Dissemination of Political News on Facebook
Ernesto de Le´
on
A thesis submitted for
the degree of Master of Science
Amsterdam, 4 June 2020
Student ID: 12156353
Contact: ernestodeleon21@gmail.com Supervisor: Dr. Damian Trilling
Second Reader: Dr. Daphne van der Pas University of Amsterdam
Engagement and Dissemination of
Political News on Facebook
Reprints and permission:
sagepub.co.uk/journalsPermissions.nav DOI: 10.1177/ToBeAssigned www.sagepub.com/ SAGE
Ernesto de Le ´
on
1 AbstractToday, citizens receive an increasing part of their daily fix of political news from social media: scrolling through Facebook, users are consistently exposed to a curated set of articles detailing the latest in the political grapevine, making it a key source of political exposure. Despite recent research showcasing the importance of emotions in the relationship between the press and politics, little work has attempted to evaluate how individuals engage emotionally with political news on Facebook. This paper brings together political communication and psychology literature to explore how Facebook users engage emotionally with political news during an election. The study addresses how Facebook users interact emotionally with different political news content, looking at how the presence of good news, bad news, and Policy Oriented and Game Strategic frames result in diverging emotional reactions, and how these reactions are in turn related to news sharing behaviour. Three main contributions are made to the state of the literature: it finds a negativity bias in emotional engagement with political news, that both Game Strategic and Policy Oriented frames elicit positive reactions from audiences, and the existence of a sadness-bias in the sharing of political news.
Keywords
Emotions, Facebook, elections, automated content analysis, news sharing, political news
1
Introduction
Social network sites (SNS) play an increasingly important role in our democracies: as a space where users from all over the world can read and discuss political ideas, sites such as Facebook have become a key source of political opinion formation. This is especially true for political news: through Facebook, individuals are both exposed to the latest political news, and as an audience, are empowered to interact and further disseminate the content to an unprecedented degree (Picone 2016). This development has directly impacted news consumption habits, making social network sites both a growing source of political news and participation (Newman et al. 2019). Because news consumption and interaction has implications for political behavior, impacting turnout (Schuck et al. 2014), political learning processes (de Vreese and Boomgaarden 2006), perceptions of political leaders (Aaldering et al. 2018), and vote choice itself (Hopmann et al. 2010), there is a growing need to understand how individuals engage with political news on social network sites.
Research bringing together politics, SNS, and news consumption have tended to focus on phenomena such as political learning processes, political interest, and exposure to new information, exploring the topics of selective exposure, filter bubbles, and ‘news finds me’ perceptions (Gil de Z´u˜niga and Diehl 2019). Much of this research has therefore taken an explicitly cognitive approach to how individuals consume and interact with political news on social networks, that, while important, often neglects the crucial emotional component of political news engagement. With previous literature suggesting that emotionally-charged media can impact how individuals understand both news and
politics (Ansolabehere et al. 1994; Marcus 2000; Neuman et al. 2018), it is important that research turns from the cognitive to the emotional, addressing the role of emotional engagement with political news on SNS. It is particularly important to explore positive and negative political news and its journalistic framing, as existing research suggests that these news types affect emotions, that in turn can impact a reader’s response on SNS. With a long lineage of psychological studies emphasising the role of emotions both in political behaviour and general human psychology, SNS research on political news consumption would be well served by taking an emotional perspective.
While literature has explored the link between emotionally charged news, news frames, and emotional states (e.g. Neuman et al. 2018), whether and how these relationships are reflected in interactions with political news on social network sites is yet to be addressed. A reason why this is lacking lies in that much of social media research, unlike the political psychology literature on emotions, is not experimental, making data collection a challenge. Facebook’s introduction of their Reactions∗feature offers researchers an opportunity to gauge emotional response on social media without turning to experiments or surveys. Through these Reactions, users can describe how a post makes them feel by clicking one of five icons (Angry, Love, Haha, Sad, and Wow)
1University of Amsterdam, NL
Corresponding author: Ernesto de Le ´on
Email: ernestodeleon21@gmail.com
∗In this study, I will distinguish between Facebook’s ‘Reactions’ feature and
corresponding to specific emotions provoked. While these emotional reactions have been used to explore Facebook user interaction with Facebook posts by political parties (Eberl et al. 2020), public engagement with published scientific literature (Freeman et al. 2019), news consumption feedback (Larsson 2018) and controversy in news (Basile et al. 2018; Sriteja et al. 2017), no research has examined how Facebook users react emotionally to political news, and much less the effect that frames and valence has on these reactions. This study therefore hopes to fill this gap by investigating how political news content leads to emotional engagement on Facebook, looking at how the presence of good news and bad news in article content result in positive and negative reactionson Facebook, and whether these patterns resemble the expectations formulated by negativity bias literature. The following research question is therefore posed:
RQ1: How is the content of political news articles related to emotional responses on Facebook?
Argued to be “the most valuable form of user engagement” (Larsson 2018, 329), a crucial metric of news engagement on Facebook is to what extent an article is ‘shared’ by users, with research asking how and why users disseminate particular news pieces over others. With findings that news sharing can extend the reach of news stories beyond a platform’s original partisan audience (Ju et al. 2014), can help level the political knowledge playing field, and can affect individuals’ feelings of involvement with politics itself (Oeldorf-Hirsch and Sundar 2015), how people disseminate political news on social media has received increasing attention. Much of this literature takes a news-value approach (Eilders 2006; O’Neill and Harcup 2009), arguing that there are structural characteristics that make a story more likely to be ‘newsworthy’, or, here, ‘shareworthy’ (Trilling et al. 2017).
Thus far the behavioral research that seeks to predict news sharing behaviour has successfully linked sharing with the content characteristics of the articles themselves. However, the role of emotions in political news sharing has yet to be explored in depth. While studies have connected article negativity to increased sharing (Trilling et al. 2017), this literature does not take into account the actual emotional reactions produced by specific frames and sentiment in an article. In other words, little research has been conducted on the mechanisms linking content characteristics and sharing behaviour - the explanations as to why negative articles are shared more often than positive articles remain in theoretically informed assumptions, and are not explicitly demonstrated by the data. This paper hopes to start the conversation on emotional responses to articles and sharing behaviour, by not only using data on content characteristics and sharing behaviour, but by including the mediating variable of how individuals react to news before sharing it. This article therefore does not only tackle how particular content results in emotional reactions, but also how these emotional reactions affect sharing behaviour. Therefore, I pose the following second research question:
RQ2: How do emotional responses to political news articles help us understand political news sharing on Facebook?
By looking at emotional reactions and sharing behaviour of political news during elections, this research speaks to
two distinct scholarly fields: the communication science literature that evaluates how emotional reactions to news results in online engagement, and the political psychology and communication literature, which is interested in how emotional reactions to political press framing leads to the diffusion of certain political content, allowing insights into how political information is distributed to an electorate during election campaigns.
2
Theoretical Framework
Emotions, Politics and Reactions
The impact of emotions on political behaviour has been a focus in political science for centuries, with research addressing how individuals act when presented with emotionally-triggering events, messages, or objects (Marcus 2000). One of the contemporary bedrocks of this literature is the conceptualisation of emotion produced by Appraisal Theory (Lerner and Keltner 2001). This theory posits that emotions surge as reactions to, and evaluations of, events and targets that are initially cognitively ‘appraised’ within milliseconds of the occurrence - this complex appraisal then dictates the emotional reaction that follows (Nabi 1999). These emotional reactions in turn guide how people rationally understand events, and can therefore influence the behaviours they adopt towards specific events, people, and objects. Within the scope of Appraisal Theory, it is impossible to overstate the importance of emotional reactions in human behaviour, making them one of the most important determinants of how people act (Frijda 1986).
Appraisal Theory therefore allows for an understanding of emotional reactions as part of the intrinsic cognitive process affecting how an individual behaves in a political world, opening doors to research interested in the effects external circumstances can have on political attitudes and behaviour, with emotions as the mediating variable. Through this framework, researchers in political science have been able to overcome the limitations imposed by rational choice models, taking a step into the field of psychology to understand individual actions determined by ‘gut’ feelings that result in people breaking from their predisposed habits. It is this understanding, taking an “emotions shape behaviour” assumption, which allows research to move beyond the view that “progress and democratic politics require less emotion and more reason” (Arkes 1993 in Marcus 2000), and opens the door to inquiries into questions where emotions play a key role, such as the effects produced by feelings invoked by political campaigning, leadership rhetoric, and most importantly to this paper, media coverage.
Much of the research evaluating the influence of emotions on behaviours measure emotions through a dual-system model (Dillard and Peck 2001), where emotions fall on a continuous scale structured by the type of arousal produced. Emotions that trigger the behavioral inhibition system (BIS) results in averse impulses to external events that may be considered a threat, punishment, or nonreward, with reactions to inhibit undesirable outcomes - these are negative affects, such as fear and anger, which are understood to be on a side of ‘tense arousal’ (Thayer 1989). On the other end of the spectrum are emotions resulting from ‘energetic arousal’, triggered by the behavioral approach
system (BAS) that responds to stimulus of reward and non-punishment, producing positive affect in the form of joy, happiness and contentment. The Discrete-Emotions Models of Affect (Lazarus 1991; Frijda 1986) take our understanding of emotions in the social world a step further by linking emotions to behaviour. While these models also conceptualize emotions as broadly pertaining to negative versus positive affect, they diverge by adding that distinct discrete emotions within these categories result in different categories of ‘action tendencies’. In other words, these models add a second axis to emotional reactions, not only typifying them by affect (positive vs negative), but also by low and high arousal (Russell 1980). For example, within positive affect emotions, happiness produces bask and bonding action tendencies, while contentment is linked to immobility; anger and sadness, both negative valence emotions, result in different action tendencies, the former leading to attack and rejection, while the latter invokes revisiting and doubt (Roseman et al. 1994; Oatley 1992; Scherer 1984).
While far from a perfect measure of emotional reactions, Facebook Reactions broadly resemble the positive-negative valence present in the majority of emotional models (Giuntini et al. 2019). These click-reactions were introduced explicitly to allow users to more clearly signal how they feel about the content presented to them beyond the default “Like” button whose interpretation can be ambiguous (Tian et al. 2017; Gerlitz and Helmond 2013). In an interview with Forbes regarding the reasons behind the introduction of Reactions, Facebook product manager Sammi Krug explained that “when people come to Facebook, they share all kinds of different things, things that make them sad, things that make them happy, thought-provoking, angry. We kept hearing from people that they didn’t have a way to express empathy” (Forbes 2016).
While the Reactions can be broadly categorized into positive (Love, “Wow”, and “Haha”) and negative (Angry and Sad), pundits have argued that some have a much clearer interpretation. Particularly, the Angry and Love emotions resemble the most clear-cut division between negative and positive valence, the former linked to sensations of antagonism and displeasure, the latter to admiration and shared interest (Eberl et al. 2020). Mapped onto the Discrete-Emotions model, the Angry Reaction is linked to anger, a high-arousal negative valence reaction, while the Love reaction can be linked to contentment, a low-arousal positive reaction. Furthermore, the Sad reaction can also be clearly evaluated under the Discrete-Emotions model, relating to sadness - similarly to anger, a negatively valence emotion, but on the low-arousal spectrum, related to doubt and inaction. Evaluating these three Facebook Reactions on the axes of valence and arousal, it is possible to address how political news produces emotional reactions on Facebook.
Valence, Frames, and Emotions in the Press
To understand how people react to political news online, one must first understand their reaction to news in a more general setting. Recent literature points to the fact that news media is becoming increasing dramatised: the growth in popularity of 24-hour broadcast news, the politicisation of media platforms, and the establishment of today’s network
society mean that news agencies of all kinds are fiercely competing for viewers. This has resulted in the rise of emotion-heavy news coverage, where emphasis on conflict, drama, fear, and anger play a central role in keeping users engaged with the content provided.
Despite the rise of emotionalized media news coverage, research in communication science has lagged behind in the exploration how particular messages in the media can result in emotional reactions, often taking the more popular cognitive approach. Nevertheless, there does exist evidence linking media messages and emotional reactions: exposure to contradictory information in the media has been shown to be accompanied by emotional responses that increase the effect of a disqualifying belief (Edwards and Smith 1996), for example. However, much of this research comes from experimental literature on the persuasion of audio-visual effects (Edwards and Smith 1996; Nabi 1999); here the role that ‘high-arousal’ content plays in emotional reactions has been explored, showing a relationship between exposure to high affect images and persuasive emotional reactions (Nabi 2003). Work on media responses to catastrophes have linked this experimental media research with political communication, exploring how citizens respond to emotionally charged news on crises, especially televised coverage of terrorism (Cho et al. 2003; Gurwitch et al. 2004; Shoshani and Slone 2008; Ofman and Mastria 1995). These studies do not only reveal that negative news induces negative emotional states, including enduring post-traumatic stress disorder (Ofman and Mastria 1995), but also that these emotional reactions have a strong impact on political attitudes, even leading to radicalization (Shoshani and Slone 2008). This literature, while mostly focusing on audio-visual effects and not on the nuances in written media, does create a basis for the understanding of the empirical effects of news coverage on emotions. Specifically, it points to how negative news indeed produces negatively valenced emotional reactions, such as anger and sadness, while positive news leads to consumers experiencing positive emotions.
Research on the effects of written news consumption has been conducted under the auspices of Framing Theory (de Vreese 2005; Lecheler and De Vreese 2011), as it provides a basis to understand how nuances in the communication of media messages can result in different effects in the audience. This understanding of news takes into account the fact that often news is presented within a frame that gives a particular understanding of the events discussed. Within this field of research, there is often a focus on the effect that the emotional valence of a news frame has on how consumers react to the information provided. Many of these studies confirm the link between valence and emotions found in the multimedia psychological literature discussed previously: valence in news has effects on emotional responses (Gross and Brewer 2007; Gross and Ambrosio 2004). Framing literature has particularly explored how positive framing in journalistic pieces results in positively valenced emotional reactions, including enthusiasm and hope (Lecheler et al. 2015), while frames that emphasise the negative aspect of a story have been shown to result in negative valence emotions and opinions (Fern´andez et al. 2013; Brader et al. 2008).
Much of this debate has focused on ‘hot’ public debates (Lecheler et al. 2015; Neuman et al. 2018; Ramasubramanian 2010), and less has evaluated how an electorate responds to political news at large, with numerous studies calling on the need to understand whether these valence effects are present outside of particular issues. This study takes up this call by addressing how an electorate responds emotionally to ‘average’ political news during an election, investigating whether the links between frame valence and emotion, as well as the negativity bias, still hold true for a large corpus of news that goes beyond typically contentious issues. I therefore formulate the following hypotheses regarding Research Question 1 on the relationship between media content and emotional reactions:
Hypothesis 1a: Negative political news is associated with negative emotional reactions, such as anger or sadness. Hypothesis 1b: Positive political news is associated with positive emotional reactions, such as contentment or happiness.
When discussing the effect of negative and positive news on reactions, it is important to note the degree to which negative and positive material impact reactions. In the cognitive literature, the existence of a negativity bias (Kahneman and Tversky 1979; Tversky and Kahneman 1992) has long been recognized, with negative content dominating attention (Neuman et al. 2018), and having a stronger impact than positive content on political behaviour (e.g. Lau and Rovner 2009; Ansolabehere et al. 1994). While less documented, work on this negativity bias confirms its existence in emotional reactions as well, with negative news leading to stronger emotional reactions than positive news (Soroka et al. 2015). These expectations are also in line with the psychological literature that views negative emotions as generally more activating than positive ones. Therefore, I also formulate a hypothesis on the comparative effect of good and bad news on emotional reactions:
Hypothesis 1c: Negative political news will elicit stronger emotional responses than positive political news. While this article focuses on the linkages between valence and emotion, news valence is far from the only content feature that affects how an audience reacts emotionally to media. Work on frames suggests news framing can have an effect on how people react emotionally to news. Particularly during elections, when there is an intense focus on political news, previous literature has suggested that how coverage portrays politics can have an impact on individuals’ appreciation of political news. Two frames stand out as being salient throughout elections. The first is the Game-Strategic frame, that focuses electoral coverage on the ‘horse race’ aspects of politics, casting the democratic process as a competition that involves the employment of distinct strategies to beat opponents, usually relying on war or sport-like commentary to describe advances in elections (Iyengar et al. 2004). The Game-Strategic frame has been studied extensively in the American and Western European contexts, partly because of its rising popularity during elections (Str¨omb¨ack and Kaid 2009), but also because of
the normative considerations that have been attached to its use. A range of academic work has argued that presenting politics as a conflict-centered horse-race has a detrimental effect on democracy, since it can steer the audience away from understanding the ‘substantive’ dimension of politics, and instead focuses on politics ‘as a game’ (Valentino et al. 2001), leading to deep cynicism about the political process (Cappella and Jamieson 1997). This study expects that these considerations linking the Game-Strategic frame to negative reactions will also be reflected in emotional reactions on social media, presented in the following hypothesis:
Hypothesis 1d: The presence of a Game-Strategic frame will result in negative emotional reactions.
A second prevalent frame in political news throughout elections is the Policy frame. This frame is often presented in contrast to the Game-Strategic frame; instead of focusing on politics as a competition between actors, it evaluates policies enacted by the government as well as those proposed by the candidates running for office. This more ’sincere’ way of covering political news that is “framed within the context of policy and leadership problems and issues” (Valentino et al. 2001, 94), has been theorized to produce positive responses in an audience as they are able to engage more substantively with politics (Patterson 1993; Zaller 1999; Gilens et al. 2007). This study explicitly tests this relationship in user interactions with Policy news on social networks in the follow hypothesis:
Hypothesis 1e: The presence of a Policy Oriented frame will result in positive emotional reactions.
Emotions and News Sharing Behaviour
With the growing importance of news sharing in the distribution of news, research has attempted to understand what characteristics of content itself leads to fluctuations in sharing. This literature takes a news-value approach (Eilders 2006; O’Neill and Harcup 2009), arguing that there are structural characteristics that make a story more likely to be ‘shareworthy’ (Trilling et al. 2017). One of the key content characteristics that has been explored, and is of interest for this work, is news valence, particularly how the presence of good or bad news relates to sharing behaviour. When it comes to sentiment, the jury is definitely still out, with some studies finding that positive valence is linked to more sharing (Bakshy et al. 2011; Berger 2012; K¨umpel et al. 2015), while others provide evidence for both negative and positive impact on sharing (Trilling et al. 2017). These effects, however, are found on news that encompass a variety of topics; when looking explicitly at political news, studies have concluded that negative news stories receive more attention on websites, with some initial evidence that this attention translates into news sharing on SNS (Trussler and Soroka 2014; Ørmen 2019; Harcup and O’Neill 2017). Because the present study focuses explicitly on political news, we expect there to be a stronger propensity for users to share negative news, due to the negativity bias discussed previously. This relationship is articulated in the following hypothesis:
Hypothesis 2a: Negative political news articles are more likely to be shared than positive political news articles.
While it is important to enquire into the relationship between content sentiment and sharing behaviour as previous studies have done, this study also aims to establish relationships between particular emotional reactions in SNS publics and sharing behaviour. According to the Discrete-Emotions Models of Affect, external stimuli do not only produce emotional reactions, but these emotional reactions are linked with respective action tendencies that help us understand how emotions guide behaviour (Frijda 1986; Lazarus 1991). Here, anger is understood to produce active states, particularly of attacking and rejecting, while positive emotions are related to passive states, such as basking and immobility (Oatley 1992; Roseman et al. 1994; Scherer 1984). I hypothesize that this relationship manifests itself on social media. These expectations provided by the psychological understandings of action tendencies are supported by empirical news value literature that proposes that people are more likely to actively engage with news that evokes negative emotions (Soroka 2009). From this literature, we expect that people share political news that is clearly negative about a candidate one does not support, such as corruption scandals, rather than positive news about a candidate one does support, such as the release of a policy plan. Based on this evidence, I formulate the following hypothesis on the relationship between emotional reactions and news sharing on social media:
Hypothesis 2b: Political news articles that produce negative emotions are more likely to be shared than political news articles that produce positive emotions.
Research on the link between news and emotions has pushed past the divide between good and bad news, with recent work identifying that within the spectrum of negative emotions, news that produces anger leads to stronger action tendencies than fear (e.g. Soroka et al. 2015; Trussler and Soroka 2014; Goodall et al. 2013; Lecheler et al. 2013). This study incorporates this approach by distinguishing between the effect of anger and sadness on sharing. Because anger has been identified as a high-arousal emotion, while sadness is low-arousal (Nabi 2002), I also hypothesize that:
Hypothesis 2c: Political news articles that produce anger are more likely to be shared than political news articles that produce sadness.
3
Method
To test the relationship between news content, emotions, and news sharing, I analyze the sharing and interaction of news from five major Mexican news sites, El Universal, El Financiero, Proceso, Milenio, and Excelsior, which are recognized to be among the top online news sources in Mexico (Reuters, 2017). With the fifth most Facebook users in the world, Mexico is an ideal case study for understanding how people react emotionally to news. Not only does a large part of the population use Facebook as a regular source of news, but over 60% of users report sharing news on a weekly basis (Newman et al. 2019). The 2018 elections are of particular interest for understanding emotional and online engagement, since it was reported to have witnessed unprecedented levels of online engagement (Glowacki et al.
2018), as well as heightened emotional polarization due to the presence of a populist candidate (Greene and S´anchez-Talanquer 2019; Camp and Mattiace 2020; Mattiace 2019).
Data
To address the research questions posed in this study, a large corpus of news articles published during the Mexican elections was required. Due to the limited availability of documented and stored Mexican news media in traditional historical newspaper databases such as Lexis Nexus, this study had to rely on alternative data collection methods. To do so, webscraping techniques were used. The data collection was restricted to the official campaigning period in Mexico, which starts the first of March and concludes the first of July when elections are held. A growing body of literature in Communication Science has relied on webscraping tools to collect news articles: these studies usually query the RSS feeds of the news sites on an automated basis, downloading articles as they are published (Trilling 2014; Trilling et al. 2017). Nevertheless, this technique only allows researchers to extract news as it is published, and not in a retrospective form. To overcome this issue, I turned to the internet archival site Archive.org, which has extensively documented thousands of news sites by taking screenshots of their homepages several times a day. The news site sample selection was therefore based on two criteria: 1. Be on the list of the top most accessed news sites in Mexico (Reuters, 2017), and 2. Have at least two daily screenshots on archive.org from March to July 2018. This resulted in five news sites: El Universal, El Financiero, Proceso, Milenio, and Excelsior.
A webscraper was coded in Python that ran through Archive.org, accessing all the homepages for the newspapers for each of the days for the selected time period. To capture within-day changes of the news sites as articles are uploaded throughout the day, the webscraper downloaded a total of 5 screenshots for each day: one in the morning, one at night, and three randomly selected time-periods in between. Once all these homepage screenshots were captured, the webscraper proceeded to archive all the article links present, accessing them and downloading the title, text, author, and date of publication, storing all this information in a dataset organized by news site. This resulted in a total of 47,341 articles from March to July of 2018.
Since the study is interested in political news articles, it was necessary to distinguish between political and non-political articles. Scraping only ‘non-political’ sections was not a viable solution, since newspapers have different organizational structures, and might have several sections that include political news, especially during the elections†. Moreover, the size of the corpus made manual coding unfeasible. Instead, the study opted for supervised machine learning classifiers to distinguish between article topic, a method that has been widely used, particularly for its positive results in distinguishing between article topics (Boumans
†For example, an article about one of the candidate’s life might appear in a
Table 1. Descriptive Statistics of All Variables in the Data Set
Variable N Mean St. Dev. Min Max
Content Variables Bad News 17,246 0.465 0.499 0 1 Good News 17,246 0.070 0.255 0 1 Policy Oriented 17,246 0.059 0.235 0 1 Game Strategic 17,246 0.020 0.139 0 1 Facebook Reactions Shares 17,246 290.645 1,586.707 0 86,212 Angry 17,246 79.299 498.230 0 18,197 Love 17,246 29.185 194.084 0 8,562 Sad 17,246 5.015 52.544 0 2,943 Likes 17,246 365.124 1,465.182 0 40,833 Haha 17,246 137.375 659.177 0 17,952 Wow 17,246 19.041 117.877 0 5,554
and Trilling 2016). To do so, a sample of 2,000 articles‡were manually annotated by topic, and used as input data to train a classifier using a Support Vector Classifier algorithm. The classifier performed above the usual quality metrics, with precision and recall above 0.85. This resulted in a total of 17,246 political articles.
To gather social media interaction with these news pieces, Facebook data was queried for each of the articles using CrowdTangle, a social media analytics company. Using the CrowdTangle Application Programming Interface (API), each article was paired with key Facebook interaction data: how many times each article was ‘Shared’, ‘Commented’, and ‘Reacted to’ on Facebook. Specifically, the querying returned information for each ‘public’ post that included the respective link, and information on how many times the post itself was interacted with by private accounts. It is important to note that this is substantially more information than that available through Facebook’s own public Graph API: for example, if an article is shared by five different public Facebook pages, the public Graph API would only return a share-count of five. The CrowdTangle API, on the other hand, provides statistics of how often each of these ‘public’ posts were then shared and reacted to by private accounts, providing much more granular data on actual Facebook activity.
Operationalization of Variables
Emotions Key to this study are emotional reactions to news, both in regards to how content variables cause emotional reactions in individuals, and how these emotional reactions then lead to news sharing. To understand how individuals respond emotionally to articles, this study takes a novel approach by operationalizing Facebook’s Reactions feature as distinct emotional reactions. Introduced in 2016, this feature allows users to react to a post beyond the ‘Like’ functionality: it presents users with five distinct ‘emojis’ that allow users to express how they feel about a post. These are the Love reaction in the form of a heart, Angry reaction in the form of an angry face, Sad reaction in the form of a crying face, ‘Wow’ reaction in the form of an amazed face, and ‘Haha’ reaction in the form of a face laughing (Tian et al. 2017).
This study focuses on three specific reactions: the Angry, Love, and Sad reactions. Out of the five Facebook reactions, these are the most unambiguous and straightforward to interpret in terms of negative and positive reactions, identified as being “the most negatively and most positively valenced Reactions that Facebook allows users to convey” (Eberl et al. 2020). Following past research (Eberl et al. 2020; Larsson 2018; Giuntini et al. 2019), Angry and Sad reactions are operationalized as negative emotional reactions, while the Love reactions are operationalized as positive emotional reactions. While far from perfect indicators of emotions, the quantity of reactions and the unambiguous nature of the Love, Sad, and Angry Reactions does allow for a general understanding of how each particular news piece created an initial feeling in the reader§. As Eberl et al. (2020) argues, “while such a click of a button shouldn’t be put on the same level with the actual and genuine emotional response to which these emojis refer, there is no denying that users understand that Reactions are affect-related and that they are able to differentiate among the various emotions presented” (p. 50).
Content Variables This study investigates how the presence of negativity (H1a), positivity (H1b), Game-Strategic (H1d) and Policy (H1e) frames in political news articles result in distinct emotional reactions in the population. These content variables therefore lie at the center of the first research question and set of hypotheses. Here, negative news is understood as news that deals with particularly negative topics, such as crime, defeat, or loss, or news that has particularly negative connotations, while positive news refers to articles that report on successes and victories. Previous studies have used sentiment in text in their research - these mostly rely on dictionary-based approaches, where words are each assigned a value corresponding to whether they are positive, negative or neutral, and an article’s overall sentiment is calculated from these values (e.g. Hutto and
‡A stratified random sample was conducted that accounted for the share of
articles by news site.
§For example, it is unlikely that an article reporting on a corruption scandal
would result in many Love reactions, and reactions would probably reflect the indignation felt by the public in a prevalence of Angry reactions.
Figure 1. Distribution of Facebook Shares and Reactions
Note. Histograms displaying the negative binomial distribution of Facebook Reactions. For better readability, the graphs are cropped at the 100 reaction count and at a frequency of 1000.
Gilbert 2014; Thelwall et al. 2010). Nevertheless, it has been shown that these sentiment dictionaries can lead to flawed results, since their broad coding of words does not allow the context in which they are used to play a role. Therefore, this study follows previous work (Boukes et al. 2019; Gonz´alez-Bail´on and Paltoglou 2015) in labelling articles as positive, negative or neutral by using manual coding and Supervised Machine Learning to classify articles by their sentiment.
The Game-Strategic frame is understood as news reporting on the fluctuations of the state of the presidential race as if it were a competition, focusing on elements such as the state of polling, who endorses which candidate, and the voter groups each candidate is targeting. Policy Oriented news is understood as news that covers the proposed policies of different candidates. Supervised machine learning classifiers were used to code the use of these frames. Manual coders used a code-book (Appendix 2) to label a total of 1,500 articles to train these classifiers. This labeling resulted in satisfactory inter-coder reliability scores as calculated by the Krippendorff’s alpha for the Game Strategic (α = 0.64) and Policy Oriented (α = 0.70) coding, as well as for Good News (α = 0.75) and Bad News (α = 0.88).
Table 1 presents summary statistics of both content-level and Reaction variables. As can be seen, there is high prevalence of Bad News in the sample, with 8,019 negative
articles, which is in stark contrast to the 1,207 articles identified as Good News. There is also a low frequency of articles containing Policy Oriented and Game-Strategic framing, with a total of 1,017 and 345, respectively. This low occurrence can be explained through the low recall of the supervised machine learning classifiers used: for Bad News, Policy Oriented framing, and Game-Strategic framing, the recall was of 0.25, 0.32, and 0.32 respectively (see Appendix 1), meaning that only between a quarter and a third of all cases were identified. Nevertheless, the high precision of the classifiers suggests that we can be certain that the majority of cases that were identified were done so correctly, meaning that estimates produced will be more conservative.
Statistical Models
To address Research Questions 1 and 2, two separate statistical models were constructed. The first addresses the effect of content variables on emotional reactions: using negative binomial regression models to account for the count distribution of the data (see Figure 1), the Facebook Angry, Sad, and Love reactions serve as dependent variables to be predicted using the independent variables of Bad News, Good News, Policy Oriented frame, and Game-Strategic Frame in a news article, while controlling for the
Figure 2. Total Count of Reactions in Sample
outlet and date of publication. For the purpose of clarity, it is important to distinguish between valence in content and valence in emotional responses. Negative and positive news are operationalized in the models as ‘Bad News’ and ‘Good News’ (content level), while “negative emotional reactions” are operationalized as Angry and Sad Reactions, and “positive emotional reactions” as the Love Reaction (reaction level).
A model was developed for each of the Reactions of interest, with Models 1 and 2 predicting Angry and Sad Reactions to examine H1a and H1d on content effects on negative reactions, Model 3 predicting Love reactions to examine H1b and H1e on positive reactions, and Model 4 to examine the relationship between content and Sharing H2a. The second set of models use these Facebook Reactions (Angry, Sad and Love) as independent variables to predict to what extent an article is shared (RQ2). Using negative binomial regression, this model predicts the numbers of Shares an article received as a function of each of the Facebook emotional reactions, controlling for the outlet and time since publication.
4
Results
Article Content and Emotional Reactions
Before addressing the results, it is important to first refer back to the Descriptive Statistics in Table 1 and Figures 1 and 2, as they allow some initial insight into how people interact with political news on Facebook in the form of Sharing and Angry, Sad, and Love Reactions. The descriptive data (Figure 1) shows broad diversity in news sharing behavior: 55% of articles were not shared at all, while others received over 50,000 shares. Figure 2 shows a similar pattern for the Angry, Sad and Love reactions, receiving no interactions in 65.7%, 62.3%, and 67.4% of the articles, respectively. Article Reactions and shares therefore follow a count distribution, specifically, a negative binomial distribution, where most articles receive little to no interactions, while some receive a large volume of shares. Lastly, Figure 2 showcases the volume of Reactions and shares: in terms of
emotional Reactions, the Angry reaction is clearly the most used reaction in political news, with a total of 1.3 million reactions in comparison to the 0.5 million Love and 86 thousand Sad reactions. These, however, pale in comparison to the total number of shares received: at over 5 million total shares, it is clear that Facebook users share political news more often than they use the Angry, Love and Sad reactions¶. This first section of the analysis addresses RQ1 by seeking to understand the effect that characteristics in political news content have on how individuals react emotionally on Facebook. Particularly, I address how news valance (good and bad news) as well as popular electoral frames (Policy Oriented and Game-Strategic) effect how individuals interact with news on Facebook. Focusing on the Facebook Reactions that have the most clearly valenced emotional interpretations as a form of emotional signalling, namely the Angry, Sad, and Love reactions, I address how these content characteristics influence emotional reactions. To do so, Table 1 presents the results of four negative binomial regression models predicting the effect of content on Reactions received by each article on Facebook. The results from these models can be interpreted in the following manner: as the incidence rate ratios (IRRs) increase by a single unit, the expected value of the dependent variable should be multiplied by the IRR. Therefore, an IRR of 0.8 means that an increase in the variable leads to 80% of expected shares, and a negative effect, while an IRR of 1.2 results in a positive effect, with 120% of expected shares.
H1a addresses the effect of negative political news on emotional reactions, predicting that this type of news will result in negatively valenced emotions, such as anger or sadness. Models 1 and 2 share insight into this hypothesis, showing that bad news has a positive relationship with both Angry and Sad reactions. Specifically, there is an effect on Angry Reactions with an IRR of 1.2002 (significant at p < 0.001), meaning that negative articles are expected to receive 20% more Angry reactions. Similarly, Model 2 shows that Bad News also has a positive effect on sadness - here, however, the effect is almost six times the magnitude, with an IRR of 2.2368 (significant at p < 0.001), leading to a 123% increase in expected shares, compared to the 20% from Angry reactions. While both these results support H1a, the relative size of the coefficients runs contrary to the expectations laid out by theory on emotional activation: anger is commonly understood to be a more stimulating emotion than sadness, which is considered deactivating. The results here indicate that when it comes to bad news, sadness is associated more with action than anger, leading to almost twice the reactions. Good news, on the other hand, has no statistically significant effect on either expressed anger or sadness.
H1b focuses on the effect of good news on Reactions, hypothesizing that good news is linked to positively valenced emotional reactions. Model 3, which displays the effect of variables on the Love Reaction, supports this hypothesis: it shows a positive relationship between good news and Love reactions, with an IRR of 1.7205 (significant at p < 0.001).
¶It is important to remember that Angry, Love and Sad are only three of six
Table 2. Negative Binomial Regressions Predicting the Number of Reactions on Facebook
Model 1: Angry Model 2: Sad Model 3: Love Model 4: Shares
Content Variables Bad News 1.283∗∗∗[1.164, 1.414] 2.374∗∗∗[2.152, 2.620] 0.696∗∗∗[0.639, 0.759] 1.366∗∗∗[1.254, 1.489] Good News 1.106 [0.917, 1.346] 1.065 [0.876, 1.303] 1.710∗∗∗[1.451, 2.028] 1.028 [0.869, 1.224] Policy Oriented 1.139 [0.937, 1.398] 1.618∗∗∗[1.334, 1.981] 1.266∗∗[1.066, 1.515] 1.418∗∗∗[1.194, 1.700] Game Strategic 1.080 [0.784, 1.540] 0.911 [0.650, 1.310] 2.333∗∗∗[1.760, 3.173] 1.740∗∗∗[1.308, 2.378] Controls El Universal 0.575∗∗∗[0.498, 0.663] 0.535∗∗∗[0.462, 0.619] 0.383∗∗∗[0.337, 0.434] 0.436∗∗∗[0.384, 0.495] Excelsior 0.662∗∗∗[0.561, 0.781] 0.526∗∗∗[0.444, 0.623] 0.558∗∗∗[0.482, 0.646] 0.446∗∗∗[0.386, 0.517] Milenio 0.827∗[0.707, 0.967] 0.850∗[0.726, 0.995] 1.067 [0.930, 1.224] 0.919 [0.800, 1.056] Proceso 4.544∗∗∗[3.868, 5.337] 2.886∗∗∗[2.461, 3.383] 1.099 [0.953, 1.268] 2.351∗∗∗[2.036, 2.715] Exposure 1.001∗∗∗[1.001, 1.001] 1.001∗∗∗[1.001, 1.001] 1.002∗∗∗[1.002, 1.002] 1.002∗∗∗[1.002, 1.002] Days since t0 1.003∗∗∗[1.001, 1.004] 1.009∗∗∗[1.007, 1.010] 1.002∗[1.000, 1.003] 1.000 [0.999, 1.002] Constant 2.650∗∗∗[2.267, 3.106] 0.219∗∗∗[0.187, 0.257] 1.706∗∗∗[1.481, 1.969] 11.256∗∗∗[9.787, 12.974] N 17,246 17,246 17,246 17,246 Log Likelihood −40,232.910 −20,478.200 −33,210.390 −56,921.280 θ 0.106∗∗∗(0.002) 0.117∗∗∗(0.002) 0.139∗∗∗(0.002) 0.132∗∗∗(0.002) AIC 80,487.820 40,978.400 66,442.770 113,864.600
Note. IRRs with confidence intervals in brackets. Values < 1 indicate a negative effect, values > 1 indicate a positive effect. Results relating to a hypothesis presented in bold. AIC = Akaike information criterion; IRRs = incidence rate ratios.
∗p < .05;∗∗p < .01;∗∗∗p < .001
This means that when an article is positive, it is predicted to receive 72% more Love reactions. On the other hand, bad news leads to a 67.4% decrease in expressed love. With these results, it is possible to confirm H1b. Having confirmed the effect of good and bad news on positive and negatively reactions, it is now time to compare the effect sizes. H1c addresses the relative strength of effects produced by good and bad news - based on research on negativity bias and the activating nature of negative emotions, bad news was expected to have stronger effects on emotional reactions than good news. Comparing the effect sizes of good and bad news across Models 1, 2 and 3, one can appreciate that on the one hand, the effect of good news on Love Reactions (1.7205) is smaller than the effect of bad news on Sad Reactions (2.2368). This effect, however, is larger than the effect of bad news on anger (1.2022). Considering that anger is understood to be the activation mechanism behind this hypothesis on the relationship between news valence and emotional reactions, the results do not produce enough evidence to support this hypothesis.
The remaining hypotheses H1d and H1e address the effects of specific journalistic frames on emotional reactions. Model 3 allows us to evaluate the effect of the Policy Oriented frame on the amount of positive reactions an article would receive. H1d expected this relationship to be positive, as users are understood to react positively to issue-based reporting: with an IRR of 1.2737 (significant at p < 0.001), it is possible to support this hypothesis, with the presence of a Policy Oriented frame increasing Love reactions by 27.37%. H1e expected news articles with a Game-Strategic framing to receive negative reactions - with effects that are statistically insignificant on both Angry and Sad Reactions, this hypothesis is rejected. The effect takes place in an unexpected, opposite direction, with a strong positive relationship (2.654) between Game Strategic frames and Love reactions.
Table 3. Negative Binomial Regressions Predicting the Number of Shares on Facebook
Model 5: Shares Reactions Angry 1.0002∗∗∗[1.0002, 1.0002] Love 0.9996∗∗∗[0.9996, 0.9996] Haha 1.0001∗∗∗[1.0001, 1.0001] Wow 1.0007∗∗∗[1.0007, 1.0007] Sad 1.0008∗∗∗[1.0008, 1.0008] Like 1.0002∗∗∗[1.0002, 1.0002] Outlets El Universal 0.6537∗∗∗[0.6511, 0.6562] Excelsior 0.3900∗∗∗[0.3878, 0.3922] Milenio 0.9549∗∗∗[0.9513, 0.9586] Proceso 3.2702∗∗∗[3.2605, 3.2798] Days since T0 0.9984∗∗∗[0.9984, 0.9984] Constant 145.4713∗∗∗[144.9965, 145.9473] N 17,246 Log Likelihood −4,993,445.0000 θ 3,612,624.0000 AIC 9,986,915.0000
Note. IRRs with confidence intervals in brackets. Values < 1 indicate a negative effect, values > 1 indicate a positive effect. AIC = Akaike information criterion; IRRs = incidence rate ratios.∗p < .05;∗∗p < .01;∗∗∗p < .001
Emotional Reactions and News Sharing
This section interprets the effect of bad news and emotional reactions on Sharing. Before addressing the relationship between emotional reactions and sharing, we first look at the effect of content and sharing. Specifically, H2a expected that negative content would be positively related to sharing: Model 4, which displays the effect of content variables on article sharing, allows us to evaluate this relationship. With an IRR of 1.366 (significant at p < 0.001) of Bad News on sharing, it is possible to see that negativity increases the
Table 4. Summary of Hypothesis Testing
Hypothesis Supported? Summary
Reactions
H1a 4 Negative news is associated with negative reactions. H1b 4 Positive news is associated with positive reactions. H1c 7 Negative news elicits stronger responses than positive news. H1d 7 The Game-Strategic frame will result in negative emotional reactions. H1e 4 The Policy Oriented frame will result in positive emotional reactions. Sharing
H2a 4 Bad news content is related to higher sharing than good news content. H2b 4 Negative emotions are related to higher sharing than positive emotions. H2c 7 Anger is associated with higher sharing than sadness.
expected number of shares an article receives by 36.6%, while Good News has no significant effect on sharing. With these results, we can confirm H2a.
H2b predicted that, because anger is an activating emotion, articles that received angry Reactions would be linked to higher sharing than articles that received love Reactions. The results in Model 1 provides evidence that supports this hypothesis: Angry reactions have a positive effect on sharing, with an IRR of 1.0002 (significant at p < 0.001), while Love reactions have a negative effect on sharing, with an IRR of 0.9995 (significant at p < 0.001). This is in line with theories regarding emotional engagement and empirical evidence suggesting that negative content that results in anger is more engaging and mobilizing than positive content resulting in passive contentment. H2c, on the other hand, expected that anger would be a stronger predictor of sharing than articles that resulted in sadness. The model does not support this hypothesis, since Sad Reactions have the strongest IRR of all reactions, at 1.0008 (significant at p < 0.001). This means that, when it comes to sharing political news on Facebook, emotions that cause sadness are more likely to be shared than those resulting in anger, or positive emotions.
When interpreting these results, it is important to note that while these IRRs might appear small when compared to the coefficients produced by binary content variables, this is in large part due to the fact that the model controls for the presence of Likes, which is highly correlated to sharing (r = 0.71). More importantly, however, these results need to be interpreted in the context of SNS, where articles receive high volumes of reactions that range into the several thousands. In this model, a 1,000 increase in Sad Reactions, for example, leads to a 122% increase in the number of expected shares (1.00081000= 2.22 IRR), a 1,000 increase in Angry Reactions
leads to 22% more shares (1.00061000 = 1.22 IRR), while
the same increase in Love Reactions leads to an article receiving only 67% of expected shares (0.99961000= 0.67
IRR). This exercise allows us to clearly appreciate the stark contrast in the effects that a same-number increase produces for different reactions.
Contextualizing these effects by evaluating the changes caused by a same Reaction increase (in this case, 1000) allows us to evaluate the drastic differences between the estimated effects. Nevertheless, a 1,000 unit Reaction increase is more common among certain reactions (Angry and Love, for example, with mean reactions per article
of 79.30 and 29.19, respectively) than others (such as the Sad Reaction, with a mean of only 5 reactions per article). To account for these drastic differences, it is important to interpret these results in relation to the relative distribution of each Reaction - to do so, it is possible to use the Standard Deviations (SD) of each. For the Angry Reaction, the model estimates that a two-SD increase in Reactions leads to an increase of 22% of shares (1.0002(498.23∗2) = 1.22); a two-SD increase in Sad Reactions leads to a 9% increase in expected shares (1.0008(52.544∗2) = 1.09); and a two-SD increase in Love Reactions results in a 14% reduction in expected shares (0.9996(194.084∗2)= 0.86). Therefore, while
the Sad reaction has a stronger effect on sharing behaviour, it is impossible to deny the dominance of Angry reactions when it comes to political news during the elections.
Lastly, Table 4 summarizes the hypotheses addressed in this section, and whether or not each hypothesis is supported. As can be observed, these results support the claims that negative news is associated with negative reactions, positive news is associated with positive reactions, that the Policy Oriented frame is associated with positive emotional reactions, that negative content is related to higher sharing than positive news, and that negative emotions are related to higher sharing than positive emotions. On the other hand, no support was found for the expectations that negative news produces stronger responses than positive news (with this hypothesis being true for sadness, but not for anger), that the game-strategic frame is associated with negative emotional reactions (instead producing positive reactions), nor that anger is associated with higher sharing than sadness.
5
Discussion and Conclusion
In this study, I have shown that both sentiment and framing found at the content-level of political news affects a public’s emotional reactions on social network sites (SNS), and consequently influences the degree to which articles are shared throughout these. By looking at a large sample of political news during an electoral campaign (N=17,246), this study has addressed how an electorate engaged emotionally with political news during the 2018 Mexican national elections on Facebook. Building on the Appraisal Theory framework for understanding emotions and Framing Theory to asses media effects on publics, this study provides evidence for (a) the existence of a negativity bias in emotional engagement with political news, (b) an emotional effect of Game Strategic and Policy Oriented frames during
an election, and (c) the existence of a ‘sadness-bias’ in the sharing of political news.
Does political news sentiment affect an audience’s emotional reactions on social media? The results presented in this paper suggest a resounding yes - positive news leads to positive emotional responses in the form of Love Reactions, while negative news is linked to negative responses, associated with both Sad and Angry Reactions. While an intuitive result, this relationship that has been theorised and tested in the psychological (e.g. Edwards and Smith 1996; Nabi 1999) and media effects literature (e.g. Lecheler et al. 2015; Soroka et al. 2015) has yet to be demonstrated in relation to political news engagement on SNS. Moreover, the bulk of studies looking at emotional effects by media sentiment and frames tend to focus on particular ‘hot-topic’ issues, such as migration and abortion, where heightened emotional responses are to be expected (Neuman et al. 2018; Ramasubramanian 2010). By using an inclusive sample of political news, we are able to address whether this relationship holds across all political news topics at an aggregate level, providing evidence that indeed, news sentiment has an effect on individuals’ emotional responses. Moreover, by using data of actual user interactions on Facebook, the study is able to circumvent desirability and lab effects that might be present in survey and experimental designs.
Comparing the effect sizes of the relationship between content sentiment and emotional reactions reveals more nuanced findings. In this study, I draw from the Discrete-Emotions Models of Affect to distinguish between emotional valence and action tendencies, where negatively valenced emotions are associated with stronger action tendencies than positively valenced emotions (Lerner and Keltner 2001; Nabi 1999; Frijda 1986; Lazarus 1991). We distinguish between activating negative emotions and non activating emotions, where anger is related to attacking and rejecting, while sadness is related to inaction. The effect sizes of article sentiment on emotional responses were expected to reflect this relationship, with negative articles leading to a stronger impact on anger than on sadness, and a larger effect size when compared to the relationship between good news and positive reactions. The results here suggest that this model of behaviour cannot fully explain how people react emotionally to political news on Facebook - the effect of both bad news on sadness and of good news on positive emotions is stronger than the effect of bad news on anger.
This does not mean that anger is not the most pervasive emotion when it comes to political news on Facebook, nor that a negativity bias does not exist: the sheer number of Angry reactions in the sample, at 1,367,583 far outnumber the other reactions: almost three times that of Love reactions at 503,316, and exceed Sad reactions by 15 times, with only 86,497 total Sad reactions. Therefore, anger definitely dominates the political news scene, suggesting that users are more likely to react with this emotion to political news. This is in line with research on emotions that show anger to be the dominating emotion the press (Soroka et al. 2015; Soroka 2009; Neuman et al. 2018; Hasell and Weeks 2016), as well as in political processes more broadly (Valentino et al. 2011; Marcus 2000). It is when we look at the effect of sentiment in the articles themselves that we find a smaller impact of
anger. The fact that the presence of negativity in articles has a stronger impact on Sad Reactions than on Angry Reactions implies a stronger association of negativity and sadness than with anger - this however, could be a product of the widespread use of angry reactions to political news. When it comes to the relationship between positivity and Love reactions, this effect could be enhanced by the fact that positive political news stories are so much rarer than negative news that when they are shared on social media, users are more likely to display a positive reaction.
Moving from the effect of article sentiment on emotions to the effect that journalistic framing has on emotions, this study provides some initial evidence of how users engage emotionally with some of the most popular electoral frames: Game-Strategic and Policy Oriented. When it comes to the Game Strategic frame, which discusses politics as a game or competition and employs war-like language, the results presented here are not in line with previous work on public engagement (Cappella and Jamieson 1997; Valentino et al. 2001). The standard understanding of game-strategic framing is that it angers and disheartens an electorate -by presenting politics as a competition instead of an issue-based exchange, consumers become cynical and develop negative emotions (Iyengar et al. 2004). The evidence presented in this study points to a relationship in the opposite direction: the presence of Game-Strategic coverage leads to an increase in positive emotional engagement, predicting a 265% increase in Love Reactions. There are several reasons that might explain this surprising relationship. Work focusing on the demand-side of journalism has argued that despite the cynicism associated with the Game-Strategic frame, its proliferation is linked to the simple fact that news consumers are interested in reading news framed in this way (Trussler and Soroka 2014; Ørmen 2019). These results could be interpreted as further support of this more recent idea that suggests that while consumers report that they dislike Game-Strategic coverage, they routinely choose it over other types of political news. It is also possible that these positive responses are a product of Facebook users wanting to show their support for a candidate who is reported as winning. The excitement that is produced by horse-race reporting on candidates during an election might be more linked to positive emotions as users see that their candidate is ahead of the pack. These results could also be linked to the fact that Game-Strategic reporting is not as widespread in Mexico as it is in places where this negative relationship has been identified: in our sample, there were only 341 articles labelled as using this frame. From these, most prominently feature headlines declaring a particular endorsement, a change of electoral strategy, or a victory in a certain poll.
Articles that contain the Policy Oriented frame, where journalists focus on particular issues or proposals, are associated both with Love and Sad Reactions. In line with the expectations laid out, articles that focus on policy produce positive emotional engagement, as audiences are presented with policy initiatives from candidates, which are seen in a favourable light. This supports literature on policy framing, which has associated policy focus on positive reactions from audiences (Valentino et al. 2001; Gilens et al. 2007; Patterson 1993; Zaller 1999). Nevertheless, this Policy Oriented frame
is also associated with sad emotional responses. I suspect that this is in part due to the highly negative conversations that occur in relation to policy proposals: when journalists cover a candidate’s proposal to solve a particular problem, this is usually contextualized by explaining how bad the problem actually is. Moreover, the Policy Oriented frame also includes articles that discuss the current (or past) government’s initiatives to address societal issues, and thus contain prolonged descriptions of pervasive societal issues.
The third contribution this article offers is in the area of news sharing. By assessing both the influence that content and emotional reactions have on the sharing of political news across Facebook, this study is able to make several contributions to this field. First is that of the existence of a negativity bias in sharing behavior, with a two-fold effect: both political articles that are themselves negative, and articles that lead to negative emotional reactions are shared more. While the former finding has been identified in previous news sharing studies, most of these limit the scope of their research to a broad measure of negativity, and do not account for the effects that different negatively valenced emotions produced by bad news have on mediating the relationship between content and sharing. Considering that the theoretical bedrock these studies rely on assumes an emotional reaction to negative/positive news that leads to a propensity (or lack of) to share, it is surprising that no previous work has dissected this relationship.
With data showing how individuals both share and react emotionally to political articles, it is possible to disentangle the effects of articles that produce sadness and anger, two different emotions that are often grouped under the umbrella of negativity. The results here show that there is indeed a difference in the relationship between sharing and the emotions of anger and sadness. While the Discrete-Emotions Models led to an expectation that anger would lead to more sharing than sadness, this was not the case: the effect of Sad Reactions on sharing is almost twice the size as that of Angry Reactions, leading to the conclusion that sadness is most closely associated with sharing. There are several explanations as to why this might be the case. One lies in the idea that users might be attempting to avoid conflict with their contacts. Since Facebook is mostly made up of ‘close ties’ (Valenzuela et al. 2018), users are less likely to publicly share political material deemed to be controversial, since it could result in negative emotions (Valenzuela et al. 2017). While reacting angrily to a political article is a low-cost, non-public action a user can take, actually sharing a political article that makes a user angry might lead to conflict with their Facebook contacts.
A second complementary explanation lies in empathy as a communal emotion. Sadness is an emotion that is much less prevalent in political science, but that has its basis in empathy. Referring directly to the articles with the most Sad Reactions, one finds headlines on Donald Trump’s child separation policy, the murder of local political candidates, and analyses of policies failing to reduce poverty. These articles, while political, are all devoid of ideological interpretations and polarization, addressing issues that are tragic in nature. Even though sadness is understood as an immobilizing emotion, users of Facebook are perhaps more inclined to share articles to collectively participate in
the grief caused by tragic political stories, than they are to showcase their outrage about a particular candidate’s position on an issue when they know this might rub a coworker, friend, or family member the wrong way. This finding suggests that when understanding the distribution of political news through Facebook, it is important to look beyond the “stuff that makes you laugh and stuff that makes you angry” (Harcup and O’Neill 2017, 1480), also considering the stuff that makes you sad.
Lastly, this study goes beyond contributing to our understanding of the effect of negative news and emotions on sharing to assessing the effect of the much overlooked positive news and emotions. While there is a debate on whether good news increases sharing or has no effect, this study suggests that positive reactions actually lead to a decrease in news sharing. While good news itself has no significant effect on sharing, an increase of positive emotional reactions to political articles is related to decreased sharing; it is in fact the only Facebook Reaction that is associated with a decrease in sharing. This finding is interesting because it is focused purely on political news during elections: studies that have found a relationship between good news and increased sharing usually address a sample containing a variety of news topics, where political news is shared less than topics such as lifestyle (Bakshy et al. 2011; Berger 2012; K¨umpel et al. 2015; Trilling et al. 2017). The results here show that when it comes to political news, audiences on Facebook are less interested in sharing stories resulting in positive emotions, opting for news that results in negative emotions. This finding further provides evidence for emotional engagement theories that argue that, when it comes to politics, positive emotions are generally deactivating.
As well as providing these novel insights on the links between emotions, politics, and SNS engagement, this article highlights several research inroads that are yet to be made. First, there is also something to be said about the use of Facebook Reactions to study emotions. Although operationalized in this study as emotional reactions, these click-reactions are definitely far from the physiological reactions that experimental research is picking up with the use of skin conductors and eye-tracking (e.g. Kruikemeier et al. 2018; Bakker et al. 2020; Bailenson et al. 2008). There is a very clear cognitive dimension to clicking Love, Sad or Angry that goes beyond the emotion itself - there are considerations of public perception and social signalling, as well as a plethora of individual characteristics that guide distinct Facebook users to click a reaction or not. Further research on this topic would do well to examine the extent to which these Reactions are representative of emotions in a political setting, as well as evaluating the cognitive and social signalling components.
The results from this study also question whether theories linking emotions and action tendencies hold up in the context of SNS - in a digital space where interactions take but a click, are emotionally-triggered action tendencies weighed differently in a user’s mind? More specifically, there is a need to further explore how the nuances within negative emotions impact how a public reacts to news on social media. While the research presented here made an argument linking negative content, Sad and Angry Reactions, and sharing, it
