Abstract
Researchers are increasingly confronting the need to examine the impacts of social media on democratic discourse. Analyzing 55,560 tweets from the official Twitter accounts of the Democratic and Republican Parties in the United States, the authors examine approaches used by political parties to encourage sharing of their content within the contemporary political divide. They show that tweets sent by the Republican Party are more likely to be predominant in the language of assessment and that tweets predominant in the language of assessment lead to more retweets. Further, this effect is reduced as political parties gain control of successive branches of government because successive increases in political power create fewer impediments to the implementation of a party’s political agenda. As impediments to action are reduced, so is regulatory fit for assessment-oriented language. Goal pursuit language shared on Twitter therefore reveals distinct approaches to obtaining and wielding power across the U.S. political system and constitutes an important tool for public policy makers to use in successfully conducting policy debates.
Political parties and their candidates need their content to be shared during campaigns and while in government. This is not only because a party’s campaign will influence whether they are likely to win an election (Farrell and Schmitt-Beck 2002; Holbrook 1996) but also because their electoral positioning and policy framing will determine the mandate on which they have to govern when in power (Shamir, Shamir, and Sheafer 2008). Moreover, unlike individual candidates, political parties must ensure that their brands outlast the short-term political impacts of elections, protest movements, and the careers of specific politicians if they are to have long-term influence over public policy. This necessitates ongoing communication between parties and their bases to maintain organizational reputations.
With traditional communication channels giving way to social media, researchers can leverage these communication efforts as data to develop an understanding of how political parties encourage their constituents to navigate power structures in the political system of the United States. In doing so, insight can be developed into the distinct differences in language that parties use to promote content within the silos of the political divide, which increasingly caters to the divergent perspectives of their supporters (Cichocka et al. 2016; Gentzkow, Shapiro, and Taddy 2019; Jones et al. 2017; Smith 2019).
When encouraging their followers to pursue electoral and public policy goals, parties use language that innately implies the goal pursuit strategies (i.e., language that implies how goals are pursued) that their followers should use to achieve them. Take, for example, posts written in the language of action (Kanze, Conley, and Higgins 2019), such as a tweet by President Barack Obama encouraging his supporters to be “Fired up! Ready to go!” during his 2012 campaign (Obama 2012). While not outlining a specific call to action for its audience, the content was designed to energize followers to become active participants in his campaign movement, thereby supporting his goal of reelection. Another common form of content is written in the language of assessment (Kanze, Conley, and Higgins 2019), such as a post written by Obama after he had left office. Here, he encouraged followers to engage in the pursuit of identifying solutions to key policy problems by deliberating on “thought-provoking” (Obama 2019) material about their causes. Subtly using goal pursuit language in social media content is a highly influential strategy for spreading messages within and between the silos of the political divide. This is because the strategies consumers use to pursue goals can be just as important a component for success as the goals themselves (Motyka et al. 2014). But, despite its importance, there is no research to indicate how parties use goal pursuit language to increase the viral strength of social media content within and across the political divide.
This paucity of research on why political posts are shared is a problem for practitioners. Content disseminated by political parties must compete for attention in an environment where voters can simply choose to not consume the information they do not like (Maarek and Wolfsfeld 2005). Moreover, there are indicators that voters are generally dissatisfied with existing political communication efforts, as 49% of Americans report feeling worn out by the number of political posts in their feeds (Anderson and Quinn 2019). In such a congested environment, political parties need to ensure that their content is well-crafted and well-targeted to be successful. However, existing research on the determinants of content virality does not offer insight beyond existing practice.
For example, empirical work shows that content is more likely to go viral when it is highly emotionally arousing (Berger and Milkman 2012) and occurs between communicators who use similar language and have close ties to each other (Herhausen et al. 2019). Yet each of these factors are already likely to be present in the content shared between political parties and their constituents. Many of the issues discussed by political parties (e.g., abortion, immigration) are already likely to be highly emotionally arousing for the audience. Political parties and their constituents are also already likely to use similar language—for example, the word “snowflake” being used among conservative representatives and voters. Finally, voters are likely to have long-term ties to a political party (Dalton 2015), with parties already communicating regularly with their constituents. Therefore, given that political communication already bears the hallmarks of successful viral content (Berger and Milkman 2012; Herhausen et al. 2019), the field cannot currently offer insights to political marketers that will enable them to streamline their content more successfully. Moreover, with political communication already crafted in ways that are likely to go viral, why does some political social media content generate more sharing than other content, and how can an examination of such content inform existing research on virality?
We address this knowledge gap by introducing goal pursuit language, referring to language that reflects distinct preferences for goal pursuit strategies identified by regulatory mode theory, as an important and overlooked factor in understanding the sharing of political content on social media. Regulatory mode theory holds that individuals have distinct preferences for goal pursuit strategies, which involve either assessing courses of action (“assessment”) or initiating action (“locomotion”) toward a goal (Kruglanski et al. 2000). Aligning language to reflect either of these predispositions creates a “regulatory fit” that resonates with individuals. Regulatory fit leads to a sense of “feeling right,” which manifests in a range of reactions, such as increased monetary value perceptions (Cesario, Higgins, and Scholer 2008; Conley and Higgins 2018; Higgins 2005) and intensified judgments of morality (Camacho, Higgins, and Luger 2003; Cornwell, Jago, and Higgins 2019). It also results in favorable responses to content (Pierro et al. 2013).
We show that people share political content that is not only reflective of their goals for engaging in word of mouth (WOM; Berger 2014) but also based on the strategies for goal pursuit embedded in the language used to craft WOM content. That is, people share content not just because it fits with the goals they have for engaging in sharing, such as to establish common ground with others (Berger 2014), but also because it fits their preferences for how to engage in goal pursuit more generally. The latter is achieved through either assessment (the evaluation of options and information) or locomotion (the initiation and continuation of action) (Kruglanski et al. 2000). For example, tweets crafted in language that invites the reader to scrutinize a policy would receive more retweets from individuals with high assessment motivations than tweets written in language encouraging followers to take action in support of that policy. In contrast, tweets written in language that encourages action would receive more shares among Twitter users with high motivations toward locomotion, compared with tweets crafted in language inspiring scrutinization of policy. Moreover, we assert that preferences for locomotion language among liberals and assessment language among conservatives explains why some political communication content generates more sharing online. In examining this, we identify whether posts crafted in the language of assessment (vs. the language of locomotion) lead to higher retweets.
Building on regulatory fit theory (Avnet and Higgins 2003; Cesario, Higgins, and Scholer 2008; Conley and Higgins 2018; Higgins 2005; Higgins and Scholer 2009; Motyka et al. 2014), we further show that the sharing of assessment or locomotion language depends on power structures inherent to the system of checks and balances in the U.S. federal government. That is, power possessed by political parties within this system can shape how their audiences engage with goal pursuit language used in their communications, thus affecting whether content is shared. For example, content written in the language of assessment, such as that encouraging readers to question a policy, is less likely to invoke motivations for assessment when posted by parties that have control over successive branches of government. This means that assessment language will be less likely to be shared when it is used by a party in power in the executive branch of government that then also gains power over additional branches of the legislative. This is because the power possessed by a party that controls successive branches of government is better suited to the language of action, as it is commensurate with its role of enacting its political agenda. We further demonstrate that fit between political power and goal pursuit language affects the likelihood that content will be shared.
As we discuss subsequently, in making these contributions, we consider regulatory mode in order to expand existing literature on WOM. We do so by first extending prior work on how sharing is impacted by consumers' goals for engaging in WOM (e.g., persuading others) (Berger 2014). We also extend research on how shared language use and closeness between senders and receivers of content can impact sharing (Herhausen et al. 2019). We do this by showing that content posted by political parties is more likely to be retweeted when parties and constituents share preferences for goal pursuit strategies and use similar goal pursuit language in their communication. Finally, we introduce the concept of regulatory fit to the field of WOM to provide a theory-driven explanation for why contextual factors—especially concentrations of power—can impact sharing (Berger et al. 2020).
To empirically examine these issues, we use Twitter as our focal social media platform for several reasons. With approximately 22% of Americans using the site (Wojcik and Hughes 2019), Twitter has been demonstrated to help candidates win elections (Bright et al. 2019; Kruikemeier 2014; LaMarre and Suzuki-Lambrecht 2013), for example, by helping increase vote share (Bright et al. 2019). The platform has also been shown to exert outsized influence over electoral politics, as content posted on Twitter can shape political coverage on other platforms, such as in the traditional news media (Conway, Kenski, and Wang 2015; Kreiss 2016; Parmelee 2014), and can also affect opinion leadership among the highly politically interested (Borge and Esteve Del Valle 2017; Park 2013). It moreover presents methodological advantages over using comparable sites such as Facebook (Murphy 2017). Twitter data can be viewed by wide audiences as well as collected and analyzed on a large scale, which is not possible on most other social media sites (Murphy 2017).
Conceptual Background
Research on WOM has identified several factors driving sharing of social media content (Berger 2014; Herhausen et al. 2019). Whether consciously or unconsciously motivated, individuals are believed to engage in online WOM to facilitate goals such as impression management, information acquisition, social bonding, emotion regulation, and persuasion of others (Berger 2014). However, no research has examined how preferred goal pursuit strategies among senders and receivers of content can impact sharing.
In addition, empirical work has more recently explored how relationships between sender and receiver, as well as their shared language use, can drive WOM (Berger 2014; Herhausen et al. 2019). For example, research has recently demonstrated that linguistic style matching is important in fostering greater virality in online firestorms (Berger 2014; Herhausen et al. 2019), as is the tie strength between sender and receiver of WOM content (Herhausen et al. 2019). However, factors that might foster or reflect the closeness and the linguistic similarity of groups, such as assessment and locomotion orientations, have not been investigated by the field. We extend previous work examining how the use of preferred language among communities, as well as ties between senders and receivers, can drive WOM by introducing goal pursuit language as a latent factor that underlies the sharing of WOM content from political parties. Finally, although it is well-understood that contextual factors (e.g., character limits, intended audience) can impact sharing (Berger et al. 2020), the factors that have been investigated by the field have been disparate and have often lacked a theory-driven basis for why they can impact sharing (Berger et al. 2020). We therefore introduce the concept of regulatory fit to the field of WOM to provide theory-driven insight into how the context in which content is sent and received can impact whether it is shared.
Regulatory Mode
Regulatory mode theory proposes that individuals have distinct preferences for the strategies that they use to pursue their goals (Kruglanski et al. 2000). One of these preferences, called assessment, concerns the comparative aspect of goal pursuit, based on choosing the right course of action (Kruglanski et al. 2000). Often conceptualized as a motivation to engage in critical evaluation, assessment orientation is summed up as a desire to “do the right thing” (Kruglanski et al. 2000). Rather than being a normative judgment, this motivation delays decisions in favor of considering other options toward goal attainment. It reflects an individual’s reservation toward progress in exchange for greater certainty in a chosen course of action (Kruglanski et al. 2000). For example, consumers who have high levels of assessment orientation are motivated to pursue goals by comparing a large assortment of different options to ensure that they make the best decision about which one they choose (Avnet and Higgins 2003; Kruglanski et al. 2000; Mathmann et al. 2017). Assessment orientation, moreover, orients an individual toward a focus on the past, as when experiencing feelings of nostalgia (Pierro et al. 2013) or fixating on past behavior (Kruglanski, Chernikova, and Jasko 2018; Pierro et al. 2008, 2018; Webb et al. 2017), thus helping assessors scrutinize options on the path to eventual action (Kruglanski et al. 2010).
The second motivation, called locomotion, relates to a preference for pursuing goals through the initiation of action toward them (Kruglanski et al. 2000). Encapsulated by the Nike slogan “Just Do It,” locomotion orientation is an impetus to initiate and sustain uninterrupted movement toward a goal (Kruglanski et al. 2000). This desire for movement is associated with an orientation toward the future (Kruglanski, Pierro, and Higgins 2016) that often manifests in positive conceptualizations of change. Locomotors generally demonstrate a positive evaluation of change (Kruglanski, Pierro, and Higgins 2016), an increased commitment to change (Scholer and Higgins 2012), and a heightened ability to cope with it (Kruglanski et al. 2007).
Further, if a consumer has higher levels of one orientation than the other, they are said to be predominant in that motivation (Kruglanski et al. 2000). For example, after an instance of interpersonal conflict, individuals predominant in assessment have been found to “dig deeper” into the conflict, thus keeping them from “moving on” (Webb et al. 2017), which would instead be facilitated by a strong predominant locomotion orientation. Within a political context, where communication material is crafted to better enable the strategic pursuit of power, it is important to consider how the use of goal pursuit language can be used to shape the success of political communication.
Regulatory Mode and U.S. Politics
The typical left versus right political divide (Jost 2017), exemplified in the United States by the Democratic and Republican Parties (Pew Research Center 2018), may reflect more than just ideology. That is, the divide potentially reflects the preferences of each party’s constituents toward different goal pursuit strategies. Progressive ideology, for example, involves a pursuit of social equality that fundamentally necessitates an orientation toward challenging tradition, embracing changes to existing social hierarchies, and initiating action toward generating change (Jost 2017; Jost, Federico, and Napier 2009). For liberals, this action is typically conceptualized as involving the use of big-government initiatives to act on social and economic issues (Jost 2017; Jost, Federico, and Napier 2009). In contrast, the conservative ideals of upholding tradition and resisting changes to existing hierarchies (Jost 2017; Jost, Federico, and Napier 2009) would necessitate a high level of scrutiny and critical evaluation of any potential changes to existing social structures. Conservatives, moreover, are typically skeptical of political action, such as that proposed by proponents of liberal ideology, which involve acts of government intervention into markets and the lives of individuals. We therefore conjecture that such motivational distinctions shape preferences for different goal pursuit strategies among constituents of the Democratic and Republican Parties. For liberals, their comfort in movement away from the past, acceptance of change to the social hierarchy, and initiation of action to produce that change, would necessitate a strong locomotion orientation. Conservatives, however, have a greater focus on the past (Robinson et al. 2015), which cultivates scrutiny about courses of action designed to foster change to traditional social structures (Jost 2017; Jost, Federico, and Napier 2009). This focus on the past and preference for scrutinizing change would necessitate an orientation toward assessment.
In the age of targeted social media, individuals can follow messaging initiated by their own party in isolation from opposing views (Barbera et al. 2015). Preferences for locomotion or assessment language reflecting the distribution of regulatory mode orientations across the political divide can therefore be revealed through citizens’ sharing behaviors. This is because, while political parties may disseminate content crafted with either kind of goal pursuit language, constituents are more likely to use and share the language that resonates with their preferred goal pursuit orientations. Therefore, although individuals crafting or sharing Twitter content are unlikely to be aware of their regulatory mode orientations, the distribution of preferred goal pursuit strategies across the political divide will have consequences for the goal pursuit language used in political communication.
The language of action, which emphasizes the initiation of movement, implies a strategy of moving forward from the past that is less impeded by deliberation. This language of action would therefore be highly reflective of locomotion orientation and would likely resonate with liberals. In contrast, the language of assessment, emphasizing skepticism and deliberation, implies a strategy of impeding action by choosing instead to spend time on evaluation of alternatives, and would therefore likely resonate with conservatives. Consequently, social media content written in the language of action, which signals locomotion-oriented goal pursuit, will align with the historically progressive stance of the Democratic Party (Blevins 2006; Sterling, Jost, and Hardin 2019) and its liberal constituents. On the other hand, social media content written in assessment-oriented language will align with the conservative ideals of the Republican Party (Rozell and Barbieri 2009) and its constituents. Building on the notion that those who favor each type of goal pursuit also share social media messages that align with their party affiliation (Babera et al. 2015), we hypothesize that
1
Situational Influences
Regulatory fit is produced when information in the environment, such as a political message, matches a consumer’s preferred goal pursuit orientation, leading to increased engagement (Motyka et al. 2014). Regulatory fit can intensify the consumer’s evaluation of their goal and goal pursuit process, thereby shaping value perceptions they have for the target of their evaluations. In turn, heightened value perceptions lead to greater engagement in environments that match a consumer’s preferred goal pursuit orientation. Accordingly, regulatory fit is not just based on chronic predispositions but may be created situationally based on an aspect of the consumer’s environment. For example, regulatory fit effects can be induced by watching another person perform a task (Motyka et al. 2014) or through exposure to language in an advertisement (Mathmann et al. 2017). Avnet and Higgins (2003) induced regulatory fit via a priming protocol asking consumers to read vignettes containing assessment or locomotion language. Mathmann et al. (2017) also used exposure to advertisements featuring assessment language to create a fit effect in consumers, who went on to value choices from larger assortments. Situational inductions of regulatory mode can therefore be induced by environmental cues (Kanze, Conley, and Higgins 2019). For example, regulatory mode orientation expressed through a company’s mission statement can create an organizationally induced regulatory mode orientation among its employees (Kanze, Conley, and Higgins 2019), affecting how they support or avoid instances of company discrimination.
Considering that regulatory fit can be situational and environmental, it is important to consider how the political context experienced by constituents might affect their propensity to share content sent from political parties. Power in the United States is distributed between the executive, legislative, and judicial branches of government, with the different branches acting as a check on the power of the others (Watts 2010). For a political party, obtaining the power to advance its political agenda therefore ideally involves gaining control of more than one branch of government. This further means that the agenda of the executive branch can be significantly constrained by strong opposition in the legislative branch, and that controlling both the executive and legislative branches of government will significantly increase a party’s political power. Given that political content is shared in an environment shaped by this system of checks and balances, we contend not only that the use of assessment-oriented language is therefore a prerogative of conservative ideologies, as described in the lead-up to H1, but also that its use reflects a deeper symptom of impediments to political action in the U.S. system.
Within this environment, a political party can increase its ability to enact its agenda when it controls successive branches of government. For instance, it would be easier for a sitting Republican president to enact their legislative agenda if the House of Representatives were held by the Republicans (vs. Democrats). This would be especially true if the Republicans held both the House of Representatives and the Senate in addition to the White House. Therefore, as a political party gains successive control of multiple branches of government, it has a greater ability to enact policy directly. This would better allow a party to pursue the goal of enacting its legislative agenda in a manner consistent with a locomotion orientation, focused on initiating and sustaining action toward a goal—in this case, the party’s political agenda—rather than questioning or delaying action, which would be consistent with an assessment orientation.
Consequently, we contend that the use of assessment-oriented language on social media is moderated by the strength of political opposition that a party faces once it has executive power. This is because political parties with control of the executive branch will have less need to deploy assessment-oriented language to prosecute their political agenda as they gain more power. For example, while a sitting Republican president would refrain from using assessment language in political communication in the White House, if the Republican Party also gained control of additional branches of government, such as winning the House of Representatives, the need to deploy assessment language would be greatly reduced. This is because a party is unlikely to encourage political communication that impedes its own legislative agenda.
These differing concentrations of power would also impact what supporters would be motivated to share online. Constituents of either party would be less motivated to share content featuring assessment language that would impede political action when their preferred party has a greater accumulation of power. This is because constituents of a party would also have a reduced motivation to share content that questions and impedes the agenda of a party that they support. For example, with Republicans in power, there would be a greater chance that the legislative agenda being pursued would be commensurate with Republican political goals (rather than Democratic political goals that Republicans would be motivated to scrutinize or delay). 2 Motivation to share content featuring assessment language will therefore be reduced among constituents as the party they support gains more power.
Thus, we conjecture that when a party has control over the executive branch of government, an increase in political power experienced by the party lowers impediments to its political agenda and reduces the challenge posed by their opposition. This will diminish regulatory fit in the political environment with assessment language used by the party and, in turn, will decrease sharing of content predominant in assessment-oriented language among its audience, leading to fewer retweets for that content. We therefore propose that an increase in political power reduces the positive relationship between the use of assessment-predominant language in political tweets and the increased likelihood of political tweets being retweeted. Thus,
Importantly, we posit that H2 will hold regardless of whether the legislative branch of government controlled by the party is measured as the House of Representatives, the Senate, or both houses of Congress combined.
Summarizing the relations suggested in H1 and H2, we arrive at the conceptual model illustrated in Figure 1. In the model, constituents sharing the Republican (vs. Democratic) Party tweets favor assessment-oriented language because it shifts attention away from the opposing party’s competing orientation toward action. We assert that this resonates with conservative constituents seeking to oppose a progressive political agenda. This role of assessment-oriented language further becomes evident when we consider the level of increased political power possessed by a party sharing assessment-oriented language. With increased political power, the need for considering the opposition is eased, lessening regulatory fit for assessment-oriented language and reducing its effect in political WOM.
Conceptual model.
Methodology
Before presenting our major analyses, we first present a pretest to assess the historical alignment of regulatory mode language among the two major political parties. To do this, we collected the inaugural addresses of all presidents from Franklin Roosevelt in 1933 to Donald Trump in 2016, as the former historical moment is considered the point when the two parties established the constituencies they largely still represent today (Chambers and Burnham 1967; Peters n.d.). Following procedures for operationalizing variables from text, which we detail in our main study, we used the regulatory mode dictionary (Kanze, Conley, and Higgins 2019) to measure the regulatory mode language of each inaugural address. We then conducted a t-test to determine differences in regulatory mode language between parties. In support of our historical conceptualizations of the two major parties, Republican presidents (M = 9.9, SD = 6.95) had a higher mean proportion of assessment-predominant language than Democratic presidents (M = 4.25, SD = 5.68), with the difference between the two reaching significance (t(17) = 2.061, p < .05).
Main Study
Next, we describe our research method for our main study, which involved data collection, text preprocessing, operationalization of variables, summary statistics and correlations, data analysis, and corrections for endogeneity. We also report an additional study to replicate our findings.
Data Collection
The first step involved collecting the data. To do this, we collected all tweets originating from the official Twitter handles of the Democratic and Republican parties (@TheDemocrats and @GOP). We downloaded all tweets disseminated by both parties from the time of their Twitter handle creation (December 2007 for @GOP and April 2008 for @TheDemocrats) until the time of our data collection in September 2019. This resulted in a rich corpus of 55,560 tweets, combining data from both handles for the 12-year period (December 2007 to September 2019), which we use in our main study. Separately, we also collected tweets from the two handles for the period between October 2019 and June 2020, which we use in our replication study. Although Twitter data have limitations, Twitter has been shown to be a rich source of data for marketing research (Berger et al. 2020; Murphy 2017). Among all social media platforms, the variables available in a Twitter data set are far more amenable to answering questions of wider research interest. In addition, Twitter’s data access policies through its application programming interface (API) are more suited for open-platform data sources.
As part of this method, we used web crawling techniques to download data from websites (Berger et al. 2020). While popular statistical software such as SAS has such functionalities available, more complex data requirements require customized coding in open-source languages. To download the tweets required for our research, we wrote specialized Python code to retrieve the Twitter data. We first searched for all the tweets that originated from the handles of the two parties using the search functionality in the Twitter API and recorded the unique identifier for each tweet that the search results returned. We then used the tweepy package in Python (Roesslein 2021) to extract the following data from the more than one hundred variables that Twitter API returns for a tweet: Created_at: Provides the date and time the tweet was created. Retweet_count: The number of times the tweet was retweeted (or shared). Text: The actual text of the tweet.
Text Preprocessing
In accordance with best practice methods for this technique (Berger et al. 2020), we preprocessed the data. Text data extracted from Twitter using the aforementioned method contains alphanumeric characters, special symbols (e.g., hashtags), URLs, and ASCII characters. These need to be removed from the tweets to achieve high-quality results. This preprocessing step uses natural language processing procedures that clean and transform the data into machine-friendly formats (Bird, Klein, and Loper 2009; Manning and Schütze 1999). The text then needs to be tokenized (fragmented into words or phrases), lemmatized (retaining the base form of the token), and tagged (Pustejovsky and Stubbs 2012). In addition, certain common words, also known as stop words (Berger et al. 2020)—such as “the,” “in,” “and,” and “with”—need to be filtered out. We filtered stop words because they would not provide valuable insight for our research. We used a combination of code in Python and R to perform these tokenization and tagging steps. We performed the lemmatization step at a later stage after filtering the tokens, as described in the next subsection. These steps provided us with a matrix with a frequency count of tokens in each of the tweets, commonly referred to as the document-term matrix (DTM). At this stage, we were left with 7,213 unique tokens across the 55,560 tweets.
Operationalizing Variables from Text
We operationalized the regulatory mode variables using the cleaned DTM and the regulatory mode dictionary provided by Kanze, Conley, and Higgins (2019). These scholars proposed and validated a regulatory mode dictionary, which they used to perform a linguistic analysis of organization mission statements to determine the degree of locomotion versus assessment language used in the statements. One advantage of this dictionary is that it provides the root of the token (e.g., “urg,” which accounts for all forms of the word including, but not limited to, urgency, urgent, urgently). We then used the stemDocument function in the tm package in R (Feinerer 2019) to retain only the regulatory mode tokens. This filtering left us with a DTM with 913 unique tokens across 55,560 tweets. To adjust for the length of the tweets in our data, we operationalize the assessment-predominant regulatory mode orientation (APL) as follows in Equation 1:
where
Operationalizing the Variables for Legislative and Executive Power
The moderating variables in our model are categorial variables indicating Y if the variable definition is satisfied and N if it is not. The executive power variable, as described in Table 1, is coded as Y if the tweet originates from the president’s party. For example, if a tweet is between January 20, 2009 and January 20, 2017 and originates from the Democratic Party’s handle, this variable is coded as Y. However, during this time period, all tweets that originate from the Republican Party’s handle are coded as N. The categorical variables for legislative power are coded in a similar manner. The categorial variable indicating the party in control of the House of Representatives is coded as Y if the tweet is from the party that controlled it and N otherwise, whereas the categorial variable indicating the party in control of the Senate is coded as Y if the tweet is from the party that controlled it and N otherwise. The categorical variable indicating the party in control of Congress is coded as Y if the party from which the tweet originates controlled both the Senate and the House of Representatives and N otherwise. The correlations among the variables, which we report in Table 2 and describe next, also provide a count of the number of tweets in each condition.
Variable Description and Summary Statistics.
Correlations Between the Variables (n = 55,560).
* Significant at the 5% level.
** Significant at the 1% level.
*** Significant at the .1% level.
Summary Statistics and Correlations
We show some examples of tweets scoring high (and low) on APL for both party handles and highlight differences in the number of retweets they garner. The following two tweets include an assessment-predominant word (true, truth) and score highly on APL. However, the tweet from the Republican handle is retweeted more (26 retweets) than the tweet from the Democratic handle (18 retweets). Republicans: As Senate Dems consider a budget for the first time in years, will the American people see their Democrats: The Republicans: If Dems cannot Democrats: Democrats
Next, we examine the correlations between the variables. These are reported in Table 2. Because some of our variables of interest are categorical (yes/no), we first report the correlations for all variables. In addition, for further detail, we report these correlations for the continuous variables in the groups formed by each of these categorical variables. The overall correlation (for all 55,560 tweets) between the retweet count and APL is not significant (r = .0034). However, when considering the coefficients split by the party from whose handle the tweet originates, the correlation coefficient is significant for both Democrats (r = .0026, p < .05) and Republicans (r = −.0128, p < .05). This correlation coefficient also shows differences in significance when the tweet is from the president’s party, the party that controls the House of Representatives, the party that controls the Senate, and the party that controls Congress. These observed variations motivate the proceeding formal tests of our conceptual model, which we report next.
Data Analysis
We test the conceptual model proposed in Figure 1 in multiple steps. The first analyses check for the mediation effect of APL on retweets. In our subsequent analyses, we include the moderation effects as a result of executive power (i.e., when the sitting president belonged to the party from which the tweet was sent) and of legislative power (i.e., when the House of Representatives, the Senate, and/or both houses of Congress were controlled by the party from which the tweet was sent). For all models, we standardize the continuous variables and include control variables to account for the year fixed effects. These control variables capture the influence of aggregate trends and help eliminate omitted variable bias caused by excluding unobserved variables that evolve over time but are constant across tweets for a particular year. We report the results and findings from each of these analyses.
First, we consider the results of two multiple regression models. Model 1 tests the main effect for the tweet’s originating handle (Republicans vs. Democrats) on APL (i.e., assessment-predominant vs locomotion-predominant language). The results are reported in Table 3. Strengthening support of H1a, and in support of H1b, Model 1 shows a significant effect of the tweet originating from the Republican handle on APL (β = .118, p < .001) and that APL has a marginally significant effect on the number of retweets (β = .005, p < .1). In addition, there is also a significant direct effect of the tweet originating from the Republican handle on the number of retweets (β = .158, p < .001), indicating a partial mediation through APL. We also bootstrap to estimate the indirect effect of the mediation as suggested by Hayes (2018, p. 585) and find a significant indirect effect (β = .001, p < .1). This supports H1b, showing that tweets originating from the Republican Party, mediated by assessment-predominant language, generate more retweets than tweets originating from the Democratic Party.
Mediating Effects of APL on Retweets Moderated by the Executive Power and Legislative Power.
† Significant at the 10% level.
* Significant at the 5% level.
** Significant at the 1% level.
*** Significant at the .1% level.
Next, we check for the moderating effects of political power on this relationship. We test for both the effects of executive power (i.e., tweets originating from the president’s party) and legislative power (i.e., tweets originating from the party that controls the House of Representatives, the Senate, and/or both houses of Congress combined). First, in Model 4, when we consider the interaction effects of executive power in the presence of interaction with legislative power (measured as the tweet originating from the party that controls the Senate), we find that executive power (β = −.036, p < .01) and legislative power (β = −.033, p < .05) both reduce the main effect of APL on retweets observed in Model 1. Further, the bootstrapped indirect effect (β = .013, p < .1), while accounting for the three-way interaction (β = .057, p < .001), is marginally significant. Also accounting for the three-way interaction, the bootstrapped indirect effect for Model 6, where legislative power is measured as the tweet originates from the party that controls the House of Representatives (β = .022, p < .05), and for Model 8, where legislative power is measured as the tweet originates from party that controls both houses of Congress (β = .029, p < .05), are both significant. Thus, the positive effect of assessment-predominant language on generating retweets is reduced as parties gain control over successive branches of government, thus demonstrating support for H2.
Importantly, as we show in Table 4, which reports the interactions along with the confidence intervals, for Model 2 we find that the interaction effect of executive power (β = −.007) with the main effect of APL on retweets observed in Model 1 is not significant. In Models 3, 5, and 7, we also find that the interaction effect of legislative power—measured as the tweet originating from the party that controls the Senate (β = −.001), the House of Representatives (β = .015, p < .1), and Congress (β = .010)—with the main effect of APL on retweets is not significant (or marginally significant for Model 5) in each model. This further demonstrates support for H2, showing that the positive affect of APL on reducing retweets does not occur when a political party obtains control of only one branch of government, but that parties need to experience an increase in political power involving successive branches of legislative government beyond the executive before the effect of APL on increasing retweets is reduced. We graphically represent these three-way interactions in Figure 2. Here, we can observe that the effect of APL on increasing retweets is reduced when the party that is tweeting has control of both the executive and legislative branches of government, and that this occurs irrespective of which branch of legislative government the party controls.
Confidence Intervals for Interactions.
Notes: Numbers in bold are significant at 95%.
Effect of APL on retweets when party controlling the executive branch controls parts of the legislative branch.
Endogeneity Correction
The previous models are built on standard premises in mediation analyses and estimated using seemingly unrelated regressions. The interpretations of the parameters in these models assume that the error terms in the outcome equation and the mediator equation are not correlated (Shaver 2005). However, this assumption is likely to be violated because managers of Twitter accounts would deliberately use language that garners maximum reach (retweets and favorites). Our previous models include the time fixed effects, which capture the influence of aggregate trends. This helps eliminate omitted variable bias caused by excluding unobserved variables that evolve over time but are constant across tweets for a particular year. However, the strategic behavior of social media managers can create endogeneity of the language used and lead to violations of the required assumptions in standard mediation analyses.
In this study, we use the latent instrument variable (LIV) approach (Ebbes et al. 2005), which introduces a binary unobserved instrumental variable that partitions the endogenous predictor (i.e., the mediator variable capturing the predominant regulatory mode orientation) into two components: one uncorrelated and the other correlated with the error term in the main equation (in this case, the retweet count model). We note that the Shapiro–Wilk’s test for normality, conducted on samples of 2,000 observations randomly drawn from the total of 55,560 of the endogenous regressor (APL), confirms the nonnormality (p < .001). This validates the use of the LIV approach (Papies, Ebbes, and Van Heerde 2017). We implement the Bayesian adaptation of this method used by Zhang, Wedel, and Pieters (2009) to account for the endogeneity introduced by the language used in the tweets in our models.
We find that the endogeneity correction using the LIV approach does not change the conclusions of our analysis. However, we find that when they are corrected for bias using the LIV approach, the coefficients are larger than the ones we report in Table 4. Thus, the coefficients reported in Table 4 are conservative estimates of the bias-corrected coefficients. We also find that the LIV component correlated with the error term in the main equation is insignificant, and we conclude that the endogeneity in the model has been addressed. We explain the Bayesian model, its estimation, results, and a sample code in OpenBUGS (Lunn et al. 2009) in Web Appendix A.
Replication Study
To demonstrate replication of our results, we now report findings from a study conducted on an additional sample of 5,012 tweets from the Democratic and Republican Party Twitter accounts for the period between October 2019 and June 2020. This method is in line with best-practice methods for conducting replications (Simmons, Nelson, and Simonsohn 2011), as it allows for the provision of an exact, rather than conceptual, replication of our findings. Importantly, as the study was conducted after the first submission of this manuscript for peer review, it further separates exploratory hypothesis generation and confirmatory testing of hypotheses (Nosek et al. 2018). We achieve these ends by collecting additional tweets from the focal Twitter handles (@GOP and @TheDemocrats) for the period between October 2019 and June 2020 and use this data for prediction.
Therefore, to begin, we collect all additional tweets and follow the text cleaning procedures described previously to calculate the desired variables. We first check for the difference in APL and find that H1a continues to be supported (MRep = −.018, SD = .037; MDem = −.016, SD = .044; t(2,512) = −1.572, p = .06). Although we cannot use this new data to replicate the entire study, as our moderating variables for executive power and legislative power do not change during this period, we run Models 4, 6, and 8 on the original data to predict the retweet count for the new data collected. We find that when legislative power is measured as tweets from the party that controls the Senate, not only is the predicted value for retweets (M = −.143, SD = .011) in the same direction as that of the actual value for retweets (M = −.001, SD = .014), but the two means are also not statistically different (t(5,011) = 8.9805, p < .001). We find similar results when the legislative power is measured as tweets from the party that controls the House of Representatives (Mpredicted = −.164, SD = .012; Mactual = −.001, SD = .014; t(5,011) = 10.067, p < .001) and when the legislative power is measured as tweets from the party that controls Congress (Mpredicted = −.174, SD = .013; Mactual = −.001, SD = .014); t(5011) = 10.587, p < .001). These findings establish a replication of the results we present for our main sample.
General Discussion
We here use 55,560 tweets from the official Twitter accounts of the Democratic and Republican Parties to examine how goal pursuit language influences retweets. In doing so, we contribute to the literature on WOM by showing that an understanding of the drivers of WOM needs to account for the preferred goal pursuit strategies of WOM recipients. We demonstrate that Twitter messages are more likely to be shared when language expressing goal pursuit aligns with the political motivations of the audience. We further argue that goal pursuit language resonates differently with Democrats and Republicans because their political agendas differ in regard to social progress. Specifically, Republicans are more likely to share tweets predominant in the language of assessment because deliberation shifts attention away from action. This aligns with a conservative ideology of restraining the progressive action typically advocated for by the Democrats.
However, we also note that political power further shapes how a party’s constituents relate to the goal pursuit language the party uses. Specifically, we show that gaining power can reduce the role of assessment-oriented language on the sharing of political content. Our interpretation is that with fewer political impediments to their political agenda, constituents become motivated to achieve their political agenda directly. While we observe this effect across Republican and Democratic Party tweets, it highlights the divergence of their political agendas. That is, Republicans are more likely to share assessment-oriented language, but as their party gains political power, the effect of assessment-oriented language is reduced. In contrast, as Democrats gain political power, their avoidance of assessment-oriented language is amplified.
Together, our contributions show that when crafting social media content, political parties should consider the impacts of goal pursuit strategies, which have so far been unexplored by the WOM literature. In designing content, parties should consider how the language used to deliver a political message frames goal pursuit for their audience. Further, by considering the extent of their political power, they can better predict how audiences will respond to their content. These contributions raise important public policy implications and provide interesting avenues for future research that should be explored. We discuss these next.
Public Policy Implications
Our findings imply that bridging political divides is essential, as it demonstrates that Republican and Democratic constituents speak different goal pursuit languages when sharing ideas on social media. A simple prescription for public policy makers in encouraging greater dialogue between opposing political sides is to learn to speak the preferred goal pursuit language of their opposition. As dictionaries of goal pursuit language are freely available (Higgins Lab 2021; Kanze, Conley, and Higgins 2019), regulatory mode language forms a concrete tool to use to reduce the alienation of political opponents on social media, which our work shows is likely to be occurring, at least in part, because of how messages are communicated. When encouraging political discourse by speaking each other’s language, public policy makers also need to consider the effect of power on the sharing of content within political systems. There are increasing efforts to educate the public about how content circulates on social media (News Literacy Project 2020). It is important that such efforts ensure that the public understands the role of the legislative and executive branches of government in acting as a check on the power of the other (Watts 2010), and that their power can shape responses to content on social media.
We find that assessment-oriented language, which is the natural language of debate, will spread further when used by parties that do not control successive branches of government. This shows that all parties have a role in contributing to public policy debates, and that their ability to contribute to these debates must be safeguarded even when there are power differences between them. Therefore, stakeholders (i.e., the press and judiciary) need to ensure that all parties are able to access high-quality information on policy issues. For example, strong Freedom of Information laws need to be maintained within democracies (Berliner 2014), and bipartisan congressional committees should provide parties with public policy detail to ensure robust debate.
Recommendations for Future Research
While our work here presents these practical implications for public policy, further research should be conducted to examine specific approaches to the programs of education we have proposed. For example, research should first evaluate which political actors should conduct education programs on the drivers of sharing on social media. Given the role of power in shaping responses to content, consumers may react differently to programs highlighting the power of goal pursuit language when they originate from government departments, political parties, or social media platforms themselves.
Particular consideration should be given to the role of social media companies in these efforts. Companies such as Facebook and Twitter are increasingly confronting the necessity of developing solutions to discourage the spread of disinformation on their platforms (Facebook n.d.; Harrison 2019). Despite these efforts, they still face substantial criticism that their solutions have not been effective enough for the sector to avoid policy intervention (Tusikov and Haggart 2019). Research should explore whether social media companies could successfully adopt programs that build awareness of goal pursuit language as part of these initiatives. In particular, it must be established whether users are still more likely to share content crafted with their preferred goal pursuit language when they have been informed about how persuasive such content is likely to be. It is likely, for example, that individuals will judge their own communications as less persuasive when crafted in the goal pursuit language of their opposition. Research should therefore be conducted to explore how this might demotivate users from adopting a specific goal pursuit language and identify methods to correct and overcome any potential demotivation that users might experience.
Finally, in examining how successive increases in power attained by a political party affects the sharing of regulatory mode language, we have theorized that reduced impediments to action for the executive also reduce regulatory fit for assessment-oriented language, thereby decreasing sharing of assessment-oriented language. Therefore, to examine how opposition to an agenda shapes communication concerning it, we have necessarily focused our theorizing on situations in which parties have the ability to dominate the political agenda through first controlling the executive branch of government. This imposes a limitation on our research, as it means that we offer limited insight into how regulatory mode language impacts sharing of content from parties with control over the legislative branch but not the executive branch. Our insights are further limited regarding communication from parties that control neither the executive nor legislative branches of government. Future research should consequently aim to answer questions of how power concerning the checks and balances inherent to the U.S. political system would affect communication by parties that do not control the executive branch and therefore do not set the political agenda to begin with. It would further be an interesting extension of the work presented here to explore differences in communication and virality when parties have no control whatsoever over the political system. This research is likely to be highly relevant to modern political communication scholarship, as there have been notable moments when one party has taken control of the executive and both branches of the legislative at the same time. Further, voters are increasingly registering support for third parties (Reinhart 2018), which by their nature must find ways of communicating with voters without control of either the executive or legislative government. While we offer limited insight into third parties, Web Appendix B presents a speculative discussion concerning these issues.
Finally, there are several avenues that future research could pursue to expand on the theoretical work presented here. First, future work on this topic could also look at virality from the perspective of regulatory focus theory, which examines whether individuals are motivated to seek vigilant or aspirational goals (Higgins 1998). By extending the existing regulatory focus dictionary (Kanze, Conley, and Higgins 2019) to enable analysis of short-form content such as tweets, researchers could also examine the effects of regulatory focus on sharing and any potential additive impacts of regulatory fit between regulatory mode and regulatory focus language in Twitter content (Cornwell, Franks, and Higgins 2019; Higgins, Nakkawita, and Cornwell 2020). In particular, such analyses could be used to investigate whether parties use locomotion language in the service of changing or maintaining the status quo. Further, considering that there is a tendency to avoid contact with opposing views on Twitter, motivations for the use of preferred regulatory mode language may be more extreme there than they would be on other platforms. Future research should examine regulatory mode usage in political speech on mediums aimed at a more general audience compared with content on mediums with more partisan audiences.
Conclusion
Using 55,560 tweets from the official Twitter accounts of the Democratic and Republican Parties, we show that tweets sent by the Republican Party are more likely to be predominant in the language of assessment (as opposed to the language of locomotion). We further find that while tweets predominant in the language of assessment lead to more retweets for Republicans, this effect is reduced as political parties gain control of successive branches of government. This is because successive increases in political power create fewer impediments to implementing a party’s political agenda. As impediments to action are reduced, so is regulatory fit for assessment-oriented language. In this way, goal pursuit language shared on Twitter reveals distinct approaches to obtaining and dealing with power across the U.S. political system. Goal pursuit language therefore presents an important tool for political parties and public policy makers to successfully conduct policy debates in the future.
Supplemental Material
Supplemental Material, sj-pdf-1-ppo-10.1177_0743915621999036 - Power and the Tweet: How Viral Messaging Conveys Political Advantage
Supplemental Material, sj-pdf-1-ppo-10.1177_0743915621999036 for Power and the Tweet: How Viral Messaging Conveys Political Advantage by Kellie Crow, Ashish S. Galande, Mathew Chylinski and Frank Mathmann in Journal of Public Policy & Marketing
Footnotes
Special Issue Guest Coeditors
Brennan David, Dhruv Grewal, and Steve Hamilton
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.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
