Controlled networking: Organizational cohesion and programmatic coherence of Swiss parties on Twitter

Introduction

In today’s established democracies, the erosion of party alignments and the growing mediatization of politics have yielded innovations in parties’ organizational and programmatic capacities that have become even more essential in election campaigns (Dalton et al., 2000; Strömbäck, 2008). A crucial innovation potential thereby lies in the deployment of new communication technologies, such as social media, to reach voters. Accordingly, social media platforms have become regularly used tools in election campaigns across established democracies (see e.g. Gibson, 2015; Grant et al., 2010; Koc-Michalska et al., 2014; Larsson and Moe, 2012; Stromer-Galley, 2014; Theocharis et al., 2016; Vergeer and Hermans, 2013).

As most previous research has focused on the intensity and content of a single politician’s social media communication, this study explores social media use at the level of political parties. More specifically, we investigate the performance of Swiss parties on the microblogging platform Twitter, during the 2015 federal election campaign, by assessing a comprehensive set of possible Twitter interactions and dynamics—follower relations as an indicator of organizational cohesion; retweets and text similarity as proxies for programmatic coherence. ¹ This approach enables us to link the analysis to established theories on party unity in election campaigns (Jungherr and Theocharis, 2017).

Switzerland is a valuable comparative case for exploring the capacity of political parties to mount effective election campaigns via Twitter. First, due to the distinctive consensus-oriented character of its political system, Switzerland has very low barriers limiting entry to the public debate (Höglinger, 2008). Relative to most other countries, a larger variety of parties can intensively engage in election campaigns. Second, election campaign managers confirm the increasing importance of social media use (Fiechter and Kohler, 2015). That said, the size of the Twittersphere in Switzerland is relatively moderate, which makes it much easier to get a comprehensive and comparable sample of Twitter accounts than it would be in larger countries, where the number of users can easily exceed several million. More precisely, our analysis relies on the networking and communication patterns of 1341 Twitter accounts, which have been subjected to bootstrapped network analyses.

Twitter is a popular microblogging service in Switzerland. During the 2015 federal elections, about 10% of Switzerland’s adult population used Twitter to become informed (NET-Metrix, 2015). Moreover, previous research on other countries has established that the political elite and journalists, that is, actors with disproportionate influence on the campaign, rely even more on microblogging than regular citizens do (Himelboim et al., 2013; Wallsten, 2010). Switzerland still lags behind; however, during the 2015 elections, about 50% of the 246 members of Switzerland’s Federal Assembly and three out of seven Federal Councilors were active on Twitter. ² Since then, however, more politicians have joined Twitter and communicating on social media has become more important (Fiechter and Kohler, 2015), which is why we can expect social media to become even more influential as a complement to more classical campaign strategies.

We use Twitter in this study, first, because it is the main social media platform for Swiss politics other than Facebook. Although Facebook has a much larger user base (Freelon, 2017), only its pages and not its more relevant personal profiles are openly accessible. By contrast, information on Twitter accounts is public by default and thereby easily accessible for scientific research (Vergeer, 2015). Also, unlike Facebook, Twitter allows nonreciprocal relationships between users, which enables the study of both interactive- and broadcast-style communication. Taken together, Twitter’s advantages allow us to conduct a much more thorough analysis of the campaign than would be possible using Facebook data. ³

Yet this study’s contribution is not restricted to Twitter; more generally, it also aims to extend existing research on election campaigns with respect to two important aspects. First, Twitter data permit us to distinguish the more structural features of party organization, such as direct relationships between party-affiliated accounts, from more programmatic ones, such as the congruence in communicative efforts between these accounts. Second, national, regional, and local politicians of varying political functions have been integrated into the analysis, which enables us to study party unity on a much broader set of politicians than previous research. Such a research strategy therefore seems generally promising for studying party unity in other strongly federalist states such as Germany or Belgium, where subnational party factions are comparatively powerful.

Our estimates provide evidence that Switzerland’s political Twittersphere is shaped by an ideological divide as well as a language-specific separation. Furthermore, it is remarkably representative of “offline” politics, with the notable exception of the Swiss People’s Party (SVP), which is heavily underrepresented on Twitter. The results also underline the importance of Switzerland’s federal structure, because it appears to be generally difficult for the national party leadership to control politicians campaigning at the cantonal and local levels. More precisely, in terms of programmatic and organizational unity, networks are denser and more strongly hierarchical at the national level. The analyses of organizational cohesion also show that larger parties are more hierarchically structured than smaller ones. Smaller parties, by contrast, exhibit higher levels of programmatic coherence, in terms of the similarity of their Twitter messages. Finally, Social Democrats stand out for their particularly high levels of organizational and programmatic unity in their networks.

Political campaigning in the digital age

Political parties fulfill crucial hinge functions in democratic processes, by activating the electorate and representing specific interests (Gibson et al., 1983). However, three longer-term processes constrain the electoral and representative functions in established democracies. First, Mair (2008) and others convincingly argue that modern governments must increasingly abide by external constraints, to the detriment of their responsiveness to voters. This is especially true for European countries, where governments are simultaneously pressured by European and international policy and market prerogatives (Pontusson and Raess, 2012). As Hellwig and Samuels (2007) and others show, the declining leeway for national policy making leads to a thinning out of the electoral linkage that makes it harder for citizens to hold governments accountable for policy outcomes. Second, political parties are challenged by the erosion of their organizational basis due to the de-alignment of citizens from previously stable party identifications (Dalton et al., 2000). Because votes have become less structured by party loyalties over the last decades, electoral volatility has increased considerably. Finally, parties are under growing pressure to adapt their campaigning styles to the prerogatives of the mass media (Thesen, 2014). This requires parties to streamline their communication strategies in accordance with the relative news value of the messages; the implications of this shift privilege contentious, sensational, personalized, and simplified messages over substantial discussions of policy positions (Esser and Matthes, 2013).

Although it is still open to debate how, exactly, international and European constraints, the weakening of party alignments and mediatization have transformed party politics over the last decades, it is clear that these factors have raised the general level of competition in electoral contests. Parties striving for votes and office must increasingly take the center stage of mass-mediated politics to attract the electorate (Kriesi and Trechsel, 2008). Because the electorate has become more skeptical and independent over the last few decades, parties must employ innovative campaign strategies. Accordingly, political communication researchers postulate the emergence of distinct changes in the style of campaigning (see Norris, 2000). On the one hand, campaign activities are increasingly marked by pressures aimed at professionalization, which implies a strengthening of the parties’ leadership and a growing importance of media and public relations consultants (Bowler et al., 1999; Gibson and Römmele, 2001). On the other hand, political campaigning is being reshaped by the continuous integration of new communication technologies (Vergeer et al., 2013). Hence, the ability to deploy new technologies, such as social media, has become a defining element of political campaigns (Esser and Matthes, 2013; Obholzer and Daniel, 2016).

The deployment of Twitter in election campaigns can be perceived in this context of technological innovations for political campaigning. Twitter is a microblogging service with a user community that has been growing since 2006. The service allows users to connect easily to other users (following) and to rapidly disseminate and share short messages of 280 characters ⁴ (tweets and retweets). Users can be identified by their @-mention (e.g. @alainberset, the most active Swiss Federal Councilor on Twitter), and trending topics are traceable by their hashtags (e.g. #WahlCH15 for the last national elections in Switzerland). There is a burgeoning literature that engages more substantively with the campaigning of political parties on Twitter. Most of this research focuses on national or regional elections (e.g. Graham et al., 2013; Golbeck et al., 2010; Obholzer and Daniel, 2016; Stromer-Galley, 2014; Theocharis et al., 2016) and does not apply a network perspective. We extend the scope of our analysis to the municipal level and to politicians active outside the parliamentary arena. Previous studies based on network analyses have focused on single aspects, such as retweets or @-mentions (Ausserhofer and Maireder, 2013; Hemsley et al., 2018).

Communication via Twitter is cheap, relative to conventional electioneering tools like leaflets, advertisements, and street events. These low costs, paired with the interactive nature of the application, might also enable parties to sustain government responsibility while being responsive, at least remotely, toward their constituencies (Obholzer and Daniel, 2016). Accordingly, early research ascribed to Twitter the capacity to engage with constituents, thereby opening the door to more citizen participation in the political process (Theocharis et al., 2016). However, most research indicates that parties’ communication on Twitter rarely ever lives up to this normative expectation (Larsson and Moe, 2012). Most politicians use the platform in a broadcasting style (Graham et al., 2013). On the one hand, such widespread noninteractivity (or nonresponsiveness) by politicians on social media is a reaction to the widespread incivility of user comments (Gervais, 2015; Theocharis et al., 2016). On the other hand, noninteractivity clearly serves the individual politicians’ interests (Grant et al., 2010). Twitter allows parties to convey political information to their members, journalists, and the broader public rapidly and directly, without bypassing the gatekeepers of traditional mass media (Graham et al., 2013).

Previous work on Twitter suggests that politicians in candidate-centered contests in countries such as the United States and the United Kingdom generally avoid interacting with each other on this medium (Hemsley et al., 2018), instead preferring to use Twitter in a way that Stromer-Galley (2014) calls controlled interactivity. Instead of using the full capabilities of Twitter to engage in a genuinely open deliberation with the public, candidates strategically craft their messages to create echo chambers in which the political orientations of their voters are reaffirmed (Colleoni et al., 2014; Conover et al., 2012; Vergeer, 2015).

Stromer-Galley’s (2014) theory of controlled interactivity was developed to study the US presidential election campaigns. These contests are highly candidate-centered and allow parties to play only the subordinate role as election platforms. Analyzing election campaigns in a multiparty system, as Switzerland exemplifies, is different. Political parties have decisive influence over the course of the contest, and previous studies have shown that there is a lot of interaction among the political elites, while users from the broader public are stranded on the network’s periphery. This has been found for Denmark, Austria, and Switzerland (Ausserhofer and Maireder, 2013; Larsson and Moe, 2013; Wueest and Mueller, 2015). In multi-party electoral contests, the control of the message begins with the parties’ control over their own members’ communications. To wit, in party-centered contests, parties must engage in controlled networking as the first step of their social media strategy, before single politicians engage in controlled interactivity with the public.

Party unity in election campaigns on twitter

Controlled networking on Twitter is, by no means, a self-propelling process. While campaigning through traditional media channels has become increasingly professionalized, more intense social media use has, ironically, reintroduced amateurism to the political process (Vergeer, 2015). Because each single politician has an individual Twitter account, communication is highly decentralized, which gives single candidates the opportunity “to engage in personal promotion outside the auspices of their parties” (Theocharis et al., 2016). Hence, it would require a considerable effort from parties to control the agenda within their own ranks and to quell messages from dissenters challenging their official campaign programs. However, as outlined further below, asserting and maintaining complete control of the network might not always be a reasonable social media strategy.

Party unity ⁵ is most commonly studied in parliaments, using different vote- or survey-based measures (Bailer, 2016; Carey and Shugart, 1995; Coman, 2015; Hix, 2004; Thorlakson, 2009). Classic studies of party unity, by contrast, emphasize the notion that party unity results from several foundational factors, including the centralization of internal decision-making processes, the amount of resources available, the degree of professionalization, and the level of programmatic activity (e.g. Gibson et al., 1983). For this analysis, we maintain that this traditional concept of party unity can be summarized along two dimensions. On the one hand, a party can show organizational unity by assuming a tightly coordinated structure. In such parties, the leadership is able to establish a high level of discipline via centralized organization and comprehensive control over the membership of its members (Carey, 2007). On the other hand, a party can exhibit unity via preferences, manifest through the substantively cohesive quality of its members’ public campaign messages, or in the congruence of policy decisions (Benoit et al., 2011). Long before the emergence of social media, “control of ‘the message’—the thematically unified collection of issues, frames, talking points, concepts, and images”—has been a key campaign objective (Freelon, 2017).

Studies focusing on parliamentary voting records generally struggle to distinguish between these organizational and programmatic pathways to party unity, as legislative party unity can originate in party discipline, as well as from the cohesive preferences of representatives (see Volden and Bergman, 2006). Twitter data, however, facilitate at least some disentanglement of these two aspects. To examine the organizational dimension, we can tap into the follower networks established by Twitter users. Generally, these follower relationships are quite stable as users must add other accounts to their personal networks only once and they rarely ever terminate these connections. To study organizational capacity, we can therefore measure the degree of coordination among accounts affiliated to the same party. Communication on Twitter, by contrast, is highly dynamic, which means that aspects such as the number of retweets and @-mentions, as well as the similarity of the tweets’ content, indicate coherence among accounts from the same party. Essentially, this means that we can consider each tweet as an instance of programmatic activity, analogous to an election leaflet (Bowler et al., 1999).

Theoretically, a party’s communication on Twitter is fragmented into as many accounts as there are party-affiliated accounts on Twitter, which represents, in our case, several hundred accounts for each of the five major Swiss parties. In reality, parties are divided into more or less coherent sub-coalitions, characterized by constant disagreement over goals and the means to pursue them (Kitschelt, 1989). One divide is particularly pronounced in the case of Switzerland. Switzerland is a fragmented country where regional (cantonal), and even local, governments exercise a considerable degree of fiscal power and jurisdiction across a wide range of policy areas (Kriesi and Trechsel, 2008).

This encourages party organization at the subnational level (Carey, 2007). As Thorlakson (2009) has established, heavily decentralized federations with low coordination requirements between federal and state-level governments are likely to be accompanied by highly autonomous state parties. Federal elections provide the most notable example for this in the Swiss case. These elections are organized at the level of cantons, leaving cantonal party sections with a decisive say in the selection of candidates and a commanding role in election campaigns in their constituencies. Parties at lower levels are thus cross-pressured. On the one hand, their programmatic activity is at least partly shaped by the specific regional conflict structure, the particular economic situation of the canton, and the constituents’ preferences (Moon and Bratberg, 2010). On the other hand, however, it is likely that the stronger national arm of a party will attempt to impose its orientation on the subnational parties’ preferences. In the case of programmatic differences among the factions of a party, unity is therefore more difficult to maintain (Lindstädt et al., 2011).

It has often been noted that the key to an effective campaign is an affirmative, consistent message that distinguishes a party from the programmatic efforts of its opponents (Shaw, 1999). This is mainly because messages emphasizing specific issues are capable of activating latent predispositions in the electorate (Gelman and King, 1993; Holbrook and McClurg, 2005). Political parties can thus have a decisive impact on the electoral outcome by shifting the center of attention toward issues on which voters perceive them to be competent (Zaller, 1992). A necessary precondition involves the exhibition of a high level of cohesiveness across party communications (Carey, 2007; Traber et al., 2014). If the campaign efforts of individual politicians are uncoordinated or only loosely coherent, this signifies to voters that the party is afflicted by internal divisions and most likely not capable of shaping policy outcomes.

Nevertheless, it still makes sense for some candidates to defect from the party line (Tavits, 2009), especially if the party is ideologically heterogeneous and must represent a variety of interests characterizing local voters or interest groups (Bailer, 2016). Due to the highly erratic nature of Twitter communication, this microblogging service might be regarded as the predestined channel for mavericks trying to win votes by deliberately departing from the party line.

We can therefore formulate two specific expectations for the analysis. On the one hand, we anticipate that the lower the level of government, the more difficult it is for parties to keep messages coherent across their accounts. In other words, we would be surprised if the national level didn’t have more organizational and programmatic coherence than the lower levels of government. This is because politicians campaigning at the regional or local level are more likely to be influenced by the regional or local conflict structure, such as particular political traditions, a specific economic situation, or distinctive voter preferences (Müller, 2013; van Houten, 2009). Well-known divisions in Switzerland exist between the language regions on questions related to European integration—with the French-speaking regions, until recently, assuming a pronounced pro-European stance. There are also fierce conflicts in the country’s largest party, the SVP, between the right-wing populist factions led by the cantonal section of Zurich and the more moderate factions led by the Berne section, which even resulted in the secession of the moderates into a new center-right party. Hence, variations in the structural conditions and preferences across units of a federation can make it reasonable for a party to mount a diverse electoral campaign (Bailer, 2016; Verge and Gómez, 2012).

The probability that a party will show unity in its Twitter campaigning also depends on demand, that is, how extensive social media usage for political activities is among the voters of this party (Daniel et al., 2019). The larger the potential audience of a party’s communication on Twitter, the greater the chances that such campaigning will have a lasting effect on party attachments (Selb et al., 2009; Stromer-Galley, 2014). A party with an already large supporter base on Twitter must therefore care about organizational and programmatic coherence. Specifically, we can formulate our second expectation as follows: the networks from parties with a large support base on Twitter send more retweets from their own ranks and send more similar messages. In Switzerland, the parties with a large support base on Twitter tend to be smaller (in terms of electoral support), newer and more leftist parties, like the Social Democrats (SPS), Greens (GPS), and Green Liberals (GLP), because their constituencies are typically both younger and more attentive to social media (Wueest and Mueller, 2015). If the audience on Twitter is small, relative to the parties’ overall voter base, the primary goal of these parties on social media should be to incite and attract voters from different societal groups than their mainstream voters (Cardenal, 2011). This is most likely the case for the more traditional, conservative, or center–right parties in Switzerland, such as the Liberals (FDP), the Christian Democratic People’s Party (CVP), SVP, and the Conservative Democratic Party (BDP), whose voters share a comparatively low level of interest in social media and are also members of an older age group.

Data

Measuring the behavior of the political elite usually involves survey data or roll-call votes (Bafumi and Herron, 2010; Bailer, 2016; Bartels, 1991; Clinton, 2012). While conducting elite surveys is costly and conceptually challenging, roll-call data are usually only available for a restricted group of politicians, including members of national parliaments. Other alternative data sources include campaign finance data (Bonica, 2014), but in some countries, including Switzerland, these data are not publicly available. We suggest that studying political communication on Twitter represents a viable alternative to tapping into the behavior of the political elite. The collection of data from this microblogging service requires comparatively little effort, because it is public to a large extent and can be collected on a large scale via application programming interfaces (API) (see Barberá, 2015). Twitter is widely used and offers rich information on political campaigns. Skeptics, however, highlight the potentially high selection bias on Twitter (Pennacchiotti and Popescu, 2011) and Jungherr et al. (2012) observe that the only way to achieve an accurate prediction from Twitter data is by accurately identifying a sample that includes the users of interest. This is why we follow a position-based approach to systematically trace Twitter users who are relevant to Swiss politics (see Marin and Wellman, 2011).

The myriad of Twitter accounts and their highly unstructured descriptions rendered the identification of relevant users the most difficult challenge associated with the data collection. We started out by hand-compiling an initial set of 157 Twitter accounts, which comprised all representatives of the Federal Assembly, all Federal Councilors, as well as the official national accounts of the seven most important parties in Switzerland, provided they had a Twitter account in 2015 (see Table A1 in the Online Appendix). In four chain-referral extension rounds, using the Rest API of Twitter, we extended this initial set to three steps for each round. First, for all previously identified accounts, the users that follow these accounts as well as the users that are followed by these accounts are collected. Second, a keyword list that contains all names, abbreviations, and paraphrases of Swiss parties, as well as all official employment titles of Swiss politicians in the three official languages of Switzerland (Italian, French, and German), is matched to the Twitter biographies of the roughly 200,000 accounts retrieved in each extension round (see Table A2 in the Online Appendix). Finally, all keyword hits are manually checked to confirm their relevance. ⁶

Because of its institutional setting, such as the mostly open party lists, and cultural diversity, such as the divide between the conservative countryside and progressive cities, Switzerland has a heterogeneous party landscape. The four parties represented in government, SVP, SPS, FDP, and CVP, also dominate the Council of States (lower house), while the party landscape in the National Council (upper house) is shaped by as many as eleven parties. Given that the network estimates require a statistically significant minimum number of accounts, we restricted our analysis to the eight parties that gained at least 2% of the votes in the election to the National Council. Ultimately, we were able to include 1341 accounts in the analysis. ⁷ Subsequently, we manually supplemented the accounts with the following data: canton of residence, party affiliation, gender, institutional level of political activity (national, cantonal, and municipal), and political function within the party (party functionaries, elected members of legislative and executive bodies, and party members with no other political function). For this annotation, we relied mainly on official sources, such as election records or protocols of public assemblies. Table A4 in the Online Appendix gives an overview of these indicators.

We also retrieved all tweets sent by these accounts for the period from August 1, 2015, to election day, October 1, 2015. The most intense phase of federal election campaigns in Switzerland begins on the 1st of August, the national holiday, on which every politician is obliged to give one or more speeches on the state of the country. The volume of communication varies greatly across the Twitter accounts, but each account had at least one tweet in this period. After filtering for the three national languages, the corpus comprises 129,271 tweets: 95,495 written in German, 30,684 in French, and 3092 in Italian.

Measurement strategy

Building on the following and communication patterns in our sample of party-affiliated Twitter users, we develop three network-based statistics to measure the organizational cohesion and programmatic coherence of Swiss parties (see Table 1 for an overview). First, we use the follower network to measure organizational cohesion. The follower network shows how well parties are able to connect affiliated accounts. Ideally, a party wants to ensure that each affiliated account follows as many other affiliated accounts as possible, because this ensures that all affiliated accounts stay abreast of each other’s status updates. Moreover, the number of followers is often used as a status symbol on Twitter, which is why having affiliated accounts that follow each other is a simple, yet effective, way of boosting parties’ Twitter reputations. For this indicator, we directly analyze the network data retrieved from Twitter.

OPEN IN VIEWER

Table 1.

Measurements and indicators used in the analysis.

Concept	Indicator
Organizational cohesion	Followers
Programmatic coherence 1	Retweets
Programmatic coherence 2	Tweet similarities
Concept	Statistic
Connectivity	Density
Hierarchy	Betweenness centrality

Second, we measure programmatic coherence by computing two indicators from the content of tweets sent by party-affiliated users. For one indicator, retweets, we use a mechanism specific to the Twitter platform. Retweets allow users to recycle the tweet of another user; to wit, the content of the tweet is copied. This mechanism allows a Twitter user to rebroadcast status updates of other users to their followers, which represents a powerful way of increasing the reach of the original message. ⁸ As the original message is not altered, the communication is maximally cohesive. For the analysis, for each user, we count which other user(s) they retweeted and how often. We standardize these counts by dividing them by the overall number of retweets by a user. ⁹ A connection in the network is then given whenever a user retweeted another user at least once. The weight—or strength—of this retweet connection is determined by the standardized retweet counts. ¹⁰

To compute the second indicator of programmatic coherence, we calculate the similarity between party-affiliated accounts’ communication as follows. First, we machine translate the tweets from all three languages (French, German, and Italian) into English. ¹¹ Second, we build a weighted bag-of-words representation ¹² from the combined tweets of every user. ¹³ In a final step, we compute the cosine similarity between the word distributions of all users.

We use two statistics from social network analysis to assess the degree of unity across the three indicators just introduced (see Table 1). All four indicators have the structure of a social network, although they consist of slightly different types of network data. The first indicator, based on the follower network, is a directed unweighted network, where directed means that the relationship between two network members can—but need not be—reciprocal. The indicator constructed from the retweets has an even more obviously directed nature, but the relationships are weighted by the fraction of retweets. Finally, the text similarities form an undirected and weighted network.

We use the density and the betweenness centralization as the main indicators of two different aspects of within-group unity. As we are interested in intraparty unity, these network statistics are computed on party-specific subnetworks. To wit, we look at each party individually, only considering the connections between its affiliated accounts and ignoring all connections to accounts of other parties. ¹⁴

The density of a network is the fraction of possible connections that are actually present (e.g. Wasserman and Faust, 1997). One important caveat to keep in mind when comparing the densities of social networks of disparate sizes is that, all else being equal, smaller-sized networks tend to have higher densities than larger-sized networks (Scott, 2017). In our interpretation of the results, we will therefore pay attention to comparing only the densities of parties with similarly sized networks. ¹⁵

The second statistic we compute is the betweenness centralization coefficient. Relative to other definitions of centrality (see Freeman, 1979), betweenness centrality is conceptually closest to the type of within-group unity we discuss above, because the betweenness centralization coefficient can be understood as a measure of hierarchy, that is, how strongly direct connections in the network depend on a small set of actors (Freeman, 1977, 39). The betweenness centralization coefficient is computed from the actor-level betweenness centrality, by averaging the differences between the most central actor and every other actor in the network. This number is then standardized to a range of one unit, such that it is 0 when every actor in the network has the same betweenness centrality and 1 when the only connections in the network are between a single actor and every other network member.

We obtain uncertainty estimates for the two network statistics from the nonparametric bootstrap. Resampling a subset of the network under study demonstrates the sensitivity of our results to the network boundaries (Costenbader and Valente, 2003; Galaskiewicz, 1991). Although we compute network statistics for party-specific subnetworks, the identification of the network sample happens at the level of the full network. Therefore, resampling is also done for the whole network. For the main analysis, we report uncertainty estimates from 1000 repetitions of resampling 95% of the network. ¹⁶

Results and discussion

Network structure

The empirical analysis starts with an initial overview of the network structure of the Swiss parties’ Twitter campaign in the run-up to the Federal Election of 2015. Figure 1 presents a visualization of the political Twitter network in play for the 2015 Swiss federal election campaign. The general arrangement of the nodes and communities is based on two nested Fruchtermann–Reingold layouts—one applied to the communities and a second applied to the nodes within their community. The communities were detected via the cluster algorithm for large networks proposed by Clauset et al. (2004). An overview of the most important characteristics of the communities is provided in Table 2.

OPEN IN VIEWER

Figure 1.

The political Twitter network in the 2015 Swiss federal election campaign. The nodes indicate the Twitter accounts and the edges show their follower relationships. The size of the nodes signifies an account’s betweenness centrality.

OPEN IN VIEWER

Table 2.

Relevant^a characteristics of the communities shown in Figure 1.

Party	BDP	CVP	FDP	GLP	GPS	SPS	SVP
Community 1	2.8	4.0	1.6	5.7	65.8	69.3	1.5
Community 2	6.9	16.6	2.0	4.1	32.3	29.1	1.9
Community 3	0	47.1	1.6	88.5	0.6	0.3	3.4
Community 4	90.3	32.3	94.7	1.6	1.3	1.4	93.2
Language	DE	EN	FR	IT
Community 1	38.9	29.3	5.5	16.3
Community 2	1.0	8.6	45.5	4.7
Community 3	23.9	19.8	1.9	7.0
Community 4	36.3	42.2	47.1	72.1

Party abbreviations: BDP: Conservative Democratic Party; CVP: Christian Democratic People’s Party; FDP: the Liberals; GLP: Green Liberal Party; GPS: Green Party; SPS: Social Democratic Party of Switzerland; SVP: Swiss People’s Party; Language abbreviations: DE: German; EN: English; FR: French; IT: Italian.

^a Only tabulations are shown for the main indicators whose relationship with the community memberships is significant in a χ² test. Frequencies in %.

The size of the vertices in Figure 1 represents the betweenness centrality of the party accounts. Generally, the inequality in this centrality reflects the rather hierarchical nature of the Swiss party network. Although roughly half of the accounts are highly central—typically party presidents, Federal Councilors, National Councilors, Councilors of State, or accounts of the national party offices—there are also many peripheral accounts with only sparse connectivity to the network.

The four communities reflect the two most salient divides characterizing the Swiss political Twittersphere. On the one hand, there is a clear ideological left–right divide within the network (see Table 2). Though the first two communities are clearly shaped by politicians from leftist parties—SPS and GPS, community 3 is almost exclusively occupied by centrist parties—CVP and GLP. Community 4, finally, mirrors the right pole of the spectrum, as the overwhelming majority of politicians from the FDP and the right-wing populist SVP are found in this community.

A second fundamental, and quite particular, characteristic of political communication in Switzerland is also reflected in the community structure displayed in Figure 1. The Swiss media system is generally separated according to language borders, which is manifested by different private media outlets, and also expressed by separate public broadcasters for every language region. The social media platform Twitter is no exception to this (see Table 2). Although Twitter accounts are not as strongly separated by language as by ideology, the decision remains significant. All other indicators for the accounts, such as gender or political function, do not significantly induce differences among the communities.

The language-related fragmentation in the Twittersphere even operates within single parties. The French-speaking accounts from most parties are, accordingly, mainly grouped into community 2, while their German- and Italian-speaking party colleagues have a stronger presence in communities 1–3.

Organizational cohesion

The analysis proceeds with an examination of the follower network in Figure 2(b), which reveals several interesting patterns for the size of the Twitter network, its density, and betweenness centrality. ¹⁷ First, it seems obvious that the cantonal and local follower networks are bigger than the national networks, simply reflecting the upwardly narrowing hierarchy of the party organizations. The local networks, however, are not always larger than the cantonal ones. Cantonal parties are as important, sometimes even more important than the local ones, for the BDP, GLP, GPS, SVP, and CVP. This result may be partly due to our sampling strategy, which started with a federal-level sample of users. However, it might alternatively reflect the importance of the cantonal level for Swiss party politics.

OPEN IN VIEWER

Figure 2.

(a) Organizational cohesion over all levels. (b) Organizational cohesion by level.

Network density provides information on the horizontal component of organizational cohesion—or in other words, on the average connectivity in the network. The results for the densities of the follower networks provide strong supportive evidence for our first expectation that it is tougher or less desirable for parties to control the campaigning of lower-level politicians. Although the network densities of the cantonal and local networks are close to the overall density across all parties, the density of the national network is clearly higher for all parties except the SVP. Hence, the national networks in the Swiss political Twittersphere tend to be much more tightly connected in organizational terms. By contrast, there seems to be more room for independent network building at the lower levels of Swiss politics.

As briefly discussed in the section on measurement strategy, it is always necessary to acknowledge the positive correlation between the density and the size of the network if the network densities are compared across parties. ¹⁸ With this in mind, the density of the SVP user network should be much higher, relative to its respectable size. In terms of the follower relationships among its accounts, the SVP is evidently only loosely organized. The clearest outlier at the other end, however, is the national network of the SPS. With a Twitter network that is very large, the Social Democrats are able to maintain a level of connectivity that rivals parties with much smaller-sized networks. This is a clear indication that the SPS is more highly organized than the other parties.

The second network measure, betweenness centralization, indicates the degree to which a party is centered around a few highly connected users. It is therefore suited to uncovering the hierarchical component of organizational cohesion. As the results in Figure 2(a) show, the larger parties in terms of their follower networks are generally more hierarchically structured than the smaller parties, especially the GPS.

Two further patterns are noteworthy. First, at the national level, the networks of the three Federal Council parties, CVP, FDP, and SPS, are among the least hierarchical. This probably reflects the consensus-oriented nature of federal politics in Switzerland, with pragmatically, rather than ideologically driven policy coalitions. Similarly interesting is the result for the national SVP, the only other party that regularly elects representatives to the Federal Council. It is a distant outlier from the general trend, which seems to confirm its exceptional position in the political system of Switzerland. The second interesting pattern is that the two parties that are farthest to the right in the political spectrum, the FDP and SVP, in general are more hierarchically structured than the other parties.

Programmatic coherence

The results for measuring programmatic coherence via retweets and tweet similarities are displayed in Figure 3(a) and (b). As for the densities of the retweet networks, the parties can be split into one large group with broadly similar network sizes, accompanied by two outliers. The BDP is the downward outlier with the smallest network. The density of its retweet network, however, is not significantly larger than the density of the retweet networks of the other parties. This indicates that Twitter users from the BDP use retweets comparatively less to refer to the content of their party-affiliated accounts.

OPEN IN VIEWER

Figure 3.

(a) Programmatic coherence: R-tweets over all levels. (b) Programmatic coherence: Retweets by level.

The large group of medium-sized parties includes the GPS, GLP, SVP, CVP, and FDP. The network sizes of all of these parties are quite similar, despite smaller differences. The densities of their retweet networks are also similar, with the minor exception of the CVP. Relative to the size of their network, the users of this centrist-conservative party are only weakly connected in horizontal organizational terms, not only (but most notably) at the local level. The upward outlier is the SPS, which has about twice as many users in its retweet networks as the parties in the large middle group. With associated densities only slightly smaller than the counterparts from this comparison group, the SPS, despite its larger-sized network, seems to be much more horizontally organized relative to those five parties.

Turning to the hierarchical structure of the retweet indicator, a general difference among the betweenness centralization in the follower networks is that it tends to be more weakly connected at the local level. In five of seven cases, the difference is significant and the national level displays higher values in terms of betweenness centralization. In the analysis on the follower networks, the local level users, if they exist, stand out as more hierarchically structured, and higher-level users have more organizational cohesion. In terms of programmatic coherence, the accumulation of centrality by a few users is more pronounced at higher political levels.

Moreover, there are basically two groups of parties exhibiting broadly comparable levels of hierarchy in their retweet networks. The first group includes the BDP and GLP, two centrist parties with a smaller network size, as well as the CVP, FDP, and SPS. The retweet networks of all these parties are characterized by a strong hierarchy at the national level and much less programmatic coherence at the cantonal and local levels. Local and cantonal users from these parties are thus substantially freer in their retweet behavior than their national counterparts, who seem to coordinate their campaigns with much more effort.

The other group of parties, the GPS and SVP, have neither high centralization scores generally nor large variation across the political levels. This indicates that there are no Twitter users in these parties setting the pace for the other party-affiliated accounts. Because there is evidently less pressure to spread the party message coherently, there is potential for single users to strategically deviate from the party line. This is somewhat surprising, as far as the SVP is concerned, since users from this party have shown exceptionally strong organizational cohesion.

Figure 4(a) and (b) displays the results from the second indicator of programmatic coherence, based on the text similarity of the tweets sent by party-affiliated Twitter users. In general, the national level of party politics is much more connected in terms of tweet similarities. As with the betweenness centrality of the retweet networks, coherence among the cantonal and local politicians is considerably weaker. This is a clear indication that the national campaign message is adapted to regional peculiarities in all parties.

OPEN IN VIEWER

Figure 4.

(a) Programmatic coherence: Tweet similarity over all levels. (b) Programmatic coherence: Tweet similarity over all parties by level.

It is worth noting the lack of an implicit relationship between the density and the size of the network in this case. However, for consistency, we do include the network size in the graph. The reason for the independence between the two statistics is that the density, in this analysis, simply reflects the average distance between two users from a given party. ¹⁹

With the ability to compare across all parties in the analysis, the smaller parties—BDP, GPS, and GLP—are revealed as having a comparatively high degree of text-similarity-based coherence. The bigger parties—SVP, CVP, FDP, and SPS—tend to have a lower degree of programmatic coherence in this regard. This difference between the two mentioned groups of parties is especially pronounced at the national level, although national users from the SPS exhibit similarly high levels of consistency, relative to the national users from the GPS, GLP, and BDP. For the other larger parties, most notably the CVP, there is more variation in the campaign message. A possible interpretation is that these parties must cater to a more diverse constituency during the election campaign. Hence, Twitter seems to be a feasible campaign platform for addressing this increased diversity.

Two of the smaller parties, GPS and BDP, tend to have a slightly more hierarchical structure in their text similarity networks, as the final graph on the betweenness centrality shows. The other parties have similarly low levels of centralization. Although this pattern does not hold for the GLP, it nevertheless seems to be the case that the communication of smaller parties across all political levels is more heavily centered on a few users. The Twitter communication of bigger parties, by contrast, is more evenly distributed across the users.

Conclusion

Existing studies on party unity in Switzerland have almost exclusively focused on representatives at the federal level (see e.g. Traber et al., 2014). We argue that these findings can only grasp part of the story, since party unity in Switzerland is constantly under strain by regional divisions across the different levels of government. Using data on Twitter accounts, we can extend considerably the number and types of politicians considered in the analysis and thus draw conclusions beyond the narrow national realm. Such a research strategy therefore seems generally promising for studying party unity in other strongly federalist states such as Germany or Belgium, where subnational party factions are comparatively powerful. Hence, we can expect similar patterns across the electoral campaigns in such comparable countries.

In a first step, our descriptive analysis largely confirmed common knowledge about “offline” politics in Switzerland. On the one hand, we found a clear congruence between the parties’ number of Twitter accounts and their strength in the national parliament. On the other hand, the general ideological left–right divide of the Swiss political system as well as the usually clear language-specific separation of the Swiss media system have become evident in how the Twitter accounts grouped into network communities. These findings are also consistent with qualitative evidence stemming from interviews with election campaign managers in the run-up to the 2015 national elections (Fiechter and Kohler 2015).

In a second step, we showed that our analysis is, at least to some extent, able to disentangle organizational and programmatic coherence. This is something that previous studies on party unity that focus only on single indicators, such as roll-call votes, cannot achieve. More precisely, we use the follower network as an indicator of organizational cohesion and the retweet and text similarity networks as indicators of programmatic coherence. Computing the density and the betweenness centralization allows us to quantify and compare programmatic and organizational unity across parties and levels of government, while the nonparametric bootstrap allows us to incorporate uncertainty to these measurements.

We find that, with respect to most statistics on organizational and programmatic unity, the national levels of the party networks tend to be more tightly connected. Hence, it is clearly tougher for parties to maintain unity in their networks at lower levels. Because cantonal and local peculiarities are often important factors in the political campaigns of lower-level politicians, it is probably even desirable for national party offices to allow them more room to maneuver. Local and cantonal users from these parties are thus substantially freer in their following, retweeting and general text messaging than their national counterparts, who seem to coordinate their campaigns with much greater effort. This evidence is almost certainly also due to the nature of the social media platform Twitter. With the brevity of its messages and high-speed dissemination, Twitter promotes a very heterogeneous communication. It might therefore be deliberately chosen by parties to leave room for dissenters, who, in turn, might be able to address voters outside their traditional constituencies.

Our comparisons across parties revealed that larger parties are generally less hierarchically structured, in organizational terms, than the smaller parties. This does also hold for the text similarity networks, where the accounts of smaller parties are significantly less equally distributed. One important exception is the national level of the SVP. This seems to confirm its exceptional position as a highly populist and disruptive force in the political system of Switzerland. Among others, this party is well-known for concentrating on a few media-savvy figureheads.

Generally, however, the Social Democrats are responsible for the most remarkable results. Despite their very large network, the Social Democrats are able to maintain a very high level of connectivity in terms of follower relations and retweets. The SPS is therefore much more organizationally and programmatically cohesive than the other parties. In these times, when Social Democratic parties are struggling in the context of election campaigns throughout Western Europe, this skillful use of Twitter by the SPS may provide a glimmer of hope.

Switzerland is not a particular front-runner with respect to the digitalization of its election campaigns, but also here, all parties except the CVP claimed to have spent considerable financial resources on Twitter during the 2015 national election campaign (Fiechter and Kohler, 2015). Moreover, each party offered training sessions and workshops on how to use social media. Since then, more politicians have joined Twitter and communicating on social media has become more important. Based on our experience from Switzerland, we can therefore expect social media to become more influential as a complement to more classical campaign strategies also in other established democracies. This offers a wide variety of opportunities for us to extend our research into organizational cohesion and programmatic coherence in further countries and other social media.

Supplemental Material