will the consequences be in the financial markets? For the ECB's own balance sheet? The conflict of interest is obvious. (Svenska Dagbladet 2013) In summary, hypothesis H2 finds support in our data. Frontex is rather clearly discussed using legislators' command arguments, and in the banking sector we find warnings against too much legislators' command in the form of politicization.
Hypothesis H3: public participation legitimacy for high-salience, high-stakes issues We hypothesized that topics of high salience would give rise to more debate along the participation dimension. This hypothesis is not confirmed by the data in our corpus. From a quantitative perspective, as we have seen, very few articles contain any of our participation words. In the qualitative phase, no (de)legitimation claims were coded as participation for the EEA. Only a handful of statements were coded as participation for the EBA and Frontex.
In general, therefore, the main impression is that participation justifications do not feature prominently, neither in positive nor negative terms, in Swedish media debates about these agencies. Hypothesis H3 does not find supporting evidence in our corpus.
Hypothesis H4: fundamental rights a prominent discourse for externally-oriented or operational agencies Lastly, we expected to find more human rights arguments in the one Migration and Home Affairs agency in our corpus: Frontex. The quantitative evidence supports this hypothesis. Only in the media coverage of Frontex are human rights invoked to any significant extent (see Figure 2).
The qualitative evidence also supports the hypothesis. Frontex is often portrayed in our corpus as a vessel for the militarization of EU borders-an executive arm of the 'Fortress Europe. ' We also find charges of hypocrisy: Politicians talk of humanitarianism while they simultaneously allow migrants to drown in the Mediterranean. One example: 'For migrants, the military operation [including Frontex] has made the journey over the Mediterranean more dangerous-a fact the EU's politicians are conscious of' (ETC Lördag 2019).
Altogether, 64 statements in our corpus are coded with the topic 'humanitarian'-all of them about Frontex. The strengthening of the agency is in many cases seen in direct opposition to the right to asylum: Because Frontex makes it more difficult to cross the external borders, it also makes it more difficult to apply (safely) for asylum in Europe. This has been a consistent issue long before the 2015 refugee crisis. Here is Amnesty International in 2009: One finds it very easy to agree on measures for border controls and more money to Frontex […]. It should have been just as easy to cooperate about the duties we have towards refugees, says Amnesty lawyer Madelaine Seidlitz. (Trelleborgs Allehanda 2009) In line with our hypothesis H4, only Frontex is discussed-and most often negativelyusing human rights arguments.

Analysis 3: A quantitative study of all EU agencies
To test the general application of our findings and hypotheses, we next gather data on all EU agencies in the period 2005-2019. To do this, we take a keyword-based approach.
Whereas the units of analysis were full-text articles in the previous section, we are now working on aggregated data. The unit of analysis here is the agency-year.

Design
Due to limitations on text downloads from Mediearkivet, we are not able to download fulltext news articles about all agencies. Instead, we leverage the portal's analysis tool, which allows us to download the yearly number of hits for a series of searches. For each agency i, we carry out five searches and download the yearly number of hits for each. The searches return the number of articles that match the following criteria, respectively: . agency i AND any word in evidence dictionary . agency i AND any word in legislators' command dictionary . agency i AND any word in participation dictionary . agency i AND any word in fundamental rights dictionary . All hits for agency i This gives us a dataset of 36 agencies by 15 years by 5 searches-2700 observations in total.

Dependent variable: dictionary hits
The dependent variable is dictionary hits as described above. A drawback with the current approach is that we are not able to give more weight to articles with more hits in a given dictionary. We only count the number of articles with any hits on a given dictionary. Comparing the search data with the analysis of full-text articles in the previous section, however, gives little reason for concern. The relative prevalence of each legitimacy dimension is not substantively different between the keyword-based and full-text approaches. See a comparison in the Appendix. For independent variables, we operationalize the concepts of salience and scientific hardness, as discussed below.

Salience
We define salience as the amount of attention given to an agency. Until now, we have used the total number of news articles about an agency as a measure of the concept. However, our dependent variable in this analysis-the number of hits on a dictionary -depends strongly on the total number of news articles about an agency. Using all articles about an agency as a measure of salience would not allow any inferences about the relationship between salience and the dependent variables. We choose to control for the total number of articles about an agency when estimating the relationship between salience and legitimacy arguments and, therefore, we opt for another measure of salience.
Several authors (Wood 2018;Koop 2011) use parliamentary attention as a partial measure of salience. Due to the above concerns, we will focus on the parliamentary dimension of salience here and count the yearly number of European Parliament 'questions for written answer' mentioning each EU agency. 5 The advantage of using parliamentary questions is that they reflect the EU-level political salience of an agency, while also giving us an 'out-of-sample' independent variable.
Since salience in principle might vary over time, there are two ways to measure its association with legitimation arguments: Between agencies and over time within agencies. For the model that compares between agencies, we calculate the sum of all variables across all years so that each observation is one agency. Since the counts are relatively high in this case, we treat the variables as ratio-scaled and fit an OLS linear model. As we have seen, all the counts are heavily positively skewedagencies differ by orders of magnitude. We therefore log-transform all variables and estimate one model for each of the four legitimacy dictionaries: where each unit i is an agency, P is the count of parliamentary questions mentioning an agency and A is the total count of articles about a given agency.
For the model that leverages variation within agencies over time, we fit an agency-level fixed-effects model with cluster-robust standard errors: Where i is an agency observed in year t and g i is the agency-fixed effects. P it and A it is the count of parliamentary questions and total articles, respectively, about an agency i in year t-both are scaled and centered. Since the counts disaggregated by year are much lower and include a large number of zeroes, we also estimate the same model specification using a model designed for discrete counts: a negative binomial model.

Hardness
A fine-grained measure of scientific hardness would be hard to defend. We therefore develop a simple ordinal three-point scale, from 'soft' to 'hard.' There is no clear answer as to where the cutoff points should be placed, but we believe the overall direction of our classification is uncontroversial. More details on the coding and every agency's classification are reported in the Appendix.
Agencies are coded as hard (3) if they deal with the natural sciences, including physics, chemistry, medicine, environment or climate science, and biology. Agencies are coded as medium (2) if they deal with economics, finance, or banking-broadly, the discipline of economics. The rationale here is that while economics is a social science, it is widely considered 'harder' or more 'scientific' than the other social sciences (e.g. Fourcade, Ollion, and Algan 2015). Agencies are coded as soft (1) if they operate in the social sciences or law, or if their operation is not clearly related to any particular expertise. An agency's hardness is not expected to vary over time. This measure is therefore coded once for each agency.
In order to explore the relationship between legitimation and scientific hardness, we work on aggregated data across years. We treat the hardness variable as categorical and calculate the proportion of articles matching each dictionary in each of the three hardness categories. To obtain a nonparametric measure of the uncertainty in the estimates, we carry out 2000 bootstrap resamples of the agency-level dataset. In each resample, we calculate the proportion of articles that match each legitimacy dictionary. We take the mean of the 2000 proportions as our point estimate and estimate a 95% confidence interval empirically from the distribution of proportions-set at the distribution's 2.5 and 97.5 percentiles.
External and operational agencies Hypothesis 4 says that the fundamental rights discourse is more prevalent in the coverage of the 6 agencies working under DG Migration and Home Affairs. Notably, the agencies dealing explicitly with fundamental rights (EIGE, FRA) are not under the Migration and Home Affairs umbrella. 6 What we aim to pick up here is not an agency's substantive focus on fundamental rights but rather their 'factual' conduct and/or external orientation. Here, too, we take a simple and flexible approach: We compare the mean score of these 6 agencies with the mean of the other 30 agencies and calculate 95% confidence intervals. Figure 3 shows the results of four OLS regressions on the aggregate data (i.e. where each observation is one agency). We see that the number of parliamentary questions is positively associated with the legislators' command dimension but not with any other dimensionwhen we control for the total number of articles about an agency. Hypothesis 1 is therefore left unsupported. We expected a negative association between the prevalence of evidencebased arguments and salience, but the estimated association is zero. The findings are, however, in line with our hypothesis 2, which expects legislators' command arguments to be more prevalent in the coverage of high-salience agencies. The findings do not support our hypothesis 3-that participation arguments are more prevalent where salience is high. In fact, participation seems negatively associated with parliamentary questions, albeit not significantly different from zero at the conventional 95% threshold. There is less supporting evidence about the effect of changes in salience over time on the prevalence of legitimacy dimensions. Table 11 in the Appendix shows the output of fixed-effects OLS and negative binomial regressions, none of which show a significant relationship between salience and any of our legitimacy dictionaries within agencies.

Hardness
Next, we examine whether an agency's hardness is associated with hits on each of our four legitimacy dictionaries.
The estimates in Figure 4 are clearly in support of hypotheses 1 and 2. In line with hypothesis 1, the evidence-based argument is significantly more prevalent in harder agencies. In line with hypothesis 2, this association is reversed for legislators' command. We also expected there to be no association between hardness and participation claims. This is supported by the evidence.

Fundamental rights and external agencies?
Testing hypothesis 4, we aim to examine whether fundamental rights discourse is more prevalent in coverage of Migration and Home Affairs agencies. Figure 5 shows that over 10 percent of articles about the MHA agencies contain at least one fundamental rights term, versus less than 4 percent of articles about other agencies. This supports hypothesis 4.

Discussion
The analysis of data on all EU agencies in Swedish news coverage largely confirms the insights generated in the smaller-N study. Hypotheses 1 and 2 are largely supported. When comparing agencies, we find that the keywords in the evidence-based dictionary are significantly more prevalent in hard-expertise agencies. Keywords in the legislators' command dictionary are more prevalent in soft-expertise agencies. These measures mirror each other, supporting our theorized notion that more evidence-based discourse gives less legislators' command discourse (and vice versa)-and that this depends on an agency's 'hardness. ' We have limited evidence that salience is associated with different legitimation discourses-the relationship is significant only for the legislators' command dictionary. And we have no evidence that salience affects legitimation discourse over time within agencies. This may suggest that the coverage of an agency is relatively stable over time. But it may also be the case that our yearly data is too coarse to pick up the relevant changes in salience and discourse. Future research should examine this hypothesis with more granular data that is better able to pick up short-term changes in salience and media discourse.
Hypothesis 3 is not supported: Participation discourse is generally low and does not vary much depending on salience or hardness. This is in line with our qualitative and article-based quantitative evidence.
Hypothesis 4 is supported: The data shows that fundamental rights terms are most often absent, but they are more prevalent in media coverage of Migration and Home Affairs agencies. If the EU's integration of core state powers continues, we might expect explicit fundamental rights (de-)legitimation of the EU and its agencies to come increasingly to the fore. Some limitations remain. First, we have focused on the media discourse in a single country, Sweden. We have argued that Sweden offers a favorable context for building our theory due to its consensus governance, EU stance, and quality and freedom of its media. It is therefore plausible that our findings will generalize to other national contexts. Nonetheless, only comparative studies can say this conclusively. Sweden is, for instance, not part of the Eurozone. Whether the public debate in Eurozone countries is different from the Swedish debate remains an interesting question for future research. Second, the use of word counts to uncover latent meanings has its inherent weaknesses. Word counts are only proxies for the underlying theoretical categories we are interested in. Any choice between different methods is a trade-off between specificity and generality, and this design is no exception (Wiedemann 2013). While the paper's quantitative parts will not capture nuance or complex meanings, they have the advantage of being general, transparent, and replicable. For instance, Study 3 is based on all articles about all agencies 2005-2019, and is designed to minimize the number of undocumented researcher choices in the counting and categorization of articles. We also mitigate the shortcomings of the quantitative approaches by combining them with qualitative insights in Analysis 2.

Conclusion
We have conceptualized and operationalized four dimensions of legitimation in this paper, firmly embedded in theoretical literature and earlier empirical studies. Our evidence suggests that there is systematic variation in the legitimation arguments used about different agencies. Technical expertise is an important (de-)legitimizing device. But, contrary to what some earlier contributions have found, expertise is not the be-alland-end-all criterion of agencies' public legitimacy. Our findings suggest that an agency's evidence base matters: Technical expertise is more central to legitimacy for agencies based on 'hard' science. The legislators' command discourse, on the other hand, is more central for agencies based on 'soft' expertise.
An agency's salience, however, does not matter as much as we expected theoretically. In line with hypothesis 2, an agency's higher salience is associated with more legislators' command terms. But neither participation nor evidence-based terms vary systematically with an agency's salience. And we find no evidence that a change in an agency's salience affects the terms used about it. Lastly, the evidence supports our expectation regarding fundamental rights. Migration and Home Affairs agencies are discussed using more fundamental rights terms than other agencies.
In summary, we make three contributions to the study of the legitimacy of non-majoritarian institutions in general, and EU agencies in particular, in this paper. We propose and operationalize a novel conceptual scheme that captures normative conceptions of legitimacy directly and demonstrate its utility in an empirical analysis. Since our theoretical argument is general and does not depend explicitly on the national context, future research may apply this scheme to new data and other contexts. They might for instance gather a larger corpus of full-text data on a wider selection of agencies, study other countries, analyze who does the speaking (agencies vs. their critics/supporters), or model support/critique more explicitly.
Our findings have theoretical implications for future empirical research on legitimacy. One is that the prominence of each type of legitimacy argument depends on an agency's access to 'hard-science' evidence and (to a lesser extent) its public salience. Future research should pay closer attention to these scope conditions for legitimacy. Another theoretical implication is that while human rights is normally in the background, it is a more important part of the (de-)legitimation of agencies that are externally oriented or 'factual' in their conduct. As the EU continues its development from a regulatory to a positive state (Genschel and Jachtenfuchs 2016), this dimension could get more important. Research on EU legitimacy should therefore theorize the role of fundamental rights more explicitly (Buchanan 2011;Sangiovanni 2019).
Finally, our analysis has implications for normative theory. To be sure, we cannot build a normative argument on empirical observations alone. Nevertheless, we maintain that 'actual' normative legitimacy depends on discursive legitimation, even if it is beyond the scope of this paper to defend the point at length. If we accept that premise for now, our analysis suggests that different types of agencies have different normative legitimacy foundations. There is no one-size-fits-all standard of democratic legitimacy against which the entire EU should be assessed. Instead, different agencies might face different legitimacy standards. This conceivably extends to non-majoritarian institutions beyond the EU system. The point is that different types of institutions are geared to do different things. Every single institution must not fulfil everything that democracy demands-so long as the system, as a whole, does. It is in this institutional view of legitimacy we expect to resolve the tension between the epistemic and participatory dimensions of democracy.