Article

 These studies offer multiple valuable insights, but one crucial aspect has largely not yet been reflected. Political communication is a rapidly changing field with new (digital) channels developing constantly. Thus, dynamics of political communication and party competition are no longer restricted to particular channels. Political actors use the ever-growing number of channels to communicate policies and connect with different audiences and social groups. Hence, party competition takes place within and (potentially) across multiple venues simultaneously, with potential implications for public discourse, voting behaviour and democratic representation. Most existing research, however, exclusively studies one particular communication channel in isolation. Comparative research is rare and limited to specific (short) time periods, such as during election campaigns (Elmelund-Præstekær Citation2011; Green and Hobolt Citation2008; Norris et al. Citation1999; Tresch et al. Citation2018). It remains largely unclear to what extent and – most importantly – why parties adapt their behaviour according to different communication channels. Do parties’ issue salience strategies change depending on the channel used? If so, why?

 Existing comparative research does offer some insights, but significant gaps remain. While Norris et al. (Citation1999), Elmelund-Præstekær (Citation2011) and Tresch et al. (Citation2018) find differences in issue salience across multiple channels, this is not the case for Green and Hobolt (Citation2008). Crucially, we still lack a coherent theoretical framework to detect the precise factors that influence parties’ issue salience strategies in different channels (Elmelund-Præstekær Citation2011). Grasping how parties use different channels is key to understanding the dynamics of party competition in the rapidly changing political communication environment of contemporary democracies.

 This article contributes to the literature by studying the extent to which parties’ issue salience changes depending on the communication channel and why. I argue that issue salience is influenced by the characteristics of communication channels. More specifically, depending on the channel, parties focus on their issue preferences (ideal agenda) to different degrees. Three factors should be important.

 First, parties reach different audiences through various communication channels. While mediated channels are primarily aimed at journalists (e.g. press releases), others allow parties to connect directly with the public (e.g. social media). Second, party communication can be centralised in the hands of the party leadership and central office (e.g. official party press releases or social media posts from party accounts) or decentralised (e.g. social media posts from individual party members). Third, communication in some channels is structured by some sort of pre-given structure or agenda (e.g. legislative agenda). Therefore, I differentiate pre-structured (e.g. parliamentary speeches) and non-pre-structured communication channels (e.g. tweets, press releases). I expect these three factors to moderate the influence of party preferences on issue salience in the respective communication channels.

 Methodologically, I use an advanced text-as-data technique to analyse a broad range of texts produced by political parties. I train a transformer-based model (BERT) on labelled manifestos and apply it cross-domain to classify press releases, parliamentary speeches and tweets from parties and individual party members into issue categories. The study covers the cases of Austria, Germany and Switzerland between January 2019 and September 2021. Overall, the data set consists of more than 41,000 parliamentary speeches and 34,000 press releases, nearly 72,000 tweets from party accounts and more than 420,000 tweets from individual party members.

 The empirical results show that political parties’ issue salience is influenced by the communication channel. I observe different issue agendas in each examined channel and find evidence that salience is moderated by specific channel characteristics. Party preferences have a greater influence on issue salience in centralised communication channels, but play a smaller role in pre-structured channels. Both observations follow the theoretical expectations. This is, however, not the case for mediated vs. direct channels; here, the results deliver no statistically significant difference.

 These findings have several implications and underscore the importance of studying different sources of party communication. First of all, this article shows that a single, unified political agenda does not exist. Political parties send different policy signals in different venues. This is driven by the nature and characteristics of communication channels. Furthermore, dynamics and patterns of party competition – such as the responsiveness to public opinion or the level of issue engagement between parties – may therefore also shift depending on the channel. This can result in different public perceptions of the parties and the competition between the parties. Interestingly, however, parties do not appear to adapt their communication significantly when the channel’s audience consists primarily of journalists. At first glance, this is a surprising and counterintuitive finding, but it actually fits hybrid media system theory. In hybrid media environments, journalists increasingly make use of alternative sources of information (e.g. social media) to learn about political processes (Chadwick Citation2017). Political actors, in turn, adapt their behaviour to this development and also address journalists in direct channels, such as on X, formerly Twitter. This modern combination of multi-channel and hybrid media environment, which simultaneously leads to a segregation and blurring of audiences, helps to explain why parties do not change their issue salience strategies considerably between mediated and direct channels.

 In the following, I will lay out the theoretical framework that captures the factors influencing party issue salience in different communication channels. Then, I will describe the data set and text-as-data approach used to study issue communication in diverse types of text. Finally, I will present the results and conclude with reflections on the broader implications of the findings.

 Political actors are subject to multiple sources of influence when it comes to communication strategies. For studying parties’ issue salience across different communication channels, two factors identified by Green-Pedersen and Walgrave (Citation2014) are especially relevant: preferences and institutions.

 First, political actors have certain preferences. In the case of parties – the unit of analysis in this article – issue preferences mainly stem from ideological and strategic sources. On the one hand, parties have certain issues that are closely connected to their ideology. The issue of the environment is, for example, at the core of Green party ideology, while Social Democratic parties have a strong ideological interest in labour and welfare state issues. Thus, parties have ideologically driven issue preferences. On the other hand, party preferences also result from strategic considerations related to issue ownership. Issue ownership theory suggests that parties ‘own’ certain issues, either because they are associated with the issue by the public or are regarded as the most competent on it (Walgrave et al. Citation2012). If the public sees a particular party as ‘better able’ to handle a specific issue than other parties, that party has ownership of that issue (Petrocik Citation1996). Thus, issue ownership scholars argue that ‘owning’ an issue brings advantages in party competition.

 Parties therefore try to raise the salience of issues ideologically or strategically important to them, while avoiding a direct issue-conflict with other parties (Budge Citation2015; Budge and Farlie Citation1983). This leads to a competition over the political agenda (Carmines and Stimson Citation1993; Green-Pedersen Citation2007). Based on such ideological and strategic preferences, parties develop a so-called ideal agenda and try to push it in the political debate. The ideal agenda reflects the importance of individual issues to a party and is best reflected in party manifestos (Budge et al. Citation1987; Norris et al. Citation1999).

 Second, political actors operate within various institutions, whose rules shape the amount of attention the actors can pay to an issue (Green-Pedersen and Walgrave Citation2014). Hence, institutions can be interpreted as structures, which influence party behaviour in issue communication. Institutions take multiple forms, and the definition of what constitutes an institution is strongly contested. One of the most influential conceptualisations argues that institutions consist of formal as well as informal rules, ranging from constitutional orders to simple conventions (Hall and Taylor Citation1996). Following from this definition, many different venues where political processes take place can be described as institutions. This also applies to political parties’ communication channels. These are venues wherein or instruments with which political parties and their members themselves communicate and discuss policies. I understand all channels that shape the public profile of parties, ranging from press releases published by central offices to parliamentary speeches and social media posts by individual party members, as party communication channels.

 While research has long focused on manifestos and press releases, recent work also points out and acknowledges the importance of other communication sources. For example, interest in social media is growing, specifically how both parties and individual politicians use it, how it transforms the relationship between parties and their members (e.g. MPs) as well as how it affects agenda setting dynamics (e.g. Gilardi, Gessler, et al. Citation2022; Peeters et al. Citation2019; Sältzer Citation2022; Silva and Proksch Citation2022). Furthermore, an increasing number of studies also finds that parliamentary speeches are another crucial avenue for parties and their MPs to send policy signals (e.g. Debus and Tosun Citation2021; Ivanusch Citation2023; Proksch and Slapin Citation2015). These studies show that several communication channels have become important tools for parties when it comes to issue communication. However, the different channels also possess distinguishing characteristics, rules and conventions that potentially influence political parties’ communication profiles (e.g. Dalmus et al. Citation2017; Elmelund-Præstekær Citation2011; Tresch et al. Citation2018). Thus, different communication channels can be viewed as institutions that create a structure governing the issue communication of parties and their members. I therefore expect party issue salience to be influenced by the communication channel and its characteristics.

 As mentioned above, parties usually aim to communicate their issue preferences, i.e. ideal agenda. However, communication channels and their characteristics provide structures that should moderate the amount political parties focus on issue preferences. What are these channel characteristics and how do the various channels influence party behaviour?

 Two relevant characteristics can be identified in the literature, namely the type of audience and the degree of control a party can exert over a given channel (Dalmus et al. Citation2017; Elmelund-Præstekær Citation2011). These theoretical considerations offer a strong foundation. Some adaptations are, however, needed. The framework set forth in the following differentiates three characteristics and postulates corresponding hypotheses (H1–H3). Figure 1 illustrates these hypotheses graphically.

 The first hypothesis relates to the differing audiences addressed by each communication channel. Certain channels allow parties or individual politicians to address the public and their followers directly, especially social media (Peeters et al. Citation2019; Popa et al. Citation2020). In such direct channels, parties can act (relatively) freely and I therefore expect them to communicate strongly according to their issue preferences. In contrast, press releases are a mediated communication channel. They are primarily aimed at journalists and rarely reach the broader public directly (Dalmus et al. Citation2017). Therefore, parties have to consider the needs and interests of journalists in press releases. This applies not only to formal criteria but also to the selection of issues addressed within a press release. Journalists are, for example, strongly interested in issues that are already salient in the media and among other important actors. In contrast, issues ‘owned’ by a party do not have a high news value (Dalmus et al. Citation2017; Meyer et al. Citation2020). Hence, I hypothesise that parties do not focus solely on their issue preferences but on a broader set of issues in mediated communication channels in order to meet the interests and needs of journalists.

 The second hypothesis is based upon communication channels varying in the degree of centralisation. In centralised channels, messages are sent by the party leadership or by the central office or at least have to pass through one or both of them. Here, the central and national organisation unit – the party in central office (Katz and Mair Citation1995) – has tight control over issue communication. In other channels this is not the case as individual party members communicate themselves. Examples of such decentralised communication channels include social media accounts of individual politicians. The degree of centralisation thereby has implications for a party’s issue communication on the whole as well as for its public profile, leading to an increasing research interest, particularly since the advent of social media. Therefore, the actual influence of centralised and decentralised communication channels on the profile of a party and its issue agenda is an important topic.

 According to Silva and Proksch (Citation2022), communication by individual party members (i.e. decentralised communication) may serve two purposes. On the one hand, decentralised communication can amplify central party messages, since individual politicians (particularly in systems with strong parties) have strong incentives to follow the party line (e.g. Kam Citation2009; Sieberer Citation2006) and parties simultaneously try to enforce unity (e.g. Proksch and Slapin Citation2015). On the other hand, decentralised communication can serve as a substitute for central party communication channels and represent an avenue to send a variety of policy signals (Silva and Proksch Citation2022).

 I argue that three factors are important here. First, strategic communication according to issue preferences requires coordination. Centralised communication in the hands of the central party office allows for better coordination and communication closer to the party line and issue preferences. Second, parties are in firm control in centralised channels, while decentralisation allows individual members to potentially circumvent partisan constrains and push their own preferences (Enli and Skogerbø Citation2013; Silva and Proksch Citation2022). Third, decentralised communication may also provide an incentive for parties to expand their appeal by focusing on a broader set of issues. Thus, decentralised channels may also be a strategic tool to focus less on issue preferences but on a variety of issues. Based on these considerations, I expect the communication of political parties to be more in line with issue preferences when communication is centralised than when it is decentralised.

Male Strippers Scottsdale

 The third hypothesis deals with a channel’s degree of pre-structuredness and the influence thereof on issue communication. In highly pre-structured channels, the topical focus is pre-given to a certain extent, and parties only have limited control over issue selection. Hence, issue communication is pre-structured. Examples of such an environment are parliamentary debates. Parliamentary debates are one of the most important arenas of political communication and a key tool for parties to send policy signals in party competition (Proksch and Slapin Citation2015). This also applies to the issues discussed in parliamentary speeches, as recent research finds that parties and their MPs use speeches to advance their issue preferences (Debus and Tosun Citation2021; Ivanusch Citation2023). However, parliamentary speech-making is also substantially influenced by the legislative agenda. Most of the time, bills or specific topics are debated in parliament (Proksch and Slapin Citation2015). Certain issues are given from the start; these in turn structure issue communication during parliamentary debates. This is the process that is encompassed by the concept of pre-structuredness. Such an environment makes it more difficult for parties and their members to communicate according to their ideal agenda. Moreover, parties may have different opportunity structures to influence the pre-structuredness of a channel. For example, partisan control over the legislative agenda often provides government parties with more influence over the issues discussed in parliamentary debates than opposition parties (Cox and McCubbins Citation2005; Döring Citation1995).

 These dynamics show that important party communication channels, such as parliamentary speeches, are subject to very specific institutional contexts. In these channels, issue communication is significantly pre-structured, restricting parties when it comes to discussing their issue preferences. Thus, although recent research shows that parties find at least some room to focus on their issue preferences in pre-structured channels (e.g. parliamentary speeches), I still expect issue preferences to be much less reflected here than in other communication channels.

 For the analysis, I draw on a data set comprising a variety of texts produced by political parties in Austria, Germany and Switzerland between January 2019 and September 2021. I rely on this case selection for three main reasons.

 First, the selected countries represent typical Western European multi-party systems while still allowing to control for potential variations stemming from different electoral systems and political cultures (e.g. government-opposition dynamics). While Germany uses a mixed electoral system, Austria and Switzerland use proportional systems. However, electoral districts and party lists play different roles in the electoral systems of Austria and Switzerland. Furthermore, Switzerland differs significantly from the other two countries when it comes to government formation and direct democracy. Switzerland has a strong tradition of consociationalism and therefore usually relies on a special formula (Zauberformel) to form a government. Additionally, Switzerland makes ample use of referenda, which is not the case in Austria or Germany.

 Second, the broad time period covered (1 January 2019–26 September 2021) allows for party issue salience to be studied at multiple points in time and in different phases of political communication. The selected period covers one election campaign per country (Austria: 2019; Germany: 2021; Switzerland: 2019) as well as ‘routine times of politics’. Furthermore, it includes several months before and during the Covid-19 pandemic to account and control for potential effects of this crisis on party behaviour.

 Third, all three countries studied are German speaking.Footnote1 As I employ quantitative computer-based text analysis, a mono-lingual analysis should ensure higher reliability and comparability. While comparing different types of texts is a significant challenge in itself, a multi-lingual analysis would create even more and yield potentially incommensurable results (Chan et al. Citation2020; Maier et al. Citation2022).

 The main text corpus used in this article consists of four types of party communication channels: press releases, parliamentary speeches, tweets from party accounts and tweets from individual party members. As an additional data source for the analysis, I use labelled manifestos from the Manifesto Project (Lehmann, Burst, Lewandowski, et al. Citation2022). This channel selection is well suited for the purpose of this study because of two main reasons.

 First, all the channels selected are important avenues for parties and their members to communicate with the public on a regular basis. Press releases are important and flexible tools for parties to inform journalists about specific issues and respond to daily developments (e.g. Dalmus et al. Citation2017; Klüver and Sagarzazu Citation2016). Furthermore, the content of press releases can potentially reach a large audience, if picked up by journalists (Hopmann et al. Citation2012; Meyer et al. Citation2020). Parliamentary speeches as well are an avenue for parties and their MPs to send policy signals (Proksch and Slapin Citation2015). Recent research shows that parties advance their issue preferences in parliamentary debates, making them an important tool in issue competition (e.g. Debus and Tosun Citation2021; Ivanusch Citation2023). How parliamentary speeches compare to other tools such as press releases is largely unclear, however. Tweets (and social media posts in general) are also frequently used by political actors to communicate with the public and to engage with or criticise political opponents (e.g. Gilardi, Gessler, et al. Citation2022; Russell Citation2018). For this case study, I choose Twitter, now X, for the social media channel because it is well suited to measure the content of broad national political debates. Previous research shows that Facebook, for example, is mainly used by political actors for (local) campaign-related purposes, whereas Twitter is the primary platform where contemporary political events are discussed on a national-level (Stier et al. Citation2018).

 Second, the chosen channels differ in a number of dimensions. Crucially, at least one channel is different to all the others for each of the three channel characteristics introduced above. These differences are displayed in Table 1. The channels are assigned values of 0 or 1 for each of the three channel characteristics. While press releases are primarily targeted to journalists (i.e. mediated channel), the other channels are aimed more directly at the general public. Although Twitter has a more ‘elitist’ audience compared to Facebook, its architecture (i.e. hashtags, retweets) facilitates the diffusion of political information across the platform (Stier et al. Citation2018; Wu et al. Citation2011). Therefore, Twitter allows political actors to communicate political information directly to a broad audience without having to rely on journalists as gatekeepers. In terms of centralisation, tweets from party accounts are fully centralised, but the platform also facilitates decentralised party communication via the accounts of individual party members. Press releases and parliamentary speeches are a special case in terms of the centralisation characteristic. While individual members (regularly) draft press releases and give the speeches in parliament, the party leadership or party office retains a certain amount of control over the content.Footnote2 In terms of pre-structuredness, parliamentary speeches stand out due to the legislative agenda structuring the context and content of parliamentary debates.

 The corpus covers the time period from 1 January 2019 to 26 September 2021 and was collected in the context of a bigger research project. The press releases contained in the corpus were published by the political parties and their parliamentary party groups (PPGs). For Austria and Germany, webscraping was used to download the press releases from a webservice of the Austrian Press Agency (https://www.ots.at/) and from the German party and PPG websites. In the case of Switzerland, the data was provided by the DigDemLab at the University of Zurich (Gilardi, Baumgartner, et al. Citation2022). The parliamentary speech data consists of an updated version of the ParlSpeech V2 data set (Rauh and Schwalbach Citation2020) for Austria and Germany and texts downloaded from the webservices of the Swiss parliament through the R package swissparl (Zumbach Citation2020). The tweets from party accounts (central office, PPG) and individual party members (party leaders, general secretary, all MPs) were collected through the Twitter Researcher API.Footnote3 Table 2 provides an overview of the complete corpus used for this study. Overall, the corpus consists of more than 571,000 individual documents.Footnote4

 Measuring issue salience in such a voluminous variety of texts is a significant challenge. In this article, I use an advanced text-as-data technique to study multiple types of text in a coherent and efficient way, namely cross-domain topic classification. Cross-domain learning is a way to reduce the necessary amount of training data and resources required for classic supervised approaches. Supervised models require labelled training data to learn about the specific task at hand, and labelling is resource intensive. With several different text types, as in this study, a very large amount of labelled training data would be needed for each individual text type. Cross-domain learning can mitigate this problem. The basic idea behind cross-domain learning is that models are only trained on a single type of text, but the trained model can also be applied to other types of text (e.g. Osnabrügge et al. Citation2021). This way, researchers only need to develop one training data set or can even use existing labelled data for training the model. Therefore, a huge potential for the application of cross-domain learning exists in political science and the social sciences in general.

 For the cross-domain topic classification, I rely on the state-of-the-art transformer-based model BERT (Devlin et al. Citation2019). BERT is elaborately pre-trained on vast amounts of unlabelled text and provides a very good general syntactic and semantic representation of words. To use BERT for a specific application, only some minor training (‘finetuning’) is necessary. For this training procedure, I adapt the approach developed by Burst et al. (Citation2023a, Citation2023b) and train the BERT model on labelled manifestos provided through the corpus of the Manifesto Project (Lehmann, Burst, Lewandowski, et al. Citation2022). The Manifesto Project (MRG/CMP/MARPOR) uses human coding to analyse party manifestos from all over the world according to a set coding scheme. The coders thereby assign each individual (quasi-)sentence from the manifestos to one specific category. Overall, the Manifesto Project coding scheme consists of 76 main codes plus one ‘NA’ category (code ‘000’). For this article, I use the labelled manifestos as training data and assign all categories from the Manifesto Project to 20 overarching issues.Footnote5 The final training data set used here consists of more than 728,000 annotated (quasi-)sentences in total.

 The annotated manifestos constitute a well-suited training data set for the BERT model.Footnote6 During training, the BERT model uses the annotated training data to learn about the specific task. In this case, BERT learns about the relationship between specific text features and issue categories via machine-learning. I then apply the trained model to the unlabelled texts of interest (press releases, parliamentary speeches, tweets from parties, tweets from individual party members). I use the model to classify each document into one of the issue categories specified in the codebook.Footnote7

 Compared to a manually coded gold standard,Footnote8 the BERT model achieves an accuracy of 58% for press releases, 59% for parliamentary speeches and 50% for tweets.Footnote9 These are comparatively good results for cross-domain topic classification of multiple categories (20 categories), especially applied to such a diverse set of texts as in this article.Footnote10

 The main goal of the analysis is to identify the effect of each individual channel characteristic on political parties’ issue salience. In order to achieve this, I use regression analysis on different samples and combinations of communication channels.

 Issue salience functions as the dependent variable in the analysis and is based on the results of the text analysis described above. It marks the percentage of attention a party devotes to a specific issue within a communication channel (e.g. press releases) over one quarter of a year. I choose to calculate issue salience by quarter as it allows a more reliable estimation than by month. Issue salience by month could be heavily influenced by external events and some channels do not produce consistent monthly communication.Footnote11 Thus, I use issue salience by quarter as the dependent variable in the regression analysis.