TWITTER AS CAMPAIGN TOOL

In analyzing the relational dynamics of negative electoral campaigning, Twitter merits special interest for three reasons. First, it is fundamentally social in conception and design. While we may also examine social dynamics in a crowded field by coding advertisements or press releases, Twitter—with its hashtags and handles, followers and retweets—makes interpersonal connections central to its very identity. Second, the lack of budgetary constraints or external gatekeepers allows candidates to employ Twitter as an unfettered signaling device. Candidates may attack multiple candidates if they feel so inclined, regardless of funds on hand. Even televised debates—quintessentially social events—are heavily restricted by moderators, who nudge top candidates to challenge one another, but rarely provide such opportunities to those performing poorly in polls. Similarly, press coverage of public statements is mediated, with interactions among lower-tier candidates far less likely to be deemed noteworthy enough to print. Third, there may be special features of social media worthy of study in their own right as performing a new function in campaigns.

Why focus on Twitter, as opposed to other social media? The most relevant factor is wide adoption, of course; the fact that all candidates are using Twitter is essential. Furthermore, Twitter is becoming what we might call a cross-over medium, serving to link social and traditional media. Tweets are not particularly flexible conduits of meaning, due to the strict limit of 140 characters per tweet. And yet, this restriction brings its own opportunities in a crowded marketplace of messages. Their very brevity allows tweets to be widely disseminated. In recent decades, politicians, among others, have had to learn to talk in sound-bites in order to fit the changing nature of television and radio news. Terse quotes are far more likely to be picked up by newsrooms that rely upon conciseness for most common formats (Flowers et al. Reference Flowers, Haynes and Crespin2003; American Press Institute 2016) and tweets are structurally required to be concise. Twitter is therefore a sound-bite medium for the sound-bite media age. On the other hand, the fact that users have only a sentence or two to make a point—and appeal for attention in a marketplace of tweets—places a premium on novelty and even outrageousness.

Twitter offers the opportunity for interactive, instantaneous, and autonomous give-and-take between candidates (Vergeer and Hermans Reference Vergeer and Hermans2013). In debates, certain candidates may not have ample opportunities; polling typically determines who is permitted to debate, the amount of attention one receives, and—during this campaign season—whether one appears in the “main” or less prominent “undercard” debate. Twitter removes the gatekeeper, promoting participation by all candidates, regardless of status and without the presence of a moderator (Graham et al. Reference Graham, Broersma, Hazelhoff and van’t Haar2013; Conway, Kenski, and Wang Reference Conway, Kenski and Wang2013). While complementing traditional means of outreach, Twitter allows candidates to circumvent the constraints associated with the traditional media (Vergeer and Hermans Reference Vergeer and Hermans2013; Conway, Kenski, and Wang Reference Conway, Kenski and Wang2013). Attack advertisements take time to produce, as do potential rebuttal ads. On Twitter, by contrast, it is not uncommon for opponents to trade attacks over the course of a single day.

In short, Twitter is a free resource through which candidates can offer easily digested, potentially newsworthy content, as they engage with one another and with followers in real time. Additionally, tweets need not only be composed with candidates’ own supporters in mind (i.e. to “rally the troops” or “preach to the converted”); they are also likely to reach followers of one’s targets, undecided voters who follow multiple candidates, and even major news networks (Conway, Kenski, and Wang Reference Conway, Kenski and Wang2013). Subsequently, news media help to further disseminate tweets to program viewers as sound bites in news stories. ^{Footnote 3} As but one example, CNN devoted two news stories to reporting on Donald Trump’s tweets about Carly Fiorina, and published the stories online (with transcripts of the tweets). ^{Footnote 4}

We have attempted to collect all candidate tweets mentioning at least one other candidate. ^{Footnote 5} After dropping tweets with neutral or ambiguous tone, as well as retweets without explicit endorsement, we are left with a total of 1,523 tweets, including 1,069 during the primary period of study (November 18, 2014—May 4, 2016). We divide this campaign period under study into six stages to assess whether the rate of tweeting increases over the duration of the campaign. Though these stages vary in length, the measurement of tweets within each (rate of tweets per candidate and per directed dyad per week) is standardized; comparing rates of tweets across stages allows for comparison of tweet behavior across qualitatively meaningful stages of the campaign.

GROWING TWEET NEGATIVITY AND INTENSITY

We code candidates’ tweet tone (or affect) as positive (praising the target), negative (criticizing or insulting the target), or neutral (unclear or lacking affect). We also code whether a tweet is policy-oriented or of a more personal nature. Prior to the start of the 2016 presidential election season, Twitter communication among eventual candidates, while limited, remained overwhelmingly amicable. Out of 454 collected tweets from this time (including retweets), just 45 reflect negative affect. In contrast to tweets during campaign season, in which twice as many tweets are personal as are policy-oriented, these pre-campaign tweets are more often policy-related (149 vs. 115 among those coded as one or the other).

In figure 1, we visualize the pre-campaign positive affect network, employing the common Fruchterman-Reingold network layout to automatically place sets of highly connected individuals close together and less connected nodes farther apart. Clustering based on edge-betweenness ^{Footnote 6} indicates that governors and senators belong to separate clusters, with fellow Floridians Marco Rubio and Jeb Bush bridging the two groups.

Figure 1 Tweets Expressing Positive Affect Among Future Candidates Prior to Nov. 20, 2014

Each arc (directed tie) represents a tweet event by one eventual candidate making positive mention of another. (Gilmore and Pataki are both isolates in this network and have been dropped from the figure). Nodes are shaded according to most recent government office held (senators in gray, governors in black, those with no government experience in white).

Figure 2 Tweet Negativity. Incidence of Negative Affect in Previous 50 Tweets Connecting Any Two Candidates (March 2013–May 2016)

The shift from mutual admiration to taunts and tirades is swift and unrelenting, interrupted only by brief moments of respite, primarily as newcomers are welcomed and dropouts are bid farewell (the troughs in figure 2). During the field’s rapid shift into attack mode during the first two weeks of July 2015, Donald Trump emerges as the clear network negativity hub (figure 3b). The overall rise in negativity is accompanied by a dramatic increase in Trump’s poll numbers. ^{Footnote 7}

The rate at which candidates send these (mostly) negative messages to each other might also be expected to increase over time. We measure this in terms of both average tweets per candidate per week and tweets per directed dyad per week ^{Footnote 8} (ddw) in order to account appropriately for the dwindling candidate pool. Tweet intensity does indeed increase dramatically as the general election date approaches (table 1), rising monotonically over the six campaign stages. By the final stage, in which only five candidates remain, candidates are tweeting at a rate of 3.44 tweets per person per week (2.681 per ddw), or roughly 100 times the initial tweet per person rate and 1,300 times the initial tweet per ddw rate.

Table 1 Tweet Negativity By Campaign Stage

In sum, as the election date approaches and the field dwindles, candidates exhibit far stronger tendencies toward negativity when engaging their opponents on Twitter. Since the vast majority of tweets are negative during election season, this finding lends additional credibility to the prominent notion in previous literature and conventional wisdom that negative campaigning will increase as the end of a campaign season approaches.

“PUNCHING UPWARDS”: ASYMMETRIC NEGATIVITY BY RELATIVE STANDING

Even a casual observer of the 2016 primaries will hardly be surprised that eventual nominee Donald Trump dominates the field as both author and target of negative tweets. ^{Footnote 9} His legendary social media vitriol has even been immortalized in the New York Times, ^{Footnote 10} which has published a “complete list” of “people, places, and things Trump has insulted on Twitter.” And yet, despite issuing some of the more jaw-dropping tweets, Trump has in fact remained on the receiving end of more barbs than he doles out, consistent with our expectation that front-runners should be attacked more than they themselves attack opponents. All together, 72% of negative tweets are directed from a lower polling candidate toward a higher polling opponent.

Figure 3 provides visualization of the separate positive and negative aggregated tweet networks. The main attractors of positive tweets (figure 3a) are candidates who leave the race early and are praised on their way out (Perry, Walker, Jindal, and Santorum), and—to a lesser extent—those who eschew negative campaigning in the early stages (Fiorina, Cruz, and Kasich). The central hubs of activity in the negative tweet network (figure 3b)—other than Trump, who dominates—include the poorest performers (Pataki and Gilmore), based entirely on their own negative tweets (out-ties).

Figure 3 Positive (3a) and Negative (3b) Tweets From Nov. 20, 2014–May 4, 2016

Nodes Shaded According to Most Recent Government Office Held, as in Figure 1

Arc widths proportional to number of positive tweets. Arc width roughly proportional to log of number of negative tweets plus one.

Given the outsized role of Trump within the dataset—he is author or target of more than two-thirds of the collected tweets—we must ask whether the observed asymmetry is driven exclusively by the Trump factor. In fact, the opposite is true, with the pattern of asymmetry appearing nearly deterministic once Trump is dropped from the data (table 2). While Trump receives more Twitter-attacks than he launches, he still strikes back at poorly performing candidates with surprising alacrity for a front-runner. Excluding Trump’s tweets, examples of downward strikes at poorer performers nearly disappear. Astoundingly, only 10 of 340 non-Trump-related tweets (3%) feature a negative tweet issued by a candidate who leads the recipient in the average of the previous five polls. Of these, all but two involve former allies: Cruz attacking Rubio and Rubio attacking Bush. ^{Footnote 11}

Table 2 Standing in Polls by Tweet Tone, Excluding Tweets Involving Trump

In order to consider all 136 dyadic Twitter relationships among the 17 candidates in a parsimonious manner, we classify dyad histories into several types. ^{Footnote 12} Because we are primarily interested in the conditions that predict asymmetry in negative affect, we emphasize these in our typology. Details are provided in table 3 and the online appendix. Dyadic histories characterized by lopsided negativity are separated into three types: snipes (a single negative tweet in one direction), sticks-and-stones (multiple negative tweets in one direction), and asymmetric warfare (attacks in both directions, but at least three times as many in one direction. In table 3, rows and columns are ordered from highest to lowest status candidate based on average polls ^{Footnote 13} during each candidate’s run; of the 60 dyadic histories characterized by lopsided negativity, all but four are in the predicted direction. (Cells in which the row candidate is the primary aggressor in one of these three types of relationships are shaded, and only four of these shaded cells fall above the diagonal). All but one of these four instances of “punching downward” involves opponents within 1% of one another in average polling, leaving only Trump’s online bullying of Ben Carson (eight vs. one negative tweet[s]) as a dramatic departure from predicted hierarchical discipline. We perform a permutation test ^{Footnote 14} to see how likely we would be to see so few instances of these asymmetric negative types in the “wrong” direction by chance and find the probability to be less than one in 100,000. The probability that at most one of 60 would involve a downward strike against someone more than 1% lower in average polling (as in the case of Trump vs. Carson) is less than one in a million.

Table 3 Dyad History Types, By Candidate Ranking in Polls From November 18, 2014 Until Either Candidate Suspends Campaign

Asymmetric dyad histories include Sticks-and-Stones (S&S), Asymmetric Warfare (AW), and Snipe; these cells are shaded if the candidate labeled “sender” is the main attacker and in bold type otherwise. Negative tweet count for “sender” vs. “receiver,” respectively, appears in parentheses. More balanced types include Mutual Enmity (ME) and Mutual Respect (MR). Negligible contact (no tweets or single positive) is indicated by n. Unrequited Love (UL) consists of one candidate in a pair sending two or more unreciprocated positive tweets. Unless indicated by “+”, all numbers are counts of negative tweets. Details on type coding appear in the appendix. If positive tweets are issued in both directions prior to any negativity beyond an isolated snipe, we append the Broken Friendship (BF) suffix.

Finally, note that the two poorest performers—Governors Gilmore and Pataki—are the sources of more asymmetric negative relationships than any others: 13 and 11, respectively. Such behavior by these apparent “spoilers” suggests that, at least in a low-cost environment such as Twitter, low-polling candidates may not fulfill previous predictions (Skaperdas and Grofman Reference Skaperdas and Grofman1995) by keeping their messaging positive. An examination of the content of their tweets indicates that they are trying, albeit unsuccessfully, to establish legitimacy by attempting to draw other candidates into acknowledging their existence.

SUPPLEMENTARY MATERIAL

To view supplementary material for this article, please visit http://dx.doi.org/10.1017/S1049096516001700.Footnote *