1. Polarizing Rhetoric on Twitter

Following Fortuna, C (Reference Fortuna2019), we conceptualize polarizing rhetoric, as a communication strategy that divides the audience into two opposing groups. Polarizing messages have two characteristics: they suggest a strong group cohesion and the existence of a common enemy. In this sense, polarizing rhetoric is more a communication style than a stable characteristic of politicians. Political actors can express opinions on volatile and divisive issues without using polarizing rhetoric and use polarizing rhetoric when discussing relatively innocuous issues.

It is important to note that polarization is not the same as polarizing rhetoric. The first one refers to the process of forming more extreme or divergent viewpoints within a society; the second one is a communication strategy that accentuates these divisions. Neither is polarizing rhetoric merely a manifestation of existing polarization. As elucidated by Ballard et al. (Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2023), polarizing rhetoric is a deliberate communicative style that seeks to create division by aligning the speaker with an ingroup and positioning them against an outgroup. This technique is rooted in social identity and the distinction between ingroups and outgroups. It fosters sentiments towards the ingroup and negative attitudes towards the outgroup. In politics, polarizing rhetoric often involves emphasizing partisan or ideological differences to confront political adversaries. In our analysis of the 2022 Colombian presidential elections, we examine when and how candidates use polarizing rhetoric, and how it affects user engagement on Twitter (Ballard et al. Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2023).

According to the literature, messages from individuals who use polarizing rhetoric are usually more emotional (Groenedyk and Banks Reference Groenendyk and Banks2014): they tend to be enthusiastic, angry, and aggressive. These messages actively seek to provoke greater partisan division by choosing a specific tone and sentiment to frame preferences on various public policies (Albertson, Dun, and Gadarian Reference Albertson, Dun and Gadarian2020). Examples of such messages include personal attacks on political opponents or disqualification of opposing political groups (Gardner and Russell Reference Gardner and Russell2022).

With some exceptions (see Gervais et al. Reference Gervais, Evans and Russell2020) existing work has found that politicians use social media and, particularly, polarizing rhetoric, to enhance their visibility. Polarizing language fosters solidarity within a group against a common external enemy (Ballard et al. Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2022). It constructs a common identity that helps political actors establish a clear identification with their followers (McGowan Reference McGowan2015; El-Shinnawy and Vinze Reference El-Shinnawy and Vinze1998) while also differentiating themselves from their opponents (Myers and Lahmm Reference Myers and Lamm1976; El-Shinnawy and Vinze Reference El-Shinnawy and Vinze1998). Using this language, politicians not only promote their image, but also denigrate their competitors. In various countries, political actors have resorted to aggressive language on Twitter to attract the attention of traditional media, gain followers, and reach audiences that are not accustomed to using digital platforms for political information (Enli Reference Enli2017).

This kind of rhetoric, however, does not come without risks. Polarizing rhetoric could alienate important sectors of the electorate or make a candidate look negative and controversial (Hong and Kim Reference Hong and Kim2016; Enli Reference Enli2017). Politicians, therefore, employ politically uncivil messages on Twitter when the gains from increased visibility outweigh potential electoral risks. Under this logic, smaller, obscure, or out-of-power parties and candidates have more to win out of polarizing rhetoric than bigger, better known, or incumbent parties or candidates. The kind of connection they can build and the fact that they have little to lose, makes polarizing rhetoric particularly appealing to these less visible contenders, who are willing to use aggressive, polarizing language to gain media attention and followers, despite the risk of alienating voters and sparking controversies (Ballard et al. Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2022; Heiss et al. Reference Heiss, Schmuck and Matthes2019).

Indeed, according to Jungherr (Reference Jungherr2016) and Ballard et al. (Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2022), for example, opposition parties and candidates are not only more likely to use Twitter than ruling parties, but they are also more likely to use polarizing rhetoric than their incumbent counterparts. Candidates with extreme ideological positions, opposition members, and electoral losers are also more likely to use social networks to communicate with citizens (Hong and Kim Reference Hong and Kim2016). Using polarizing rhetoric, they denounce corruption, abuse of power, and unpopular policies to increase visibility and attract followers, highlighting differences between parties (Bennett and Livingston Reference Bennett and Livingston2018). Thinking about these issues in the specific context of the Colombian presidential race (which we characterize in detail later on) we expect that candidates who are trailing during the campaign to be more likely to resort to polarizing rhetoric than candidates who are ahead.

It is unclear, however, if this polarizing rhetoric provides the results desired by those who use it. In general, scholars agree that polarizing language enhances visibility in the digital realm. The question is, does polarizing rhetoric increase engagement? Previous research has shown that negative, uncivil messages can translate into greater interaction with social media content, increasing likes, comments, or different modes of sharing (Rega and Marchetti Reference Rega and Marchetti2021). Social media users are more likely to support and disseminate messages with polarizing content (Browniatowski et al. Reference Broniatowski, Jamison, Qi, Al Kulaib, Chen, A. Benton and Dredze2018) and Twitter interactions seem to increase when politicians’ messages contain negative content toward political outgroups (Enli Reference Enli2017). Twitter’s algorithm also seems to favor tweets with higher engagement, potentially further enhancing their visibility and interaction rates (Tenemboin Reference Tenenboim2022). Indeed, Ballard et al. (Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2022) show that polarizing language in tweets can enhance the number of retweets garnered by US politicians. Likewise, using a related concept (populist rhetoric) Cassell (Reference Cassell2021) suggests that incivility also increases engagement, particularly in Latin America. Based on this, we hypothesize that:

A key feature driving engagement, and one closely connected with polarizing rhetoric is the tone of the message. Following many works in the political communication literature, and in particular recent research evaluating tone on Twitter (Brie and Dufresne Reference Brie and Dufresne2020; Maurer and Diehl Reference Maurer and Diehl2020; Theocharis et al. Reference Theocharis, Barberá, Fazekas and Popa2020), we focus on tone as a general positive or negative sentiment that is conveyed through word choice. The relationship between polarizing rhetoric and negative tone in tweets and messages is a complex and multifaceted issue that has been explored in various studies. While these two elements are distinct, they often intersect in political communication, particularly on social media platforms.

Following Maurer and Diehl (Reference Maurer and Diehl2020) we sketch a distinction between polarizing rhetoric and tone. As we have stated, a key feature of polarizing rhetoric is that it involves a clear distinction between the ingroup and the outgroup, while, in contrast, tone is a more general sentiment that can be directed towards any target, not necessarily an opposing group. In other words, polarizing messages need not to be negative, and negative messages need not to be polarizing. That said, there is research that suggests a strong correlation between polarizing rhetoric and tone (Fridkin and Kenney Reference Fridkin and Kenney2011). Polarizing rhetoric often employs a negative tone when referring to the opposing group as a tool to create divisions and rally opponents (Enli Reference Enli2017); for example, “mudslinging” between candidates on Twitter has been found to intensify polarization (Evans et al. Reference Evans, Smith, Gonzales and Strouse2017).

Though not all polarizing rhetoric is negative and not all negative rhetoric is polarizing (Ballard et al. Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2023), negative and polarizing political messaging on Twitter have been shown to increase engagement in general (Lee and Xu Reference Lee and Xu2018; Lazarus and Thornton Reference Lazarus and Thornton2021). There is no clear consensus (see for example Stieglitz and Dang-Xuan Reference Stieglitz and Dang-Xuan2013), but the literature on psychology and political communication acknowledges a “negativity bias” according to which negative stimuli tend to elicit a stronger response. With that in mind, we expect a negative tone to increase engagement in the context of the Colombian presidential campaign:

Given the correlation between tone and polarizing rhetoric, we also expect polarizing rhetoric to enhance the effect of negative tone. Therefore, we hypothesize that:

2. Twitter Data and Methodology

To assess these hypotheses, we use data from Colombia’s 2022 election. With the help of a research assistant, we used Twitter’s API to download and analyze tweets composed by four presidential candidates: Gustavo Petro, Rodolfo Hernández, Federico Gutiérrez, and Sergio Fajardo.

We downloaded all the tweets produced by these four candidates between March 1 and June 21. For each tweet we collected the number of likes and retweets, as well as followers at the time of the tweet. We also recorded the exact date and time of the tweet. This dataset contains 7,425 observations distributed as shown below (Table 1). Each observation has on average 5,892 likes and 1,403 retweets (see Table A1 in the Appendix for more details).^{Footnote 1}

Table 1.

Distribution of Tweets by Candidate

Because we are interested in analyzing both the content and tone of the candidate’s messages, as well as the distribution of and interaction with (de)polarizing messages, we hired additional RAs^{Footnote 2} to carry out content analysis, manually classifying these candidates’ original tweets. These RAs did three types of coding. First, they coded the topics candidates tweeted about inductively into 31 categories (see Table A3 and Figure A2 in the Appendix). Overall, they were able to classify 4,394 tweets (97% of all original tweets). Two thousand two hundred and thirty-six (50.9%) messages were classified as “campaign” related, that is, tweets that discussed the campaign itself (invitations to rallies, distribution of campaign information, etc.). Two hundred and sixty-five (6%) were classified as corruption related, that is, they criticized corruption. Two hundred and ninety-two (6.7%) of all original tweets focused on other candidates (2.9% on Petro, 1.6% on Hernández and Gutiérrez, 0.27% on Fajardo and 0.20% on a combination of presidential hopefuls). The rest dealt with various policy issues: mostly the economy (494, 11.24%) and education (212, 4.8%), followed by issues related to security and the armed conflict (255, 5.8%), health care (49, 1.1%) and social security (67, 1.5%), as well as rural politics (including land reform) (111, 2.5%). Besides these topics, candidates also tweeted about personal topics (103, 2.3%), gender (128, 2.9%), the environment (90, 2.1%), and others.

In addition to classifying tweets by topic, we asked our RAs to categorize them according to whether they deployed (or not) a polarizing rhetoric. Inspired by Ballard et al (Reference Ballard, DeTamble, Dorsey, Heseltine and Johnson2022), we coded as polarizing tweets those that fulfilled two conditions: they clearly identified an ingroup and an outgroup, and the candidate tweeting was part of the ingroup. We did not classify tweets as polarizing if they contained criticisms towards political adversaries, unless these criticisms were directed towards specific, defined groups. For example, on June 5, Gustavo Petro tweeted the message below. This tweet clearly identifies an ingroup (the non-fascists, not Uribistas) and an outgroup (fascists, Uribistas).

By the same token we identified depolarizing tweets as those messages that promoted pluralism, explicitly accepted democratic rules, or called for reconciliation, thereby reflecting a commitment to dialogue and cooperation in the political space. For example, on June 24, Gustavo Petro tweeted:

We identified as “neither” those tweets that were not regarded as polarizing or de-polarizing. This group includes those messages that failed to identify an ingroup and an outgroup (or did so but the candidate tweeting was not part of the ingroup), as well as those messages that did not reflect a commitment to dialogue and cooperation by promoting pluralism, explicitly accepting democratic rules, or calling for conciliation. An example of a neutral tweet would be the following one by Rodolfo Hernández on May 21:

As shown in Table 2, most tweets produced by candidates (92%) did not use polarizing or depolarizing rhetoric. These “neither” tweets provided logistical campaign information (46%); referenced the candidate’s policy proposals (37%); described and commented on places visited during the campaign (11%), or signaled famous individuals’ support for the candidate (11%). Only 6% of the messages were polarizing, and 2% were depolarizing.

Table 2.

Original Tweets by Tone and Polarizing Rhetoric

Lastly, we analyzed the tweets’ tone (sentiment). Tweets that appealed to negative feelings (i.e., sadness, rage, anxiety, etc.) were deemed as having a “negative” tone. Tweets that appealed to positive feelings (i.e., happiness, hope, etc.) were deemed as having a “positive” tone. Tweets that did not appeal to any feeling (positive or negative), were deemed as “neutral.” As shown in Table 2 above, most original tweets (62.5%) have a positive tone; a third of all tweets (28%) have a negative tone; and very few of them (9.4%) are neutral.

Not surprisingly, given the literature discussed above, there is a strong correlation between tone and polarizing rhetoric. Most polarizing tweets are negative (89%), while most depolarizing tweets are positive (83.7%). Neither polarizing nor depolarizing tweets have messages with a “neutral” tone.

We use this data to analyze the effect of polarizing and depolarizing rhetoric on engagement. We operationalize engagement using the natural logarithm of likes and retweets received by each message.^{Footnote 3} We measure polarizing and depolarizing rhetoric using a categorical variable that scores 1 if the tweet uses polarizing rhetoric, 3 if it uses depolarizing rhetoric, and 2 if it does not use any of these rhetorics.

Aside from these two indicators, we use five controls. First, we control for the candidate to account for visibility. Since some individuals are better known than others on Twitter, who writes the tweet can have an important impact on engagement. With that in mind, we introduced a four-category variable (user) that scores 1 if the tweet was produced by Petro, 2 if it was produced by Hernández, 3 if it was produced by Gutiérrez, and 4 if it was produced by Fajardo.

We also control for topic. In any given campaign, some topics are bound to be more popular than others. We classified topics through a 10-category variable grouping each of the 31 topics identified by the RAs into larger groups (see Table A3 in the Appendix for details). In a similar vein, we also control for the stage of the campaign. As we will see below, candidates’ Twitter activity changed throughout the duration of the race. For example: the primaries and the runoff saw less Twitter activity than the first round. With that in mind, we add a variable that scores 1 if the tweets were published during the primaries (on or before March 7, 2022), 2 if they were published during the first round of elections (between March 8 and May 22 of 2022), and 3 if they were published during the runoff (between May 23 and June 14, 2022).

Lastly, we control for a couple of platform tools: hashtags and tags. The literature on social media suggests that hashtags and tags can increase engagement. With that in mind, we introduced two dichotomous variables that score 1 if the tweet has a hashtag or is tagging another user, and 0 otherwise.

3. The 2022 Presidential Elections in Colombia

The 2022 presidential elections were historic, as they brought to power the first leftist president in Colombia’s modern history. Though the race started with more than 20 candidates, we focus on the four main contenders: the two establishment candidates—Sergio Fajardo and Federico Gutiérrez—and the two anti-establishment candidates—Rodolfo Hernández and Gustavo Petro.^{Footnote 4} Hernández and Petro faced off against each other in the runoff, indicating voters’ readiness for change amidst a representation crisis.

The classification of Rodolfo Hernández and Gustavo Petro as anti-establishment candidates stems from their distinct political positioning and campaign narratives, which markedly diverged from traditional party politics and the established political discourse in Colombia. The term “anti-establishment” here refers to their stance against the conventional political system and the traditional political elite. Both candidates positioned themselves as alternatives to the entrenched political order, resonating with voters’ desires for change and disillusionment with the status quo. Petro, with his leftist ideology, represented a progressive alternative, challenging conventional political structures (Laclau Reference Laclau2005). Conversely, Rodolfo Hernández adopted a distinctly populist strategy, eschewing a clear ideological stance. His campaign was characterized by a lack of coherent ideological positioning, aligning instead with the broader trends of populist mobilization (Zanotti and Botero Reference Zanotti and Botero2023). This approach distinctly positioned Hernández not only as a unique and non-traditional figure but also as one who actively distanced himself from the established political class, underscoring his anti-establishment stance (Mudde and Rovira Kaltwasser Reference Mudde and Rovira Kaltwasser2017). When the race cramped down to the two anti-establishment candidates, it foreclosed the possibility that any of the traditional forces that dominated Colombian politics for decades might reach the presidency.

Petro’s victory was an expression of the seismic changes that have taken place in Colombian politics in recent decades. Peace negotiations with the FARC guerrillas allowed for new concerns and cleavages to emerge in the national arena, where most debates and campaigns had hinged, for decades, on how to handle the armed conflict (Gamboa Reference Gamboa2019; Wills Otero Reference Wills Otero, Luna and Rovira Kaltwasser2014). The persistence of armed guerrillas had also weakened leftist democratic political forces. With the FARC’s demobilization starting in 2016, the stigma on leftist militants decreased and new socioeconomic concerns took center stage in electoral discussions. As Botero and Gamboa (Reference Botero and Gamboa2022) note, these developments, together with the long cycle of protests between 2019–2021, the economic difficulties unleashed by the pandemic and low satisfaction with democratic institutions, came together in late 2021 to configure a crisis.

The presidential elections became a pivotal point in that process, and our tracking of the topics that dominated the digital discussion as put forth by the four candidates confirms it. Corroborating the seismic shift in cleavages at the national level (Gamboa Reference Gamboa2019; Botero et al. Reference Botero, García-Montoya, Otero-Bahamón and Londoño2023) —aside from campaign tweets (logistics, etc.)—the top three issues across all four candidate tweets were the economy, corruption, and education and only 5.8% of all tweets discussed security issues in 2022 (see Figure A8 in the Appendix). Candidates also used Twitter to attack each other. In the case of Gutiérrez and Hernández, Petro was their third most frequent topic. Likewise, for Petro, Hernández was his third most frequent tweeting topic. Fajardo placed less importance on his contenders, but Petro still appears sixth in his list.

Our tracking happened during the most important months of the campaign. The race began in earnest in February 2022 with a very crowded field. Over the next few months, the number of candidates whittled down as the process advanced through three distinct stages. First, there were the primaries among the party coalitions. At that point, three broad coalitions took form, with approximately five candidates vying for the possibility to run under each of the three coalition labels: left, center, and right. On the left, Gustavo Petro (candidate for the Pacto Histórico) emerged victorious. Sergio Fajardo (running as an independent) won the primary for the center coalition (Centro Esperanza), and Federico Gutiérrez (also running as an independent, but with the backing of the government party and, more broadly, Uribista forces) became the candidate for the right-wing camp (Equipo por Colombia). Outside of these coalitions, two candidates were also in the competition: Ingrid Betancourt (who was running for the recently revived Oxígeno Verde political party) and Rodolfo Hernández, a newcomer to national politics, who ran as an independent.

The second stage of the campaign corresponded to the months between the primaries and first round on May 29. Petro was the clear frontrunner throughout this period. This was his third time running for president and the first time a leftist candidate was the contender to beat in a presidential election. The race was understood by most analysts as a confrontation between the left (Petro) and the right (Gutiérrez), with Fajardo and Hernández in third and fourth place, far away from the top two and with little possibilities to come in second. Petro always stood at least 10 points ahead of any other candidate in the polls, but the lead was never sizable enough to confidently indicate that he could avoid the runoff.

The four leading candidates became more active on Twitter during this second stage of the campaign. As we can see in Figure 1, the total number of unique tweets (not including retweets) by Petro, Gutiérrez, Hernández, and Fajardo increased as the campaign advanced, peaking right before the first round.

Figure 1.

Daily Tweets

This does not mean that all candidates tweeted the same amount or at the same pace. Petro and Fajardo tweeted the most, with 41% and 32% of the tweets respectively. Hernández and Gutiérrez tweeted significantly less, with 12% and 15% of the total tweets respectively (see Figure A4 in the Appendix). Petro, however, was more prone to retweeting (Figure 2), while all other candidates (Fajardo included) produced more original content.

Figure 2.

Candidate Tweets and Retweets

As shown above in Figures 3 and 4 below, the candidates tweeted with different intensities at different points in the campaign as well. Petro’s tweets are distributed evenly over time, Hernández, on the contrary, increased his tweeting rate as the campaign advanced. Gutiérrez and Fajardo picked up their pace in May and April respectively.

Figure 3.

Distribution of Candidates' Original Tweets by Month

Figure 4.

Candidates' Daily Original Tweets as a Percentage of All Daily Original Tweets

Hernández’s growing use of Twitter maps well onto his dramatic rise and entrance into the runoff. In the final three weeks before election day for the first round (May 29), Hernández started to climb in the polls and upended the race. With no political party or other structures to back him, and with little nationwide recognition, Hernández took the country by surprise. His anti-corruption message as well as his novel social media strategy captivated many center-right citizens, particularly those who were unwilling to vote for Petro, but wanted a change from the status quo (see Barrenechea and Otero Bahamón Reference Barrenechea and Otero Bahamón2023). Petro came in first with 40% of the vote, and Hernández second with 28%.

The third and final stage of the campaign were the three weeks between late May and June 19, when the runoff took place. The dynamics shifted dramatically in this last month. The hitherto favorite, Petro, lost his clear lead and the two anti-establishment candidates faced off against each other in an extremely tight race. Social media became particularly important during this final phase. Both candidates were social media savvy, but with very different strategies and strengths. According to Velez (Reference Vélez2022), Petro had a solid and somewhat hierarchical structure of social media influencers promoting his proposals and campaign videos. A veteran politician, he had high numbers of followers on his accounts, particularly on Twitter (5,592,016 by May 30, 2022), which he personally used as frequently as his megaphone. Hernández had significantly fewer followers than Petro on Twitter (406,120 by May 30, 2022). Still, social media was a crucial part of his campaign, the tool he most consistently used to become more widely known. Relying heavily on Tik Tok, but also harnessing Twitter, Hernández’s communication team created original and trendy content that marketed the candidate and his anti-corruption message in simple and very effective terms (Buitrago Reference Buitrago2022).

4. Analysis

Below we discuss the tone and rhetoric of these messages. We analyze the candidates’ use of rhetoric and tone, and the extent to which these strategies enhanced engagement.

4.1 Polarizing and Depolarizing Tweets in Colombia

In Figure 5, we chart candidates’ positions in the race according to four leading polling firms.^{Footnote 5} Petro led the polls until June 2022, when Hernández closed the gap. Fajardo, Gutiérrez, and Hernández began at the bottom of the race. Polls on March 20, 2022 put Gutiérrez behind Petro, 13 points over the other two. Hernández did not emerge as a front runner until late May 2022; it was a surprise when he won the runoff. Fajardo’s campaign never took off. He trailed most of the campaign. We believe it is safe to assume that candidates knew where they stood in the race based on these surveys, which were widely publicized.

Figure 5.

Polls Published between March and June 2022

As mentioned above, most of the tweets produced by the four candidates did not use polarizing or depolarizing rhetoric. The 356 tweets categorized as polarizing and depolarizing are unevenly distributed across candidates. Contrary to our expectation that trailing candidates would be more likely to resort to polarizing rhetoric, the undisputable leading candidate, Petro, had, by far, the most polarizing Twitter timeline: he tweeted 41% of all polarizing tweets. He was followed by the less known (though ultimately successful candidate) Hernández who authored 31% of all polarizing tweets. The least polarizing tweeters were Fajardo and Gutiérrez (left out of the runoff) who tweeted 16% and 11.5% polarizing tweets respectively.

Interestingly, the inverse relationship does not hold for depolarizing tweets. Gutiérrez produced more than half of all depolarizing tweets (51%). The rest were distributed somewhat evenly between Fajardo (18.3%), Petro (17.3%) and more distantly, Hernández (13.5%). When we look at the divide between polarizing and depolarizing tweets by candidates (Figure 6), however, it is clear that the two candidates that used polarizing rhetoric the most—Petro and Hernández—have an over-representation of polarizing tweets vis-à-vis depolarizing tweets. Discounting tweets that use neither rhetoric, the former represents 85% of Petro’s and Hernández’s messages whereas the latter represents 14.9% and 15.5%. The least polarizing candidates—Gutiérrez and Fajardo—have, on the contrary, a more even distribution of polarizing and depolarizing tweets. The former represent 43.6% and 60.4% of all of Gutiérrez’s and Fajardo’s polarizing/depolarizing tweets. The latter represent 56.4% and 39.6%. For the anti-establishment and runner-up candidates, there is a statistically significant difference (70%) between polarizing and depolarizing messages (p < 0.000); for the establishment candidates that difference is radically smaller (13–20%) and statistically insignificant (p < 0.2198 and p < 0.1575 respectively).

Figure 6.

Candidates' Original Tweets over Rhetoric

Despite the correlation between polarizing rhetoric and tone, the distribution of negative and positive messages across candidates does not entirely correspond to the distribution of polarizing and depolarizing tweets. As discussed in the previous section, the campaign revolved around the need for structural change, and the candidates’ tone on Twitter appears to have channeled this mood: most candidates were, overall, positive in their Twitter messages. That being said, Fajardo and Gutiérrez were proportionally more positive. They wrote 56.5% (34% and 22.5%) of all positive tweets, but only 29% (19.8% and 8.94%) of the negative tweets. Petro and Hernández—the leading anti-establishment candidates—were on the contrary proportionally less positive. Together, they wrote 71% of all negative tweets (39.8% and 31.4%) and only 43.5% (30% and 13.4%) of the positive tweets.^{Footnote 6}

The most negative candidate was Hernández: 48.3% of his tweets were negative, 45.8% were positive. Though the vast majority of Petro’s tweets were positive (59.7%), he had a non-trivial percentage of negative tweets (35.5%). In clear contrast: the vast majority of Fajardo’s and Gutiérrez’s tweets were positive (67.5% and 75.2%, respectively). Only a few of their messages were negative (17.7% and 13.4% respectively).

Rhetoric and tone also behaved differently across time. As shown in Figure 7, polarizing, negative and positive tweets grew incrementally up until the first round. Forty-two percent of all polarizing tweets, 37% of all negative tweets, and 40.2% of all positive tweets were posted in May (right before the first round); preceded by 24.6%, 28.3% and 29.6% in April and 17%, 19.9% and 18.9% in March. Depolarizing tweets on the contrary remained fairly stable throughout March and April (22% and 23% respectively) and grew significantly in May, with 47.1% of all depolarizing tweets posted that month. All types of tweets declined significantly in June, with the final round of voting being held on the 19th.

Figure 7.

Tone and Rhetoric by Month

Polarizing, depolarizing, negative, and positive tweets also varied by topic. As we saw earlier, the top two substantive issues in this digital campaign were the economy and corruption. Interestingly, though negative, the economy was not particularly portrayed with polarizing or depolarizing rhetoric. Aside from the campaign itself, the top polarizing and negative issue was corruption (where 50% of the tweets are polarizing; and 82% are negative) and the two finalists themselves: Petro (33% of the tweets about him were polarizing; 93% were negative) and Hernández (23.3% of the tweets about him were polarizing; 86% were negative).^{Footnote 7} The polarizing negative rhetoric on Twitter thus centered on campaign issues such as corruption and, in no small part, on personalistic jibes against Petro and Hernández.

What topics candidates chose to use polarizing rhetoric on or talk negatively about also varied. Corruption was the key substantive polarizing negative issue; looking at Figure 8, we can note that Hernández dominated the polarizing negative rhetoric surrounding this issue. He penned 62% and 48% of the polarizing and negative tweets related to corruption. This breakdown also shows that the personalistic polarizing tweets were mostly attacks from Petro against Hernández as well as Gutiérrez and Hernández against Petro (with 60.6% and 30.3% of all polarizing tweets and 33% and 35% all negative tweets in this issue respectively). This behavior does not mirror the one of depolarizing or positive tweets. There were only 2 depolarizing and 10 positive tweets on Hernández; and a total of 4 and 9 for Petro. Corruption had barely one depolarizing tweet, and only 40 (18%) positive tweets.

Figure 8.

Polairizing and Negative Tweets by Topic

4.2. Engagement

Now that we have analyzed how presidential candidates used Twitter and polarizing rhetoric in Colombia, the next question is: did it work? Did polarizing rhetoric increase their visibility and engagement during the campaign?

In line with the literature outlined above, and our Hypothesis 1, polarizing rhetoric seems to have increased engagement. Without accounting for other variables, tweets that used polarizing rhetoric generated more engagement than regular tweets that did not use any of these rhetorics. As shown in Figure 9, the difference in mean response between polarizing tweets vis-à-vis tweets that engage in neither rhetoric, is large and statistically significant. Tweets that engaged in polarizing rhetoric received on average 8,996 likes and 2,635 retweets. Tweets that engaged in neither polarizing or depolarizing responses received almost half of that: 5,628 likes and 1,331 retweets.

Figure 9.

Mean Likes and Retweets over Rhetoric

Surprisingly, depolarizing rhetoric seems to have increased engagement as well. Tweets with unifying and conciliatory messages received on average 9,109 likes and 2,063 retweets. The difference in mean response between depolarizing tweets vis-à-vis tweets that engage in neither rhetoric is large and statistically significant (see Figure 9).

Likes and retweets behave differently. There is no statistically significant difference between the average number of likes to polarizing and depolarizing tweets (Figure 9), but a small and marginally statistically significant difference for retweets. The mean difference in likes between polarizing and depolarizing tweets is 23 (p < 0.9861); the mean difference in retweets is 572 (p < 0.077). In line with studies that have highlighted how likes and retweets, though correlated, can reflect different expressive values (Meier et al. Reference Meier, Elsweiler and Wilson2014), this suggests that polarizing rhetoric may have different effects on people’s willingness to highlight or share a particular piece of information.

T-tests are great to provide a general overview of the data but cannot account for confounders. To do that, we use an OLS regression with the logged transformation of likes and retweets and controls for candidate, topic, stage of the campaign, hashtags and tagging.^{Footnote 8} Similar to what we saw with the t-tests above, the OLS regression results (Figure 10) suggest that both polarizing and depolarizing tweets increase engagement (vis-à-vis tweets that do not engage in either rhetoric). Polarizing tweets increase likes by 0.14 points (p < 0.078) and retweets by 0.28 points (p < 0.000), roughly 14.5% and 32.7%. Depolarizing tweets increase likes and retweets by 0.61 and 0.62 points (p < 0.000), between 85% and 86.6%. The results also suggest that depolarizing tweets increase engagement when compared with their polarizing counterparts. Messages using depolarizing rhetoric increase likes by 0.48 points (p < 0.000) and retweets by 0.34 points (p < 0.011), approximately 61.5% and 40%. Once again, we observe a differential effect of polarizing and depolarizing rhetoric across types of engagement. When compared with “neither” the effect of polarizing tweets is larger for retweets; when compared with each other (i.e., polarizing vs. depolarizing tweets) the effect is larger for likes.

4.3 Is It the Tone of the Tweet or Is It Its Rhetoric?

As explained earlier, polarizing rhetoric and tone are closely related and intersect, particularly with regards to the study of political communication in media platforms. To tease out the effect of these two variables, we assessed tone separately. Interestingly, tone behaves slightly differently from polarizing rhetoric. Like polarizing and depolarizing tweets (and in line with Hypothesis 2), both, negative and positive messages increase engagement (vis-à-vis neutral tweets). As shown in Figure 11, the difference in mean likes between negative and neutral tweets is 5,067 (p < 0.000); the difference in mean retweets between positive and neutral tweets is 3,317 (p < 0.000). Different from polarizing rhetoric—but in line with existing literature—negative tweets seem to increase engagement vis-à-vis depolarizing tweets. The mean difference in likes between positive and negative tweets is 1,749 (p < 0.000); the mean difference in retweets is 839 (p < 0.000).

The OLS regression confirms these results (Figure 12). Both, negative and positive tweets have a positive and statistically significant effect on engagement (vis-à-vis neutral tweets). Negative tweets increase likes by 1.20 points (p < 0.000) and retweets by 1.35 points (p < 0.000), roughly 237% and 286%. Positive tweets increase likes by 0.97 points (p < 0.000) and retweets by 0.95 (p < 0.000), an enhancement of 163% and 156.6%. Negative tweets, however, seem to have the upper hand. When compared to positive tweets, they increase the number of likes by 0.25 points (p < 0.000) and the number of retweets by 0.41 points (p < 0.000), that is 28.3% and 50.5%. When compared to neutral tweets the effect of positive and negative tone is similar across different types of user tools. When compared to each other the effect of positive and negative tone seems much stronger for likes than retweets.

Figure 10.

Effect of Rhetoric on Engagement

Figure 11.

Mean Likes and Retweets over Tone

Figure 12.

Effect of Tone on Engagement

It appears therefore, that though both tone and rhetoric have a significant effect on engagement, this effect is slightly different for each. In order to further tease out this relationship (Hypothesis 3) we ran another set of regressions with both IVs (Figure 13). The results are illuminating. First, the impact of tone seems stronger than the impact of polarizing rhetoric. When compared with neutral toned messages, both negative and positive tweets increase likes by 1.21 (p < 0.000) and 0.95 (p < 0.000) points, 234% and 157.5%, respectively; and retweets by 1.32 (p < 0.000) and 0.92 (p < 0.000) points, 276% and 150.6%, respectively. On the contrary—once we account for tone—only depolarizing tweets seem to have a significant effect on engagement. Tweets using depolarizing rhetoric increase likes by 0.52 points (p < 0.000) and retweets by 0.54 points (p < 0.000)—69% and 71.7%, respectively.

Figure 13.

Effect of Rhetoric and Tone on Engagement

When compared with each other, the results are the exact opposite. Negative tweets increase engagement, bumping up likes by 0.26 points (p < 0.000) and retweets by 0.40 points (p < 0.000), roughly 29.6% and 50%. On the contrary, polarizing tweets decrease engagement (vis-à-vis depolarizing tweets). They depress likes by 0.53 points (p < 0.000) and retweets by 0.46 points (p < 0.000), approximately 41.4% and 37%.

The relationship between tone and polarizing rhetoric becomes clearer once we analyze the interaction of polarizing rhetoric and tone. For tweets that have a negative tone, (de)polarizing rhetoric is irrelevant (Figure 14).^{Footnote 9} Negative tweets with polarizing rhetoric do not generate more engagement than negative tweets using depolarizing rhetoric, or negative tweets that use neither polarizing nor depolarizing rhetoric. For tweets that have a positive tone, (de)polarizing rhetoric becomes more important. It seems that positive tweets that also use a unifying and conciliatory message increase likes by 0.64 points (p < 0.000) and retweets by 0.68 points (p < 0.010)—roughly 90 and 98.3%—compared to positive tweets that use neither depolarizing nor polarizing rhetoric. They also seem to increase likes by 0.39 points and 0.17 points (vis-à-vis tweets polarizing tweets) though this impact is insignificant for retweets and only marginally significant for likes (p < 0.095).

Figure 14.

Effect of Rhetoric by Tone

4.4 Who Tweets?

To further investigate these surprising results, we analyzed the effect of (de)polarizing rhetoric and tone on engagement across candidates. We ran a similar OLS regression for each of the candidates in the race. The results are in line with what we have discussed so far. Negatively and positively toned tweets increase engagement (vis-à-vis neutral tweets) for all candidates (Figure 15). Negative toned tweets also increase engagement (vis-à-vis positive tweets) for all candidates, except for Hernández. The relationship with polarizing rhetoric is more complex. Depolarizing rhetoric (vis-à-vis no-rhetoric) increases engagement for all candidates except Fajardo. For Petro, Hernández, and Gutiérrez, uniting messages increase likes by 0.52 (p < 0.009), 0.42 (p < 0.022), and 0.60 (p < 0.069) and retweets by 0.55 (p < 0.071), 0.35 (p < 0.004), and 0.34 (p < 0.008) points. Polarizing messages increase Fajardo’s retweets by 0.50 (p < 0.031) points.

Figure 15.

Effects of Rhetoric and Tone by Candidate (vs. Neither/Neutral)

Figure 16.

Effects of Rhetoric and Tone by Candidate (vs. Depolarizing/Positive)

When compared to each other, the differences between the behavior of tone and polarizing rhetoric remain. As shown in Figure 16, For all candidates, except Hernández, negative tweets increased engagement, though the effect size varied. Polarizing rhetoric however, had a selective negative effect decreasing Petro’s likes and retweets by 0.55 points (p < 0.012) and 0.40 points (p < 0.046), respectively, and Hernández’s by 0.66 (p < 0.067) and 0.56 (p < 0.089) respectively. Depolarizing messages had no statistically significant effect for Gutiérrez or Fajardo.

Altogether, this analysis suggests that a negative tone might be more important than a divisive message to increase engagement on Twitter in Colombia. Once a tweet is negative, polarizing rhetoric will not have any effect on engagement. It also shows, however, that when tweets are positive, depolarizing rhetoric can have an important impact on engagement. If the tweet expresses feelings of hope and happiness, it increases user interaction; but if on top of that it contains a uniting and pluralistic message, this effect might be even bigger. The results also suggest that the effect of both, polarizing rhetoric and tone might be different based on who tweets. For tone it might be a matter of size (in terms of number of followers). The effect of a negative tone was drastically bigger for a candidate like Fajardo (with 1.7 million followers at the time) than it was for a candidate like Petro (with 5.6 million followers at the time).

For (de)polarizing rhetoric, these differences go deeper. It seems that polarizing rhetoric helped candidates like Fajardo, more than it did candidates like Petro or Hernández. Likewise, it seems that depolarizing rhetoric helped candidates like Hernández, and particularly Petro, more than it did Fajardo. One explanation for this could be the public’s perceptions of these candidates. Hong and Kim (Reference Hong and Kim2016) have warned about the differential effects of polarizing rhetoric on engagement for different types of political leaders. In this case, the differences could align with how the public perceives these politicians. Petro, with a long history as an opposition politician and ideologically more extreme, is perceived as more polarizing. It is reasonable to expect that the Twitter audience would be more likely to reward depolarizing tweets from him. Fajardo, a self-defined “centrist” and even reluctant to take a position on crucial issues is seen as less polarizing. Therefore, the public might have been more likely to reward polarizing tweets from him.