1. Are politicians able to shape public discourse?

Despite a growing body of work on the usage of moral appeals by political elites (Jung, Reference Jung2020; Bos and Minihold, Reference Bos and Minihold2022; Kraft and Klemmensen, Reference Kraft and Klemmensen2023), research has yet to address how societal actors influence one another to moralize. We argue that politicians wield significant rhetorical power to shape public discourse, first, because they occupy leadership roles offering greater weight to politicians' voices than to those of many other (individual) actors in society. Second, despite social media platforms such as Twitter democratizing public expression (Papacharissi, Reference Papacharissi2004), politicians generally have larger followings, amplifying their rhetorical impact. Third, while many members of the public might not consciously try to affect other people, let alone politicians, the fact that politicians are on a mission to do just that likely makes their language more carefully crafted with rhetorical persuasion in mind (Triadafilopoulos, Reference Triadafilopoulos1999).

As mentioned, we distinguish between moral and moral-emotional rhetoric. While both make appeals to notions of fundamental right and wrong, the latter overlays these appeals with potent feelings, such as anger, hate, or pride. Within each of these types, we further consider whether the appeals are positive or negative. Prior work suggests that moral-emotional language is more effective in increasing retweets than moral language alone, with mixed evidence for the relative power of positive and negative variants (Brady et al., Reference Brady, Wills, Jost, Tucker and Van Bavel2017). As our study examines the diffusion of rhetorical styles, we remain open as to whether H1 applies more strongly to some styles than others.

2. Are challenger parties more rhetorically influential than others?

Exploring the rhetorical influence of different party types, we focus on the distinction between mainstream and challenger parties (De Vries and Hobolt, Reference De Vries and Hobolt2020). Challenger parties, defined as political entrepreneurs without government experience, aim to disrupt mainstream parties' dominance, employing a dual strategy consisting of policy- and rhetorical innovation. While policy innovation, particularly from the radical right on immigration, has been studied extensively, less attention has been paid to rhetorical innovation. We extend De Vries and Hobolt's (Reference De Vries and Hobolt2020) argument that challenger parties use anti-establishment rhetoric to weaken the competence advantage of established parties by considering how moral appeals can challenge “ordinary” political talk.

Thus, while moral rhetoric is part of standard political vocabulary, used by parties across the board (Jung, Reference Jung2020), challenger parties may stand out by combining moral rhetoric with emotional appeals that contribute to “firing up” moral judgments (Garrett, Reference Garrett2019). Because of norms of civility and temperance, mainstream parties are likely to be restrained in their use of strong, especially negative, moral-emotional language (Widmann, Reference Widmann2021). In turn, since this rhetoric will seem out of place when used by mainstream parties, they are unlikely to be successful in gaining resonance with it among the public. These dynamics create an opening for challenger parties to disrupt “ordinary” political talk by provocatively adding emotional weight to moral statements. Considering its potency in drawing boundaries between groups (Brady et al., Reference Brady, Wills, Jost, Tucker and Van Bavel2017; Reference Brady, McLoughlin, Torres, Luo, Gendron and Crockett2023), we expect negative moral-emotional rhetoric to be a particularly powerful tool for challenger parties to influence public discourse.

3. Case: German immigration discourse

We investigate the diffusion of moral and moral-emotional appeals in German immigration discourse. Typically considered a highly moralized topic, immigration should allow us to observe the rhetorical phenomenon of interest. At the same time, research demonstrates fluctuations over time regarding parties' use of moral language on immigration, as well as the degree to which the public sees it in a moral light (Simonsen and Bonikowski, Reference Simonsen and Bonikowski2022; Simonsen and Widmann, Reference Simonsen and Widmann2023). This variation suggests that morality is not a stable characteristic of the issue but something that can be affected by political (or other) dynamics. In Germany, immigration has remained central in political and public debate since 2015/2016 where the country received high numbers of migrants and refugees. During the period of study, national and international media portrayed Germany as a moral leader on the topic in Europe (Deutsche-Welle, 2016), raising questions about the relative rhetorical influence of different actors.

The research period runs from September 2017 until the end of 2021 (covering the 19th legislative period of the German Bundestag). This was a period of continued debate about immigration; however, it does not include the height of the “refugee crisis” of 2015/2016. Earlier periods are excluded as the German radical-right challenger party AfD (Alternative für Deutschland) is central to our investigation of H2, and the party did not enter parliament until 2017 (the study's politician population is limited to elected parliamentarians). Although the “refugee crisis” might have provided momentum to moralize the topic, the focus of this study on the succeeding years offers insights into rhetorical dynamics outside of the immediate crisis moment, which arguably characterizes most political and public debate. We believe, therefore, that the results of our study have the potential to generalize to other salient topics.

4. Data and topic identification

Data from Twitter allow us to analyze public, media, and political rhetoric through a single source. The high granularity of the data (with daily communication) enables us to closely track rapidly changing temporal rhetorical patterns. Consequently, we can precisely assess the extent to which different actors employ moral rhetoric in response to changes in other actors' rhetorical styles.

Data were collected using the academictwitteR package by Barrie and Ho (Reference Barrie and Ho2021) and include three samples of societal actors: the political sample (all available accounts of MPs and parties in parliament, 29,073 tweets), a sample of the attentive public (10,634 tweets), and a media sample (13,278 tweets). For more details, see online Appendix A. To examine H2, we split the political sample into two, comparing the radical right challenger AfD to all other parties. With the topic of immigration, AfD is considered a challenger party because it alone represented a disruption of the pro-immigrant policy consensus characterizing German politics at the time (Gessler and Hunger, Reference Gessler and Hunger2021). Although the radical-left party Die Linke had also never held government power, we group it with the other mainstream parties because of its policy position.

Using a dictionary approach to identify the topic of immigration, we created a keyword list to filter tweets including immigration-related terms (see online Appendix B). After filtering, our study includes a total of 53,019 tweets from 1648 different accounts.

5. Computational text analysis

To measure moral and moral-emotional appeals in political text, we use the transformer-based classifier fine-tuned by Simonsen and Widmann (Reference Simonsen and Widmann2023). The model can reliably classify moral appeals, emotional appeals, and valence (positive or negative) in up to 100 languages. Online Appendix B provides details on the transformer model and reports performance metrics.

We use this transformer model to classify all immigration tweets in the study. Based on the model classification, we create the variables of interest. These include moral appeals (i.e., tweets that refer to fundamental beliefs about right and wrong without concomitant emotional content), and moral-emotional appeals (i.e., tweets where references to fundamental beliefs about right and wrong are combined with emotional appeals), both of which are subdivided into positive and negative. The distinction between positive and negative variants relates to whether the tweet emphasizes moral virtues (positive) or vices (negative). This means that the valence of a tweet is conceptually and empirically distinct from the emotionality of the tweet, which instead relates to the expression of emotions such as anger, sadness, joy, or pride. Daily proportional scores ranging from 0 to 1 serve as endogenous variables in our SVAR analysis.

6. Structural vector autoregression analysis

The vector autoregression (VAR) model analyzes a set of stationary time series where α represents the constant, β coefficients capture the effects of each rhetorical style Y _i,t, representing the proportion of daily moral(-emotional) appeals for each group i (politicians, public, and media) on day t. We design the VAR model to delineate the autoregressive and endogenous relationships among these variables, formulating each variable Y _i as a function of its own lags and those of other variables. The main model can be formally expressed as follows:

$$Y_{i, t} = \alpha + {\rm \;}\mathop \sum \limits_{\,p = 1}^P \beta _pY_{i, t-{\rm p}} + \varepsilon _{i, t}$$

We use the VARselect function of the vars R package (Pfaff, Reference Pfaff2008) to determine the optimal lag structure based on two selection criteria (see online Appendix C). Online Appendix E replicates the analyses with a different lag structure.

After calculating different VAR models, we proceed to fit SVAR models based on the VAR models. In comparison to VAR models (Pfaff, Reference Pfaff2008), which account for the endogenous variables solely through their own history, SVAR models enable the explicit modeling of contemporaneous (i.e., same-day) interdependence among the variables, which is crucial in dynamic social media contexts.

Running a model for each rhetorical style, the simple SVAR models (Figure 1, testing H1) include three actor groups: politicians and parties, attentive public, and news media (as a control variableFootnote ¹). For the complex SVAR models (Figure 2, testing H2), we split the politics group into two according to their belonging to the radical right challenger party AfD. Following Pickup and Kellstedt (Reference Pickup and Kellstedt2023), we investigate the balance of our time series equations to ensure valid statistical inferences. Online Appendix F shows that our models result in a balanced vector autoregression with white noise residuals.

Figure 1.

Simple, cumulative IRFs: predicted rhetorical appeals across the public and politics group. IRFs (with 1000 simulation runs) are based on four SVAR models (one for each rhetorical style) with a one-day lag structure, illustrating the predicted, cumulative change over the seven subsequent days in public discourse after a unit shock on day 0 in political discourse (left panel) and in political discourse after a unit shock on day 0 in public discourse (right panel).

Figure 2.

Complex, cumulative IRFs: predicted rhetorical appeals across different actor groups. IRFs (with 1000 simulation runs) are based on four SVAR models (one for each rhetorical style) with a one-day lag-structure, illustrating the predicted, cumulative change over the seven subsequent days in public discourse after a unit shock on day 0 in different parties' discourse (left panel) and in parties' discourse after a unit shock on day 0 in public discourse (right panel), i.e., y-axis denotes the impulse group and the title denotes the response group.

The results of the SVAR models are best expressed using cumulative impulse response functions (IRFs). These cumulative IRFs indicate how a “unit shock,” i.e., increase of one standard deviation of the disturbance terms of the different variables by one group predicts the change in rhetorical style of other groups, cumulatively over seven days. We calculate and report orthogonal, cumulative IRFs based on 1000 simulation runs.