Users are likely to argue over wide ideological divides, and are increasingly likely to engage with those who are different from themselves

Why Keep Arguing? Predicting Engagement in Political Conversations Online. Sarah Shugars, Nicholas Beauchamp. SAGE Open, March 27, 2019. https://doi.org/10.1177/2158244019828850

Abstract: Individuals acquire increasingly more of their political information from social media, and ever more of that online time is spent in interpersonal, peer-to-peer communication and conversation. Yet, many of these conversations can be either acrimoniously unpleasant or pleasantly uninformative. Why do we seek out and engage in these interactions? Who do people choose to argue with, and what brings them back to repeated exchanges? In short, why do people bother arguing online? We develop a model of argument engagement using a new dataset of Twitter conversations about President Trump. The model incorporates numerous user, tweet, and thread-level features to predict user participation in conversations with over 98% accuracy. We find that users are likely to argue over wide ideological divides, and are increasingly likely to engage with those who are different from themselves. In addition, we find that the emotional content of a tweet has important implications for user engagement, with negative and unpleasant tweets more likely to spark sustained participation. Although often negative, these extended discussions can bridge political differences and reduce information bubbles. This suggests a public appetite for engaging in prolonged political discussions that are more than just partisan potshots or trolling, and our results suggest a variety of strategies for extending and enriching these interactions.

Keywords: politics, social media, interpersonal communication, deliberation, polarization, natural language processing, topic models, sentiment

---

Introduction

Digital communication plays a critical role in our political infrastructure. Online platforms have expanded the reach of “kitchen table conversations,” as people increasingly turn to social media as a primary news source (Bakshy, Messing, & Adamic, 2015; Lee & Ma, 2012; O’Connor, Balasubramanyan, Routledge, & Smith, 2010) and elected officials use digital channels to communicate with their constituents (Farina, Epstein, Heidt, & Newhart, 2013; Kavanaugh et al., 2012). Such interactions are often modeled as one-shot games or as evidence of long-term links in a social network (Feng & Wang, 2013; Myers & Leskovec, 2014). Yet, much online activity consists not of single-shot, unidirectional interactions with elites, but repeated interactions among peers. These iterated interactions—conversations—have important implications for political theory: while conventional wisdom claims that brief social media interactions have little effect on subsequent online behavior, a number of recent experiments have shown modest but real effects of single-shot interactions (Friggeri, Adamic, Eckles, & Cheng, 2014; Munger, 2017). Deliberative theory suggests that repeated interpersonal interactions where individuals engage in extended conversation may have even more substantial effects (Axelrod, 1987; Bednar & Page, 2007; Habermas, 1984).
In this article, we focus not on persuasive outcomes but on the more fundamental question of what leads people to engage in extended online conversation and argument in the first place. Existing work in this area has generally had a more practical bent, focusing on tweet- or conversation-level recommendation and aiming to predict user interest in conversation threads to better curate and recommend targeted content. Such work has looked at user engagement in various forms of online conversation (Chen et al., 2012; He & Tan, 2015; Vosecky, Leung, & Ng, 2014; Yan, Lapata, & Li, 2012), as well as via retweeting (Feng & Wang, 2013; Hong, Doumith, & Davison, 2013) and re-entry back into existing conversations (Backstrom, Kleinberg, Lee, & Danescu-Niculescu-Mizil, 2013). Our work here is closest to the latter: We are more interested in the extended dynamics of conversation, particularly the decision to re-engage or exit, than in the initial decision to interact. We focus on users who have already made that initial participation, and seek to understand and predict whether and when they re-engage based on user, tweet, and thread-level features.
While ongoing deliberative conversation is the substantive focus here, this framing also turns an impossible problem—predicting initial responses to a tweet out of the entire pool of twitter users—into a practicable prediction task—predicting re-participation of users whom we know are already part of a conversation. This approach also conditions out the even harder problem of explaining the origins of an initial tweet or conversation, particularly given the immense variety of motivations behind those first moves. Instead, we focus on existing conversations—at least a first move followed by a response—and model the processes that lead to extended and branching conversations among existing participants. Twitter might seem less suited to such models than traditional online forums, but Twitter in fact produces immense quantities of impromptu extended, branching conversations, and by focusing only on re-entry by existing participants, we can study what causes individuals to continue an argument or drop out, bracketing the question of initial engagement.
By conditioning on existing user interaction, we aim to get more deeply at the question of why people bother arguing online. What brings them back to a repeated argument? What factors contribute to an individual returning to or abandoning an argument? While in face-to-face settings, social etiquette suggests that a comment will most likely be greeted with a response, there is no a priori reason to expect a response to the vast majority of online posts. While we expect to find many types of conversations occurring online, we might expect that more extreme content (positive or negative) will increase engagement, as trolls successfully incite arguments and partisan allies reinforce each other’s positions (Cheng, Danescu-Niculescu-Mizil, Leskovec, & Bernstein, 2017). Between the extremes lies a more productive and interesting mode of engagement: true deliberative argument, in which participants exchange content in a genuine attempt to persuade or inform. Such behavior is not as uncommon as skeptics might assume, and is prevalent in knowledge-sharing platforms such as StackOverflow, Yahoo! Answers, and other such forums, where users may be motivated to some degree out of a general sense of community (Adamic, Zhang, Bakshy, & Ackerman, 2008; Anderson, Huttenlocher, Kleinberg, & Leskovec, 2012; Oktay, Taylor, & Jensen, 2010) even as they argue over better or worse solutions to shared problems.
We find evidence for all of these behaviors in our data and, in particular, show that while many of these engagements are negative, conversations often cover a range of emotions and go on far longer than a single-shot attack or mutual trolling might suggest. While we leave for later the ultimate question of persuasive effect, we establish here that even a medium as apparently unpromising as Twitter is full of complex, extended political conversations, and that individuals’ decision to repeatedly re-engage in those conversations is surprisingly systematic.

Related Work

Much of the theoretical work around questions of conversational dynamics has been done within the literatures of deliberation and persuasion. While both these approaches focus their attention on back-and-forth conversations, they vary in their characterization of those conversations. The persuasion literature looks broadly at how people convince others and “win” arguments, while the deliberative ideal imagines thoughtful participants reasoning together to generate public opinion centered on the common good (Cohen, 1989; Habermas, 1984). “Reasons” may constitute factual arguments or emotional appeals (Mansbridge, 2015), but ideal deliberation is often taken to be free of persuasion, coercion, or other forms of instrumental action (Habermas, 1984). Contra persuasion models, ideal deliberators should engage in rational speech acts—aiming to honestly express themselves and truly trying to understand the other. Huckfeldt, Mendez, and Osborn (2004) argue that ideal citizens “are those individuals who are able to occupy the roles of tolerant gladiators—combatants with the capacity to recognize and respect the rights and responsibilities of their political adversaries” (p. 91) If political debate serves to sharpen our own understanding and build our collective knowledge, then we owe it to our interlocutors to press them on their positions, to find the holes in their armor and encourage refinement of beliefs. The process of debate makes us all better—thus allowing tolerant gladiators to walk away as friends. Citizens who silence their discussants, seek to coerce others, are easily persuaded by false beliefs, or who otherwise refuse to engage in rational argument, therefore, do a disservice to themselves and to their communities.
Experience tells us, however, that such a lofty deliberative ideal is rarely met in political conversation. Sunstein (2002) advances the “law of group polarization,” finding through numerous empirical studies that “deliberation tends to move groups, and the individuals who compose them, toward a more extreme point in the direction indicated by their own pre-deliberation judgments” (p. 175). Sunstein argues this polarization is the natural result of the social context, which serves as a significant driver of individual actions and opinions. Hearing friends express a view makes a person socially inclined to express the same view. In other words, deliberating groups tend toward extremism in the direction of the predeliberation median because nobody wants to take the social risk of expressing an unpopular view. Sanders (1997) similarly argues that the broader context of power dynamics frequently has a debilitating but under-recognized effect on deliberation, as marginalized individuals feel silenced and unable to share their true opinions. Importantly, the majority of participants may mistakenly assume that such power effects are negligible if “deliberation appears to be proceeding.”
Another line of work has tackled conversational dynamics from the perspective of the data-processing problem of platform curation, for example, trying to predict which posts will be popular for the purpose of highlighting those posts for users. Much of this work focuses on post-level engagement, predicting engagement as a function of topics (Hong et al., 2013) or social network structures (He & Tan, 2015; Pan, Cong, Chen, & Yu, 2013). Much of this work has considered “popularity” as a raw aggregate of engagement with an initial post, finding, perhaps unsurprisingly, that the popularity of a user’s past content is a strong predictor for the popularity of their future content (Artzi, Pantel, & Gamon, 2012). Backstrom et al. (2013) break the task into related subtasks: length prediction and re-entry prediction. Intuitively, these subtasks indicate distinctive types of threads: threads which are long because a high number of users chime in a small number of times—to offer congratulations or condolence, for example—while other threads are long because a small number of users contribute a large number of times in a back-and-forth conversation. Supporting this theory, Backstrom et al. (2013) find that the number of distinct users in long threads follows a bimodal distribution. Using data from Facebook and Wikipedia, Backstrom et al. (2013) find the identities of recent commenters is most predictive of conversation re-entry.
This lattermost line of work is largely atheoretical and not particularly concerned with normative issues. While the current study borrows many of their methods, we are also fundamentally interested in the dynamics of online conversation from a deliberative perspective. Thus, we are interested less in conversation recommendation or modeling engagement in conversations per se, and more focused on how individual speech acts (tweets) lead existing discussants to re-engage with each other or abandon a conversation. Regardless of the outcome of a conversation, it is important to understand what sustains conversations—particularly acrimonious ones—and keeps mutual opponents or supporters engaged with each other. As we will see, this engagement can take more or less productive forms, but simply understanding the deliberative dynamics is an important first step.