Establishment Responses to Populist Radical Right Challenges: the Politics of Moral Distancing ∗

Establishment Responses to Populist Radical Right Challenges: The Politics of Moral Distancing ∗

Frederik Hjorth University of Copenhagen†

This version: September 10, 2020

Ascendant right-wing populist parties sometimes challenge foundational norms of established political systems. In the face of such challenges, establishment actors face a choice: whether to employ a strategy of moral distancing, i.e. seeking to portray the challenger as morally objectionable. Existing research into this choice examines only party system- or party-level variation. I challenge this approach, using text as data methods applied to around 31,000 paragraphs of legislative speech to measure responses at the individual level. Revisiting an oft-studied case, responses to the entry of the right-wing populist Danish People’s Party, I uncover systematic individual-level, within-party variation. Speciﬁcally, I ﬁnd that more electorally secure and less issue constrained legislators engage in more moral distancing. The results suggest a previously overlooked role of individual-level factors in explaining establishment responses to populist challenges. This in turn may explain why system-wide exclusionary strategies against such challenges sometimes fail.

∗Special thanks to Mathias Osmundsen, Martĳn Schoonvelde, Sophia Hunger, Silvia Decadri, Sven-Oliver Proksch, Matthew Denny, Daniel Bischof, Charles Crabtree, Simon Munzert, Joshua Kertzer, and seminar attendants at the Center for Social Data Science at the University of Copen- hagen for comments on earlier versions of this article. †Øster Farimagsgade 5E, 1353 Copenhagen K, (+45)26272441, [email protected]

1 Political history is littered with phrases later regretted. Consider three pertinent examples: At a fundraiser on September 10, 2016, then-nominee for President Hillary Clinton offered the following assessment of the candidacy of Donald Trump: “you could put half of Trump’s supporters into what I call the basket of deplorables. Right? The racist, sexist, homophobic, xenophobic, Islamophobic — you name it” (Chozick, 2016). In an April 2006 interview, UK’s Tory party leader David Cameron dismissed members of the populist radical right UK Independence Party (UKIP) as “fruitcakes, loonies, and closet racists” (Ford and Goodwin, 2014, p. 71). Lastly, in a 1999 parliamentary debate, Denmark’s social democratic Prime Minister Poul Nyrup Rasmussen, speaking directly to Pia Kjærsgaard, the leader of the then-ascendant populist radical right Danish People’s Party, declared that “no matter how much effort you exert, in my eyes you will never be housebroken” (Downs, 2002). The targets of the three statements by Clinton, Cameron, and Rasmussen all enjoyed subsequent political success. Two months after Clinton’s speech, Donald Trump had won the election, with the “deplorables” remark having become a rallying cry among his supporters. A decade after Cameron’s comment, UKIP led the Leave campaign in the UK’s Brexit referendum, setting in motion UK’s exit from the European Union. And in the first election after Rasmussen’s dismissal of the DPP as “not housebroken”, the DPP surged to 12 percent electoral support, effectively keeping the Social Democrats out of power for a decade. But the three remarks share a more important feature. They all represent instances of a particular political moment: a right-wing populist actor appears on the scene, and political elites respond by denouncing the newcomer as violating widely shared democratic norms. In this article, I study how individual establishment actors choose this type of response, a rhetorical strategy I call “moral distancing”. I build on a rich existing comparative party politics literature dedicated to this question at the level of parties or party systems, but contribute to the literature by developing a novel approach with which to characterize response strategies with a previously unseen level of granularity. Focusing empirically on an oft-studied case, the rise of the right-wing populist Danish People’s Party (DPP), I uncover considerable within-party variation in

2 response strategies. Specifically, I find that within parties, moral distancing is more common among legislators with high electoral security and among party leaders and non-members of the government cabinet. I interpret the latter two differences as reflecting fewer constraints in terms of which issues to emphasize. The effects of electoral security and issue constraint persist when examining only within-party variation, demonstrating that the existing party-level perspective is incomplete. These results jointly indicate that in spite of its ostensibly moral quality, moral distancing is sensitive to the constraints and opportunities afforded by legislators’ political characteristics. The article contributes to a number of ongoing research agendas. First, I contribute to the existing literature on responses to the populist radical right by highlighting the role of individual- level dynamics in shaping establishment responses to populist challenges, a dynamic previously examined only at the level of parties or party systems. In doing so, I connect two otherwise disjointed strands of scholarship, namely the literature on party responses to right-wing populist challenges (Meguid, 2005; Bale et al., 2010; van Spanje, 2010; Heinze, 2018) and the literature on legislator behavior and representational style (Proksch and Slapin, 2012; Grimmer, 2013; Back, Debus and Muller, 2014; Slapin et al., 2018; O’Grady, 2019). Second, these results in turn add to our understanding of why system-wide exclusions of radical right challengers, so-called ‘cordons sanitaires’, sometimes break down or fail to emerge (Downs, 2002). The existing literature has generally explained cordons sanitaires with reference to party system-level strategic interaction. When radical right parties are small and ideologically distant, mainstream parties tend to erect cordons sanitaires (van Spanje, 2010); conversely, when other majority options are unavailable, mainstream right parties have an incentive to break the cordon (Bale et al., 2010). However, these arguments all presume parties are able to pursue their strategically optimal response. This article calls attention to the role of collective action problems within parties: with legislators facing different incentives and constraints, parties may fail to coordinate on their preferred response strategy. Lastly, I develop and apply a set of methods with applicability to a wide range of other substantive questions. By applying machine learning methods to speeches from the parliamentary record, I am

3 able to characterize behavior at the legislator level which has previously been conceptualized and observed solely at the level of parties or party systems. With the increasing availability of databases of parliamentary speech (e.g., Rauh and Schwalbach, 2020) and growing use of machine learning methods in political science (Hindman, 2015), this approach allows for peering “inside the black box” of parties (Müller, 1997). In doing so, the approach enables researchers to revisit a host of theories of party politics to directly compare party strategy and legislator behavior.

1 Understanding Establishment Responses to Populist Challenges

Several political systems in the Western world have seen the sudden rise of populist movements in recent decades. Among these, arguably the most prominent is a party family often referred to as ‘populist radical right’ (PRR) (Ivarsflaten, 2008; Mudde, 2010), members of which include France’s Front National, UK’s United Kingdom Independence Party, and Germany’s Alternative für Deutchland (Akkerman, de Lange and Rooduĳn, 2016). PRR parties are typically defined as parties combining populist, nativist, and authoritarian ideas (Mudde, 2007). This ideational mix has proven to be a strikingly politically popular recipe. In a 2017 tally of the political performance of PRR parties, they received around 16 percent of the overall vote in the most recent elections across Europe, up from just 5 percent two decades earlier (Tartar, 2017). Political science has responded to this trend by producing an abundance of research into the causes and consequences of the political success of PRR parties. The bulk of this research has examined the electoral causes of PRR success, with Betz (1993) famously identifying “losers of modernization” as the key PRR constituency, a theme that recurs in contemporary scholarly debates on economic (e.g., Colantone and Stanig, 2018) contra cultural (e.g., Norris and Inglehart, 2018) causes of PRR support (Golder, 2016). While studies of voter behavior have thus dominated research explaining the success of PRR parties (Arzheimer, 2016), a distinct comparative party politics literature studies how political systems respond to those parties once they gain political support. In an influential early study, Downs (2001) outlines a “fundamental choice” for establishment parties faced with an ascendant

4 PRR party: engage or disengage with the newcomer. This broad distinction has recurred in the subsequent literature in varying forms. Reviewing theories of establishment party response, Heinze (2018) identiﬁes a total of six subtypes of the disengagement strategy and two subtypes of the engagement strategy. Since the literature uses varying labels for some of these subtypes, the total number of terms used to characterize party strategies is even greater.

1.1 Party-level responses to PRR success

This conceptual clutter notwithstanding, the literature has settled on some stylized facts about establishment responses to PRR success. First of all, in spite of initial resistance, mainstream parties tend to accommodate the PRR position over time. For example, Meguid (2005) argues that mainstream parties respond to the entry of niche (including PRR) parties either by repositioning or changing the salience or issue ownership of the niche party’s key issue. As a result, even in cases where mainstream parties successfully defuse the PRR’s support, they do so by elevating its signature political issue, immigration, to the top of the political agenda. Similarly, in a comparative case study of Denmark, the Netherlands, Norway, and Austria, Bale et al. (2010) argue that from the 1980’s to the 2000’s, Denmark’s Social Democrats went from an initial strategy of holding firm on its initial pro-immigration stance over trying to defuse the issue to eventually adopting an anti-immigration position. Most recently, Abou-Chadi and Krause (2018) use a regression discontinuity design to demonstrate a rightward shift on immigration among mainstream parties in response to PRR parties gaining representation in parliament. Another stylized fact emerging from the literature is that stigmatizing PRR parties generally appears to ineffective. Rather than merely trying to defuse issue conflict with PRR parties (Bale et al., 2010), establishment parties can choose a strategy of stigmatization (van Spanje and Azrout, 2019), sometimes labeled ‘ostracization’ (van Spanje, 2010), proclaiming the PRR party to be “beyond the pale” and hence outside the scope of legitimate political coalition-building. A typical consequence of stigmatization strategies is the erection of a “cordon sanitaire” around PRR parties, excluding their seats from legislative coalition-building (Downs, 2002). Although the cordon

5 sanitaire strategy effectively blocks PRR parties from obtaining legislative influence, the empirical record suggests stigmatized parties suffer little or no electoral costs (Spanje and Brug, 2007; Akkerman and Rooduĳn, 2015), although there is some evidence they lose votes when mainstream parties simultaneously mimic their policy positions (van Spanje, 2018). These stylized facts represent important insights on establishment responses to PRR entry. However, in highlighting these aspects, the literature inevitably brackets others. Specifically, a unifying feature of this literature is that it consistently theorizes responses to PRR entry in terms of party-level strategic concerns. As a consequence, this theoretical focus discounts within-party variation between individual legislators. Here, I argue why we should expect important variation within parties, effectively overlooked in the existing literature.

1.2 Moral distancing as an individual response strategy

There are good reasons to expect individual-level variation in response strategies. Most importantly, the existing literature presents persuasive evidence that legislators’ representational styles vary in meaningful ways within parties. Previous work using text as data methods finds rich variation in representational style in the US Congress (Grimmer, 2010; Grimmer, Messing and Westwood, 2012), and within-party variation in representational styles has also been found in European parliamentary systems (Tavits, 2009). While this suggests that individual legislators’ response strategies are likely to vary, it does not imply that those strategies will merely mirror those of the parties they represent. On the contrary, the opportunities and constraints faced by individual legislators means that their set of relevant response strategies is entirely different from that of parties. Most significantly, in a party- centric system, individual legislators will have little sway in setting the party line with respect to parliamentary cooperation and non-cooperation (Hug, 2010). The engagement or disengagement strategies summarized in Heinze (2018) are the purview of the party, not the individual legislator. Instead, I focus on an arena in which legislators do have individual discretion: legislative speech. Although legislative speech broadly speaking reflects party strategy, the former is not a one-to-

6 one expression of the latter (Proksch and Slapin, 2012). Because legislators speak individually, legislative speech offers a window into individual-level variation, and has been shown to reveal meaningful within-party variation in legislators’ political positions (Lauderdale and Herzog, 2016). This in turn raises the question of how to conceptualize response strategies to PRR challenges in the context of legislative speech. Here, I conceptualize a specific response strategy which I call moral distancing. As a starting point, consider the regularity with which political rhetoric frames responding to PRR parties as a morally charged decision. Establishment politicians often explicitly define PRR challengers as morally objectionable – a salient recent example being Hillary Clinton’s “deplorables” remark. One indicator of the strong moral dimension in views of PRR actors is that rhetoric attacking PRR parties is infused with cleanliness and contamination metaphors, a common correlate of moral cognition (Chapman et al., 2009; Tybur et al., 2013). Consider for example Danish prime minister Poul Nyrup Rasmussen’s remark that the DPP is not “housebroken”, a thinly veiled characterization of the DPP as filthy. In another speech from the data analyzed in this article, a legislator rejects a policy proposal from the DPP as “disgusting”. Conversely, distancing oneself from PRR parties is associated with cleanliness: Downs (2001) cites a Belgian city councillor considering whether to cooperate with the PRR party Vlaams Belang weighing his the goal of political success against a desire to have “clean hands”. Even the scholarly term “cordon sanitaire”, used to denote a strategy of excluding PRR parties from influence (Downs, 2002), is itself a metaphor casting the PRR party as a contamination risk to the body politic. The common denominator of these and similar portrayals of PRR parties is that they portray recognition of and cooperation with PRR parties as an act of moral debasement. In various ways, these statements make clear that PRR parties, due to their political positions and/or rhetoric, exist at a lower moral plane than established, mainstream parties. I refer to this rhetorical strategy as moral distancing. I define moral distancing as a type of speech portraying another party, its positions, or its rhetoric as morally objectionable. Moral distancing can be understood as a speech act (Searle, 1969) simultaneously constituting its target as morally illegitimate (and, by implication, the speaker as legitimate).

7 The concept of moral distancing draws on insights from existing empirical scholarship on the role of morality in elite political communication. Recent work has shown how elite framing of political issues in moral terms can prime citizens’ moral foundations (Clifford and Jerit, 2013), strategically signal moderation or extremity (Lipsitz, 2018), or mobilize copartisans (Jung, forthcoming). However, whereas these studies conceptualize moralization as a type of issue framing, moral distancing relates to the status of political actors themselves, irrespective of the political issue at hand. This focus on the moral quality of actors themselves rather than any one political issue links moral distancing to “moral grandstanding”, which Tosi and Warmke (2016) describe as “moral talk that attempts to get others to make certain desired judgments about oneself, namely, that one is worthy of respect or admiration because one has some particular moral quality”. Yet in contrast to moral grandstanding, the focus of moral distancing is on a target political opponent, not (in the first instance) to signal qualities of the speaker. While moral distancing is by definition negative, it is not merely an arbitrary expression of negative sentiment. Theoretically speaking, moral rhetoric is not reducible to extreme valence, but consists of appeals to the listener’s moral intuitions (Clifford and Jerit, 2013). Moral distancing is also empirically clearly distinct from negative rhetoric. Whereas almost all speech directed toward the DPP in this study is negative, only a small subset can be considered moral distancing. In the section on validation below, I conduct discriminant validation of this article’s measure of moral distancing, demonstrating that it is empirically distinct from negativity. Given this conceptualization, a natural question is whether moral distancing reflects a strategic calculus by the speaker. On the one hand, the literature on party-level response strategies typically explains mainstream parties’ decisions of whether to engage or disengage with PRR parties in terms of mainstream parties’ strategic environments (Bale et al., 2010). Accordingly, moral distancing may reflect individual legislators’ optimization within this broader context. On the other hand, the moral quality of moral distancing suggests it may be divorced from strategic considerations. Evidence from political psychology shows that moralized political questions are

8 particularly unlikely to be decided in terms of a strategic calculus (Ryan and Brader, 2017). In the next section, I present a set of hypotheses positing that moral distancing is in fact responsive to the strategic constraints faced by the individual legislator.

1.3 Hypotheses

I highlight two factors likely to affect a legislator’s willingness to engage in moral distancing. First, I expect moral distancing to be responsive to electoral security, a characteristic to which legislators generally tend to be acutely sensitive. In Mayhew’s (1974) famous characterization, legislators are “single-minded seekers of reelection”: individual legislators are attuned to electoral security because it defines their career prospects. Consistent with this notion, existing research finds that legislator behavior is highly sensitive to electoral security (Eggers and Spirling, 2014; Dropp and Peskowitz, 2012). Electoral security is plausibly relevant to moral distancing insofar as it may represent an opportunity cost to electorally insecure legislators, crowding out constituency-directed efforts. Consistent with this tradeoff, Grimmer (2013) finds that more electorally secure representatives dedicate more rhetoric to ideological position-taking and less to constituency service. By the same token, electorally secure legislators should be more willing to engage in moral distancing, as they are able to assume the potential electoral costs of doing so, whether direct or as opportunity costs. Building on this, I expect that more electorally secure legislators engage in more moral distancing (Hypothesis 1). Second, I focus on another characteristic that varies between legislators, namely issue constraint. Legislators vary in terms of how much latitude they are allowed in choosing which issues to emphasize, providing less issue constrained legislators with more opportunities to engage in moral distancing. I focus here on two factors which plausibly affect legislator issue constraint. The first of these is cabinet membership: whereas the government is forced to respond to issues on the party system agenda, the opposition can choose issues relatively freely (Green-Pedersen and Mortensen, 2010). I expect this difference to recur within government parties: legislators who are also members of the cabinet government will be particularly issue constrained, effectively tied to

9 the agenda of their cabinet portfolio. Hence, I expect that legislators holding cabinet positions will engage in less moral distancing (Hypothesis 2a). Issue constraint is also a product of a legislator’s position in the party hierarchy. In systems with high levels of party discipline, party leaders will have more agenda control than backbenchers, and will in turn have more discretion to engage in moral distancing (Proksch and Slapin, 2012; Skjæveland, 2001). Moreover, party leaders are more likely to feature in high-level, non-speciﬁc debates in which morally charged speech is more likely to occur (Giannetti and Pedrazzani, 2016). Hence, based on their greater organizational and situational latitude, I expect that party leaders will engage in more moral distancing (Hypothesis 2b).

2 Methods and Data

One likely reason for the paucity of individual-level perspectives is a lack of methodological tools for studying behavior at the level of individual legislators. I tackle this problem by applying machine learning methods to speeches from the parliamentary record. This constitutes a far more granular level of measurement than the existing literature, which measures strategies at the level of parties varying by decade (Bale et al., 2010), parties (van Spanje and Weber, 2017) or even entire party systems (Akkerman and Rooduĳn, 2015). These previous studies rely on either qualitative case knowledge or expert surveys to assess party strategies. In contrast, the speech-level measurement in this article allows for ﬂexibly aggregating measurement to the level of arbitrarily granular time periods, parties, individual legislators, or combinations thereof. Measuring moral distancing at the speech level requires ﬁrst obtaining a large data set of relevant political speeches and then a method for classifying speech as moral distancing. Here, I describe each of these two steps in turn.

2.1 Obtaining and pre-processing speech data

I obtain speech data from the parliamentary record of the Danish parliament, Folketinget. I do so by extracting text available in an online interface of scanned transcripts of parliamentary speech from

10 1890 to 2008.1 From this full set of transcripts, I obtain transcripts of all speeches between 1990 and 2008 and break these speeches down into paragraph-like chunks of text of 3 to 5 contiguous sentences each. The paragraph-level corpus consists of 1.1 million paragraphs in total. Appendix A presents additional details on this data retrieval stage. Of the 1.1 million paragraphs in the total corpus, only a subset constitute speech directed at the DPP. Other paragraphs represent speech directed at other political actors, or speech from DPP representatives, neither of which are in and of themselves informative of responses to the DPP. In order to focus the data set on speech directed at the DPP, I subset the corpus of paragraphs to paragraphs mentioning the DPP or the name of any of its legislators, and which are not spoken by a member of the DPP. I also limit the time period covered to speeches made in the first decade of the DPP’s existence, i.e. 10 years starting October 6, 1995, since preliminary analyses indicate that moral distancing recedes over time and is near-nonexistent after 2005. This focused corpus, consisting of 31,056 paragraphs in total, is thus an approximation of all the political speech directed at the DPP in the party’s first decade of existence. Figure A1 in the appendix shows how the paragraphs are distributed over time. The number of paragraphs increases over time, reflecting increasing attention to the party as it grows in size and gains political power after becoming the support party for the conservative government after 2001.

2.2 Measuring moral distancing

Manually assessing whether each of the 31,056 paragraphs reﬂects moral distancing would be prohibitively time-consuming. Moreover, since the texts are not in English, distributed crowd- based coding is infeasible in practice (though see de Vries, Schoonvelde and Schumacher, 2018). In order to use information from the full data set while keeping coding costs within feasibility, I use machine learning methods to characterize the content of the full set of paragraphs based on information from a manually coded subset. To develop a classiﬁer for the full set of paragraphs, I rely on the “machine learning-augmented 1Available at https://beta.folketingstidende.dk/da/e_folketingstidende

11 dictionary method” presented in Siegel et al. (2019). I begin by developing a dictionary of terms relevant for moral distancing based on a manual coding of 1,000 randomly sampled paragraphs. This dictionary is presented in Table B1 in the appendix. I then extract all 3,277 paragraphs in the full corpus matching at least one term in the dictionary. The reason for this initial step is that the base rate of moral distancing is very low. In the random sample of paragraphs, the rate of moral distancing is only around 9 percent. This would translate into either a training set with very little information or the need to code a very large random sample. Sampling using a custom dictionary results in a training set with a higher rate of moral distancing and thus more information available to the classiﬁer. From this smaller set of 3,277 paragraphs matching one or more dictionary terms, I then sample 1,000 paragraphs and code them according to whether they express moral distancing or not. Of the 1,000 coded paragraphs, 400 are coded by myself, and 200 by each member of a team of three research assistants. Each of the research assistants also codes a balanced sample of 50 paragraphs from my initial coding. This shared subset of paragraphs allows for assessing intercoder reliability

across the set of all four coders, which is reasonably high (Krippendorf’s α=.67). As the final step, I then use this set of 1,000 coded paragraphs to predict moral distancing in the full set of paragraphs. I construct a document-feature matrix with the 31,056 paragraphs in the rows and 7,632 word features in the columns. To reduce sparsity, I remove stopwords, numbers, and punctuation, lowercase all words, apply a word stemmer, and restrict the document-feature matrix to words mentioned at least 10 times across the entire text corpus. Alternative preprocessing choices, such as using lemmatizing instead stemming or including bi- and trigrams, yield substantively similar results (for a more elaborate discussion of consequences of text preprocessing, see Denny and Spirling, 2018). The machine learning literature on supervised learning offers several useful methods for classifying text based on known classifications from a subset of manually coded texts (Grimmer and Stewart, 2013). Here, I follow the approach in Theocharis et al. (2016) and use regularized regression for classification. The output of interest from the classifier is the predicted probability of a

12 each paragraph expressing moral distancing. In Appendix C I provide additional discussion of key choices with respect to classifier selection and tuning. In order to assess how well the classifier performs out of sample, I conduct hold-out validation, fitting the model to a subset of the labeled data and evaluating how well the model predicts labels in the held-out set. I do so in two ways. In section C.2 I present ‘separation plots’ based on the approach presented in Greenhill, Ward and Sacks (2011). In the held-out sample, known cases of moral distancing are assigned significantly higher predicted probabilities (z = 3.83, p < .001). Section C.3 presents a ‘receiver operating characteristic’ (ROC) curve, showing how sensitivity (i.e., true positive rate) and specificity (i.e., true negative rate) of the predictions of the regularized regression model vary across cutoffs for the predicted probability. The figure also shows that the chosen regularized regression model outperforms a Naive Bayes classifier. At a threshold of .11, the classifier has .94 precision and .56 recall. The consequence of this limited out-of-sample classification accuracy is that in paragraphs not included in the manually coded sample, moral distancing is observed with more measurement error. Since this measurement error is on the dependent variable, it does not lead to attenuation bias, but does lead to larger standard errors. In the analyses to follow, I examine how moral distancing varies across individual legislators by aggregating paragraphs up to the relevant units of analysis. One potential pitfall of this approach is that units of analysis with very few observations (e.g., legislators with very few speech paragraphs in the data) will have very high variance, masking the meaningful variation in the data. As Monroe, Colaresi and Quinn (2008) observe, one way to solve this overfitting problem is through shrinkage, i.e. imposing a conservative prior on the model, thereby pulling sparse averages toward the grand mean. Here, I follow Park, Gelman and Bafumi (2004) and apply shrinkage by fitting intercept-only random effects models and using the group-level predictions as measurements.

2.3 Validating the measure of moral distancing

Separate from the question of whether the classiﬁer captures information in the uncoded set of paragraphs is whether the measure of moral distancing captures a substantively meaningful

13 phenomenon to begin with. I evaluate this question here, following Adcock and Collier (2001) in distinguishing between content, convergent, and discriminant validation. Section D in the online appendix presents the tables and figures referenced here. As content validation, section D.1 presents the paragraph with the highest predicted probability of moral distancing (.93) along with an English translation. In the paragraph, a legislator chastises the conservative government for relying on the support of the DPP, “a party in conflict with the very most fundamental and basic democratic principles”. The paragraph corresponds well to the theoretical concept of moral distancing, suggesting the measure reflects the theoretical content of moral distancing. In section D.2 I present convergent validation of the measure. Although there are no existing measures of moral distancing, the politics of moral distancing from the DPP is closely linked to political conflict over immigration. One would therefore expect pro-immigration parties to express more moral distancing. Figure D1 presents party-level average moral distancing plotted against parties’ position on “multiculturalism” based on data from the Comparative Manifesto Project (Volkens et al., 2018), a measure used in other studies to proxy parties’ immigration stance (e.g., Abou-Chadi and Krause, 2018). As one would expect, more pro-immigration parties are significantly more likely to engage in moral distancing (r = .81, p < .01). Lastly, in section D.3 I present discriminant validation of the moral distancing measure, demonstrating that the moral distancing is not merely capturing negative speech. Figure D2 plots moral distancing for each legislator-term against a measure of speech sentiment calculated using AFINN, a widely used sentiment dictionary (Nielsen, 2011). As one would expect, speeches high in moral distancing are recognized as more negative (r = −.15, p < .01), but the correlation is nonetheless weak, indicating that moral distancing is distinct from mere negativity.

2.4 Measuring legislator characteristics

I measure the focal variable in hypothesis 1, electoral security, using data from electoral returns. Because of Denmark’s party-centric electoral system, there is no existing database on individual-

14 level election returns. To obtain these data, I retrieve the official election reports made available by the Ministry for Economic Affairs and the Interior and write a custom script to extract vote counts at the individual and district level. Using these data, I measure electoral security as each legislator’s personal vote count as a fraction of their party’s district-specific total vote count in the most recent general election. Because seats within parties are allocated based on candidates’ within-party personal vote count rank, candidates high in this measure are the least vulnerable to declines in party vote share and thus the most electorally secure, and vice versa. For the focal variable in hypothesis 2a, cabinet membership, I exploit the fact that cabinet members are labeled as such (e.g., Minister of Finance or Minister of Foreign Affairs) in the speech transcripts, making it simple to extract information about cabinet membership for each speech. Lastly, I measure the focal variable in hypothesis 2b, party leadership, by merging data on parties’ leadership histories to create a binary indicator for whether each legislator is the leader of her party in a given term. When parties change leadership during a parliamentary term, I assign leadership status to the person holding the position for the majority of the term. I also obtain measurements of ‘standard demographics’ of each legislator, i.e. age, gender, and level of education. Since no publicly available data on demographic characteristics of identifiable Danish legislators exists, I compile a novel data set. I do so by first collecting data from the EveryPolitician API, a service providing basic info on thousands of members of legislatures around the world. The EveryPolitician API includes information about the name and gender of all Danish legislators elected between 1994 and 2001. Crucially, for each of these the API also provides their identifier on Wikidata, an online database maintained by the Wikimedia foundation. For each legislator in the data, I query the Wikidata API to get their date of birth. Lastly, to obtain data on educational background, I use the Wikipedia API to obtain the full text of each legislators’ biography on Wikipedia. Using these biographies, I construct a simple dichotomous measure of education level: if a legislator’s biography contains a mention of a university within the first 2,000 characters, I code them as having some university education, and none otherwise. Although this measurement strategy is crude, a subsequent check of 10 randomly selected cases revealed zero

15 misclassiﬁcations, suggesting measurement error is minimal. For the few remaining legislators with missing data, I manually search for biographical details to get complete data on gender, date of birth, and educational background.

3 Results

This section presents two sets of results. First, present an overall comparison of between-party variation in moral distancing, the dominant perspective in the existing literature, vis-à-vis within- party, legislator-level variation. I then turn to the question of how speciﬁc legislator characteristics predict moral distancing, presenting a set of models evaluating hypotheses 1, 2a, and 2b.

3.1 Within- vs. between-party variation

Figure 1 plots the estimated level of moral distancing for each legislator-term in the data along with party-level averages. Table 1 presents a more direct comparison of different types of variation in the data, showing variance and standard deviation at each level as well as residual variation when including legislator and party as group-level intercepts in a random effects model. Table 1: Within- vs. between-party variance in a random effects model

Level Variance SD 1 MP 0.0002 0.015 2 Party 0.00004 0.006 3 Residual 0.005 0.068

The speciﬁc legislator-level estimates provide some additional validation of the measure. The legislators with highest estimated levels of moral distancing, Margrethe Auken (Socialist People’s Party) and Sonja Albrink (Centre Democrats), were both known at the time as vocal critics of the DPP, providing additional face validation to the measure. These speciﬁcs aside, the key insight from Figure 1 and Table 1 is that between-party variation only accounts for a subset of the total variation, and there is considerable within-party variation. For example, variation in moral distancing within the Social Democrats (S) spans almost the entire observed range of moral distancing.

16 Progress Party

Liberals

Social Democrats

Conservative People's Party

Christian People's Party Party

Red−Green Alliance

Social Liberals

Socialist People's Party

Centre Democrats

.075 .1 .125 .15 .175 Estimated level of moral distancing

Figure 1: Legislator-term-level estimates of moral distancing by party. The plot shows the distribution of estimated levels of legislator-term-speciﬁc moral distancing for each party in the data. Vertical bars represent party-level averages.

17 Numerically speaking, variation between legislators after parsing out party-level variation (SD = .015) is more than twice the between-party variation (.06). To be sure, between-party variation in moral distancing remains signiﬁcant. But this decomposition of variation shows that even in a system characterized by strict party control, individual legislators have enough leeway to select strikingly diﬀerent rhetorical strategies.

3.2 Tests of hypotheses about legislator characteristics

Table 2 shows results from a series of OLS models regressing moral distancing on the variables specified in hypotheses 1, 2a, and 2b along with varying sets of controls. Since most legislators are observed in more than one electoral term, standard errors in all models are clustered by legislator- term. Models 1-3 regress moral distancing on electoral security, cabinet membership, and party leadership respectively, all including demographic controls. Model 4 includes all three variables jointly and adds party fixed effects. Model 5 adds fixed effects for parliamentary terms. To preserve space, party fixed effects are not shown in Table 2, but the full results are shown in Table E2 in the appendix. The models do not include time as a continuous covariate, but in appendix F I present estimates by month, showing that moral distancing recedes over time. Notably, the associations do not change appreciably when including party fixed effects, and thus only considering within-party variation. Figure 2 visualizes the association of interest in hypothesis 1, based on the estimate in Table 2, model 5. As shown, electoral security is significantly and positively associated with moral distancing such that more electorally secure legislators are more likely to engage in moral distancing. Figure 3 shows predicted levels of moral distancing for the variables of interest in hypotheses 2a and 2b, cabinet members vs. non-members (panel a) and party leaders vs. non-leaders (panel b). Like in Figure 2, the estimates are based on Table 2, model 5. In both cases, the differences are of the expected sign and statistically significant, lending support to the notion that less issue constrained legislators (i.e., non-government and party leaders) are more likely to engage in moral

18 Table 2: Models predicting moral distancing

(1) (2) (3) (4) (5) Intercept .106∗∗∗ .109∗∗∗ .108∗∗∗ .116∗∗∗ .116∗∗∗ (.003)(.003)(.003)(.007)(.007) Electoral security .012∗∗∗ .011∗∗ .012∗∗∗ (.003)(.003)(.003) Cabinet member −.013∗∗∗ −.016∗∗∗ −.016∗∗∗ (.002)(.002)(.002) Party leader .007∗ .007∗ .007∗ (.003)(.003)(.003) Age .000 .000 .000 .000 .000 (.000)(.000)(.000)(.000)(.000) Gender: male −.002 −.002 −.002 −.002 −.002 (.001)(.001)(.001)(.001)(.001) University education .001 .003∗ .002 .001 .001 (.001)(.001)(.001)(.001)(.001) Party FE XX Term FE X R2 .058 .100 .028 .231 .246 Adj. R2 .049 .091 .019 .206 .219 Num. obs. 423 444 444 423 423 RMSE .013 .013 .013 .012 .012 N Clusters 411 432 432 411 411

∗∗∗ p < 0.001; ∗∗ p < 0.01; ∗ p < 0.05

19 .125

.120

.115 Moral distancing Moral

.110

0 .25 .5 .75 1 Electoral security

Figure 2: Association between moral distancing and electoral security (hypothesis 1). Plot shows predicted levels of moral distancing across the observed range of electoral security. The dark and light bands around each prediction line represent 90 and 95 pct. conﬁdence intervals. distancing. In substantive terms, moving across the observed range of political characteristics is associated with an increase in moral distancing from 9.5 percent (government non-leaders with minimal electoral security) to 13 percent (non-government party leaders with high levels of electoral security). To be sure, this is a somewhat modest change in absolute terms. However, recall that legislator- term-level estimates are ‘shrunk’ towards the overall mean in order to reduce the noise in estimates for legislators with few observations. This model-imposed shrinkage constrains the observed range of moral distancing by design. Hence, it is more informative to evaluate the change in moral distancing associated with all three political characteristics relative to the entire observed range of moral distancing. This comparison reveals that electoral security, cabinet membership, and party leadership jointly predict variation in moral distancing corresponding to around 34 percent of the observed range of moral distancing. Hence, once accounting for the observed range of moral distancing, the association between political characteristics of legislators and moral distancing is both statistically and substantively signiﬁcant.

20 0.115 0.125

0.110 ● 0.120 0.105

Moral distancing Moral 0.100 distancing Moral 0.115 ●

0.095

Cabinet member Non−member Party leader Non−leader

(a) Predicted levels of moral distancing for (b) Predicted levels of moral distancing for cabinet members vs. non-members. party leaders vs. non-leaders.

Figure 3: Association between moral distancing and issue constraint (hypotheses 2a and 2b). Both sets of estimates based on Table 2, model 5. Thick and thin error bars represent 90 and 95 pct. conﬁdence intervals.

The robust and significant associations for electoral security, cabinet membership, and party leadership are in stark contrast with the role of demographic characteristics. The results in Table 2 indicate that moral distancing is linearly unrelated to the age, gender, and educational background of legislators, even without conditioning on any political legislator characteristics. All demographic variables are statistically insignificant in all specifications. Even taking the point estimates at face value, the variation in moral distancing predicted by demographic characteristics corresponds to only only 6.6 percent of the observed range of moral distancing. In sum, then, speech-level data reveal considerable within-party variation in moral distancing. For some parties, the within-party range in moral distancing is close to spanning the entire observed range across parties. Moreover, this variation is highly predictable in terms of political characteristics, with moral distancing significantly more common among highly electorally secure non-government party leaders. Conversely, demographic characteristics are unrelated to moral distancing, whether considered or their own or jointly with political characteristics.

21 4 Conclusion: The Politics of Moral Distancing

Politics in Western democracies is long past the era of ‘frozen’ party systems (Lipset and Rokkan, 1967). In recent years, many Western democracies have witnessed the rise of ‘challenger parties’ unsettling the order of the established political system (Hobolt and Tilley, 2016). Among challenger parties, right-wing populist parties have made remarkable electoral gains in several countries. Faced with an insurgent right-wing populist party, establishment actors find themselves having to decide whether to treat the challenger as a legitimate political opponent or, alternatively, to portray the challenger as morally illegitimate. I introduce the concept of moral distancing to characterize this latter strategy as an individual-level phenomenon. The concept adds to a rich existing literature in comparative party politics studying this response at the level of parties or party systems. I have presented an approach using text as data methods to move this research agenda to the individual level. Using supervised machine learning to measure moral distancing towards the Danish People’s Party (DPP) in around 31,000 paragraphs of parliamentary speech, I uncover a wide range of response strategies at the individual level, even after parsing out party-level variation. I find that within parties, moral distancing is strongly associated with political characteristics, namely electoral security, leadership status and cabinet membership. In contrast, moral distancing shows weak or no associations with legislators’ demographic characteristics. These results are informative of the extent to which moral distancing is affected by electoral security and issue constraint at the level of individual legislators. However, irrespective of the relative role of these specific characteristics, the joint importance of individual legislator characteristics holds important implications for the study of establishment responses to populist challenges. Specifically, the importance of individual-level factors can help explain why parties or party systems sometimes fail to pursue collectively optimal strategies. Facing their own sets of incentives and constraints, legislators may deviate from the party line in engaging in moral distancing or abstaining from doing so. This in turn implies that establishment parties face an even greater challenge than previously

22 recognized in responding to radical right challenger parties. In addition to settling on a coordinated response at the system level, parties must solve the problem of unifying their own legislators around a response strategy. Indeed, the considerable within-party variation demonstrated in this article is itself evidence that parties often fail at this important coordination problem. Some caveats are in order. For one, the individual-level findings presented here are purely associational. Since individual characteristics of legislators are not randomly assigned, these associations do not credibly identify causal effects. Hence, while this article uncovers individual- level variation, it makes only limited headway in explaining it. Equally importantly, the single country case nature of this study leaves open the question of whether the findings generalize to other cases of political systems responding to right-wing populist challengers, such as those mentioned above. A natural extension of this article would be to apply the methodology presented here to those cases. Another question of generalizability is whether and how moral distancing occurs in response to left-wing challenges. In the case studied here, moral distancing is closely intertwined with the anti-racism norm, itself an important factor in explaining right-wing populist support (Ivarsflaten, Blinder and Ford, 2010). Since left-wing parties are not generally accused of violating the anti-racism norm, moral distancing with respect to left-wing challenges will in all likelihood look very different. An interesting conjecture is that establishment responses to left-wing challenges will tend to take form of appeals to ‘binding’ moral concerns such as ingroup loyalty and respect for authority (Graham, Haidt and Nosek, 2009). Characterizing moral distancing in response to left-wing challenges is an important topic for future research. These caveats notwithstanding, this article has shown how scholars of individual political behavior can study establishment responses to populist challenger parties, a topic previously the sole purview of comparative party politics. By the same token, the individual-level variation uncovered here challenges the implicit premise in the existing literature that establishment responses to populist challenges are exclusively – or even predominantly – a party-level phenomenon. This in turn suggests that scholars looking to understand establishment responses to populist challenges cannot solely consider which strategies are optimal from the perspective of parties. We also need

23 to consider how legislators facing divergent constraints and opportunities coordinate around a common strategy – or fail to do so. Lastly, for political behavior scholars, this study raises important questions about how the politics of moral distancing operate at the mass level. Simply put, is moral distancing of right- wing populist challenger parties an eﬀective way to dissuade or demobilize otherwise likely voters of those parties? Or does it fuel resentment which may further entrench support for right-wing populists? Qualitative evidence provides some support for the latter notion: some Trump supporters report feeling resentful at disparaging elites (Hochschild, 2016), and moral distancing may even fuel contrarian political engagement among otherwise apolitical individuals (Nagle, 2017). In an era of political upheaval and intense conﬂict over the delineation of morally legitimate positions, these should be questions of vital interest to political science.

24 References

Abou-Chadi, Tarik and Werner Krause. 2018. “The Causal Eﬀect of Radical Right Success on Mainstream Parties’ Policy Positions: A Regression Discontinuity Approach.” British Journal of Political Science pp. 1–19.

Adcock, Robert and David Collier. 2001. “Measurement Validity: A Shared Standard for Qualitative and Quantitative Research.” American Political Science Review 95(03):529–546.

Akkerman, Tjitske and Matthĳs Rooduĳn. 2015. “Pariahs or Partners? Inclusion and Exclusion of Radical Right Parties and the Eﬀects on Their Policy Positions , Pariahs or Partners? Inclusion and Exclusion of Radical Right Parties and the Eﬀects on Their Policy Positions.” Political Studies 63(5):1140–1157.

Akkerman, Tjitske, Sarah Leah de Lange and Matthĳs Rooduĳn. 2016. Radical Right-Wing Populist Parties in Western Europe: Into the Mainstream? (Paperback) - Routledge. Routledge.

Arzheimer, Kai. 2016. Electoral Sociology–Who Votes for the Extreme Right and Why–and When? In The Populist Radical Right: A Reader. Taylor and Francis.

Back, H., M. Debus and J. Muller. 2014. “Who Takes the Parliamentary Floor? The Role of Gender in Speech-Making in the Swedish Riksdag.” Political Research Quarterly pp. 1065912914525861–.

Bale, Tim, Christoﬀer Green-Pedersen, Andréa Krouwel, Kurt Richard Luther and Nick Sitter. 2010. “If You Can’t Beat Them, Join Them? Explaining Social Democratic Responses to the Challenge from the Populist Radical Right in Western Europe.” Political Studies 58(3):410–426.

Betz, Hans-George. 1993. “The New Politics of Resentment: Radical Right-Wing Populist Parties in Western Europe.” Comparative politics 25(4):413–427.

Chapman, H. A., D. A. Kim, J. M. Susskind and A. K. Anderson. 2009. “In Bad Taste: Evidence for the Oral Origins of Moral Disgust.” Science 323(5918):1222–1226.

25 Chozick, Amy. 2016. “Hillary Clinton Calls Many Trump Backers ‘Deplorables,’and GOP Pounces.” The New York Times .

Cliﬀord, Scott and Jennifer Jerit. 2013. “How Words Do the Work of Politics: Moral Foundations Theory and the Debate over Stem Cell Research.” The Journal of Politics 75(3):659–671.

Colantone, Italo and Piero Stanig. 2018. “Global Competition and Brexit.” American Political Science Review 112(2):201–218.

de Vries, Erik, Martĳn Schoonvelde and Gĳs Schumacher. 2018. “No Longer Lost in Translation: Evidence That Google Translate Works for Comparative Bag-of-Words Text Applications.” Political Analysis 26(4):417–430.

Denny, Matthew J. and Arthur Spirling. 2018. “Text Preprocessing For Unsupervised Learning: Why It Matters, When It Misleads, And What To Do About It.” Political Analysis 26(2):168–189.

Downs, William M. 2001. “Pariahs in Their Midst: Belgian and Norwegian Parties React to Extremist Threats.” West European Politics 24(3):23–42.

Downs, William M. 2002. “How Effective Is the Cordon Sanitaire? Lessons from Efforts to Contain the Far Right in Belgium, France, Denmark and Norway.” Journal of Conflict and Violence Research 4(1).

Dropp, Kyle and Zachary Peskowitz. 2012. “Electoral Security and the Provision of Constituency Service.” The Journal of Politics 74(1):220–234.

Eggers, Andrew C. and Arthur Spirling. 2014. “Electoral Security as a Determinant of Legislator Activity, 1832–1918: New Data and Methods for Analyzing British Political Development.” Legislative Studies Quarterly 39(4):593–620.

Ford, Robert. and Matthew J. Goodwin. 2014. Revolt on the Right : Explaining Support for the Radical Right in Britain. Taylor and Francis.

26 Giannetti, Daniela and Andrea Pedrazzani. 2016. “Rules and Speeches: How Parliamentary Rules Aﬀect Legislators’ Speech-Making Behavior: Rules and Speeches.” Legislative Studies Quarterly 41(3):771–800.

Golder, Matt. 2016. “Far Right Parties in Europe.” Annual Review of Political Science 19(1):477– 497.

Graham, Jesse, Jonathan Haidt and Brian A Nosek. 2009. “Liberals and Conservatives Rely on Diﬀerent Sets of Moral Foundations.” Journal of Personality and Social Psychology 96(5):1029– 1046.

Green-Pedersen, Christoﬀer and Peter B Mortensen. 2010. “Who Sets the Agenda and Who Responds to It in the Danish Parliament? A New Model of Issue Competition and Agenda- Setting.” European Journal of Political Research 49(2):257–281.

Greenhill, Brian, Michael D. Ward and Audrey Sacks. 2011. “The Separation Plot: A New Visual Method for Evaluating the Fit of Binary Models.” American Journal of Political Science 55(4):991–1002.

Grimmer, J. and B. M. Stewart. 2013. “Text as Data: The Promise and Pitfalls of Automatic Content Analysis Methods for Political Texts.” Political Analysis 21(3):267–297.

Grimmer, Justin. 2010. Representational Style: The Central Role of Communication in Represen- tation Ph.D. Dissertation Harvard University.

Grimmer, Justin. 2013. “Appropriators Not Position Takers: The Distorting Eﬀects of Electoral Incentives on Congressional Representation.” American Journal of Political Science 57(3):624– 642.

Grimmer, Justin, Solomon Messing and Sean J. Westwood. 2012. “How Words and Money Cultivate a Personal Vote: The Eﬀect of Legislator Credit Claiming on Constituent Credit Allocation.” American Political Science Review 106(04):703–719.

27 Hastie, Trevor and Robert Tibshirani. 2009. The Elements of Statistical Learning - Data Mining, Inference, and Prediction. Second ed. New York: Springer-Verlag.

Heinze, Anna-Sophie. 2018. “Strategies of Mainstream Parties towards Their Right-Wing Populist Challengers: Denmark, Norway, Sweden and Finland in Comparison.” West European Politics 41(2):287–309.

Hindman, Matthew. 2015. “Building Better Models: Prediction, Replication, and Machine Learning in the Social Sciences.” The ANNALS of the American Academy of Political and Social Science 659(1):48–62.

Hobolt, Sara B. and James Tilley. 2016. “Fleeing the Centre: The Rise of Challenger Parties in the Aftermath of the Euro Crisis.” West European Politics 39(5):971–991.

Hochschild, Arlie Russell. 2016. Strangers in Their Own Land: Anger and Mourning on the American Right. New Press, The.

Hug, Simon. 2010. “Selection Eﬀects in Roll Call Votes.” British Journal of Political Science 40(1):225–235.

Ivarsﬂaten, E. 2008. “What Unites Right-Wing Populists in Western Europe?: Re-Examining Grievance Mobilization Models in Seven Successful Cases.” Comparative Political Studies 41(1):3–23.

Ivarsﬂaten, Elisabeth, Scott Blinder and Robert Ford. 2010. “The Anti-Racism Norm in Western European Immigration Politics: Why We Need to Consider It and How to Measure It.” Journal of Elections, Public Opinion and Parties 20(4):421–445.

Jung, Jae-Hee. forthcoming. “The Mobilizing Eﬀect of Parties’ Moral Rhetoric.” American Journal of Political Science .

Lauderdale, Benjamin E. and Alexander Herzog. 2016. “Measuring Political Positions from Leg- islative Speech.” Political Analysis 24(3):374–394.

28 Lipset, Seymour and Stein Rokkan. 1967. Party Systems and Voter Alignments. Free.

Lipsitz, Keena. 2018. “Playing with Emotions: The Eﬀect of Moral Appeals in Elite Rhetoric.” Political Behavior 40(1):57–78.

Mayhew, David R. 1974. Congress: The Electoral Connection. Vol. 26 Yale University Press.

Meguid, Bonnie M. 2005. Competition between Unequals: The Role of Mainstream Party Strategy in Niche Party Success. Vol. 99.

Monroe, Burt L., Michael P. Colaresi and Kevin M. Quinn. 2008. “Fightin’ Words: Lexical Feature Selection and Evaluation for Identifying the Content of Political Conﬂict.” Political Analysis 16(4):372–403.

Montgomery, Jacob M. and Santiago Olivella. 2018. “Tree-Based Models for Political Science Data.” American Journal of Political Science 62(3):729–744.

Mudde, Cas. 2007. Populist Radical Right Parties in Europe. Vol. 22 Cambridge University Press.

Mudde, Cas. 2010. “The Populist Radical Right: A Pathological Normalcy.” West European Politics 33(6):1167–1186.

Müller, Wolfgang C. 1997. “Inside the Black Box: A Confrontation of Party Executive Behaviour and Theories of Party Organizational Change.” Party Politics 3(3):293–313.

Nagle, Angela. 2017. Kill All Normies: Online Culture Wars From 4Chan And Tumblr To Trump And The Alt-Right. Winchester, UK ; Washington, USA: Zero Books.

Nielsen, Finn Årup. 2011. “A New ANEW: Evaluation of a Word List for Sentiment Analysis in Microblogs.” arXiv:1103.2903 [cs] .

Norris, Pippa and Ronald Inglehart. 2018. Cultural Backlash. Cambridge University Press.

29 O’Grady, Tom. 2019. “Careerists Versus Coal-Miners: Welfare Reforms and the Substantive Representation of Social Groups in the British Labour Party.” Comparative Political Studies 52(4):544–578.

Park, David K., Andrew Gelman and Joseph Bafumi. 2004. “Bayesian Multilevel Estimation with Poststratiﬁcation: State-Level Estimates from National Polls.” Political Analysis 12(4):375–385.

Proksch, Sven-Oliver and Jonathan B. Slapin. 2012. “Institutional Foundations of Legislative Speech.” American Journal of Political Science 56(3):520–537.

Rauh, Christian and Jan Schwalbach. 2020. “The ParlSpeech V2 Data Set: Full-Text Corpora of 6.3 Million Parliamentary Speeches in the Key Legislative Chambers of Nine Representative Democracies.”.

Ryan, Timothy J. and Ted Brader. 2017. “Gaﬀe Appeal A Field Experiment on Partisan Selective Exposure to Election Messages.” Political Science Research and Methods 5(04):667–687.

Searle, John R. 1969. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.

Siegel, Alexandra, Evgenii Nikitin, Pablo Barbera, Joanna Sterling, Bethany Pullen, Richard Bon- neau, Jonathan Nagler and Joshua A Tucker. 2019. “Trumping Hate on Twitter? Online Hate in the 2016 US Election and Its Aftermath.” p. 30.

Skjæveland, A. 2001. “Party Cohesion in the Danish Parliament.” The Journal of Legislative Studies 7(2):35–56.

Slapin, Jonathan B., Justin H. Kirkland, Joseph A. Lazzaro, Patrick A. Leslie and Tom O’grady. 2018. “Ideology, Grandstanding, and Strategic Party Disloyalty in the British Parliament.” American Political Science Review 112(1):15–30.

Spanje, Joost Van and Wouter Van Der Brug. 2007. “The Party as Pariah: The Exclusion of

30 Anti-Immigration Parties and Its Eﬀect on Their Ideological Positions.” West European Politics 30(5):1022–1040.

Tartar, Andre. 2017. “How the Populist Right Is Redrawing the Map of Europe.” Bloomberg.com .

Tavits, Margit. 2009. “The Making of Mavericks: Local Loyalties and Party Defection.” Compar- ative Political Studies 42(6):793–815.

Theocharis, Yannis, Pablo Barberá, Zoltán Fazekas, Sebastian Adrian Popa and Olivier Parnet. 2016. “A Bad Workman Blames His Tweets: The Consequences of Citizens’ Uncivil Twitter Use When Interacting With Party Candidates.” Journal of Communication 66(6):1007–1031.

Tosi, Justin and Brandon Warmke. 2016. “Moral Grandstanding.” Philosophy & Public Aﬀairs 44(3):197–217.

Tybur, Joshua M., Debra Lieberman, Robert Kurzban and Peter DeScioli. 2013. “Disgust: Evolved Function and Structure.” Psychological Review 120(1):65–84. van Spanje, Joost. 2010. “Parties beyond the Pale: Why Some Political Parties Are Ostracized by Their Competitors While Others Are Not.” Comparative European Politics 8(3):354–383. van Spanje, Joost. 2018. Controlling the Electoral Marketplace: How Established Parties Ward Oﬀ Competition. Political Campaigning and Communication Palgrave Macmillan. van Spanje, Joost and Rachid Azrout. 2019. “Tainted Love: How Stigmatization of a Political Party in News Media Reduces Its Electoral Support.” International Journal of Public Opinion Research 31(2):283–308. van Spanje, Joost and Till Weber. 2017. “Does Ostracism Aﬀect Party Support? Comparative Lessons and Experimental Evidence.” Party Politics .

Volkens, Andrea, Werner Krause, Pola Lehmann, Theres Matthieß, Nicolas Merz, Sven Regel, Bernhard Weßels and Wissenschaftszentrum Berlin Für Sozialforschung (WZB). 2018. “Mani- festo Project Dataset.”.

31 Online Appendices

Appendix A Paragraph-level data 2

A.1 Preprocessing paragraph-level data ...... 2

A.2 Number of paragraphs over time ...... 4

Appendix B Dictionary 5

B.1 Dictionary terms used for dictionary-based sampling ...... 5

Appendix C Text classiﬁcation 6

C.1 Model selection and tuning ...... 6

C.2 Separation plots ...... 6

C.3 Receiver-operator characteristic (ROC) curve ...... 9

Appendix D Validation 10

D.1 Highest predicted probability paragraph ...... 10

D.2 Party-level estimates and manifesto data ...... 11

D.3 Is it just negativity? Moral distancing and speech sentiment . . . . 12

Appendix E Full regression tables 13

Appendix F Trend in moral distancing over time 15

1 Appendix A Paragraph-level data

A.1 Preprocessing paragraph-level data

The server at https://beta.folketingstidende.dk/da/e_folketingstidende hosting the scanned transcripts also hosts a corresponding set of XML ﬁles indexing all words on each page and their positions in order to make the online interface searchable. I obtained a copy of these XML ﬁles from the Folketinget IT department and wrote a custom script to extract every speech from 1990 to 2008. For each speech, I extract data on date, speaker, and speaker party from the text patterns in the transcript.

This initial data collection yields the full text of around 295,000 speeches.

Since the text data is obtained in the form a set of speeches, a natural starting point would be to consider speeches the unit of analysis. However, a signiﬁcant drawback of such a strategy is that speech lengths are highly heterogeneous. The shortest speeches in the corpus are single-word utterances like "yes", "no", or

"sorry"; the longest speech in the corpus consists of 6,835 words distributed across

206 sentences. Because of this heterogeneity, treating all speeches as comparable units of analysis would be unreasonable.

One way to address this heterogeneity would be to consider individual sentences instead of whole speeches as the unit of analysis. Breaking the corpus down to the sentence level yields a corpus of 3.7 million sentences. However, standalone sentences are often ambiguous outsider their natural context, which makes manual coding diﬃcult.

2 What is needed, then, is a unit of analysis at an intermediate level between whole speeches and individual sentences. To get such an intermediate level, I construct a corpus of ’synthetic paragraphs’ consisting of speeches broken down into sequences of 3 to 5 contiguous sentences each. Each speech is broken down into paragraphs the number of which depends on which number of paragraphs divides the speech most evenly. Speeches consisting of 5 or fewer sentences are kept as is. The paragraph-level corpus consists of 1.1 million paragraphs in total.

3 A.2 Number of paragraphs over time

1200

900

600

300 No. of paragraphs in data of paragraphs No.

1996 1998 2000 2002 2004 2006

Figure A1: Number of paragraphs in data by month.

4 Appendix B Dictionary

B.1 Dictionary terms used for dictionary-based sampling

Table B1: Dictionary used for initial sampling

Original term English translation 1 nedsætte derogatory 2 usympatisk unlikable 3 svinehund swine 4 hetz witch hunt 5 nazi nazi 6 racis racist 7 moral moral 8 anstændig decent 9 niveau level 10 skam shame 11 nationalist nationalist 12 nationalisme nationalism 13 frygtsom fearful 14 lokum crapper 15 menneskesyn view of humanity 16 retssyn view of law 17 skrækkampagne fear campaign 18 skræmmekampagne intimidation campaign 19 forstem oﬀend 20 verdensbillede world view 21 krænke oﬀend 22 smålig petty 23 splid division 24 holdninger views 25 værdig decent 26 ubehagelig uncomfortable 27 pigtråd barb wire 28 hudfarve skin color 29 syndebuk scapegoat 30 instinkt instinct 31 reaktionær reactionary 32 kynisk cynical 33 lavest lowest 34 lavmål low bar 35 menneskerettigheder human rights 36 stueren house trained

5 Appendix C Text classiﬁcation

C.1 Model selection and tuning

I predict moral distancing in the full set of paragraphs using regularized regression.

One advantage of regularized regression is its simplicity: regularized regression is essentially a standard regression with an penalty for model complexity included in the loss function (Hastie and Tibshirani, 2009). This makes regularized regression more intelligible than relatively more complicated methods such as random forests or neural networks (Montgomery and Olivella, 2018). I ﬁt a regularized regression model with the hand-coded values as labels and the word features as predictors, using cross-validation to select the hyperparameters α and λ.

A standard concern in machine learning applications is the risk of overfitting, i.e. the classifier relying on arbitrary features in the hand coded sample which are not informative of out-of-sample relationships. To evaluate the out-of-sample accuracy of the classifier, I use hold-out validation, designating a random subset of 75 pct. of the hand coded paragraphs as a training set and the remaining 25 pct. as the test set. By fitting a model only to the training set data, I can use the model’s ability to predict the hand-coded values in the held-out test set to assess how well the full model classifies uncoded text.

C.2 Separation plots

Figure C1 shows separation plots for the training set and test sets. In each plot, paragraphs are ordered from lowest to highest predicted probability. Paragraphs

6 manually coded as moral distancing are shown in dark color (green), and the remaining paragraphs are shown in light color (yellow). Hence, the separation plot visualizes the predictive power of the model by showing how neatly predicted probabilities separate paragraphs without moral distancing in the leftmost part of each plot from paragraphs with moral distancing in the rightmost part.

1.00

0.75

0.50

0.25 Predicted probability

0.00

(a)

1.00

0.75

0.50

0.25 Predicted probability

0.00

(b)

Figure C1: Separation plots of predicted probabilities vs. true values in training set (panel a) and test set (b). The bars represent coded paragraphs ordered by predicted probability of moral distancing from low to high, The model nearly perfectly separates categories using within the texts used to ﬁt the model, but has weaker accuracy on out-of-sample texts. Even so, predicted probabilities are signiﬁcantly associated with true values in the test set (z = 3.89, p < .001).

Figure C1 shows clear evidence of overﬁtting: in the top panel, showing a separation plot for the training set, predicted probabilities nearly perfectly separate paragraphs. In contrast, as shown in the bottom panel, predicted probabilities are

7 only weakly associated with true instances of moral distancing speech in the test set.

The pattern suggests that the out-of-sample classiﬁcation accuracy of the model is far from perfect. Still, predicted probabilities are signiﬁcantly associated with true instances of moral distancing in the test set (z = 3.89, p < .001), indicating that the model does predict out-of-sample variation in moral distancing.

8 C.3 Receiver-operator characteristic (ROC) curve

ROC curve for classification methods

1.00

0.75

Classification method

0.50 Naïve Bayes Regularized regression Sensitivity

0.25

0.00

0.00 0.25 0.50 0.75 1.00 1−Specificity

Figure C2: Receiver-operator characteristic (ROC) curve, showing how sensitivity (i.e., true positive rate) and 1 - specificity (i.e., true negative rate) of the predictions of the regularized regression model vary across cutoffs for the predicted probability. The plot is based on predictions for observations in the test set (25 pct.) from a model fitted to observations in the training set (75 pct.).

9 Appendix D Validation

D.1 Highest predicted probability paragraph

Table D1: Paragraph with highest predicted probability of moral distancing

Original English translation Nogle gange kan det jo under en folketingsdebat Sometimes during a parliamentary debat it can være vanskeligt at tro sine egne ører. Jeg troede, be hard to believe your own ears. I thought Ms. Torsdag den 4. oktober 2001 (R 1) 211 fru Pia Pia Kjaersgaard began her speech by trying to dis- Kjærsgaard begyndte sit indlæg med at forsøge på tance herself from what she had said, but she does ligesom at tage lidt afstand fra det, hun selv havde want to combat Islam. That is deeply reprehen- sagt, men hun vil altså bekæmpe islam. Det er sible, and it is deeply reprehensible to note that a dybt forstemmende, det er dybt forstemmende at party leader, a party in Folketinget, which really måtte konstatere, at der er en partiformand, et wants a religious war, and I think the conserva- parti her i Folketinget, som i realiteten ønsker en tive Denmark has a colossal problem on its hands religionskrig, og jeg synes altså, at det borgerlige after today’s debate. I think is should give rise to Danmark står med et kolossalt problem efter de- a more fundamental debate for Mr. Anders Fogh batten i dag. Jeg synes, at det bør give anledning Rasmussen, for Mr. Bendt Bendtsen, about what til en mere grundlæggende debat hos hr Anders to do with the problem that one is basing a gov- Fogh Rasmussen, hos hr Bendt Bendtsen, om, ernment on a party in conﬂict with the very most hvad man skal stille op med det problem, at man fundamental and basic democratic principles on har tænkt sig at basere en regeringsdannelse på et which a society is based, and Ms. Pia Kjaersgaard parti, som er i modstrid med de mest fundamen- has stepped up a moment ago and acknowledged tale og grundlæggende demokratiske principper, that she does not respect democracy, does not som et samfund er baseret på, og fru Pia Kjærs- respect freedom of religion. I really hope Mr. gaard er gået op her på talerstolen lige for et øje- Anders Fogh Rasmussen and Mr. Bendt Bendt- blik siden og har vedstået, at hun ikke respekterer sen will now make it clear that one cannot have demokratiet, ikke religionsfriheden. Jeg håber government cooperation with or build one’s par- virkelig, at hr Anders Fogh Rasmussen og hr liamentary majority on such a foundation... Bendt Bendtsen nu vil gå op og gøre det klart, at man ikke kan have et regeringssamarbejde eller basere sig parlamentarisk på et sådant grundlag...

10 D.2 Party-level estimates and manifesto data

0.02

CD ● 0.01 RV ●

SF ●

● EL 0.00 ●

● KF

● FP SD

Avg. level of moral distancing of moral level Avg. ●

V −0.01 ●

−5.0 −2.5 0.0 2.5 5.0 Net multiculturalism score in party manifestos

Figure D1: Net multiculturalism score in party manifestos and party-level estimates of moral distancing. The net multiculturalism score for each party is calculated as the percentage of quasi- sentences in the "Multiculturalism: Positive" category (per607) minus the percentage of quasi- sentences in the "Multiculturalism: Negative" category (per608), averaged across all manifestos beteen 1994 and 2005. The dashed line represents the linear best ﬁt. The correlation between the two measures is signiﬁcant (r = .81, p < .01).

11 D.3 Is it just negativity? Moral distancing and speech sentiment

0 Avg. paragraph−level sentiment paragraph−level Avg.

−2

0.075 0.100 0.125 0.150 0.175 Avg. level of moral distancing

Figure D2: Average level of moral distancing for each legislator-term and average paragraph- level sentiment. Sentiment is calculated using the AFINN sentiment lexicon available at https: //github.com/fnielsen/afinn. The dashed line represents the linear best fit. The correlation between the two measures is significant (r = −.15, p < .01).

12 Appendix E Full regression tables

13 (1) (2) (3) (4) (5) Intercept .106∗∗∗ .109∗∗∗ .108∗∗∗ .106∗∗∗ .116∗∗∗ (.003)(.003)(.003)(.003)(.007) Electoral security .012∗∗∗ .013∗∗∗ .012∗∗∗ (.003)(.003)(.003) Cabinet member −.013∗∗∗ −.017∗∗∗ −.016∗∗∗ (.002)(.002)(.002) Party leader .007∗ .006 .007∗ (.003)(.003)(.003) Age .000 .000 .000 .000 .000 (.000)(.000)(.000)(.000)(.000) Gender: male −.002 −.002 −.002 −.002 −.002 (.001)(.001)(.001)(.001)(.001) University education .001 .003∗ .002 .002 .001 (.001)(.001)(.001)(.001)(.001) Party: Red/Green Alliance −.005 (.007) Party: Conservatives −.009 (.006) Party: Christian Democrats −.011 (.007) Party: Social Liberals −.002 (.007) Party: Social Democrats −.006 (.006) Party: Socialist People’s Party −.003 (.007) Party: Liberals −.010 (.006) Term: 1998 −.001 (.002) Term: 2001 −.004∗∗ (.001) Party FE XX Term FE X R2 .058 .100 .028 .182 .246 Adj. R2 .049 .091 .019 .171 .219 Num. obs. 423 444 444 423 423 RMSE .013 .013 .013 .012 .012 N Clusters 411 432 432 411 411

∗∗∗ p < 0.001; ∗∗ p < 0.01; ∗ p < 0.05

Table E2: Models predicting moral distancing with political and demographic characteristics, full set of results

14 Appendix F Trend in moral distancing over time

Figure F3 shows predicted values of moral distancing by month, beginning with the ﬁrst speeches in the data in October 1995 and ending with the last speeches in

October 2005.

●

● 0.150

●

● ● ● ● ● ● 0.125 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.100 ● ● Avg. level of moral distancing of moral level Avg. ● ● ● ●

0.075 1996 1998 2000 2002 2004 2006

Figure F3: Average predicted paragraph-level probability of moral distancing by month. The gray dashed line shows the average level across all months. The darker and lighter error bars correspond to 90 and 95 pct. conﬁdence intervals respectively.

Two features of Figure F3 stand out. First of all, there is a weakly negative slope, indicating that moral distancing becomes less prevalent over time. The linear association between days since the DPP’s founding and speech-level moral distancing is negative and signiﬁcant (t = −2.1, p = .03).

The other notable feature is that the downward trend is interrupted by a few spikes of unusually high levels of moral distancing. The spikes correspond mean- ingfully to political events in which the debate over the legitimacy of the DPP

15 was particularly heated: ﬁrst, a contentious debate in early 1997, then the wake of the PM’s "housebroken" speech in late 1999, and lastly the month of the general election in 2001, where immigration policy dominated the political agenda.