According to some observers, America has not been as divided since the Civil War (Brownstein, 2021; Manchester, 2018). Democrats and Republicans favor their co-partisans and express hostility toward the opposition (see Mason, 2018). This favoritism toward one’s own party is also known as in-party affinity or positive partisanship, and the hostility toward the opposing party is dubbed out-party animosity or negative partisanship. The growing gap between how ordinary voters feel toward their own versus the opposing party reduces trust in the government run by the opposition (Hetherington & Rudolph, 2015), shapes individual policy preferences (Druckman et al., 2020), and influences attitudes and behaviors in non-political contexts, such as evaluations of college applicants (Iyengar & Westwood, 2015) or another person’s physical attractiveness (Broockman et al., 2020; Nicholson et al., 2016).

Despite the burgeoning research, it is unclear whether biased attitudes and behaviors mentioned above are caused by positive or negative partisanship. If the admission committee awards a scholarship to the candidate who shares their partisanship, is that primarily because they prefer and want to help the in-party member or dislike and want to disadvantage the out-party candidate? Some scholars argue that positive partisanship is a stronger driver of individual attitudes and behaviors (Amira et al., 2021; Lelkes & Westwood, 2017), others find that negative partisanship matters more (Abramowitz & Webster, 2016; Iyengar & Krupenkin, 2018), and yet other work suggests that their effects on political behavior depend on the behavior itself (Bankert, 2021; Caruana et al., 2015), as we detail below.

This important evidence on the intricacies of positive and negative partisanship exclusively focuses on “ordinary” citizens and their self-reported attitudes or experimentally observed behaviors. In contrast, our pre-registered project uses online behavioral data from Twitter to examine whether elite expressions on social media are mainly motivated by positive or negative partisanship (i.e., are American politicians more likely to praise their in-group or criticize the out-group?). We systematically investigate how 564 politicians, including members of the 116th Congress, Presidential and Vice-Presidential nominees from 2000 to 2020, and members of the Trump Cabinet, talk about their own and opposing parties as well as about other politicians in their social media posts. We also test whether these online expressions vary by ideological extremity of the politician and by whether their party is in power or in opposition.Footnote 1 Although these behaviorally tracked expressions may not always capture the actual feelings that politicians have toward the in-party and/or the out-party, our approach has important advantages, as we detail below.

In addition to examining the prevalence of and the factors influencing positive and negative partisanship in online expressions of political elites, we investigate what role that positive and negative partisanship play in citizen engagement with elite messages (i.e., are Twitter users liking and sharing elite posts that praise the in-group over those that criticize the out-group?).

In short, we study elite expressions and people’s reactions together in one study. We utilize behavioral trace data from Twitter given the platform’s important role in American politics. Twitter is a key platform for elite expressions—almost all members of Congress have a Twitter account (Quorum, 2021) and politicians are more active (Pew, 2021c) and have more followers (Pew, 2020) on Twitter than on other social media platforms. These expressions, furthermore, have direct effects on citizens’ attitudes and behaviors. Because journalists use politicians’ tweets in news reporting, these messages ultimately reach a larger audience and exert influence beyond Twitter (Parmelee, 2014).

In addition, Twitter is an important channel that citizens use to get political information (Pew, 2022), which affects their political beliefs and participation (Pew, 2021b; Vaccari et al., 2015). It is also a key outlet for individuals to express their political opinions and engage in political activities (Pew, 2022). These tweets and engagement metrics are often treated as proxies of public opinion by journalists and campaign strategists (McGregor, 2019, 2020), set political agendas that politicians follow (Barberá et al., 2019), and influence what journalists and media report (Hanusch & Tandoc Jr, 2019; Nelson & Tandoc Jr, 2019).

Relying on tweets sent by an extensive list of American politicians, ranging from representatives and senators to presidential Democratic and Republican candidates, including Presidents Donald Trump and Joe Biden, our pre-registered analyses find that elite expressions on social media are primarily motivated by positive partisanship—politicians praise their own party over criticizing the other party, although the latter also accounts for a large proportion of their online expressions. We find two exceptions to this aggregate pattern: politicians who are more ideologically extreme (as determined by two validated approaches, namely DW-NOMINATE (Poole & Rosenthal, 2007) scores and Barberá’s (2015) Bayesian spatial following model) as well as those whose political party is in the opposition (i.e., Democrats in our study) are more negative toward the out-party than their more ideologically moderate counterparts and those whose party is in power.

Shedding important light on how citizens react to these elite expressions, we show that users engage more with negative than positive partisanship, falling into the well-documented negativity bias. Users engage more with elite tweets that are negative toward the opposition, both when it comes to liking and sharing these tweets. These robust patterns emerge regardless of how many followers a politician has, across numerous issues, ranging from the economy to civil rights, and across different times of the day a tweet is posted. In short, contrary to popular concerns about inter-party hostility among the elites, most American politicians, most of the time, promote their own party on social media, even though—naturally—they also frequently bash the opposition. In turn, it is the biases in social media users themselves (namely disproportionate attention to hostility toward the opposing party) that may be additionally contributing to (perceived) political polarization in American politics.

Partisanship and Elite Expression

Partisanship has two essential and distinct components: one’s affinity for the in-party, i.e., positive partisanship, and one’s animosity toward the out-party, i.e., negative partisanship. Positive partisanship, or the psychological attachment to one’s political party (Campbell et al., 1980), is an affirmational identity, which is defined by what we represent (e.g., Democrats) and leads to greater assimilation to the in-group (e.g., the in-party in our case) than differentiating from the out-group (e.g., the out-party; Zhong et al., 2008a, 2008b). Thus, the in-group is psychologically primary, leading to ingroup favoritism. Because this does not necessarily imply prejudice against the outgroup, individuals with a positive group identity are motivated to help the in-group rather than to attack the out-group (Brewer, 1999).

In turn, negative partisanship, or “an affective repulsion from that party” (Caruana et al., 2015, p. 772), is a negational identity, which is defined by who we are not (e.g., not Republicans; Zhong et al., 2008a, 2008b). With negational identities, out-groups become the core reference point, encouraging individuals to differentiate themselves from the out-group to a greater extent than to assimilate to the in-group. Therefore, individuals with negative group identity tend to harm out-group members rather than benefit their in-group (Zhong et al., 2008b). Importantly, positive and negative partisanship are not necessarily mutually exclusive: although sometimes they work independently, they can also occur in parallel, such as when one favors and wants to benefit the in-group while disliking, and wishing to harm, the out-group (Anderson et al., 2021).

Some scholars claim that American politics is characterized by positive partisanship. The vast majority of partisans identify more with their own party rather than against the out-party (Theodoridis, 2019). This in-party affinity is also found to have greater effects on vote choice (Bankert, 2021; Caruana et al., 2015; Medeiros & Noël, 2014) and what news partisans want to disseminate (i.e., recommending to publish an article favoring the in-party vs. an article criticizing the out-party; Amira et al., 2021; suppressing an article hostile toward the in-party vs. promoting an article attacks the out-party, Lelkes & Westwood, 2017). Other work, however, suggests that negative partisanship exceeds positive partisanship (Finkel et al., 2020) and drives cognitions and behaviors. Out-party animosity is a stronger predictor of voting and other forms of political participation (Abramowitz & Webster, 2016; Iyengar & Krupenkin, 2018) and also of one’s satisfaction with the functioning of democracy (Ridge, 2022).

Yet other studies suggest that both positive and negative partisanship are strong among American citizens. The majority, 64%, of U.S. partisans displays high levels of both positive and negative partisanship (Bankert, 2021) and most partisans affiliate with their party because they support its policies and simultaneously oppose the other party’s policies (Pew, 2018b). Whether positive or negative partisanship drives political activities also varies by behavior. The former encourages engagement advancing one’s party (e.g., volunteering for one’s own party), whereas the latter exerts greater influence on opposition-oriented attitudes and activities such as anti-bipartisanship and protesting (Bankert, 2021; Caruana et al., 2015).

This literature sheds important light on the complexities of partisanship. Nevertheless, this work has two major limitations. For one, it focuses on rank-and-file partisans and overlooks politicians (but see Pew, 2018a). Are their political expressions primarily driven by positive or negative partisanship? Needless to say, members of Congress and the Cabinet and Presidential and Vice-Presidential nominees have a much greater influence over the political process than the majority of citizens. Elite communication shapes citizens’ opinions and behaviors. It can activate partisan identities and stereotypes, thus intensifying affective polarization (Druckman et al., 2013; Iyengar et al., 2012), distort citizens’ political preferences, such that party attachments outweigh substantive information (Mullinix, 2016; Nicholson, 2012), and mobilize political engagement, ranging from voting (Jackson & Carsey, 2007) to political discussions (Marquart et al., 2020). Studying elite communication on social media is especially crucial because politicians produce a large amount of political content and dominate political discussions on social media (Barberá et al., 2019). Also, social media platforms afford politicians the most direct and unfiltered channel of communication with the voters, which can have large scale democratic consequences.

Second, existing research relies heavily on survey self-reports to measure positive and negative partisanship. Self-reports have well-known limitations, such as response biases, which may overestimate or underestimate positive and negative partisanship due to normative pressures (Iyengar et al., 2019) and question wording (Druckman & Levendusky, 2019; Kingzette, 2021). Although experimental designs address this problem, their findings often lack external validity. Experimental settings are quite different from the real world, particularly when participants are forced to make decisions, such as how much money to afford to another player (Lee et al., 2022).

We address both gaps, studying elite messages online as well as individual reactions to these messages side by side in one project. The ability to collect online trace data allows researchers to study actual expressions and engagement in a naturalistic environment, and—when combined with machine learning approaches to automatically and at-scale classify content—can shed light on the role of positive and negative partisanship in motivating elite expressions and citizen engagement. We see these tracked expressions and engagements as proxies for positive and negative partisanship. If these behaviors are primarily driven by positive partisanship, politicians will be more likely to praise their own party than bashing the opposition, and messages praising the in-party will get more likes and shares, and vice versa for negative partisanship. Thus, expressions and engagement are advantageous over other frequently studied behaviors, for which the effects of positive partisanship are often entangled with those of negative partisanship. For instance, it is difficult to know whether someone voted for Joe Biden in 2020 mainly because they liked the Democratic Party or disliked the Republican Party.

It should be noted, however, that our approach is not free from limitations. Tweets may not always capture the actual feelings that the analyzed politicians have toward the in-party and/or the out-party and other factors inevitably play a role in elite expressions on social media. For instance, these expressions could simply reflect policy (dis)agreements, not negative/positive affect. Yet, policy (dis)agreements may well be motivated by affect when politicians attack out-party policies because they do not like the out-party, not because they truly disagree with its proposals (e.g., GOP Rep. Hinson condemned Biden’s infrastructure bill as “a partisan, socialist spending spree,” but claimed credit for it when the bill started to benefit her constituents). In addition, it is not possible to disentangle positive/negative partisanship from strategic considerations, which also motivate politicians’ online expressions. Nevertheless, in/out-party feelings in these expressions may be more consequential than the underlying strategizing and the private feelings. For example, Donald Trump may attack Joe Biden for strategic reasons, but the negativity in Trump’s tweet has significant effects on American democracy, greater than whether in private he actually likes Joe Biden.

Given the difficulties of directly measuring politicians’ feelings, attitudes, and beliefs, inferring them from actual behaviors is a common practice and the best possible approach in naturalistic settings (e.g., it is not feasible to conduct surveys or experiments among national political elites). For example, the widely used DW-NOMINATE score (Poole & Rosenthal, 2007) infers legislators’ ideological positions based on their voting records despite the fact that lawmakers’ voting decisions are influenced by many factors (e.g., pressures from party leadership or public opinion). In short, although imperfect, our approach is the only feasible for politicians and in fact—may provide more direct proxies than the approaches used in extant work on positive and negative partisanship.

Theoretical Expectations

We expect that elite expressions are more strongly driven by negative partisanship. This expectation is based on several literatures. First, scholarships on media coverage and on negative campaigning suggest that politicians and their party are frequently challenged and attacked by pundits, journalists, other politicians, and citizens from the opposite side (e.g., Berry & Sobieraj, 2013; Geer, 2008, 2012; Le et al., 2017; Puglisi & Snyder Jr, 2011). Under such circumstances, promoting their own party may not be enough, and politicians need to attack back to win the game (Amira et al., 2021). For example, politicians who are attacked employ more negative campaigning, responding to each attack with an attack (Iyengar, 2018; Lau & Pomper, 2001), although there is little evidence that negative campaigns are effective (see Lau & Rovner, 2009; Lau et al., 2007 for meta-analyses). In addition, the work on party polarization suggests that political elites are more ideologically divided than in the past (Layman et al., 2006; Theriault, 2008). In a divided system, the opposition becomes a greater threat to individual political identities, which leads to more negative attitudes toward the other side. In fact, research confirms that greater ideological division leads to stronger negative partisanship (Mason, 2015; Medeiros & Noël, 2014; Webster & Abramowitz, 2017). Third, the work on inter-party competition relatedly shows that the competition between the two parties is becoming fiercer than ever before (Lee, 2014). This motivates the parties and their politicians to distinguish themselves from and oppose the other side and also—as research shows—enhances negative partisanship (McGregor et al., 2015). In a study directly relevant to ours, Pew (2018a) examined lawmakers’ Facebook posts mentioning Obama, Clinton, Trump, the Democratic Party, and the Republican Party, finding that they are more likely to express opposition toward the out-party than support for their in-party. Integrating these works, we propose:

H1:

Politicians are more likely to publish posts negative toward the opposite party than posts positive toward their own party on social media.

In addition to this aggregate effect, we expect heterogeneous effects dependent on two factors. First, ideologically extreme politicians should exhibit stronger negative partisanship, in that they reject the characteristics of the out-party most strongly (Ridge, 2022). Indeed, previous work suggests that ideological extremity leads to more negative attitudes toward the out-party (Mason, 2015; McGregor et al., 2015; Medeiros & Noël, 2014; Webster & Abramowitz, 2017). In the tested contexts, politicians who are more extreme should be more likely to attack the out-party than moderate politicians. Offering some suggestive support for our expectation, Pew (2018a) found that the 10% most liberal and 10% most conservative legislators are more likely to attack the opposition than moderates (i.e., the 20% in the middle) on Facebook. We test ideological extremity in two validated ways, as noted below.

Second, negative partisanship should be more pronounced among members of the oppositional party, or Democrats in this case. This expectation is consistent with extant theorizing. For instance, Huddy (2013) notes that out-group hostility is especially likely to arise when a group is threatened (Brewer & Caporael, 2006) and that (perceived) threat activates the link between in-group identity and out-group hostility (Brewer, 2007). Recent work shows that partisans are more likely to develop a negative partisan identity if their party is portrayed as losing (Bankert, 2021). Inasmuch as electoral loss threatens one’s partisan identity (West & Iyengar, 2022), negative partisanship should be heightened. Furthermore, political psychology suggests that electoral losses activate anger (Huddy et al., 2015), which is likely to lead to a stronger desire to lash out at the out-party. Supporting these ideas, evidence shows that challengers are more likely to engage in negative campaigning and sponsor attack ads than incumbents (Gelman et al., 2021; Lau & Pomper, 2001) and politicians are more likely to attack the other party when their own party loses the election or is trailing in polls (Haber, 2011; Lau & Pomper, 2001; Russell, 2018). Directly germane to our argument, negative partisanship was a strong motivator for Democratic voters in the 2018 House election while positive partisanship had no impact (Bankert, 2021). Because the Democratic Party was the minority party in the Senate and was not in the White House from 2016 to 2020 (the timeframe from which we collected the data), Democrats’ partisan identity was threatened, which should have activated the processes described. Our next hypotheses predict:

H2:

Negative partisanship is stronger among politicians who are ideologically extreme such that they are more likely to publish posts negative toward the opposite party than those who are ideologically moderate.

H3:

Negative partisanship is stronger among Democratic politicians such that they are more likely to publish posts negative toward the opposite party than Republican politicians.

Partisanship and Citizens’ Reactions

In addition to looking at the prevalence of positive and negative partisanship in elite communications, our second objective is to examine how users react to politicians’ messages on social media. We note that Twitter users are not representative of the general population: they are younger, more educated, and more likely to be Democrats (Pew, 2019b). Among Twitter users, only a small share cares about politics (Pew, 2021a). Nevertheless, although small, this group is highly consequential. They produce the majority of engagements with elite tweets (Wojcieszak et al., 2022) and political content on Twitter (Pew, 2022), set political agendas (Barberá et al., 2019), and are more likely to engage in political activities online and offline than the general public (Pew, 2022). Tweets produced by this small group of politically active users are often used to represent public opinion (McGregor, 2019, 2020). If those users engage more when politicians express out-party animosity than in-party affinity, more users would encounter these posts, perceive the political climate as polarized, and develop a more negative attitude toward the opposition as a result (Levendusky & Malhotra, 2016). Moreover, if elite posts attacking the opposition receive more likes and shares, politicians would be encouraged to continue to attack the opposition, resulting in yet more toxic political climate and greater out-party hostilities.

Which messages by politicians should be most liked and shared? On one hand, according to the negativity bias theory, humans are evolutionarily more likely to respond to negative than positive environmental stimuli (Rozin & Royzman, 2001; Soroka & McAdams, 2015). This pattern, demonstrated across contexts and situations, also applies to media consumption. Research shows that, relative to positive content, negative messages are more likely to be selected (Trussler & Soroka, 2014), generate stronger and more sustained reactions (Soroka et al., 2019), and have stronger influence on information processing (Ito et al., 1998). Therefore, users should be more attentive and attracted to expressions of negative partisanship (i.e., posts attacking the out-party) than to expressions of positive partisanship (i.e., posts praising the in-party), which would result in greater engagement (i.e., likes and shares).

On the other hand, tweets praising the in-party may generate greater engagement. Citizens mostly follow congenial accounts on social media and discuss politics with their partisan in-group (e.g., Cinelli et al., 2021; Himelboim et al., 2013; Wojcieszak et al., 2022), thus avoiding threats to partisan identity. Therefore, their negative feelings toward the out-party may not be salient, which could lead to the desire to benefit the in-party (e.g., liking and sharing posts praising the in-party) rather than to hurt the out-party (e.g., liking and sharing posts attacking the out-party, Amira et al., 2021). Indeed, when their partisan identities are not threatened, Democrats and Republicans prefer to publish an article (make it seen by others, akin to sharing and liking) praising their own side rather than the one attacking the out-party (Amira et al., 2021). Given these distinct possibilities, we advance the following question:

RQ1:

Will posts positive toward one’s own party or posts negative toward the opposite party get more likes and shares?

Methods

Data

To test our hypotheses, a pre-registered studyFootnote 2 was carried out in October, 2020. We used the Twitter API to collect the timelines of then members of the House and Senate (i.e., the 116th Congress) as well as other influential politicians (i.e., Presidential and Vice-Presidential nominees of the two major parties from 2000 to 2020 and then members of the Trump Cabinet).Footnote 3 We started with an existing dataset of official accounts for members of the 116th Congress, maintained by the @unitedstates project (https://github.com/unitedstates/congress-legislators). A research assistant manually added official Twitter accounts of the remaining politicians and personal/campaign accounts of all national political actors (e.g., for Rep. Alexandria Ocasio-Cortez, both her official account @RepAOC and personal account @AOC were included).Footnote 4 The most recent 3200 tweets sent by each of those accounts were collected, excluding retweets and quote tweetsFootnote 5 as well as those posted before the 2016 election.Footnote 6 The final sample has 1,195,844 tweets sent by 1018 accounts owned by 564 elites with the first tweet sent on November 9, 2016 and the last tweet on October 29, 2020.Footnote 7

Measures

To examine whether politicians attack their opponents or praise their own party on Twitter, we first need to determine whether a tweet discussed the two major parties or party elites. A dictionary approach was adopted, such that for each of the two major parties, we created a list of keywords consisting of full names, nicknames, and Twitter handles of politicians from that party and the party itself and searched tweets mentioning these keywords within our dataset.Footnote 8 That is, if a tweet included any of the Republican keywords such as “Donald Trump”, “drumpf”, “GOP”, or “Republican,” we assumed that it was directed at the Republican Party (n tweets discussed the Democratic Party = 109,840; n tweets discussed the Republican Party = 151,429).Footnote 9 Tweets that did not mention either party or its key politicians were excluded from analysis. We then converted the target of the tweets from the Democratic and/or Republican Party to in-party and/or out-party, based on the consistency of politicians’ party affiliation with the party they tweeted about (e.g., in-party if a Republican senator’s tweet discussed the Republican Party).

The next step is to establish the sentiment of the tweets: positive toward the in-party (i.e., positive partisanship) or out-party, neutral toward the in-party or out-party, and/or negative toward the in-party or out-party (i.e., negative partisanship). To do so, two trained coders manually annotated a random sample of tweets (n = 10,000, with 50% tweets targeting each party) for whether they were negative, neutral, or positive toward the targeted party.Footnote 10 We used these labeled data to train five machine learning models separately for posts targeting the Democratic and the Republican Party: (a) a K-Nearest-Neighbor model, (b) a Decision Tree, (c) a Gradient Boosting Machine, (d) a Support Vector Machine, and (e) a Random Forest model.Footnote 11 We transformed the text to lowercase, removed URLs, numbers, punctuations, stopwords, and non-ASCII characters, and lemmatized the remaining tokens to create an IF-TDF matrix with a sparsity of .998.Footnote 12

The accuracy of the algorithms was tested using fivefold cross-validation and an 80/20 train-test split. Figure 1 shows the overall accuracy of each of the five classifiers. Moreover, we also assessed the ability of each classifier to predict each target category. Figure 2 shows the precision, recall, F1, and balanced accuracy at the target class level for each of the classifiers. The Random Forest achieved high overall accuracy and a better balance between precision and recall—meaning the probabilities of false positive and false negative are roughly equal—for all three categories. Thus, we used the Random Forest classifier to predict the tone of tweets. The findings are robust when using predictions produced by alternative models (see Online Appendix B).

Fig. 1
figure 1

Overall out-of-sample accuracy of the five machine learning classifiers trained to predict attitudes of the tweets (top panel for tweets discussing the Democratic Party and bottom panel for tweets discussing the Republican Party)

Full size image
Fig. 2
figure 2

Target-class-level precision, recall and F1, and balanced accuracy for the five machine learning classifiers trained to predict attitudes of the tweets (top three panels for tweets discussing the Democratic Party and bottom three panels for tweets discussing the Republican Party). The horizontal red lines indicate the proportion of tweets coded as negative, neutral, and positive in the test set (Color figure online)

Full size image

We used two different approaches to measure ideological extremity to test its effects on the expressions of positive and negative partisanship in the tweets. First, we used the dataset maintained by Voteview (Lewis et al., 2020) to get the DW-NOMINATE scores for members of the House and Senate (other politicians do not have the score). DW-NOMINATE is an estimate of ideology that places lawmakers on a liberal (− 1) vs. conservative (+ 1) scale based on their voting decisions. We created a continuous ideological extremity variable by (1) calculating the average DW-NOMINATE score for all actors, (2) calculating the difference for each actor between the average score and the score of the actor, (3) extracting the absolute value for this difference, and (4) rescaling it between 0 and 1 (M = 0.46, SD = 0.17). Second, we also used Barberá’s (2015) validated Bayesian Spatial Following model, which estimates one’s ideological position on a continuous scale by examining elite accounts one follows. We created an extremity variable following the steps above (M = 0.33, SD = 0.13).

In order to test RQ1, we also collected the number of likes and retweets using the Twitter API. Because these measures were skewed (e.g., some tweets only had five likes while some of Donald Trump’s tweets had thousands), we transformed the data by taking the log of (1 + x).

In order to ascertain that the tested patterns are robust and not due to some confounders (e.g., users are more likely to engage with tweets about certain policies; the tweets by politicians with more followers are liked or reshared more frequently), our models included stringent covariates. Topic of the tweet. We trained a Convolutional Neural Net (CNN) to predict the presence of topics from the Comparative Agendas Project in the tweets (see Online Appendix C). This variable was used as a control variable when testing RQ1 as citizens may react differently to tweets discussing different topics. Additionally, the number of followers, the number of accounts one follows, and time of the day the tweet is posted may influence the way politicians discuss politics, the reach of their tweets, and consequently, the number of likes and retweets. We controlled for these factors when testing H2, H3, and RQ1. Because the number of followers and accounts followed were skewed, they were transformed by taking the log of (1 + x). Time of the day the tweet was posted was categorized into six levels (00:00–03:59, 04:00–07:59, 08:00–11:59, 12:00–15:59, 16:00–19:59, 20:00–23:59).

Results

In total, 260,605 tweets sent by 1008 accounts owned by 563 elites discuss the two parties or their key politicians (39% of these tweets are from 2020, 32% from 2019, 18% from 2018, 10% from 2017, and 1% from 2016). 526 Democratic accounts send 144,319 (55%) tweets while 482 Republican accounts send 116,286 (45%) tweets. The majority of politicians (99%) post fewer than 1000 relevant tweets during the 4 years analyzed (see Fig. 3).

Fig. 3
figure 3

The distribution of accounts based on the number of tweets posted

Full size image

Our first expectation was that politicians would predominantly exhibit negative partisanship in their social media posts, resulting in more posts criticizing the opposition than posts positive toward their own party. We compared the number of tweets negative toward the out-party to the number of tweets positive toward the in-party, as determined by our classifier. As seen in Fig. 4, politicians, as a whole, post 19% more tweets positive toward the in-party than tweets negative toward the out-party. This is a non-trivial difference. Although not pre-registered, we calculated the ratio of tweets positive toward the in-party to tweets negative toward the out-party for each account. The result shows that the majority of elite accounts (68%) post more tweets praising one’s own side than tweets attacking the opposition.Footnote 13 The same pattern emerges for 12 accounts owned by eight most powerful politicians (i.e., Donald Trump, Mitt Romney, Mike Pence, Mitch McConnell, Joe Biden, Barack Obama, Hillary Clinton, and Nancy Pelosi): as a group, they post 39% more positive in-party tweets than negative out-party tweets. Individually, eight of these 12 accounts (or six out of eight politicians if we combine the official and personal accounts) post more tweets positive toward the in-party than tweets negative toward the out-party. Donald Trump and Joe Biden are the only two exceptions, a finding we return to in the discussion. Thus, H1 is rejected.

Fig. 4
figure 4

Number of tweets by sentiment grouped by target

Full size image

Among all elite tweets discussing the two parties and key politicians, 35% are positive toward the in-party and 29% are negative toward the out-party. Almost a quarter are neutral toward either the in-party (15%) or out-party (4%). Only on rare occasions do politicians criticize the in-party (7%; e.g., Doug Collins (R-GA): “Kelly won’t answer the question because she doesn’t want Georgians to know the truth: @KLoeffler and @MittRomney are cut from the same Never Trump cloth—two self-funders putting self before country.”) and praise the out-party (9%; e.g., Jim Langevin (D-RI): “Congratulations to my good friend & GOP colleague @CongressmanGT for his work on reauthorizing the #PerkinsAct…”).

H2 predicted that ideologically extreme politicians will exhibit stronger negative partisanship, posting tweets negative toward the opposing party, than the moderates. We first tested it at the tweet level by estimating two binary logistic regression models, predicting whether tweets mentioning the out-party were negative toward the out-party versus neutral or positive (tweets about the in-party were excluded). The first model tested the extremity of the politician as determined by the DW-NOMINATE score. We see a consistent pattern supporting our expectation. Controlling for the number of followers, the number of accounts followed, and the time of the day the tweet is sent, a one-unit increase in extremity, assessed on a normalized DW-NOMINATE scale, leads to a 730% increase in the odds of a tweet being negative toward the out-party (vs. positive or neutral toward the out-party). Clearly, the more extreme actors exhibit greater negative partisanship than those who are more moderate (see Table 1).

Table 1 The effects of ideological extremity on elite negative partisanship
Full size table

We estimated a parallel model using ideological extremity based on Barberá’s method. The results are nearly identical—for a one-unit increase in extremity on a normalized scale from moderate to most extreme, the odds of a tweet being negative toward the out-party (vs. positive or neutral toward the out-party) increases by 701% (see Table 1). For example, @replucymcbath (owned by Rep. Lucy McBath), the most moderate account according to the Bayesian Spatial Following model, post 20 tweets discussing the out-party or its key politicians and only five (25%) are negative. In contrast, @bryansteil (owned by Rep. Bryan Steil), one of the most extreme accounts, post 50 tweets about the opposition and 31 (62%) are negative.Footnote 14

We additionally tested H2 by examining the proportion of tweets negative toward the other side. For each account, we divided the number of tweets negative toward the out-party by the sum of tweets negative toward the out-party and positive toward the in-party (i.e., out-party negative/(out-party negative + in-party positive)) and regressed it on a politician’s ideological extremity, controlling for the number of followers and the number of accounts followed. A one-unit increase in the normalized DW-NOMINATE based ideological extremity leads to a 24% increase in the proportion of tweets negative toward the out-party. For instance, Sen. Susan Collins, a famous moderate Republican, post 1 tweet attacking the Democrats and 53 tweets praising her own party (2%) using her official account @SenatorCollins; in contrast, Sen. Ted Cruz, a very conservative Republican, post 354 and 38 tweets attacking the Democrats and 241 and 15 tweets praising his own party (59%; 72%) using his official account @sentedcruz and personal account @tedcruz, respectively. In a parallel model, ideological extremity based on the Bayesian Spatial Following model has no significant influence on the proportion of tweets negative toward the out-party, but the effects are in the predicted direction (see Table 1).Footnote 15 In short, although there is one exception, H2 is supported in general.Footnote 16

Furthermore, we tested whether members of the opposition party (i.e., Democrats in our data) exhibit stronger negative partisanship than politicians whose party is currently in power (i.e., Republicans; H3). We estimated a binary logistic regression model predicting whether a politician’s tweets were negative versus neutral or positive toward the out-party (tweets about the in-party were excluded). Consistent with H3, we find that the odds of being negative toward the out-party (vs. positive or neutral) are 56% higher for Democrats than Republicans.Footnote 17 Also, the proportion of tweets negative toward the out-party (vs. positive toward the in-party) is 8% higher for Democratic politicians, compared to that of Republicans (see Table 2). Although not pre-registered, we also tested how Democrats winning the House back in the 2018 midterm election influenced negative partisanship. Results show that Democrats regaining the House indeed makes them less negative and Republicans more negative toward the out-party on Twitter, offering additional support to H3 (see Table 3).Footnote 18

Table 2 The effects of ruling/opposition status on elite negative partisanship
Full size table
Table 3 The effects of 2018 midterm election on elite negative partisanship
Full size table

Before examining whether users engage most with tweets representing positive or negative partisanship, we describe the engagement data. In the aggregate, the median number of likes on elite posts is 32 and the median number of retweets is 10 (mean likes = 1798, mean retweets = 481). Tweets sent by Democratic politicians have a median of 34 likes and 11 retweets (mean likes = 1850, mean retweets = 430) while tweets by Republicans have a median of 28 likes and 9 retweets (mean likes = 1735, mean shares = 543). 82,570 (32%) tweets receive more than 100 likes and 51,860 (20%) tweets get more than 100 retweets.

To address RQ1, we created a two-level “tweet type” variable by combining the target and sentiment (i.e., positive toward the in-party vs. negative toward the out-party; other tweets excluded). We find clear patterns of greater engagement with elite messages reflecting negative partisanship. Whereas tweets positive toward the in-party have a median of 29 likes and 9 retweets, elite tweets negative toward the out-party have 57 likes and 22 retweets (see Online Appendix D for other types of tweets). To assure that these patterns hold when accounting for potential confounders, such as elite partisanship (i.e., elite tweets are liked and shared not only because of their party affiliation), the number of followers and accounts followed, the topic of the tweet, and the time of the day (i.e., tweets posted at late night may get fewer likes and retweets), we included these in models that regressed the number of likes and retweets on tweet type. Table 4 clearly shows that compared to tweets positive toward the in-party, those that are negative toward the out-party are liked and retweeted more even after accounting for these stringent covariates.Footnote 19

Table 4 The effects of positive/negative partisanship on the number of likes and retweets
Full size table

Although not pre-registered, we tested whether the number of likes and shares depends on elite partisanship and ideological extremity. Tweets attacking the opposition get more likes and shares if they are sent by Republican than Democratic politicians. In contrast, tweets praising the in-party are liked and shared more if they are posted by Democrats than Republicans. When it comes to elite ideological extremity, tweets positive toward the in-party are liked and shared more and tweets negative toward the out-party are liked and shared less when they are posted by ideological extremists, as indicated by Bayesian Spatial Following model based extremity score (but models using DW-NOMINATE based extremity score show an opposite pattern for likes and null effects for shares; see Online Appendix D for details).

Discussion

Our project examined social media expressions of American politicians as well as users’ engagement with these tweets to offer comprehensive evidence on the effects of positive and negative partisanship on elites and citizens. We offer three noteworthy findings. First, American politicians are more likely to express support toward their own party than to speak out against the opponents, indicating that the overall charge of partisanship for these elites is positive when they discuss politics on social media. This robust pattern emerges when comparing the total number of tweets praising the in-party with that of tweets criticizing the out-party and also calculating the proportion of elite accounts/elites that post more tweets expressing in-party favorability than those expressing out-party negativity. Although many scholars observe that negative partisanship is on the rise in the US (e.g., Abramowitz & Webster, 2016; Iyengar et al., 2012), our evidence suggests that positive partisanship still dominates political expressions of American politicians on Twitter, so—at least in this context—our findings are more optimistic than the general observations. However, we note that a large share of elite tweets attacks the opposition, an issue we address in more detail below. Because we analyzed an exhaustive set of 564 political elites, identified all their tweets mentioning the in-party and the out-party, and employed validated classifiers to analyze those tweets, we are confident that our results are a robust and accurate representation of political expressions by American politicians on Twitter.

Second, in an exception to this general pattern, we find that politicians who are ideologically more extreme are more likely to express negative partisanship than their more moderate counterparts: the former group is most motivated to attack the out-party. This largely holds for two measures of extremity and is consistent with prior literature that ideological polarization contributes to negative feelings toward the opposition (Abramowitz & Webster, 2018; Webster & Abramowitz, 2017). Combined with the fact that negative partisanship does not trail by a large margin, it implies that if the ideological division between the two parties keeps widening, negative partisanship may increase and may outstrip positive partisanship in elite opinion expressions. Moreover, in line with previous evidence that threats to party identity strengthen negative partisanship (Amira et al., 2021), members of the party in the opposition—i.e., Democrats—show greater negative partisanship than those of the ruling party and they become less negative toward the out-party after winning the House back in the 2018 midterm election. This finding suggests that negative partisanship may be used by the opposition as a tool to unite the party and against the common enemy (Bankert, 2022), yet more research on the temporal variations in and factors influencing positive and negative partisanship among politicians is needed to shed more light on this consequential finding.

Third, despite the overall dominance of positive partisanship among the elites, it is negative partisanship that drives ordinary citizens’ liking and sharing of elite messages. Tweets that attack the out-party receive more likes and retweets than those that favor the in-party. Although this contradicts previous finding that citizens penalize politicians for attacking the out-party (Costa, 2021), the inconsistency could be due to different methodologies and different demographics of the general public vs. (politically active) Twitter users. That said, we suspect our finding is not constrained to Twitter. Extensive work has established that “negativity bias” holds across studies, cultures, samples, and in divergent contexts (e.g., Ito et al., 1998; Soroka et al., 2019; Trussler & Soroka, 2014). More specific to our work, although the most politically active Twitter users follow in-party accounts, are more sorted, and have colder feelings toward the out-party than average Twitter users (Pew, 2019a), which could partly explain our finding, the very same characteristics apply to politically active and interested citizens in general (Levendusky, 2009), whether on Twitter or not. Lastly, a recent study (Rathje et al., 2021) found similar patterns across Twitter and Facebook, suggesting that greater engagement with out-party animosity is not unique to Twitter users.

When interpreting the results, several limitations need to be noted. First, as the data were collected before the 2020 election, it is unclear if these patterns would emerge after Joe Biden became the president. It is possible that Twitter suspending the account of Donald Trump and other events may have changed the nature of elite expressions. Second, as the American political system, media system, and party structures are unique, the findings of our study should not be generalized to other countries. More international and comparative research should be done to test whether politicians and citizens in different countries behave similarly or differently.

Third, we focus on tweets that explicitly mention the two parties and key politicians but do not account for policies, media organizations, or other factors associated with a party. It is very challenging to construct a comprehensive and unbiased list of political topics discussed from 2016 to 2020, and assessing how politicians talk about policies and other things is an important task for future work.

Fourth, we do not know the partisan affiliations of those users who engaged with elite tweets. We are interested in the more foundational evidence about elite expressions and citizen engagement; the questions surrounding political homophily are secondary and cannot be tested in our data. We speculate that users are more likely to engage with messages attacking the out-party because they primarily follow and interact with likeminded others (e.g., Eady et al., 2019; Mosleh et al., 2021; Wojcieszak et al., 2022). But regardless of this ideological consistency, our evidence shows that negativity toward the out-party, per se, is engaging and gets more attention (see also Rathje et al., 2021). In short, although the specific mechanisms behind these engagements cannot be disentangled in our data, we suspect it is the mere negativity combined with ideological consistency. It is likely that tweets negative toward the out-party infuriate out-party members, yet there are no reasons for those users to like and share these tweets (i.e., the behaviors we examined). Future work should explore whether out-party members are activated by such posts and fight back by commenting and quote tweeting.

Despite these limitations, our findings have important implications for American politics. To begin with, although elite expressions are mainly positive, the difference between the strength of positive and negative partisanship is not substantial. Some of the most powerful and most widely followed politicians, such as Trump and Biden, are also the ones who are most negative toward the out-party. Donald Trump post 440% (162 vs. 30) and Joe Biden post 294% (563 vs. 143) more tweets negative toward the out-party than tweets positive toward the in-party, respectively. Therefore, out-party negativity may actually reach a larger audience than in-party positivity. If we simply use the number of followers as a rough indicator of readership (e.g., if a politician has one million followers and they post two out-party attacks, then we assume that these tweets are read two million times), overall, tweets attacking the out-party are read 27% more than tweets praising the in-party, although the former is outnumbered by the latter.Footnote 20

As importantly, negative partisanship may be amplified through citizen-elite interactions and through other means of communication. In our data, citizens reward politicians for attacking the opposition with more likes and shares, encouraging politicians to express out-party hostility more fiercely; and more exposure to elite tweets attacking the out-party may, in turn, generate greater citizen engagement and enhance out-party hostility among citizens. In fact, tweets sent by extreme in-party elites—who are most negative toward the out-party—are also more likely to be shared (Wojcieszak et al., 2022). Other work also shows that news organizations are more likely to cover extreme politicians than moderate politicians (Wagner & Gruszczynski, 2018). This amplification, by social media users, journalists, and news organizations, may further increase the general perception that American elites are hostile toward the other side. Questions as to whether users’ preferences for out-party negativity on social media are further exacerbated by Twitter’s algorithmFootnote 21 or whether elite communication as mediated through mainstream media is more or less negative compared to social media are important directions for future work.

These two factors—the most powerful politicians are very negative toward the out-party and elite negative partisanship may be amplified by users and media—may ultimately create “illusion of polarization” in that citizens encounter a greater share of negative than positive elite information and expressions. Again, however, our over-time evidence systematically shows that this is an illusion, in that most elites, most of the time, post information and opinions that are not negative towards their political opponents. After accounting for the number of followers, the modal tone of the political elites is positive—politicians who mostly post positive tweets, as a whole, have 58% more followers than politicians who are mainly negative. Simply, the positive majority is not as influential and popular as the negative minority (e.g., an average negative politician has 43% more followers than an average positive politician).

We hope these findings inspire further research on the political use of social media, research that investigates the dynamics between politicians, media, and citizens using actual behavioral data and ideally, across different platforms. Only then will we be able to have a complete overview of the social media ecosystem, correctly identify problems, and come up with solutions.