Twitter Users’ Displays of Affect in the Global Warming Debate

Abstract

This article engages with recent discussions in the field of technical communication that call for climate change research that moves beyond the believer/denier dichotomy. For this study, our research team coded 900 tweets about climate change and global warming for different emotions in order to understand how Twitter users rely on affect rhetorically. Our findings use quantitative content analysis to challenge current assumptions about writing and affect on social media, and our results indicate a number of arenas for future research on affect, global warming, and rhetoric.

Keywords

quantitative content analysis Twitter global warming digital media affect

Introduction

As scientific evidence of the rise of the average global temperature mounts, the terms global warming and climate change are often used interchangeably, but any scientist will tell you that each definition differs in important ways. NASA suggests the confusion lies in the differences between weather and climate: Weather refers to atmospheric conditions that occur on a local level over short periods of time, and climate refers to the long-term regional or global average of temperature over the span of seasons, years, or even decades. This distinction provides clarity regarding the difference between global warming and climate change. Global warming is defined as the long-term heating of Earth’s climate system due to human activities, primarily the burning of fossil fuels. Climate change is defined as the long-term change in the average weather patterns that define Earth’s local, regional, and global climates (NASA, 2008). In other words, humans cause global warming, and global warming, along with other natural processes, cause climate change.

Ceccarelli (2011) notes that while the Intergovernmental Panel on Climate Change agrees that global warming is real and causes climate change, a “majority of Americans believe there is a lot of disagreement among scientists over whether or not global warming is happening” (p. 205). Thus, a deeper interrogation of these two terms, and the public debate surrounding them, seems particularly pressing for technical communicators. After global warming resulted in a polar vortex that sent arctic blasts through Chicago and much of the United States in January of 2019, our research team noticed that CNN was not using the term global warming throughout their coverage of the event and was instead incorrectly using the term climate change. We wondered, then, if news agencies were using the terms interchangeably (and incorrectly), then how was the public using and understanding these two distinct terms?

In an attempt to discern the ways these two terms are understood by the general public, this study began with the goal of understanding how the debate surrounding global warming on Twitter might provide insight into the general public’s usage of these two terms. Over the course of 3 consecutive weeks, six coders worked in pairs to code 900 tweets that included either the phrase climate change or global warming. Based on the ways we saw the terms used on social media, we coded tweets into categories of affect, hypothesizing that Twitter users would choose different terms when expressing different emotions. For the purposes of our study, we understand affect via the American Psychological Association, which defines affect as “any experience of feeling or emotion, ranging from suffering to elation, from the simplest to the most complex sensations of feeling, and from the most normal to the most pathological emotional reactions” (n.p.). In addition, affect is usually described as positive or negative, and both mood and emotion are considered affective states of mind. That said, we posited that global warming would be used with angry, humorous, or sarcastic affect, whereas climate change would be used with reasonable or fearful affect.

To make sense of the data, we used quantitative content analysis (QCA), which Boettger and Palmer (2010) explain is useful “for predefined terms or phrases and [then used] inferential statistics to make conclusions about their presence” (p. 346). Our results ultimately challenge the perception that negative emotions are prevalent on Twitter (Goodwin, 2019; Hasan et al., 2014; Waterloo et al., 2018). In addition, we found no significant difference in the ways the terms global warming and climate change were used on Twitter, and we conclude that this is partially because most Twitter users either do not understand or recognize the difference between these two terms, but this finding may be a result of both the technology we used to pull our data and Twitter’s use of hashtags (most tweets in our data set included both #globalwarming and #climatechange). Thus, additional research into the use of these two terms is necessary.

Twitter and Affect

Interest in the public discourse surrounding climate change has been steadily increasing in Technical and Professional Communication (TPC) and related disciplinary research. In an interdisciplinary review of climate change research, Cagle and Tillery (2015) found that much of this research can be organized around five broad categories: “specific influences on behavior related to climate change; global and local aspects of perceptions of global warming; problems with public understanding; beliefs about and perceptions of science and technical experts; and media coverage of climate change” (p. 150). These broad categories demonstrate the importance of rhetorical concerns to climate change research, such as “how people’s opinions are shaped, how those opinions direct behaviors, and what influences can change their opinions” (p. 150). For example, multiple studies have demonstrated the presence of rhetorical frames in climate change discourse, such as technical documents and corporate news releases (Tutt, 2009), congressional discourse (Majdik, 2019), and popular news coverage (Foust & Murphy, 2009). Foust and Murphy (2009) found that news articles frame climate change within either a tragic apocalyptic frame, which suggests that climate change is a result of fate and out of one’s control, or a comic apocalyptic frame, which indicates that humans have contributed to, and therefore can help mitigate, climate change. Interestingly, the authors suggest that the terms climate change and global warming may be used differently depending on the chosen frame; climate change, associated with the tragic frame, often leads to public divisiveness, passivity, and skepticism toward scientific experts, while global warming, associated with the comic frame, can lead to increased public awareness and desire to take action.

These studies have demonstrated the ways that various communication outlets shape the public’s social values and attitudes; however, recent research has called for an increased focus on digital communication tools and the specific user-centered strategies already employed by the public to circulate their views on climate change (Cagle & Tillery, 2015; Koteyko et al., 2015). In response, current TPC research on climate change has focused specifically on public discourse on social media platforms. For example, multiple studies have demonstrated how social media can serve to enable climate change deniers and allow them to disseminate misinformation and circulate that information widely enough to persuade others (Bloomfield & Tillery, 2019; Matthews, 2015). This work indicates that the internet has turned into a “public battleground over climate science” (Bloomfield & Tillery, 2019, p. 24), where digital media users are divisively categorized as either climate change believers or skeptics. While understanding the rhetoric of climate change denialism is crucial to developing messages that can disrupt those beliefs, Cagle and Herndl (2019) argue that studies focused exclusively on a believer/denier dichotomy may miss some of the more nuanced exchanges happening online. In their analysis of the Reddit subforum “Change My View,” the authors found tense and unproductive exchanges but also found that “contrary to widespread perception, collaborative deliberation about climate change does exist online” (p. 23). While Cagle and Herndl (2019) found that Reddit’s guidelines for participation helped create productive dialog, our study investigates whether productive rhetorical exchanges can also occur in a public, unmoderated space such as Twitter.

Recent studies of emotion and affect use data from Twitter because of its popularity as a platform, and scholars in multiple fields interested in digital communication have found a vast and continuously expanding pool of informative data within the social media application. Sometimes referred to in this research as microblogs, tweets are extremely short in comparison to regular blog posts due to the character limit, making the tweet a quick and effective method for broadcasting different types of information (Bollen et al., 2011). Twitter is a useful data source due to its accessibility to diverse communities and interactive hashtags, allowing keyword searches for easy data collection on Twitter. As a delivery system, Twitter tends toward quickly drafted and short arguments, with its 280-character limit, and the most common length of a tweet being 33 characters (Perez, 2018). Twitter users connect to conversations using hashtags, which are brief keywords or abbreviations following the symbol “#” in order to make tweets more easily searchable amidst a broad network of tweets (Bruns & Stieglitz, 2013).

In terms of methods, many studies (Fersini et al., 2016; Fung et al., 2016; Gaspar et al., 2016; Mohammad & Kiritchenko, 2015; Saif et al., 2016) use keyword counting and manual categorization to find user emotions and sentiment on Twitter. Other studies (Do et al., 2016; Hasan et al., 2014; Q. Li et al., 2020) rely on computational methods, and some use a combination of both computational and manual methods (Chung & Zeng, 2020). In most of this research, there is a strong focus on a positive and negative categorization of emotions (X. Li et al., 2015), largely because the research is centered on significant events, such as a natural disaster, community crisis, or a wide-spreading disease. That said, sentiment analysis techniques are increasingly used across multiple disciplines to study reactions from social media users related to unexpected and potentially stressful social events (Gaspar et al., 2016). According to Jin (2010) and Liu et al. (2011), public reactions to crises comprise four main coping methods: rational thinking (cognitive), emotional venting (emotional), instrumental support, and action (conative). These subcategories of positive and negative emotions are fairly consistent throughout current research involving sentiment analysis.

In one example Li et al. (2020) used computational methods across multiple platforms (Twitter, local microblogs, Facebook, and news blogs) to track public emotion during COVID-19 using the contents of microblogs and to extract the details of dominant events. The researchers divided their data into three phases and determined prevalent emotions varied according to different phases of COVID-19. Across the three phases, love (positive) was the most frequent and stable emotion during the first phase, and anger (negative) was the most stable emotion during the second phase. Transitions between sadness (negative) and disgust (negative) were relatively high during the third phase. However, the researchers in another study (Do et al., 2016) also used computational methods but coded their data as positive, negative, or neutral. They determined fear and anger (negative) dominated tweets categorized as emotional during the initial spread of the MERS virus in South Korea in 2015, while 80% of tweets were categorized as neutral (nonemotional). In addition, they determined fear was largely directed toward the virus itself while anger was mostly directed toward certain government officials for their slow response to the virus.

Also contrary to the common emotional categorizations of positive or negative in affective studies, Gaspar et al. (2016) conducted a qualitative analysis of affective expressions on Twitter collected in Germany during the 2011 EHEC food contamination incident. In this study, affective expressions of coping during this crisis were found to be diverse not only in terms of valence but also in terms of the adaptive functions they served. The authors of this study argue affective results reach beyond a positive or negative tone because some people perceived the outbreak as a threat while others viewed it as a challenge to cope with. In other words, the intent behind each emotional expression complicated whether or not they perceived an emotion as merely positive or negative. The authors also argue human-based methods can help discern beyond positive and negative emotions on Twitter. For example, expressions of anger, which may be seen as simply irrational, could have actually been an attempt to improve the situation rather than lash out at a particular person or situation. Often, these expressions were not unspecific “bursts” of anger, but rather targeted at institutions perceived as responsible for mitigating the threat (p. 18). In addition, affective expressions in their Twitter data set varied not only in terms of positive or negative valence but also in terms of (a) the form in which it was expressed (e.g., worrying for other people) and (b) the function it may have served (e.g., anger toward the authorities). These studies illustrate nuances important to the study of global warming, digital communication, and affect and influenced the ways we approached this study, particularly with regard to the necessity of human interpretation.

Research Methods

With the ever-expanding and thus overwhelming nature of data sets in the digital world, and particularly on Twitter, social media researchers lean toward computational methods of content analysis as the preferred method of analysis as opposed to manual or traditional approaches. These computational methods use large, complex data sets and employ algorithmic or computational solutions to generate patterns and inferences from data (Shah et al., 2015). Technical communication scholars have expressed caution toward these “big data” approaches, emphasizing that while this work may reinvigorate the field by requiring increased attention to new coding methods, TPC researchers should not abandon their unique ability “to produce coherent and meaningful narratives from data” (Pflugfelder, 2013, p. 19). However, Zamith and Lewis (2015) recognize the benefits of both methods and propose a “hybrid” approach to content analysis. They focus on four key processes of QCA: (a) the development of the coding protocol and sheet; (b) the specification of the population and, if applicable, the sample; (c) the establishment of intercoder reliability; and (d) the coding of content (p. 309). Despite the reality that algorithmic approaches are exponentially faster, we followed Zamith and Lewis’s hybrid model because the manual coding approach allows researchers to make sense of data sets that contain ambiguity in the sense that human coders are more likely to adapt to unusual manifestations in the data and can adjust the coding scheme accordingly.

In addition to Zamith and Lewis’s (2015) four processes, this study also took Knudsen and Stage’s (2016) work on affective methodologies into consideration. They outline the nature of affective research as an innovative strategy for “(1) asking research questions and formulating research agendas relating to affective processes, (2) for collecting or producing embodied data, (3) and for making sense of this data in order to produce academic knowledge” (p. 1). Considering the nuances of human emotion, a sole reliance on algorithmic methods for content analysis can be problematic when analyzing mood, emotion, or affect because machines cannot (yet) do the interpretation necessary to analyze the presentation of emotions, and humans are more capable of adjusting in response to intricate social nuances.

Thus, considering Knudsen and Stage’s suggestions, this study used QCA to analyze the data after it was pulled from Twitter using algorithmic methods. QCA is considered a beneficial method for technical communication research by both Boettger and Palmer (2010) and Thayer et al. (2007). Thayer et al. (2007) note that QCA is “one of the most powerful yet least understood methods” in TPC research, while Boettger and Palmer (2010) argue that QCA offers an empirical framework “for making replicable and valid inferences from texts (and other meaningful matter) in the contexts of their use” (p. 346). As opposed to QCA, which relies on researchers’ abilities to determine relevant themes that emerge from the data, QCA identifies meaning through mutually exclusive categories determined prior to coding, uses interrater reliability in order to increase measures of validity and reliability, and utilizes inferential statistics to answer a question or prove/disprove a hypothesis. In addition, QCA requires increased rigor during the initial phase of research, which leads to sustainability, durability, and portability in the research process, as other researchers can either build on or more accurately challenge these results (Meloncon & St. Amant, 2019; St. Amant & Graham, 2019).

Keeping the inherent complexity of the study of affect in mind, this study blends computational methods and manual methods in order to gather and organize data from Twitter. We chose to use computational methods to gather the data but manual methods to code and analyze the data for two reasons. First, computational data collection methods allowed us to gather data from Twitter in a reliable manner that can be replicated by others, and such gathering methods are available open access. Second, we manually coded the data with a research team in response to research that indicates the coding of emotions is complex and thus requires human understanding and the ability to interpret social nuance. With six coders, we were able to manually code a large data set without sacrificing complexity. In the following sections, we demonstrate how we completed each of these steps during our research process before discussing our results.

Research Questions and Sample Selection

We defined our research questions based on our hypothesis that Twitter users would use the terms global warming and climate change incorrectly and that global warming would most often be associated with affective responses such as anger and humor while climate change would more often be associated with reasonable or information-centered affect. We settled on the following research questions:

How, if at all, are the terms global warming and climate change used differently on Twitter?

Are the terms global warming and climate change used more frequently when a Twitter user is expressing a specific emotion?

To test these questions, we could not simply select a bank of tweets and expect generalizable results. Twitter users often write in waves, and certain terms will grow popular and fade within a few days. In “Contextualizing Experiences: Tracing the Relationships Between People and Technologies in the Social Web,” Liza Potts and Dave Jones (2011) explain that Twitter should be understood within the context of Actor Network Theory. Actor Network Theory provides a framework that helps researchers identify and understand the wide variety of agents operating within an instance of writing production. Potts and Jones point out that within Twitter, other tweets, as well as the networks and environments of the participant, will impact the nature of the tweet produced in dramatic ways. Writing is only a part of a wide set of agents and activities within a network. A study like ours that deals with social media, then, should account for time in order to create a more diverse network and not skew the affective results. For example, something as simple as a nice, temperate day could dramatically influence the way that Twitter uses terms like global warming or climate change, and a tweet or gaff from a political figure could trend the results one way or another. Therefore, our study collected samples of tweets over a period of time, making the set of topics, trends, and uses of language around our search terms more varied. As a result, we chose to select a wide set of tweets over the course of 3 separate weeks. As previous studies concluded that political events were more likely to influence media coverage than temperature events (Cagle & Tillery, 2015), we selected 3 consecutive weeks for our data collection, and these weeks included both a political event (the introduction of the Green New Deal) and a number of temperature events (including record low temperatures in California, record high temperatures in Alaska, and a record single day snowfall of 35.9 inches in Flagstaff, AZ). These weeks took place in 2019: February 15 to 21, February 22 to 28, and March 1 to 6.

Data Collection

To collect our data, we used a Google Sheet template called TAGS, developed by Martin Hawksey (2014). This template runs automated data collection on Twitter based on relevant keywords, identifying tweets that are tagged into categories and exporting the tweets into a Google Sheet. These tweets covered 6 to 9 days before the creation of the sheet (Hawksey, n.d.), so we created a new set of tweets after each week to obtain our sample. Each data set covered instances of the phrases climate change and global warming.

TAGS collects tweets by scraping from Twitter’s Application Program Interface (API). An API serves as the intermediary between the server and the computer or mobile device trying to access the server. You don’t access Twitter’s servers when you log on, you interact with Twitter’s API. Twitter’s API won’t show every tweet, and when TAGS scrapes through Twitter’s API, it can only interact with tweets that Twitter keeps on the API. Therefore, if a tweet used a key phrase (global warming or climate change) and was included on Twitter’s API, it was added to our data set. Twitter’s API prioritizes tweets that are interacted with by other users. Our data set then consisted only of tweets that were successfully part of an interaction, meaning the tweet has been “liked,” retweeted, or replied to. Based on the way Twitter’s API functions, the more a tweet is interacted with, the more likely that tweet will be included in the API list, a factor that could inherently bias the collection (González-Bailóm et al., 2014). Thus, based on our use of TAGS and its use of API biases, we can only make conclusions about tweets that are interacted with, eliminating tweets that elicit no reaction, which excludes large ecosystems of spam bots. This limitation, however, also serves as a benefit to what we are researching. This privileging of interaction centers our analysis around rhetoric that works on Twitter, if we are defining rhetoric that works as rhetoric that interacts with others by encouraging some form of response (be it retweet, a like, etc.). As the tweets selected by TAGS successfully provoked a response, the elimination of tweets that did not provoke interaction also eliminated irrelevant items from our data set. As a result of this bias in API data sets, our data set limits our conclusions to tweets that are actively engaging with others, a limit that both helps our study have more refined results, but also limits what we are studying to tweets that garner significant interaction.

Coding Scheme and Coding

For each selected week, TAGS identified a number of tweets tagged for climate change and global warming. To get a diversity of tweets, we chose to analyze three separate instances, separated by about a week each. These tweets were collected into data sets, and the below chart indicates the number of Tweets archived each instance for each term.

After collecting this data, we organized it into a master coding sheet using Google Sheets. Because the number of tweets TAGS returned varied widely between terms and weeks, we chose to select our sample size from the full data corpus (Figure 1). As TAGS identified a total data corpus of 32,999 tweets (including both tagged terms over the full 3 weeks), we selected a sample size of 900 tweets, which allowed us to reach a confidence interval of 95% with a 4% margin of error. Selecting this number meets Boettger and Palmer’s (2010) and Thayer et al.’s (2007) recommendation to code 10% of the overall corpus in order to determine interrater reliability. Furthermore, this number helped us evenly split the sample size over the 3-week period, with 300 tweets to analyze each week.

Coding for Affect

Using our previous knowledge of Twitter, we developed preliminary coding schemes (as directed by QCA) with codes for the affect presented in each tweet (humor, anger, fear, sarcasm, and reasonability). Six coders trained on the coding scheme were then grouped into pairs, with each pair assigned a week’s worth of data. To test the codes, each pair tested these initial categories on 50 tweets per tagged term per week (300 total tweets). Each coder worked individually and then compared the coding results with their partner in order to determine the interrater reliability of their 100 tweets. After comparing their initial coding, each pair produced an interrater reliability of 90% or higher using a simple percentage agreement. The six coders and the rest of the research team then met as a group, discussed some of the problems they encountered during coding, and made revisions to the categories based on the sample, leading to the final categories seen in Figure 2.

Figure 1.

Tweet Numbers by Week.

Figure 2.

Coding Categories and Examples.

While “the most effective method of reliability measurement remains up for debate” (Thayer et al., 2007, p. 277), we used both percentage agreement and Cohen’s Kappa to determine interrater agreement. Two of the pairs coded at greater than 90% interrater reliability, and one pair coded at more than 80% interrater reliability. The mean average Cohen’s Kappa of the three groups of coders was .601. For the 10% to 20% of tweets that were not coded identically, a third outside coder broke the tie. The final coding was then converted into one master spreadsheet so results could be tallied.

When coders approached the tweets in their data set, they needed to determine the affect used in that tweet. Thus, before we could quantify our data using QCA, coders had to analyze each tweet for affect. During the initial coding stage, when each coder worked through and coded 100 tweets for each week, our research team realized that some tweets were presenting more than a single affect or emotion, and thus, we needed to determine how to approach these tweets systematically, as QCA does not allow datum to be classified into multiple categories in the way a qualitative content analysis would. While some tweets clearly expressed anger or only provided information (such as a link) and thus made it easy to classify them as reasonable, other tweets expressed multiple emotions at once or blurred the lines between emotions such as sarcasm and anger. As such, this section outlines how the research team created our code book and decided to code tweets for affect. These decisions were made after encountering a number of tweets that presented similar challenges in the initial coding phase. Once the coding categories were finalized, each coder analyzed the full 300 tweets from their weekly data set.

Humor

Humor is often used as a means to a particular end, so disentangling the goal of the tweet from the affect was critical. While a tweet may make the viewer more aware of global warming, an outcome that may be associated with reasonability, if the tweet used humor as the affective means to achieve this reasonable goal, the tweet was coded as humor. For example, this tweet uses humor in a simulated conversation:

[consoling friend after break up]

me: don’t worry there’s plenty of other fish in the sea

global warming: like hurry tho. (@continentlbkfst, 2019)

While this tweet illustrates the immediacy of global warming and shows how global warming may disrupt the normal life plans of a generation, our coders concluded that the primary affect of this tweet was humor, not another affective category. In other tweets affect was not as easy to identify. The following tweet was written about a time when author Rutger Bregman was a guest on conservative Fox News host Tucker Carlson’s show: “Rutger Bregman, who confronted Davos zillionaires for climate-change denialism, used facts to kick Tucker Carlson’s ass up & down the block. So Mr. Hurty Feelings went ballistic in the unaired segment, which Bregman just leaked. NOW THAT’S ENTERTAINMENT!” (@HeWhoLovesWords, 2019). While anger could be read in this tweet, coders decided that certain word choices, such as “Mr. Hurty Feelings,” and the use of capitalization to create emphasis ultimately produced a humorous affect. While it is possible to see humor as a tool for achieving a different goal or persuasive aim for a tweet, as we were looking for affect and not persuasive aim, humor emerged as the dominant affect for these tweets.

Sarcasm

Sarcasm is similar to humor as the two are often related, but considering sarcasm as an affective response, rather than as a rhetorical technique, helped coders differentiate between the codes. For example, the following tweet, while sarcastic, is not primarily attempting to evoke humor: “But I thought weather wasn’t climate change . . . they’re going to start counting polar vorte1es (sic) as global warming, just watch” (@KatiePavlich, 2019). This tweet’s use of irony demonstrates a clear example of sarcasm as affect. Another example of this is the following tweet articulating disappointment at the then current administration’s environmental policy:

Pulling out of a global accord to fight climate change,

Nominating an oil and gas lobbyist as Secretary of the Interior,

Putting a climate denier in charge of a commission on climate,

Welcome to the Administration’s Dirty Old Deal. (@AdamSchiff, 2019)

While this could be considered something similar to anger that targets criticism at the current administration, the tweet’s use of irony, especially the reframing of “Green New Deal” to the “Dirty Old Deal,” led our coders to understand the primary affect as sarcasm, highlighting the ironic characteristics of the administration’s decisions. Thus, irony and word play were useful identifiers for the coding of sarcasm.

Fear

Fear presented fewer difficulties for coders. Some tweets were clearly relying exclusively on fear: “Man made climate change is beyond reversable [sic]. We are living in the last days civilisation [sic]. We need to make the most of our time with loved ones” (@GrellierStephen, 2019). This tweet expresses fear of loss and death, and this fear is by far the dominating affect. In another example, this tweet expresses the user’s fears explicitly: “I’m a 15 year old who is terrified of the future under Trump and I care about whether or not I get shot in school, whether or not I have a future after climate change #ImTheRadicalLeft” (@abbysanderssss, 2019). Like many of the tweets coded into the fear category, this tweet lists multiple fears and focuses on those fears rather than on blaming someone for the user’s fear.

Anger

While fear was often as explicit and central as in the above example, differentiating between fear and anger was sometimes complicated as fear often provokes anger. While the above tweet mentions fear of a future under Trump, other tweets identified and explicitly blamed a source for their anger, focusing on the past or present. Thus if coders could identify a subject of blame or a target the tweet was disagreeing with or attempting to provoke a response from, coders placed these tweets in the anger category. Indeed, anger as affect on Twitter often focuses on a target, and this helped coders categorize tweets. For example, @FranklyPhysiol (2019) wrote,

@AdamSchiff: This week, scientists reached ‘gold standard’ confidence that climate change is man-made, and catastrophic. Also this week, Trump concocted a panel of climate skeptics to undermine the scientific consensus. That’s why Congress must lead on climate change and think big.

This tweet was coded as anger because @FranklyPhysiol’s expresses frustration with the contradiction between scientific confidence and Trump’s response. @FranklyPhysiol’s anger also has two recipients of blame, Trump and California Congressman Adam Schiff. As the tweet is directed at Schiff and ends with a call for Congress to act in response to climate change, this tweet was coded for anger. Coders determined the tweet was expressing frustration and anger at both the contradictions between what scientists and Trump’s climate panel had to say about climate change and the failure of congress to do anything in response to either climate change or Trump.

We see this blending of blame and a call to action through the affect of anger in other tweets as well, and this again helped coders differentiate between anger and fear, as tweets with a primary affect of fear were less likely to include calls to action. @seven_thenumber (2019) wrote:

this hurts my heart so much.

reduce your usage

recycle

stop using fucking plastic

reduce your intake of animal products (animal agriculture is the leading cause of climate change and water usage) be a little bit compassionate, once this world is dead we all die too.

The rhetorical usage of the word fucking in “stop using fucking plastics” illuminates @seven_thenumber’s anger. This anger seeks to spur action in the daily lives of Twitter followers while also blaming the general population for not doing enough to curve global warming and not being “compassionate enough.” While the post ends with the fear that “once this world is dead we all die too,” this tweet was coded as anger as the anger precedes the presentation of fear and dominates the tweet (and it is attempting to invoke anger more than demonstrate the fear of the user).

While the above posts both name the target of their anger (Trump/general Twitter users) and end with a call to action, other tweets rely more on anger as an affective tool to place blame, which is consistent with Gaspar et al.’s (2016) findings concerning pointed anger in tweets rather than irrational “bursts” of anger. For example, @erin_michaela98 (2019) tweeted,

@kafanabebo: Stop acting like individual consumption is the leading cause of climate change when the energy sector contributes 72% of global emissions. People switching to reusable water bottles isn’t going to make nearly as much as a dent as corps that consistently evade regulation.

Like other angry tweets, this tweet includes a call to action to “stop” and places blame on the corporations who “consistently evade regulations” rather than explicitly evoke change.

Reasonable

While users often presented multiple affects in a single tweet, reasonability often appeared without other affective displays, making it easier to code. Some reasonable tweets simply served as conduits for information rather than providing opinions or analysis. For instance, @paigeey_p (2019) tweeted, “BBC News—Wetland mud is ‘secret weapon’ against climate change,” and @grist (2019) tweeted, “One professor set out to design an easy-to-grasp, easy-to-replicate visual of our planet’s warming temperatures over the past century. The result: these stripes, each one representing the average global temperature for a single year.” Posts such as these provide information, explanations of graphs, or links to studies and do not elicit or display an emotional response. As such, they were coded in the reasonable category.

While our research team could have made different decisions about how to code these tweets, particularly with regard to how to differentiate between humor and sarcasm and anger and fear, we concluded that prioritizing affect and looking for blame and calls for action helped us code in consistent ways. In addition, while we discussed including other affect categories, such as happy, coders found that because of our topic and the negativity surrounding that topic, all the tweets we encountered could be coded into the five categories of fear, anger, humor, sarcasm, and reasonable.

Results

To analyze our quantitative data, we used a binary logistic generalized linear model. Because the data set had two variables (the use of the phrase climate change or global warming and the affect used), we chose to use a binary model. Because the dependent variable (whether the tweet used the phrase climate change or global warming) was a binary outcome, we could not perform a standard linear regression. There can be no normal distribution with a binary outcome. Instead, we chose to rely on a logistic generalized linear model. The model we chose to use is one specifically designed to compare a series of binary options (yes, no; global warming, climate change). This model examines the likelihood of each term being used with a certain affect and then compares the likelihoods.

As affect is measured categorically, not numerically, we needed a term against which each other data points could be measured. We chose to use sarcasm as the reference point because of its low frequency in the data. Accepting 0.05 as our standard for statistical significance, Figure 3 demonstrates that there is no statistically significant relationship between the use of the terms global warming and climate change as it relates to the coded affect related to the tweets associated with the corresponding term. As a result, we accept the null hypothesis that the use of the phrase global warming and climate change is not significantly impacted by the affective situation of the tweet. A significant barrier to finding a statistically significant difference was that many tweets were tagged using both the terms climate change and global warming. Consequently, we cannot reasonably infer from the data that this term use is significant. However, while inferential statistics ultimately disproved our hypothesis about affect, using descriptive statistics to examine the frequency of the use of these terms produced more useful information for the purposes of this article.

Figure 3.

Global Warming Versus Climate Change.

Figure 4 indicates that while there is not a relationship between the use of the terms global warming or climate change, our data show that, in tweets chosen by the TAGS system, 43% are using what our coders deemed a “reasonable” affect, and 36.7% were coded as “anger.” These two categories take up the lion’s share of the tweets, leaving fear, humor, and sarcasm with a cumulative 20.3%. The anger and reasonable categories dominate the conversation surrounding these terms in our data set, as demonstrated by Figure 5.

Figure 4.

Affect Frequencies.

Figure 5.

Affect Pie Chart.

The third week, beginning on March 1, had a significant increase in tweets that mentioned our keywords. This week also was coded by our coders as having an increased number of “angry” tweets. While February 15 had a total of 14 tweets coded as angry, and February 22 had 54, March 1 was coded as having 134 angry tweets. This bump in activity seems to be primarily because the Green New Deal was being discussed by Representative Alexandria Ocasio-Cortez and Senator Ed Markey, a discussion that occurred over multiple days in our data set. Many tweets were angrily directed at these two members of congress or their opponents because of their environmentally focused legislation. This jump in tweets is a reason why separating our coders into three groups, each covering a set of tweets from different weeks, was so important. Twitter’s use of particular terms rises and falls along certain trends, and an analysis of how affect works on Twitter must include multiple instances to catalogue how multiple events impact the results.

The affective category that proved to be the most present in the TAGS-related tweets, then, was the set of tweets the coders identified as “reasonable.” Thus, our research team’s hypothesis that anger would not only be the most dominant emotion in the corpus but that anger would also be tied to either global warming or climate change was not demonstrated by the inferential data.

Discussion

As TAGs favors tweets with high interaction, and using multiple hashtags increases the likelihood that tweets will be seen and interacted with, tweets were often repeated in both columns of our data sets. While we set out to determine how the terms global warming and climate change were used differently on Twitter, we are unable to make conclusions about their use with regard to affect because individual tweets often included hashtags for both terms at the end of the tweet. This frequent use of both terms in individual tweets could indicate that the terms are simply not used differently—that, as we hypothesized, Twitter users and the general public are using the terms interchangeably despite their different meanings. Yet while our results could mean that users do not understand the differences between the two terms, the frequent occurrence of both hashtags could also mean that Twitter is not the correct forum for a study on closely related terms and their use. Thus, in order to determine whether the general public conflates these terms through ignorance of their meaning, a desire to increase engagement, or semantic change that has rendered the terms indistinguishable in common use, further research is necessary.

When our research team met after the initial coding phase, coders indicated they were overwhelmed by the amount of anger that appeared in the tweets. Before we calculated affect frequency, our research team was certain that anger dominated the affect category. Interestingly, the results from the QCA challenged our coders’ responses. The results demonstrated that tweets in our data set were actually more reasonable than angry, though only slightly so. Our coders’ impressions of the data are important to note because they illustrate how researchers are impacted by the emotions they perceive in what they read. In other studies of digital communication and affect, researchers may want to consider using quantitative frequency data alongside qualitative methods as a way to account for how negative emotions such as anger can be perceived as dominant because a negative occurrence is subjectively more potent and of higher salience than its positive counterpart (Rozin & Royzman, 2001). In other words, a single angry tweet can have more resonance and impact on the reader than multiple reasonable tweets, and checking the frequency of coded data, even when not doing quantitative research, can help researchers examine and account for our affective responses to data.

In addition, tweets coded as “reasonable” tended to pass information along or retweet another person’s claim. This kind of tweet, primarily transmitting information, had a less significant impact on our coders’ perceptions of the data, as illustrated by our research meeting discussions. This indicates that it can be difficult to perceive the passing along of information as a rhetorical action connected to agency or belief. When someone retweets another person’s statement or article, we process this act as neutral or even nonrhetorical, whereas angry recapitulations of other’s ideas are seen as individual, intentional, and significant. While these reasonable tweets do not have the same emotional impact on the researcher, the users have both rhetorical and affective intentions when they make the rhetorical choice to share information.

Indeed, our results show that when discussing global warming and climate change, Twitter users are more reasonable than any other affect. That said, while Li et al. (2020) found affective studies to span beyond the positive and negative valence and thus included a neutral code, our reasonable tweets may appear neutral but the appearance of neutrality itself is a rhetorical choice to establish ethos. However, the exact intentions behind a retweet or the sharing of an informative study were largely invisible to our coders. These results challenge conceptions about the ways discussions surrounding global warming and climate change are framed. Unlike the findings of Foust and Murphy’s (2009) study on media outlets, our results indicate that the frames they identified do not carry over to Twitter. Furthermore, our assumption about the divisive nature of internet communication, particularly via Twitter, was challenged by our results. This finding indicates Twitter users in this study were more often reasonable when presenting their arguments than any other affect, and consequently, Twitter users, and digital communications more generally, may be more reasonable than researchers expect. Therefore, these findings add to those presented by Cagle and Herndl (2019), and we echo their call for increased nuance regarding discussions of climate change on social media.

In previous studies on emotion and digital communication, researchers often focused on specific events, and our results indicate that focusing on single events may skew results. In particular, in Week 3, the discussion focused on the Green New Deal, and this dramatically increased the frequency of angry tweets. We chose 3 different weeks for our study in order to ensure that our findings were about discussions of global warming and climate change on Twitter more generally and not focused on a specific hashtag or event being discussed on Twitter (such as a snowstorm or political act). Had we only studied a single day or hashtag during Week 3, then anger would have been the most prominent affect coded, misrepresenting the broader affective landscape surrounding global warming and climate change on Twitter. Thus, while previous studies have focused on the responses to significant events on Twitter, our 3-week study indicates that such events and the affect they produce may not be representative of the affect present in sustained conversations on Twitter. As such, while our study indicates that Twitter conversations may be more reasonable than expected, a more longitudinal study than ours is necessary in order to understand the ways events intervene in these sustained Twitter discussions. Such a longitudinal study could better investigate the ways terms and affect are used on Twitter. That said, by examining, testing, and adding to the ideas and approaches described in this study, researchers across multiple disciplines can contribute to the expanding research on affective studies, particularly in digital contexts.

Footnotes

Acknowledgments

The authors would like to thank Molly Kessler for feedback and suggestions on early drafts of this article, as well as the members of our Spring 2019 Research Methods course, who helped them code and develop this research project: Alyssa Radtke, Kaitlin Graves, Allison Preslar, Jasmine Ross, and Rachael Summers-Thompson.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article

ORCID iD

Shanna Cameron

Author Biographies

Shanna Cameron is a doctoral candidate and a first-year writing instructor at the University of Memphis. Her research interests focus on technical and professional communication, digital rhetorics, and the rhetoric of health and medicine.

Alexandra Russell is a PhD student at the University of Memphis and a first-year writing instructor. Her research interests include digital feminist and activist rhetorics.

Luke Brake is a doctoral candidate and a first-year writing instructor at the University of Memphis. His research focuses on religious rhetoric.

Katherine Fredlund is an associate professor and Director of the First-Year Writing Program at the University of Memphis. Her research is forthcoming in CCC and has appeared in College English, Rhetoric Review, Peitho, and elsewhere.

Angela Morris is a doctoral candidate at the University of Memphis where she teaches first-year writing, literature, and technical communication. Her research has previously appeared in Spark: A 4C4Equality Journal and The Peer Review.

References

@abbysanderssss. (2019, February 20). I’m a 15 year old who is terrified of the future under Trump and I [Tweet]. Twitter.

@AdamSchiff. (2019, February 28). Pulling out of a global accord to fight climate change, nominating an oil and gas [Tweet]. Twitter.

@Adrianaaaana. (2019, February 28). They ain’t mood swings, it’s called climate change [Tweet]. Twitter.

@AshishBhalla. (2019, February 28). Most firms remain blind to the physical threats of severe climate change. They should start [Tweet]. Twitter.

Bloomfield

E. F.

Tillery

(2019). The circulation of climate change denial online: Rhetorical and networking strategies on Facebook. Environmental Communication, 13(1), 23–34.

Boettger

R. K.

Palmer

L. A.

(2010). Quantitative content analysis: Its use in technical communication. IEEE Transactions on Professional Communication, 53(4), 346–357.

Bollen

Mao

Pepe

(2011). Modeling public mood and emotion: Twitter sentiment and socio-economic phenomena. Computers & Society [conference presentation], Proceedings of the Fifth International AAAI Conference on Weblogs and Social Media.

Bruns

Stieglitz

(2013). Towards more systematic Twitter analysis: Metrics for tweeting activities. International Journal of Social Research Methodology, 16, 91–108.

Cagle

L. E.

Herndl

(2019). Shades of denialism: Discovering possibilities for a more nuanced deliberation about climate change in online discussion forums. Communication Design Quarterly, 7(1), 22–39.

10.

Cagle

L. E.

Tillery

(2015). Climate change research across disciplines: The value and uses of multidisciplinary research reviews for technical communication. Technical Communication Quarterly, 24(2), 147–163.

11.

Ceccarelli

(2011). Manufactured scientific controversy: Science, rhetoric, and public debate. Rhetoric & Public Affairs, 14(2), 195–228.

12.

Chung

Zeng

(2020). Dissecting emotion and user influence in social media communities: An interaction modeling approach. Information & Management, 57(1), 103108. https://doi.org/10.1016/j.im.2018.09.008.

13.

@continentlbkfst. (2019, February 19). [consoling friend after break up] me: Don’t worry there’s plenty of other fish in the sea global warming: Like hurry [Tweet]. Twitter. https://twitter.com/continentlbkfst/status/1097915715087204352

14.

Lim

Kim

Choi

(2016). Analyzing emotions in Twitter during a crisis: A case study of the 2015 Middle East respiratory Syndrome outbreak in Korea [conference presentation]. In 2016 international conference on big data and smart computing (BigComp), pp. 415–418.

15.

@erin_michaela98. (2019, February 26). @kafanabebo Stop acting like individual consumption is the leading cause of climate change [Tweet]. Twitter.

16.

Fersini

Messina

Pozzi

(2016). Expressive signals in social media languages to improve polarity detection. Information Processing & Management, 52(1), 20–35.

17.

Foust

C. R.

Murphy

W. O.

(2009). Revealing and reframing apocalyptic tragedy in global warming discourse. Environmental Communication, 3, 151–167.

18.

@FranklyPhysiol. (2019, February 25). @AdamSchiff: This week, scientists reached “gold standard” confidence that climate change is man-made, and catastrophic [Tweet]. Twitter.

19.

Fung

Duke

Finch

Snook

Tseng

Hernandez

Tse

(2016). Ebola virus disease and social media: A systematic review. American Journal of Infection Control. https://www-sciencedirect-com-443.web.bisu.edu.cn/science/article/pii/S0196655316304977

20.

Gaspar

Pedro

Panagiotopoulos

Seibt

(2016). Beyond positive or negative: Qualitative sentiment analysis of social media reactions to unexpected stressful events. Computers in Human Behavior, 56, 179–191.

21.

@ginadivittorio (2019, February 24). Do rich people know climate change will end their ski trips? Have we emphasized that [Tweet]. Twitter.

22.

González-Bailóm

Wang

Rivero

Borge-Holthoefer

Moreno

(2014). Accessing the bias in samples of large online networks. Social Networks, 38, 16–27.

23.

Goodwin

(2019, July 07). How to deal with all the negativity on twitter. https://www.searchenginejournal.com/twitter-negativity/292245/#close

24.

@GrellierStephen. (2019, February 28). @AOC Man made climate change is beyond reversible. We are living in the last days civilisation [Tweet]. Twitter.

25.

@grist. (2019, March 6). One professor set out to design an easy-to-grasp, easy-to-replicate visual [Tweet]. Twitter.

26.

Hasan

Rundensteiner

Agu

(2014). EMOTEX: Detecting emotions in twitter messages [conference presentation]. ASE BIGDATA/SOCIALCOM/CYBERSECURITY Conference, Stanford University.

27.

Hawksey

(2014). About. https://tags.hawksey.info/about/

28.

Hawksey

(n.d). FAQ. https://tags.hawksey.info/help/faq/

29.

@HeWhoLovesWords. (2019, February 20). Rutger Bregman, who confronted Davos zillionaires for climate-change denialism, used facts to kick Tucker Carlson's ass up & down the [Tweet]. Twitter.

30.

Jin

(2010). Making sense sensibly in crisis communication: How publics’ crisis appraisals influence their negative emotions, coping strategy preferences, and crisis response acceptance. Communication Research, 37(4), 522–552.

31.

@KatiePavlich. (2019, February 28). But I thought weather wasn’t climate change . . . they’re going to start counting polar vortexes as [Tweet]. Twitter.

32.

Knudsen

Stage

(2016). Affective methodologies: Developing cultural research strategies for the study of affect. Palgrave MacMillan.

33.

Koteyko

Nerlich

Hellsten

(2015). Climate change communication and the internet: Challenges and opportunities for research. Environmental Communication, 9(2), 149–152.

34.

Wei

Dang

Cao

Liu

(2020). Tracking and analyzing public emotion evolutions during COVID-19: A case study from the event-driven perspective on microblogs. International Journal of Environmental Research and Public Health, 18, 17.

35.

(2015). A global optimization approach to multi-polarity sentiment analysis. PLoS One, 4, 10.

36.

Liu

B. F.

Austin

Jin

(2011). How publics respond to crisis communication strategies: The interplay of information form and source. Public Relations Review, 37(4), 345–353.

37.

Majdik

Z. P.

(2019). A computational approach to assessing rhetorical effectiveness: Agentic framing of climate change in the congressional record, 1994–2016. Technical Communication Quarterly, 28(3), 207–222.

38.

Matthews

(2015). Why are people skeptical about climate change? Some insights from blog comments. Environmental Communication, 9(2), 153–168.

39.

Meloncon

St. Amant

(2019). Empirical research in technical and professional communication: A 5-year examination of research methods and a call for research sustainability. Journal of Technical Writing and Communication, 49(2), 128–155.

40.

@michaeldaleyMP. (2019, February 27). Labor believes climate change is real. A Daley Labor Government will introduce NSW’s first renewable [Tweet]. Twitter.

41.

Mohammad, S.M. & Kiritchenko, S. (2015). Using hashtags to capture fine emotion categories from tweets, Computer Intelligent 31(2); 301–326.

42.

NASA. (2008). Global climate change: Evidence. http://climate.nasa.gov/evidence/

43.

@paigeey_p. (2019, March 6). BBC News—Wetland mud is 'secret weapon' against climate change [Tweet]. Twitter.

44.

Perez

(2018, October 30). Twitter’s doubling of character count from 140 to 280 had little impact on length of tweets. Tech Crunch. https://techcrunch.com/2018/10/30/twitters-doubling-of-character-count-from-140-to-280-had-little-impact-on-length-of-tweets/

45.

Pflugfelder

E. H.

(2013). Big data, big questions. Communication Design Quarterly Review, 1, 18–21.

46.

Potts

Jones

(2011). Contextualizing experiences: Tracing the relationships between people and technologies in the social web. Journal of Business and Technical Communication, 25(3), 338–358.

47.

@Redbaiternz. (2019, February 28). Notice @ScottMorrisonMP @LiberalAus climate change policy was met with scorn from both Labor & Conservative voters [Tweet]. Twitter.

48.

Rozin

Royzman

E. B.

(2001). Negativity bias, negativity dominance, and contagion. Personality and Social Psychology Review, 5(4), 296–320.

49.

Saif

Fernandez

Alani

(2016). Adapting sentiment lexicons using contextual semantics for sentiment analysis of twitter [conference paper]. Lecture Notes in Computer Science, pp. 54–63.

50.

@seven_thenumber. (2019, February 24). This hurts my heart so much [Tweet]. Twitter.

51.

Shah

Cappella

Neuman

W. R.

(2015). Big data, digital media, and computational social science. The ANNALS of the American Academy of Political and Social Science, 659, 6–13.

52.

St. Amant

Graham

S. S.

(2019). Research that resonates: A perspective on durable and portable approaches to scholarship in technical communication and rhetoric of science. Technical Communication Quarterly, 28(2), 99–111.

53.

Thayer

Evans

McBride

Queen

Spyridakis

(2007). Content analysis as a best practice in technical communication research. Journal of Technical Writing and Communication, 37(3), 267–269.

54.

Tutt

(2009). Frames in reports and in reporting: How framing affects global warming information in the Intergovernmental Panel on Climate Change’s “summary for policymakers” and in documents written about the report. Journal of Technical Writing and Communication, 39(1), 43–55.

55.

Waterloo

S. F.

Baumgartner

S. E.

Peter

Valkenburg

P. M.

(2018). Norms of online expressions of emotion: Comparing Facebook, Twitter, Instagram, and WhatsApp. New Media & Society, 20(5), 1813–1831. doi:10.1177/1461444817707349

56.

Zamith

Lewis

S. C.

(2015). Content analysis and the algorithmic coder: What computational social science means for traditional modes of media analysis. Annals of the American Academy of Political and Social Science, 659(1), 307–318.