Social Media Influence and Electoral Competition

Abstract

Do social media platforms help or hinder democracy? Internet enthusiasts posit that social media could have a democratizing effect by lowering the costs of promotion, while skeptics argue that these platforms replicate or even exacerbate preexisting inequalities. We inform this debate by combining campaign finance and electoral outcome data from the Federal Election Commission with Twitter metrics of candidates who ran in the 2016 U.S. congressional elections. We find that poorer candidates, who spent less than their competitor, performed better if they had indirect influence on Twitter—getting their tweets shared by users whose own tweets are widely shared. The effect of indirect influence on election outcomes was more pronounced in races with larger financial inequities between candidates or fewer total expenses across candidates. Moreover, poorer candidates with indirect influence saw smaller vote gaps than their party’s candidate in the same district (in House races) or state (in Senate races) in 2014.

Keywords

campaign spending indirect influence political elections Twitter two-step flow

Social media platforms facilitate information exchange between voters (Bond et al., 2012) and provide new ways for politicians to engage with their constituents (Golbeck et al., 2010). Because messages can travel quickly and broadly on social media (Nahon & Hemsley, 2013), these platforms have become a widely used tool for political campaigning (Gulati & Williams, 2013). In the 2016 U.S. presidential race, for example, both Hillary Clinton and Donald Trump spent considerable efforts to infiltrate social media with their respective messaging. By many measures, Trump had the upper hand on social media despite his relative disadvantage in terms of campaign funds raised (Allison et al., 2016), money spent on television ads, and time devoted to face-to-face canvassing (Easley, 2016). He maintained a 30% larger following on Twitter than Clinton, received twice as many Facebook likes (10 million compared to Clinton’s 5.2 million), and had over 20% more followers on Instagram than Clinton (2.2 million vs. 1.8 million for Clinton). Aware of this discrepancy, Trump had this to say in an interview for the CBS News program 60 min, just days after the election: “I think that social media has more power than the money [Clinton’s campaign] spent, and I think maybe to a certain extent, I proved that” (Morin, 2016).

While Trump was likely unaware of it, in making this statement he was taking sides in the long-standing debate about whether online platforms equalize or normalize inequalities. On the one hand, social media lower the costs of promotion, which can help poorly financed groups gain a more prominent voice (Earl & Kimport, 2011). However, managing an effective social media campaign can itself be costly. The learning curve for using these tools effectively can thus serve to replicate or even exacerbate preexisting inequalities (Schradie, 2018). We contribute to this ongoing debate by combining campaign finance and electoral outcome data from the Federal Election Commission (FEC) with Twitter metrics of candidates who ran in the 2016 U.S. congressional elections. These metrics were constructed using data obtained with Twitter’s Application Programming Interface (API) and a snowball sampling method that lets us identify users who retweeted (i.e., shared) candidates’ tweets in the months before the election. We explicitly consider how electoral outcomes varied with Twitter metrics of financially worse- and better-off competitors, the extent of financial inequity between them, and the total spending by candidates in a given race.

As expected, financially equal races tend to have close electoral outcomes, while unequal races see large outcome gaps (about 35 percentage points, on average). We investigate how this relationship between financial and electoral equality varies with Twitter metrics of the competing candidates, finding that it is significantly moderated when the poorer candidate in a race has indirect influence on Twitter—getting their tweets shared by users whose own tweets are widely shared. Indirect influence thus helps to explain when financially unequal races exhibit relatively close electoral outcomes. The effect persists after controlling for other Twitter metrics, state-level fixed effects, and several election and candidate-specific variables, including exposure on traditional media. Moreover, the relationship between money and votes does not depend on the influence possessed by financially better-off candidates (direct or indirect), adding to the body of evidence showing that social media can have equalizing effects (e.g., Samuel-Azran et al., 2015).

To explain why the effect of indirect influence is more pronounced in financially unequal races, we draw on work by political scientists on “reactive” spending (e.g., Jacobson, 1978). Competing candidates often reach financial equality when political parties attempt to outspend each other in races perceived to be close. Empirically, such reactive spending results in higher spending overall. We argue that the role of indirect influence in financially unequal races can thus be explained, at least in part, by the fact that total spending in these races is typically low—implying that candidates have less to spend on traditional advertising. Indeed, we show that the effect of indirect influence is more pronounced in lower cost races. Finally, we compare election outcomes in 2016 to those in 2014, finding that poorer candidates with indirect influence see smaller electoral gaps than their party’s candidate in the same district (in House races) or state (in Senate races) in the previous election year. Our results thus do not stem from geographic variation in social media use.

In the next section, we discuss previous research on congressional campaigning and online influence, framing our work as a test of the normalization versus equalization argument (e.g., Gibson et al., 2014). In the third section, we discuss the details of our data and empirical models and describe our novel method of using the Twitter API to snowball out from candidate accounts to measure their indirect influence. We discuss the results of our analyses in the fourth section, providing graphical support for our findings and describing candidates and tweets that garnered the most indirect influence. We conclude in the fifth section with implications of our findings for contemporary elections and discuss limitations of our study and directions for future research.

Congressional Campaigns and Online Influence

Congressional members were initially slow to adopt social media as a viable campaign resource, but even early scholarship views social media use by congresspeople as a “vehicle for self-promotion” (Golbeck et al., 2010). In other words, social media present yet another venue for campaigning. More recent work on online campaigning has noted a steady increase in the use of social media by congressional campaigns (Gulati & Williams, 2013). Today, it appears that members of congress have realized the potential of social media platforms for engaging with their constituents and have begun to fully embrace its use. For example, in the 115th Congress of the United States (active from 2017 to early 2019), only four congressional members did not have a Twitter account (Littman, 2017).

While candidates may have come around to using social media platforms for political campaigning, the extent that behaviors on social media are associated with political outcomes is debated. DiGrazia et al. (2013) find that when members of congress are mentioned on Twitter, they reap dividends in their general election vote share. However, Jungherr et al. (2017) show that the relationship between Twitter mentions and electoral outcomes is tenuous, arguing that mentions are a better proxy of attention toward politics rather than political support. Even if a relationship between Twitter mentions and electoral success were to exist, mentions are not directly tied to a candidate’s own campaign activities. The implications of such a relationship for political campaign strategy thus remain unclear, even though the implications for polling may be significant. Moreover, Twitter mentions represent just one class of metrics on a single platform, and the findings of Jungherr et al. (2017) may or may not translate to other metrics or platforms.

Can political candidates leverage social media to influence electoral outcomes? While scholars of technology and politics acknowledge the import of this question (Gibson et al., 2014), so far research relating social media influence to electoral success remains scarce. One notable exception is research by Bode and Epstein (2015) who use a unique data collection strategy that relies on Klout—an online service that compiles data from several platforms to construct a single “influence score.” They find that a candidate’s Klout score can help predict their vote share, even after controlling for other candidate- and race-specific variables. This work provides important evidence that social media influence may play a role in shaping electoral outcomes, but the Klout score is a blunt instrument that does not reveal which specific social media activities are associated with electoral success. Here, we build on the work of Bode and Epstein (2015) by employing a novel snowball sampling method that allows us to construct several classes of metrics that can each be separately evaluated for their relationship with electoral outcomes. Importantly, these are all metrics that campaigns could feasibly construct, in real time and for free, using Twitter’s API.

To motivate the classes of metrics we develop, we leverage previous research on the flow of political information. Hilbert et al. (2017) distinguish between one-, two-, and network-step flows on Twitter. While one-step flows capture messages that reach the broader population directly from the source, two-step flows are mediated by the so-called opinion leaders (Katz & Lazarsfeld, 1955). Network-step flows, meanwhile, are more intricate processes that are tied to the underlying interconnectivity between people. Choi (2015) finds evidence of two-step flows in political discussions on Twitter, noting that “those whose messages were frequently retweeted were highly likely to be opinion leaders” (p. 705). The idea that messages may rely on several “steps” before reaching the broader population helps us reconcile how Twitter, a platform used by fewer than a quarter of Americans (Wojcik & Hughes, 2019), could have effects that go beyond the platform itself. For example, in their study of political campaign websites, Norris and Curtice (2008) find evidence of two-step flows to explain how online campaign information can reach people who do not visit candidate websites. While few people visit these sites, those who do are more likely to discuss politics with people that they know. Many voters can thus discover political information indirectly, without participating in online discussions.

Reflecting on this work, we posit that social media could benefit candidates by way of all three of these flows. First, social media could help via one step by allowing candidates to engage directly with more people. By building a large audience, candidates can gain celebrity-like status (Wheeler, 2013), which could have electoral effects. Second, social media can help via two steps by allowing candidates to garner the attention of opinion leaders. The diffusion of information on social media is usually driven by large broadcasts (Goel et al., 2016), implying that candidates can gain indirect influence by having their messages shared by influential users. Finally, social media could help candidates via a network step. Klar and Shmargad (2017) show that networks which are structurally diverse—providing access to many different social network regions—can accelerate the diffusion of messages from low-resourced groups. We thus test for the effects of social media influence via one-, two-, and network-step flows.

Beyond the general relationship between social media influence and electoral outcomes, there is also debate concerning how the benefits of social media campaigning are distributed across better- and worse-off candidates. Scholars who have contributed to this debate, sometimes called the “normalization versus equalization argument” (Gibson et al., 2014), have tended to focus on candidates’ initial adoption of online campaigning rather than metrics that capture the effectiveness of online campaigns. For example, in a special issue of the Journal of Information Technology and Politics alone, scholars investigated the adoption of online campaigning in Denmark (Hansen & Kosiara-Pedersen, 2014), France (Koc-Michalska et al., 2014a), Germany (Marcinkowski & Metag, 2014), and Poland (Koc-Michalska et al., 2014b). Our work helps to inform this debate by analyzing how specific metrics are associated with electoral success, extending prior work that has primarily focused on the initial adoption of online campaigning.

Scholars of social movements have noted that the success of modern movements depends on their ability to leverage-specific affordances of the internet. In particular, they cite low costs of mobilization and asynchronous participation as prime opportunities that the internet affords (Earl & Kimport, 2011). In congressional politics, social media can give candidates with few resources the ability to enter into a political arena where monetary costs are low, the transmission of ideas is relatively unfiltered, and name recognition can be built quickly. This is in contrast to traditional media outlets, like television, where costs of promotion run high. For example, the costs of running a political ad on a local television station are estimated at $200–$1,500 per 30 s ad, per run. For those with little money to build name recognition through traditional channels, social media can provide a cheaper alternative. Campaigning on social media can thus help marginalized candidates attract attention, broaden exposure to their messages, and build name recognition.

In this light, the benefits of social media can be viewed as being qualitatively similar, in a mild sense, to the advantages that are bestowed upon political incumbents. Political incumbents—officials who currently hold the office for which they are running—enjoy several benefits over those who challenge them including free publicity (Cain et al., 1987) and wider name recognition (Kam & Zechmeister, 2013). Social media platforms, given their democratic premise that anyone’s ideas can go viral (Nahon & Hemsley, 2013), could similarly provide marginalized candidates with the prospect of gaining high levels of exposure. We return to this analogy when describing the setup of our empirical analysis that shows qualitatively similar patterns for incumbency and social media influence.

We acknowledge that the benefits of social media could help well-resourced candidates as well. However, well-resourced candidates have less to gain by using social media because they already tend to have name recognition, established donor bases, and alternate means of unfiltered communication with their constituents. Well-resourced candidates are thus more likely to be at a point of saturation, wherein they have already expended resources to get information out to voters (Jacobson, 1990). Social media platforms can thus serve as a useful tool, one among many, that well-endowed candidates employ, but is not a game changer. On the other hand, for candidates with few financial resources, social media could decrease the promotional resource gap, thereby narrowing the gap in vote shares.

Data and Models

Data for this project were collected with the help of 10 research assistants. Each assistant was assigned one or several U.S. states and was tasked with collecting the Twitter handles of the top two candidates (in terms of vote share) competing in 2016 House and Senate races in those states. Data about the vote shares of candidates in each electoral race were obtained from the FEC and were combined with the FEC candidate financial summaries. The combined FEC data include several useful data points about the congressional candidates, including their incumbency status and total disbursements (i.e., expenses) during the campaign season. Data from the FEC were collected for electoral races in 2014 and 2016, so that comparisons could be made across election years.

Before discussing the data that we gather from Twitter, we relate financial and electoral equality across election years. We exclude uncontested races (i.e., those with a single candidate) and any candidate who came in third or worse. We measure the financial equality of a race with the ratio of the poorer to richer candidates’ expenses. A ratio of 1 thus implies that the race was financially equal, while a ratio close to 0 implies that the richer candidate outspent the poorer candidate by a considerable amount. We measure the electoral equality of a race by subtracting the vote share of the poorer candidate from that of the richer candidate. We use the two-party vote share, which is the percentage of votes that a candidate receives divided by the total number of votes that went to the top two candidates. All of the analyses in this article use the two-party vote share, though the results are nearly identical when using the raw vote percentages instead.

In addition to investigating financial and electoral equality across years, we classify races by the incumbency status of the candidates. In particular, we designate races as open (i.e., no incumbent present), richer incumbent, or poorer incumbent. Figure 1 depicts financial and electoral equality across years and incumbency. In general, the patterns are similar in open races and those that feature the richer candidate as the incumbent. In particular, financially unequal races (i.e., those in which the ratio of expenses between poorer and richer candidates is near 0) have large electoral outcome gaps. This is evident from the large y-intercepts in the left two columns. However, as races get more financially equal the gaps tend to zero, implying that poorer and richer candidates see similar vote shares.

Figure 1.

Financial and Electoral Equality by Year and Incumbency.

Panels in the rightmost column of Figure 1, which depict the relatively few races where the poorer candidate was also the incumbent, reveal a different pattern. In particular, electoral outcome gaps do not appear to vary, on average, with the extent of financial inequality between candidates. The intuition behind this pattern is simple: One candidate is richer, while the other is the incumbent, and in the end neither type of candidate consistently gets elected in these races. Empirically, the rightmost column shows a flattening, or moderation, of the relationship in the two leftmost columns. We similarly search for such flattening when considering the Twitter metrics that render financial inequities irrelevant, or at least less pertinent, for electoral equality. More directly, we are interested in Twitter metrics that exhibit qualitatively similar patterns as poor incumbency.

Twitter Data

The final step of the data collection procedure took place in March 2017 and relied on Twitter’s API to extract data from candidates’ accounts. Often, candidates maintained more than one account, and research assistants were instructed to keep every handle that they could locate in such cases. The assistants then each ran a script, written in the programming language R, to collect data capturing candidates’ tweets, users who retweeted (i.e., shared) candidates’ tweets, tweets of users who retweeted candidates’ tweets, and users who retweeted tweets by candidate retweeters. The data about candidates were thus detailed enough to let us create each candidate’s ego social network that includes information about which users retweeted a candidate’s tweets and the extent that these users retweeted each other.

Figure 2 depicts our Twitter data collection strategy and the metrics that we construct to capture one-, two-, and network-step flows. At one step, we include the number of retweets and likes that candidates received on their tweets, on average. At two steps, we include the average number of followers that candidates’ retweeters had at the time of data collection (in 1,000s), and the number of retweets and likes their tweets received, on average. At the network step, we include two measures of structural diversity: the clustering coefficient (Watts & Strogatz, 1998) and social network constraint (Burt, 1992). These measures capture the extent that a candidate’s retweeters are located in different social network regions—where a low clustering coefficient and network constraint both imply high levels of structural diversity. Both measures range from 0 to 1, and we multiply them by 100 to ease the interpretation of our model estimates.

Figure 2.

Data Collection Strategy and One Step, Two Step, and Network Step Metrics.

We restrict data collection to the 3 months leading up to the morning of Election Day, November 8, 2016. Practically, this means that the script requested tweets with ID numbers greater than 760836150760050689 (a tweet by Clinton on August 3, 2016) and smaller than 795954831718498305 (a tweet by Trump on the morning of November 8, 2016). To constrain the time spent collecting data, the script requested at most 100 tweets from each candidate and retweeter account. Twitter’s API restricts data collection to the most recent 3,200 tweets per account, including replies and retweets. Since data collection took place more than 6 months after the start of the period of interest, it is possible that some tweets from some candidates were not retrieved. In Figure 3, we depict the total number of tweets collected and a count of the earliest tweet collected by candidates, broken down by day and political party.

Figure 3.

Volume of Total and Earliest Candidate Tweets by Day and Political Party.

Twitter’s API also restricts the number of retweeters that can be retrieved for each tweet to the most recent 100, implying that retweeter data were incomplete for tweets that were widely retweeted. In Figure 4, we depict the retweeter retrieval rate by the average retweets and political party of the candidates. The average retweeter retrieval rate across candidates was just over 90%, though the retrieval rate drops off steadily for candidates who received more than 100 retweets, on average. Finally, while analyzing our data, it became clear that one of the research assistants did not properly complete the task, so that Twitter metrics are missing for candidates in Illinois, Maryland, Oklahoma, New Mexico, and West Virginia. Still, the full set of Twitter metrics was available for 400 candidates, and 142 electoral races included the full set for both candidates. We restrict our main analysis to the 284 candidates in these races.

Figure 4.

Retweeter Retrieval Rate by Average Retweets and Political Party.

In Table 1, we report summary statistics and a correlation matrix of the Twitter metrics. On average, candidates’ tweets received about 20 retweets and 30 likes. Retweeters, on average, saw significantly lower levels of engagement with their tweets than candidates, receiving about 2 retweets and just over 3.5 likes per tweet. The correlation matrix reveals that some of the Twitter metrics are highly correlated. In particular, correlations were high between the metrics capturing retweets and likes that candidates received, as well as between metrics capturing the followers, retweets, and likes that candidate’s retweeters received. We revisit this correlation table when discussing the metrics that we can comfortably include in the same model. We now turn to describing the empirical models that we use to test how Twitter metrics were associated with the outcomes of the 2016 U.S. congressional races.

Table 1.

Summary Statistics and Correlation Matrix of Twitter Metrics.

	Mean	SD	X ₁	X ₂	X ₃	X ₄	X ₅	X ₆
One-step metrics
Average retweets (X₁)	18.3	81.3
Average likes (X₂)	31.3	140	.976
Two-step metrics
Average followers (X₃)	5.56	7.29	−.048	−.037
Average retweets (X₄)	2.17	2.95	−.054	−.054	.654
Average likes (X₅)	3.59	4.52	−.059	−.054	.705	.973
Network-step metrics
Clustering coefficient (X₆)	2.35	6.27	−.067	−.069	−.092	−.054	−.051
Network constraint (X₇)	7.07	9.80	−.140	−.141	−.181	−.182	−.210	.211

Note. Number of candidates = 284.

Empirical Modeling

We start by modeling the relationship between financial and electoral equality depicted in Figure 1. To formalize the financial equality of a given electoral race, we construct the expense ratio to be the ratio of expenses between the poorer and richer candidates, constrained to the top two candidates in terms of vote share. For candidate expenses, we follow Margolin et al. (2016) and use the TTL_DISB column from the 2016 FEC candidate summaries.

Expense Ratio = \frac{Expenses of Poorer C and}{Expenses of Richer C and} .

We then define the vote gap in an electoral race to be the difference in vote counts of the richer and poorer candidates divided by the total vote count across these two candidates:

Vote Gap = \frac{Votes to Richer Cand - Votes to Poorer Cand}{Votes to Richer Cand + Votes to Poorer Cand} .

We first employ a simple ordinary least squares (OLS) regression to relate the vote gap to the expense ratio:

Vote Gap = α_{1} + α_{2} Expense Ratio + ∊ .

Model 1

Our next specification adds a vector of Twitter metrics and control variables, X, an interaction between this vector and the expense ratio, and state-level fixed effects s_i:

Vote Gap = α_{1} + α_{2} Expense Ratio + Β_{1} X + Β_{2} X \times Expense Ratio + s_{i} + ∊ .

Model 2

The coefficients in $Β_{1}$ capture the effects of Twitter metrics and control variables in financially unequal races (where the expense ratio is zero), while the coefficients in $Β_{2}$ capture the effects of Twitter metrics and control variables as races tend toward financial equality. The flattening we observe for poor incumbency in Figure 1 would thus be exhibited by a negative coefficient in $Β_{1}$ and a positive coefficient in $Β_{2}$ .

While the state-level fixed effects control for variation across states, they do not control for variation across congressional districts. If certain districts consistently see larger vote gaps, and these same districts are more likely to feature particular Twitter activities among candidates, then the Twitter metrics in X would be endogenous. In order to control for time-invariant district-level characteristics, we match candidates in 2016 with vote shares of their party’s candidate in the same electoral race in 2014. Formally, we define the vote gap in 2014 as:

{Vote Gap}_{2014} = \frac{{Votes to Richer Cand’s Party}_{2014} - {Votes to Poorer Cand’s Party}_{2014}}{{Votes to Richer Cand’s Party}_{2014} + {Votes to Poorer Cand’s Party}_{2014}} .

We then model the difference in vote gaps between 2016 and 2014 as:

{Vote Gap}_{2016} - {Vote Gap}_{2014} = α_{1} + α_{2} Expense Ratio + Β_{1} X + Β_{2} X \times Expense Ratio + s_{i} + ∊ .

Model 3

If certain Twitter metrics are associated with smaller vote gaps in 2016 than 2014, we attribute the shrinking across years to these metrics.

Finally, we discuss the control variables that we also include in X. These are the total disbursements across both candidates in the race (in millions), an indicator equal to 1 for Senate races and 0 for House races, the total number of candidates in the race, an indicator equal to 1 when an incumbent is present and is the richer candidate, an indicator equal to 1 when an incumbent is present and is the poorer candidate, an indicator equal to 1 if the richer candidate is affiliated with the Republican party, exposure of both candidates in traditional media, measured as the number of mentions candidates received in CNN transcripts (DiGrazia et al., 2013) over the same time period that we used for Twitter data collection, and candidates’ engagement on Twitter measured as the average number of days that passed between two consecutive candidate tweets. Next, we describe the results of estimating several specifications of the models presented.

Results

We start by establishing the empirical relationship between financial and electoral equality and the moderating effect of poor incumbency (as depicted in Figure 1). In Table 2, we report estimates from several models that use FEC data from electoral races in 2014 and 2016. The left column shows that financially unequal races see large vote gaps—about 35 percentage points, on average. However, vote gaps tend to zero as races get more financially equal (i.e., as the expense ratio approaches 1). In the two right columns of Table 2, we add indicators for poor and rich incumbency and fixed effects for state and year. Under both specifications, incumbency of the poorer candidate is associated with a statistically significant moderation of the relationship between financial and electoral equality. Specifically, the indicator for poor incumbency has a negative baseline coefficient $(Β_{1})$ and positive interaction with the expense ratio $(Β_{2})$ .

Table 2.

OLS Models Relating Financial and Electoral Equality.

	Model 1		+ Incumbency		+ Fixed Effects
	Coefficient	SE	Coefficient	SE	Coefficient	SE
Constant	34.6**	.698	31.0**	2.26	31.1**	2.41
Expense ratio	−42.8**	1.98	−42.4**	4.50	−43.0**	4.75
Poorer incumbent
Baseline			−46.8**	13.5	−51.6**	13.9
× Expense ratio			55.6**	16.9	63.4**	17.4
Richer incumbent
Baseline			3.68	2.38	3.29	2.49
× Expense ratio			4.25	5.17	4.27	5.43
State fixed effects	No		No		Yes
Year fixed effects	No		No		Yes
Adjusted R²	.403		.423		.430
Observations	694		694		694

Note. OLS = ordinary least squares.

*p < .05. **p < .01.

Next, we test for moderating effects of the Twitter metrics. We begin by estimating Model 2, which includes state-level fixed effects and the full set of control variables at both the baseline level and as an interaction with the expense ratio. Since some of the Twitter metrics are highly correlated (see Table 1), we start by adding them separately into the model. Each metric enters the model in four ways: at both the baseline level and as an interaction with the expense ratio, and constructed for both the poorer and richer candidates in a given race. In Table 3, we report estimates of the seven metric-specific specifications. Only two of the seven metrics show statistical significance: The average number of retweets and likes received by poorer candidates’ retweeters. Notably, none of the metrics shows statistical significance for the richer candidate in the race. Moreover, for the poorer candidate, indirect influence appears to play a more significant role than their direct influence (i.e., number of retweets or likes on their own tweets).

Table 3.

OLS Models With Twitter Metrics Included Separately.

Model 2	Poorer Candidate				Richer Candidate				Adjusted. R²
	Baseline		× Expense Ratio		Baseline		× Expense Ratio
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
One step
Average retweets	−0.110	.165	−0.021	0.486	.035	.022	−.219	0.266	.500
Average likes	0.003	.077	−0.041	0.133	.025	.013	−.263	0.180	.504
Two step
Average followers	−0.293	.382	0.236	1.33	.156	.296	−.109	0.795	.462
Average retweets	−2.69**	.975	3.17*	1.42	−.161	.576	.529	2.22	.501
Average likes	−2.09**	.707	2.44*	1.03	−.159	.387	.477	1.35	.507
Network step
Clustering coefficient	−0.300	.279	1.44	1.50	.410	.388	−.765	1.32	.468
Network constraint	0.082	.163	.688	0.950	−.661	.652	.762	1.56	.476

Note. Number of observations = 142. Each row includes estimates from a different model. State-level fixed effects included. All control variables included at both the baseline and interaction levels.

*p < .05. **p < .01.

These two metrics, average retweets and likes among poorer candidates’ retweeters, exhibit qualitatively similar empirical patterns as poor incumbency. In particular, they show a significant negative baseline coefficient and positive interaction with the expense ratio—a flattening of the relationship between financial and electoral equality. Indirect influence thus helps to explain financially unequal races that see relatively close outcomes. However, we are careful not to attribute causality to these effects. The specifications in Table 3 do not account for variation across congressional districts. It is possible that high engagement candidates compete in districts that tend to lean strongly toward one political party, thereby exhibiting larger vote gaps. In Model 3, we address the alternative explanation that our results stem from geographic variation in social media use. In particular, we control for district-level variation by comparing races in 2016 to the prior congressional election year of 2014.

The next analyses include multiple Twitter metrics in the same specification. We exclude the metric capturing followers across a candidate’s retweeters since it is highly correlated with retweeters’ retweets and likes (see Table 1). Moreover, unlike retweets and likes, the follower metric had no significant effects on its own (see Table 3). This null result is consistent with prior work suggesting that follower counts are a poor measure of a user’s influence on Twitter (e.g., Cha et al., 2010). Since retweeter retweets and likes are also highly correlated with each other, we exclude the metric capturing average likes across a candidate’s retweeters. The results are nearly identical when excluding the number of retweets instead. We also exclude the metric capturing likes that candidates themselves received, though specifications that exclude retweets instead show nearly identical results. Alternative specifications, which include average likes across candidate and retweeter tweets, are available in the Online Appendix.

In Table 4, we report estimates from Models 2 and 3 in the left and right panels, respectively. Both specifications include state fixed effects, all control variables, and four Twitter metrics. We do not report estimates of the control variables here for the sake of brevity, but they are available in the Online Appendix. Notably, the number of observations is lower in Model 3 than in Model 2. This discrepancy is due to (1) some states holding Senate races in 2016 but not in 2014, (2) some races in 2014 being uncontested so that vote gaps could not be calculated, and (3) some races featuring a third-party candidate in 1 year but not in another (so that races could not be matched by party across years). Across both specifications, we find strong support for the role of indirect influence. In financially unequal races, poorer candidates see smaller vote gaps in 2016 and compared to their party’s candidate in 2014, when they have indirect influence. The effect of indirect influence is thus not an artifact of geographical variation in social media use.

Table 4.

OLS Models With Interactions Between the Twitter Metrics and Expense Ratio.

	Model 2: Vote Gap 2016				Model 3: Vote Gap 2016–2014
	Baseline		× Expense Ratio		Baseline		× Expense Ratio
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
Constant	44.5**	10.7	−65.6*	25.2	15.4	13.0	−61.3	37.5
Poorer candidate
Average retweets	−0.121	0.179	0.296	0.540	0.137	.423	−0.727	1.94
Average retweets of retweeters	−2.37*	1.02	2.48	1.52	−2.34*	.900	2.00	1.54
Clustering coefficient	−0.227	0.321	0.376	1.97	−0.112	.296	−0.420	1.93
Network constraint	0.112	.174	0.896	1.27	0.201	.212	1.55	1.39
Richer candidate
Average retweets	0.038	0.025	−0.409	0.302	−0.005	.062	−0.216	0.735
Average retweets of retweeters	0.013	0.601	−0.747	2.70	0.307	.501	−2.28	2.65
Clustering coefficient	0.764	0.432	−0.815	1.43	1.11	.665	−1.94	1.81
Network constraint	−0.811	0.816	−0.030	1.89	−0.493	.862	−2.32	2.22
State fixed effects	Yes				Yes
Adjusted R²	.507				.458
Observations	142				111

Note. All control variables included at both the baseline and interaction levels. Estimates of control variables omitted here for brevity but are available in the Online Appendix.

*p < .05. **p < .01.

Total Spending

Next, we address the following question: Why is the effect of indirect influence more pronounced in races that feature large financial inequity between candidates? To address this question, it helps to understand how such inequity arises in the first place. Political scientists have long argued that candidates’ spending decisions reflect their perceived chances of winning (e.g., Jacobson, 1978). As Jacobson (1990) writes, “campaign spending may affect the vote, but the (expected) vote affects campaign contributions, and thus spending, because potential donors give more to candidates in races that are expected to be close” (p. 335). Donations thus pour into races that are perceived to be competitive as political parties attempt to outspend each other. The result of such reactive spending, paradoxically, is that competitors end up spending roughly the same amount. Indeed, in our data set, total spending in a race is highly correlated with the extent of financial equity (i.e., the expense ratio) in the race (ρ = .429).

Financial inequity is thus more likely to arise in races that feature less spending overall. In these races, candidates have less to spend on traditional forms of advertising, opening the door for the relatively inexpensive promotional capabilities provided by social media. We thus expect that social media will be more effective in races that feature lower spending. In our final model specifications, we investigate whether social media mattered more in low spending races. If the effect of indirect influence is more pronounced in low spending races, this would help to explain, at least in part, why such influence is also more pronounced in races that feature larger financial inequity between candidates.

In Table 5, we report estimates of Models 2 and 3 modified to include interactions with total spending instead of the expense ratio. As in Table 4, both specifications include state fixed effects, all control variables, and four Twitter metrics. We do not report estimates of the control variables here for the sake of brevity, but they are available in the Online Appendix. Across both models, we find strong support for the role of indirect influence in low spending races. In low spending races, poorer candidates see smaller vote gaps in 2016, and compared to their party’s candidate in 2014, when they have indirect influence. Moreover, the effect of indirect influence tends to zero as total spending increases, though the interaction only reached marginal significance in the latter specification (p = .053).

Table 5.

OLS Models With Interactions Between the Twitter Metrics and Total Expenses.

	Model 2: Vote Gap 2016				Model 3: Vote Gap 2016–2014
	Baseline		× Total Spent		Baseline		× Total Spent
	Coefficient	SE	Coefficient	SE	Coefficient	SE	Coefficient	SE
Constant	26.8**	9.83	−.602	2.06	12.5	12.4	−11.6**	3.87
Poorer candidate
Average retweets	−0.165	0.190	.013	0.027	−0.012	0.443	0.088	0.125
Average retweets of retweeters	−3.48*	1.44	.528*	0.250	−4.49**	1.34	0.468	0.235
Clustering coefficient	0.037	0.325	−.128	0.238	0.091	0.385	−0.330	0.326
Network constraint	0.149	0.184	.047	0.109	0.061	0.243	0.338*	0.136
Richer candidate
Average retweets	0.024	0.036	−.005	0.016	0.012	0.102	−0.047	0.060
Average retweets of retweeters	−0.637	0.727	.233	0.333	0.197	0.631	−0.022	0.350
Clustering coefficient	0.387	0.636	−.017	0.330	2.28*	0.847	−1.07**	0.380
Network constraint	−0.175	0.890	−.159	0.266	−1.52	0.976	0.453	0.338
State fixed effects	Yes				Yes
Adjusted R²	.505				.468
Observations	142				111

Note. All control variables included at both the baseline and interaction levels. Estimates of control variables omitted here for brevity but are available in the Online Appendix.

*p < .05. **p < .01.

The estimates in Table 5 also point to the role of network influence in how candidates performed relative to their party’s candidate in 2014. In low spending races, richer candidates performed better (i.e., saw larger vote gaps) when their social networks were highly clustered. This effect reverses for high spending races, wherein richer candidates perform better when their networks are structurally diverse (i.e., less clustered). In high spending races, poorer candidates also perform better when their networks are structurally diverse, as is evident from the larger vote gaps associated with a larger network constraint. While we did not theorize about the effects of social media in high spending races, our results suggest that more intricate network dynamics may also underlie how social media shape political competition (Klar & Shmargad, 2017).

Graphical Support

In Figure 5, we depict our findings graphically. In the top panels, the y-axis reflects vote gaps in 2016, while in the bottom panels, it reflects differences between 2016 and 2014. The left and right panels capture the expense ratio and total spending on the x-axis, respectively. Each point represents a different electoral race, with its size reflecting how many retweets retweeters of the poorer candidate in the race received, on average. Gray points capture races where poorer candidates did not have indirect influence, while black points capture races where they did. This split roughly approximates the median of this metric (1.83 across all 142 races; 1.99 across the 111 races where comparisons to 2014 could be made). The lighter and darker bands reflect 95% confidence intervals of the best fitting line for the gray and black points, respectively.

Figure 5.

Vote Gaps by Poorer Candidate’s Indirect Influence.

In the top-left panel, the flattening of the relationship between financial and electoral equality is evident. Indirect influence helps to explain when poorer candidates see smaller vote gaps in races with large financial inequity (i.e., low expense ratios). In the bottom-left panel, we see that these races also saw smaller vote gaps in 2016 than 2014. In the right panels, it appears that no high spending races feature a poorer candidate with indirect influence. Moreover, all of the 11 rightmost points in the top-right panel are gray, each representing a Senate race. The Online Appendix includes a figure that excludes Senate races, where the statistical support is similar across races that feature low and high levels of indirect influence for the poorer candidate.

Still, in the top-right panel, we find visual confirmation that, in low spending races, poorer candidates with indirect influence see smaller vote gaps. In the bottom-right panel, we also find evidence that indirect influence helps to explain when vote gaps were smaller in 2016 than 2014, though the bands do overlap, possibly because there are fewer races here and thus less statistical power. Overall, the patterns of the raw data conform to the estimates of our empirical models, which point to the equalizing effect of indirect influence in elections the feature large financial inequity or low overall spending. They suggest that social media may be particularly useful for candidates who, in some sense, have been forgotten my political parties and donors. These candidates spend significantly less than their competitor and tend to compete in races that do not receive much funding more generally.

Descriptive Analysis

To conclude this section, we discuss candidates and tweets that achieved high levels of indirect influence (i.e., were retweeted by users whose own tweets were widely retweeted). In Table 6, we list the 10 candidates who had the most indirect influence but were outspent by their competitor. The list is roughly split across party, with six Democrats and four Republicans. Candidates from both parties can thus achieve relatively high levels of indirect influence. Of the 10 candidates, only two won their race: Brian Mast and Val Demings, one Republican and one Democrat, who both ran for House seats in Florida. The most influential retweeters for these two candidates, respectively, were @Braveheart_USA (an activist) and @donnabrazile (a political strategist and commentator). Other influential retweeters include then Speaker of House Paul Ryan (@SpeakerRyan) and a political action group that emerged out of Bernie Sanders’ presidential primary campaign (@OurRevolution). Shmargad (2018) provides a detailed account of the full social network of influential retweeters.

Table 6.

Poorer Candidates With the Most Indirect Influence.

Candidate	State–District	Party	Expense Ratio/Total Spent	Vote Gap/Win or Loss	Indirect Influence	Most Influential Retweeter
Morgan Carroll	CO–6	Dem	0.876/6.71 M	8.88/Loss	23.3	@OurRevolution
Brett Murdock	CA–39	Dem	0.021/3.74 M	14.5/Loss	9.33	@CA_Dem
Brian Mast	FL–18	Rep	0.264/13.7 M	−10.9/Win	9.21	@Braveheart_USA
Frank Guinta	NH–1	Rep	0.994/3.10 M	1.54/Loss	7.76	@SpeakerRyan
Jim Reed	CA–1	Dem	0.167/0.944 M	18.1/Loss	6.85	@_HorsesForLife
Kai Degner	VA–6	Dem	0.076/2.16 M	33.6/Loss	5.50	@OurRevolution
Jim Lawrence	NH–2	Rep	0.048/2.28 M	4.65/Loss	5.37	@rpollockDC
Brady Walkinshaw	WA–7	Dem	0.646/4.91 M	12.0/Loss	4.97	@latinovictoryus
Val Demings	FL–10	Dem	0.284/0.065 M	−29.7/Win	3.82	@donnabrazile
Scott Garrett	NJ–5	Rep	0.924/9.08 M	4.51/Loss	3.73	@SpeakerRyan

Note. A candidate’s indirect influence is the average number of retweets their retweeters received on their own tweets. A candidate’s most influential retweeter is that who received the most retweets on their own tweets, on average. All of the candidates in this list competed in House races. Patty Judge of Iowa was the Senate candidate with the most indirect influence (average retweets of retweeters = 2.14). M = million.

In Table 7, we report the most influential tweets from the candidates listed in Table 6. For each candidate, we list all of their tweets that received an influence score of at least 10, implying that their retweeters received at least 10 retweets on their own tweets, on average. For candidates who did not have tweets with an influence score of at least 10, we report their tweet with the most indirect influence. We also include the text of the most influential tweet by each candidate listed. Republicans often pronounced their complaints with the current administration, while Democrats posted about events and encouraged people to vote. Of course, this is a small selection of tweets from a much larger corpus (see Figure 3) and thus may not represent candidate tweets overall. Future work could investigate textual characteristics that are associated with achieving indirect influence. While outside of the scope of the current article, such an analysis would complement existing studies that typically apply measures of direct influence (e.g., number of retweets).

Table 7.

Influential Tweets From Poorer Candidates With the Most Indirect Influence.

Candidate	Tweet ID	Indirect Influence	Text of the Most Influential Tweet Recovered
Morgan Carroll	794237520150073344	76.9	EXCITING NEWS: Senator Bernie Sanders will host a rally for us THIS SATURDAY! Link for the event here—RSVP now! (https://act.myngp.com/Forms/-2459545269794961408)
Brett Murdock	776852891990499328	9.33	Hey @RepEdRoyce, when are you going to debate @Brett_Murdock1?
Brian Mast	780520118082166784	52.7	Shameful. When our Veterans are desperate for our help, the @DeptVetAffairs doesn’t answer the phone (https://apnews.com/0c2c10e0c6b14b28bd5920790884f5d1 #FL18 #sayfie)
	777988548016222208	41.9
	789506829021749248	22.6
	786243524102545408	18.5
Frank Guinta	794932832678412288	69.0	There’s a #BetterWay to fight poverty (https://iqconnect.lmhostediq.com/iqextranet/view_newsletter.aspx?id=101415&c=NH01FG) Read abt @HouseGOP plan in my latest “Weekend Read”
	790968356069371904	46.0
	788006023101607936	10.8
Jim Reed	793519489249792000	6.85	I posted a new video to Facebook (http://fb.me/6bvdpwc08)
Kai Degner	791707858463236096	38.3	NO TWEETS RECOVERED
Kai Degner	793129772767408128	10.6	NO TWEETS RECOVERED
Jim Lawrence	783445782783463424	27.0	I said two years ago that #Obamacare was going to raise costs and it has. We need market-driven solutions. #nh02 #nh02debates #nhpolitics
	791068400122490880	23.0
	783447447406247936	13.7
	794679638039678976	13.6
	794685942455996416	13.6
	794680227901374464	13.6
	794683872453820416	10.8
Brady Walkinshaw	792027330608635904	9.40	Vote today; the climate can’t wait! Learn more about Brady’s environmental leadership (http://bradywalkinshaw.com/environment/)
Val Demings	790519281909846016	10.7	It is the first day of in-person early voting! Join @HillaryforFL @SMurphyCongress & I for a rally in Orlando this am
Scott Garrett	773561242724229120	14.7	I proudly voted for this bill to make sure every survivor of sexual assault is empowered to seek justice (Quoting a tweet by @SpeakerRyan: Last night, the House unanimously passed a bill to protect sexual assault victims. @RepMimiWalters @RepZoeLofgren http://spkrryan.us/2cy6XUK)

Note. A tweet’s indirect influence is the average number of retweets its retweeters received on their own tweets. Only tweets with an indirect influence score of at least 10 are included in this list, unless none exist for the candidate in which case their tweet with the most indirect influence is included.

Discussion

In this article, we investigate how the benefits of social media campaigning are distributed across competitors with financial inequities. While social media provide an inexpensive way for groups to promote their messages (Earl & Kimport, 2011), there is a significant worry that these platforms merely serve to recreate or even exacerbate preexisting inequalities (Schradie, 2018). We contribute to this debate, sometimes called the “normalization vs. equalization argument” (Gibson et al., 2014), by analyzing Twitter use among 2016 U.S. congressional candidates. We find that a specific class of Twitter metrics helps to explain when financially unequal races see relatively close electoral results. These metrics capture the extent that the poorer candidate in the race builds what we call indirect influence by having their tweets shared by users whose own tweets are widely shared. We argue that social media platforms may thus serve an equalizing role by letting candidates broaden their dissemination via two-step flows (Katz & Lazarsfeld, 1955).

These results have implications for how we go about measuring the success of social media campaigns. The success of a campaign is typically evaluated based on the number of times its posts have been liked or shared. Our findings point to another important aspect of a campaign that informs its success—the extent that it achieves indirect, rather than direct, influence. Indeed, electoral outcomes did not vary with metrics tied to candidates’ own tweets, such as the number of times they were retweeted. The metrics of indirect influence we use are simple and could be feasibly constructed by campaigns, in real time and for free, using Twitter’s API. Moreover, they reflect the findings of previous work (e.g., Bakshy et al., 2011), which suggest that successful diffusion arises from support among many users of moderate influence rather than a few highly influential ones. Future research should investigate the types of messages that garner high levels of indirect influence, as these appear to be indicative of a political campaign’s success.

Our findings also have important implications for the competitiveness of U.S. elections. Electoral outcomes in the United States are seldomly close, which have generated some concern about the “vibrancy of American democracy” (Fraga & Hersh, 2018). Our results highlight that social media platforms can provide lower resourced candidates with an edge, especially in races that feature large financial inequity or lower spending overall. In some sense, these races are the ones that have been overlooked by political parties and donors, and social media may serve as an alternative means generating attention in these cases. Of course, it is possible that political tweets which achieve such influence have normatively negative qualities, such as being ideologically extreme or even undemocratic (Tucker et al., 2017). Electoral competition is just one measure of a vibrant democracy, and social media could hurt democracy in other ways. For example, if tweets containing misinformation or disinformation are more likely to be propagated, social media platforms would broaden exposure to these messages as well (e.g., Vosoughi et al., 2018).

There are two limitations to our study that are worth noting. First, while we compare across election years to control for geographic variation in social media use, it is possible that some candidates are simply better at attracting attention and would do so even in the absence of social media. However, given that the influence we uncover requires messages to be shared by many, moderately influential users, the networked nature of social media likely makes them more conducive to this kind of decentralized influence. Second, while prior research suggests that the diffusion of information is usually driven by large broadcasts (Goel et al., 2016), the mechanisms that connect broadcasts to electoral outcomes need further investigation. For example, does indirect influence merely broaden exposure to a candidate’s messages or does it add legitimacy to these messages through social signals? Future work should seek to inform our understanding of the mechanisms that convert social media attention to voting behavior.

Footnotes

Authors’ Note

The code used for snowball sampling and a list of candidate Twitter handles used are available here: .

Data Availability

Data on electoral outcomes, incumbency, and party affiliations of candidates in 2014 and 2016 are available at .

Data on the campaign finance activities of candidates in 2014 and 2016 are available at .

Twitter and other data are available from the first author upon request, subject to Twitter’s Developer Policy ().

Acknowledgments

We are grateful for the feedback we received from participants at the 2017 North American Social Networks conference (NASN), the 2018 Political Networks conference (PolNet), the 2018 International Conference on Computational Social Science (IC2S2), the 2018 APSA pre-conference on Politics and Computational Social Science (PaCSS), the Summer Institute in Computational Social Science at the University of Colorado Boulder, and the brown bag series at the Department of Political Science at the University of Colorado Boulder.

Declaration of Conflicting Interests

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

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Software Information

R was used to collect data from Twitter (with the twitteR package), analyze candidate networks (with the igraph package), and visualize data (with the ggplot2 package). Stata MP Version 13.1 was used for model estimation. A Stata data file and do file replicating the regression analyses are available from the first author.

Supplemental Material

Supplemental material for this article is available online.

References

Allison

Rojanasakul

Harris

Sam

(2016). Tracking the 2016 presidential money race. Bloomberg Politics. Retrieved December 9, 2016, from https://www.bloomberg.com/politics/graphics/2016-presidential-campaign-fundraising/

Bakshy

Hofman

J. M.

Mason

W. A.

Watts

D. J.

(2011, February, 9–12). Everyone’s an influencer: Quantifying influence on twitter [Conference session]. Proceedings of the 4th International Conference on Web Search and Data Mining, Hong Kong, pp. 65–74.

Bode

Epstein

(2015). Campaign Klout: Measuring online influence during the 2012 election. Journal of Information Technology & Politics, 12(2), 133–148.

Bond

R. M.

Fariss

C. J.

Jones

J. J.

Kramer

A. D. I.

Marlow

C. C.

Settle

J. E.

Fowler

J. E.

(2012). A massive scale experiment in social influence and political mobilization. Nature, 489(7415), 295–298.

Burt

(1992). Structural holes: The social structure of competition. Harvard University Press.

Cain

Ferejohn

Fiorina

(1987). The personal vote: Constituency service and electoral independence. Harvard University Press.

Cha

Haddadi

Benevenuto

Gummadi

K. P.

(2010, February 9–12). Measuring user influence in twitter: The million follower fallacy [Conference session]. Proceedings of the 4th International Conference on Web Search and Data Mining, Hong Kong, pp. 10–17.

Choi

(2015). The two-step flow of communication in twitter-based public forums. Social Science Computer Review, 33(6), 696–711.

DiGrazia

McKelvey

Bollen

Rojas

(2013). More tweets, more votes: Social media as a quantitative indicator of political behavior. PLoS One, 8(11), e79449.

10.

Earl

Kimport

(2011). Digitally enabled social change: Activism in the internet age. MIT Press.

11.

Easley

(2016). Exit poll: Clinton bests Trump with ground game. Morning Consult. Retrieved November 8, 2016, from https://morningconsult.com/2016/11/08/clinton-bests-trump-ground-game-per-exit-poll-data/

12.

Fraga

B. L.

Hersh

E. D.

(2018). Are Americans stuck in uncompetitive enclaves? An appraisal of U.S. electoral competition. Quarterly Journal of Political Science, 13(3), 291–311.

13.

Gibson

Römmele

Williamson

(2014). Chasing the digital wave: International perspectives on the growth of online campaigning. Journal of Information Technology & Politics, 11(2), 123–129.

14.

Goel

Anderson

Hofman

Watts

D. J.

(2016). The structural virality of online diffusion. Management Science, 62(1), 180–196.

15.

Golbeck

Grimes

J. M.

Rogers

(2010). Twitter use by the US Congress. Journal of the American Society for Information Science and Technology, 61(8), 1612–1621.

16.

Gulati

G. J.

Williams

C. B.

(2013). Social media and campaign 2012: Developments and trends for Facebook adoption. Social Science Computer Review, 31(5), 577–588.

17.

Hansen

K. M.

Kosiara-Pedersen

(2014). Cyber-campaigning in Denmark: Application and effects of candidate campaigning. Journal of Information Technology & Politics, 11(2), 206–219.

18.

Hilbert

Vásquez

Halpern

Valenzuela

Arriagada

(2017). One step, two step, network step? Complementary perspectives on communication flows in twittered citizen protests. Social Science Computer Review, 35(4), 444–461.

19.

Jacobson

G. C.

(1978). The effects of campaign spending in congressional elections. American Political Science Review, 72(2), 469–491.

20.

Jacobson

G. C.

(1990). The effects of campaign spending in house elections: New evidence for old arguments. American Journal of Political Science, 34(2), 334–362.

21.

Jungherr

Schoen

Posegga

Jürgens

(2017). Digital trace data in the study of public opinion: An indicator of attention toward politics rather than political support. Social Science Computer Review, 35(3), 336–356.

22.

Kam

C. D.

Zechmeister

E. J.

(2013). Name recognition and candidate support. American Journal of Political Science, 57(4), 971–986.

23.

Katz

Lazarsfeld

P. F.

(1955). Personal influence: The part played by people in the flow of mass communications. Transaction Publishers.

24.

Klar

Shmargad

(2017). The effect of network structure on preference formation. The Journal of Politics, 79(2), 717–721.

25.

Koc-Michalska

Gibson

Vedel

(2014a). Online campaigning in France, 2007–2012: Political actors and citizens in the aftermath of the web 2.0 evolution. Journal of Information Technology & Politics, 11(2), 220–244.

26.

Koc-Michalska

Lilleker

D. C.

Surowiec

Baranowski

(2014b). Poland’s 2011 online election campaign: New tools, new professionalism, new ways to win votes. Journal of Information Technology & Politics, 11(2), 186–205.

27.

Littman

(2017). 115th U.S. congress tweet ids. Harvard Dataverse. https://doi.org/10.7910/DVN/UIVHQR

28.

Marcinkowski

Metag

(2014). Why do candidates use online media in constituency campaigning? An application of the theory of planned behavior. Journal of Information Technology & Politics, 11(2), 151–168.

29.

Margolin

D. B.

Goodman

Keegan

Lin

Y.-R.

Lazer

(2016). Wiki-worthy: Collective judgment of candidate notability. Information, Communication and Society, 19(8), 1029–1045.

30.

Morin

(2016). Trump says social media was key to victory. Politico. Retrieved November 12, 2016, from http://www.politico.com/story/2016/11/donald-trump-social-media-231285

31.

Nahon

Hemsley

(2013). Going viral. Polity Press.

32.

Norris

Curtice

(2008). Getting the message out: A two-step model of the role of the internet in campaign communication flows during the 2005 British general election. Journal of Information Technology & Politics, 4(4), 3–13.

33.

Schradie

(2018). The digital activism gap: How class and costs shape online collective action. Social Problems, 65(1), 51–74.

34.

Samuel-Azran

Yarchi

Wolfsfeld

(2015). Equalization versus normalization: Facebook and the 2013 Israeli elections. Social Media + Society, 1(2), 1–9.

35.

Shmargad

(2018). Twitter Influencers in the 2016 US Congressional Races. Journal of Political Marketing, https://doi.org/10.1080/15377857.2018.1513385

36.

Tucker

J. A.

Theocharis

Roberts

M. E.

Barberá

(2017). From liberation to turmoil: Social media and democracy. Journal of Democracy, 28(4), 46–59.

37.

Vosoughi

Roy

Aral

(2018). The Spread of true and false news online. Science, 359(6380), 1146–1151.

38.

Watts

D. J.

Strogatz

S. H.

(1998). Collective dynamics of ‘small-world’ networks. Nature, 393(6684), 440–442.

39.

Wheeler

(2013). Celebrity politics. Polity Press.

40.

Wojcik

Hughes

(2019). Sizing up twitter users. Pew Research Center. Retrieved April 24, 2019, from https://www.pewresearch.org/internet/2019/04/24/sizing-up-twitter-users/