676646

research-article2016
NMS0010.1177/1461444816676646new media & societyFreelon et al.

Article

new media & society

Quantifying the power and

2018, Vol. 20(3) 990­–1011
© The Author(s) 2016
Reprints and permissions:
consequences of social [Link]/[Link]
DOI: 10.1177/1461444816676646
[Link]
media protest [Link]/home/nms

Deen Freelon
American University, USA

Charlton McIlwain
New York University, USA

Meredith Clark
University of North Texas, USA

Abstract
The exercise of power has been an implicit theme in research on the use of social media
for political protest, but few studies have attempted to measure social media power
and its consequences directly. This study develops and measures three theoretically
grounded metrics of social media power—unity, numbers, and commitment—as wielded
on Twitter by a social movement (Black Lives Matter [BLM]), a counter-movement
(political conservatives), and an unaligned party (mainstream news outlets) over nearly
10 months. We find evidence of a model of social media efficacy in which BLM predicts
mainstream news coverage of police brutality, which in turn is the strongest driver of
attention to the issue from political elites. Critically, the metric that best predicts elite
response across all parties is commitment.

Keywords
Black Lives Matter, computational methods, connective action, protest, social
movements, Twitter

Corresponding author:
Deen Freelon, School of Communication, American University, 4400 Massachusetts Avenue NW,
Washington, DC 20016, USA.
Email: dfreelon@[Link]
Freelon et al. 991

Social media (broadly defined) have become essential tools for 21st-century social
movements. Accordingly, the use of social media for political protest is a thriving
research area, with studies applying both qualitative and quantitative methods to under-
stand the nature and magnitude of the phenomenon. Most researchers in this area agree
that social media can be consequential for social movements and their protests in at least
some contexts (Bennett and Segerberg, 2012; Earl and Kimport, 2011; Shirky, 2011).
All successful social movements must exercise power to help bring about their chosen
social goals. Movements have traditionally done so by a number of means, including
protests, petitions, and directly lobbying politicians. Contemporary social movements
such as Black Lives Matter (BLM), which we examine here, consider social media an
important component of their overall strategies. But existing studies have not fully
explored how movements harness power through social media. In particular, they have
not adequately accounted for the fact that social movements are not alone in social
media: other parties interested in the same topic almost always emerge to wield their own
power alongside, against, or orthogonally with respect to the movement.
This article introduces a new methodology to address this reality. It defines several
forms of social media power that are particularly relevant to social movements, proposes
accompanying techniques to measure them, and tests the extent to which they predict a
key movement outcome—elite responses. Critically, non-movement parties may also
wield these forms of power, which are rooted in Charles Tilly’s concept of WUNC (wor-
thiness, unity, numbers, commitment; Tilly, 1999; Tilly and Wood, 2013). Using 40.8 mil-
lion tweets about police shootings of unarmed Black people in 2014 and 2015, we
demonstrate that the digital manifestations of three of WUNC’s four components can be
measured quantitatively for both movement and non-movement constituencies. Our
analysis of the relationships between these metrics and elite response suggests that cer-
tain of the former probabilistically cause the latter.

Social movement power

In an influential article, Diani (1992) outlines four essential components of social move-
ments: “a) networks of informal interaction; b) shared beliefs and solidarity; c) collective
action on conflictual issues; d) action which displays largely outside the institutional
sphere and the routine procedures of social life” (p. 7). This definition would seem to
admit a wide range of structures and tactics, but from the development of resource mobi-
lization theory in the late 1970s until very recently, formal social movement organiza-
tions (SMOs) have been considered all but essential for social movements (Earl and
Kimport, 2011; McCarthy and Zald, 1977). This study draws on contemporary frame-
works that take digitally enabled collective action seriously, in particular Bennett and
Segerberg’s connective action typology (Bennett and Segerberg, 2013; Earl and Kimport,
2011). We contend that connective movements are social movements, sharing all of
Diani’s definitional characteristics despite differing tactics and hierarchies.
It perhaps goes without saying that power is an indispensable resource for social
movements, but as is often the case, the obvious warrants clarification. We define
“power” for the purposes of this article as the capacity to bring about desired changes in
society. This is consistent with the views of a broad range of scholars who view power as
992 new media & society 20(3)

fundamental to all social systems (Bennett, 2003; Castells, 2012; Couldry and Curran,
2003; Giddens, 1987). Giddens (1987), for example, refers to power as “the capability to
intervene in a given set of events so as in some way to alter them” (p. 7). We are con-
cerned primarily with what is often labeled “media power” (Couldry and Curran, 2003),
that is, non-coercive power that flows through various forms of media.
Media power is especially important for connective movements. The assumption that
shifts in discourse may eventually lead to broader social changes underlies every social
movement’s communication efforts. In some cases, changing the conversation about the
issue in question is the ultimate goal. In others, movement-led discussions of social
issues on social media are not ends in and of themselves, but rather one means of address-
ing a larger problem. This is particularly true of movements like BLM whose goals
involve institutional policy change (see Movement for Black Lives, 2016). Among other
uses, social media allow activists to interact with lawmakers directly, given that many if
not most of the latter have Twitter and Facebook accounts (at least in the United States).
While some recent connective movements, most notably Occupy and the Egyptian and
Tunisian revolutions, have explicitly avoided engaging politicians directly (Castells,
2012), doing so is essential to fulfill policy-related goals. Movements pushing for insti-
tutional changes must seek the attention of those in charge, the same as any formal
interest-based organization (Button, 1989: 6; Tarrow, 1998: 34). Elite attention is a key
outcome of power in such cases.
Anyone who has ever observed or participated in a connective movement as it has
pressed its case online knows that it does not operate in a vacuum. Movements fortunate
enough to attract substantial public attention online quickly find themselves among
allies, opponents, journalists, celebrities, curious onlookers, and would-be entertainers
seeking to capitalize on the latest trend. Almost invariably, similar groups of individuals
tend to cluster together in social media, communicating about the topic at hand mostly
within like-minded communities (Adamic and Glance, 2005; Conover et al., 2011;
Hargittai et al., 2008). Each of these communities is involved in a power competition
with the others, whether its participants are aware of it or not (Aouragh, 2012; Kahn and
Kellner, 2004). The simple act of sharing one side’s message rather than another’s is a
key component in this process.
This suggests that when researchers analyze social movements’ power online, they
should not focus solely on the movement. Instead, they should include other collective
interests so that they may be compared. Aside from BLM, two additional interests will
be analyzed here. First, movements with controversial or radical aims often attract coun-
ter-movements dedicated to thwarting them.1 Although social media make confronting
one’s ideological adversaries easier than ever before, few studies have examined online
counter-movements directly (exceptions include Croeser and Highfield, 2014; Jensen
and Bang, 2013). Second, the mass media typically cover movements that achieve a
certain threshold of popularity. True to their ostensibly objective principles, they usually
align neither with movements nor counter-movements consistently and are best consid-
ered “unaligned,” for lack of a better term. While US mainstream news (MN) outlets
exhibit their own distinct ideology (see, for example, Barnhurst, 2005; Reese, 1990),
they do not consistently favor the left or the right. On social media, most high-visibility
movements will likely attract both counter-movements and unaligned observers. Because
Freelon et al. 993

all these communities are embedded with one another in a system of digitally mediated
power relations, they can all potentially command the same forms of power.

Measuring social media power

One of this article’s central claims is that movement power as exerted through social
media can be quantified. Previous studies have found social media to be important for
information sharing, frame building, and/or offline protest facilitation (Bastos et al.,
2015; De Choudhury et al., 2016; González-Bailón et al., 2013; Theocharis, 2013;
Tufekci and Wilson, 2012; Valenzuela et al., 2014). This study differs from this work in
two important ways. First, it demonstrates how abstract concepts of power developed for
offline social movements manifest and change over time in social media. Second, it pre-
sents evidence that these forms of power can help further movements’ policy goals
directly, as opposed to solely facilitating communication among activists.
Our power metrics are based on digital traces of social media activity such as retweets,
hashtags, and screen names. But we will not simply assume without justification that
particular traces signify particular theoretical constructs, as some studies have done
(Freelon, 2014). Instead, we will argue that certain trace-based metrics can be considered
indicators of Tilly’s concept of WUNC (Tilly, 1999; Tilly and Wood, 2013). WUNC is an
acronym whose letters signify worthiness, unity, numbers, and commitment, all essential
elements for social movements to wield adeptly. Tilly describes WUNC as both a set of
defining characteristics of social movements and as a source or index of social movement
power. He associates its elements with movement “strength” and notes that they “increase
the plausibility of the implied threat that the claimant will use its weight to enter, realign,
or disrupt the existing polity” (Tilly, 1999: 262; Vliegenthart and Walgrave, 2012). Thus,
it is no major conceptual leap to consider WUNC as power by our definition.
Tilly conceives of WUNC as a measurable set of properties. He writes of “high” and
“low” values of its four components (Tilly, 1999), which clearly imply possibilities for
quantification. Yet, most empirical applications have been qualitative, with authors
describing how specific social movements’ characteristics fit the WUNC framework
(Agbaria and Mustafa, 2012; Bennett and Segerberg, 2012). For example, in developing
their theory of connective action, Bennett and Segerberg (2012) write that “digitally
mediated action networks often seem to be accorded higher levels of WUNC than their
more conventional social movement counterparts” (p. 742). Again we see a clear sugges-
tion that WUNC can be measured—and in digital contexts no less—but it is followed by
no methodological suggestions as to how.
Of WUNC’s four components, we propose to measure only the latter three. While
quantifying worthiness may be possible, it seems to us prohibitively difficult compared
to unity, numbers, and commitment. Tilly and Wood (2013) give the following offline
examples of worthiness: “sober demeanor; neat clothing; presence of clergy, dignitaries,
and mothers with children” (p. 5).2 Demeanor, clothing, and religious identity on social
media could perhaps be judged by human coders, but not at scale. And because not all
mothers identify themselves as such online, it would likely be impossible to reliably code
social media profiles for motherhood. Fortunately, the remaining three components of
WUNC are much more empirically tractable.
994 new media & society 20(3)

Unity
As a theoretical construct, unity makes a much smoother transition to social media con-
texts than worthiness. Tilly (1999) cites the “wearing or bearing of common symbols
[and] direct affirmation of a common program or identity” (p. 261) as signifiers of unity,
among others. For movements that use social media extensively, few common symbols
are as emblematic as their best-known hashtags. They are the digital analogues of hand-
held signs at street protests. #Jan25, #Occupywallstreet, and #Blacklivesmatter are three
iconic examples that instantly identify their corresponding movements. Creating hashtags
based on victims’ names after police killings is a common practice within BLM, so much
so that participants sometimes speak of their fear of “becoming a hashtag” (Moodie-
Mills, 2015). The names of the most famous victims become metonyms for the everyday
fears of many Black Americans.
Empirically, unity can be expressed through social media as a tendency for a given
community to use a small number of movement-related hashtags disproportionately more
often than others. This indicates that participants are conveying a unified message, par-
ticularly when the hashtag in question expresses a normative claim (e.g. #Blacklivesmatter).
A lack of consensus in hashtag use suggests at a minimum a corresponding lack of unity
in social media messaging, and perhaps also in deeper tactical or philosophical view-
points. Like the other two metrics, hashtag inequality can be measured at the community
level, thus permitting quantitative comparisons.

Numbers
Of WUNC’s four elements, numbers is probably the most straightforward to measure in
social media. Doing so is much easier than in offline protests, where journalistic and
activist estimates of attendance frequently diverge (Mann, 1974). While overall counts
of social media users over time are important, we are more concerned with the specific
numbers of users associated with movements, counter-movements, and unaligned par-
ties. We describe and implement a novel method of doing so in the “Data and Methods”
section below. This method relies on a network analysis technique known as community
detection to categorize users based on their retweeting behavior. We use network com-
munities as the main unit of analysis throughout this article because they intuitively
approximate participants’ tendency to congregate with ideological allies.
Once a set of communities has been identified and labeled, the participants in each
can be counted just as easily as for the entire dataset. Importantly, our method allows us
to aggregate community user counts per day so that longitudinal changes may be
observed. It is perhaps self-evident that, other things being equal and barring purchased
followers, “bots,” and other obfuscatory shenanigans, numbers signify power.

Commitment
Tilly (1999) defines commitment as, among other things, “declarations of readiness to
persevere” (p. 261).3 The longitudinal nature of social media data allows us to improve
upon this operational definition and directly observe perseverance itself. Having first
Freelon et al. 995

disaggregated a social media conversation into multiple communities and then reconsti-
tuted those communities on each individual day, it becomes possible to measure how
committed each community’s participants are. We propose a simple method of doing so:
computing the proportion of participants in a given community on any given day who
tweet at least once during the following 3 days.4 Note that participants do not need to
appear in the same community on the first day as in the next three—they simply need to
post at least one relevant message in the latter.
Comparing this repeat participation rate between communities allows us to deter-
mine which are most and least committed. High proportions indicate that many partici-
pants from a given community are returning to continue promulgating its point of view.
Low proportions, in contrast, indicate a high turnover rate and therefore a less committed
and less stable community. Commitment as expressed in this way sends the message that
movements and their interactants will not disperse (digitally speaking) when the next
trending topic emerges.

BLM
We apply these three power metrics to nearly 10 months of Twitter conversations started
by the BLM movement. Rising to prominence in late 2014, BLM is a loosely coordi-
nated, nationwide movement dedicated to ending police brutality. It takes its name from
a hashtag started by three Black feminist activists—Patrisse Cullors, Alicia Garza, and
Opal Tometi—but the movement and the hashtag are not synonymous. BLM has achieved
national prominence through their online and offline organizing, obtaining extensive
news media coverage and widespread public recognition (Pew Research Center, 2016).
Participants have cited the importance of social media in helping them pursue their goals
(Jackson and Welles, 2016; Stephen, 2015).
BLM is important to study for several reasons. First, it qualifies as an “organization-
ally enabled network” in Bennett and Segerberg’s (2013) typology of connective action.
It operates both online and in the streets, with much of the coordination being handled by
formal organizations such as Million Hoodies, the Black Youth Project, and Ferguson
Action. But these organizations do not directly control the movement—rather, they are
among many groups and individuals that help plan and organize protests and activist
messaging. Second, the movement has succeeded in shifting police brutality from the
margins of American politics to a much more prominent position. Our analysis strongly
suggests that the movement and the news media, rather than the elites who usually con-
trol the political agenda, drove this shift. Third, BLM serves as an apt case to test the
influence of social media activism on policy goals. Unlike the Arab Spring uprisings and
Occupy, which were short on policy demands, BLM’s core demand is simple: “stop kill-
ing us” (Kang, 2015). And while other policy-oriented movements such as the anti–
SOPA/PIPA (Stop Online Piracy Act/ PROTECT IP Act) campaign have used social
media heavily (Benkler et al., 2015), many of these are relatively short-term affairs.
Finally, this study adds to a small but growing collection of studies analyzing BLM and
recent anti-police brutality protests in the United States (Anderson and Hitlin, 2016;
Bonilla and Rosa, 2015; De Choudhury et al., 2016; Gallagher et al., 2016; Jackson and
Welles, 2015, 2016; Kelley, 2015; LeFebvre and Armstrong, 2018; Olteanu et al., 2015).
996 new media & society 20(3)

Research questions
This article will undertake two empirical tasks: (1) measuring social media power using
the metrics described above and (2) testing for associations between them and elite atten-
tion to police killings of unarmed Black citizens. Our dataset features three communities:
one connective social movement (BLM), one counter-movement (Political Conservatives
[PC]), and one unaligned community (MN). There is little theoretical basis for predicting
how these communities are likely to differ from one another on each individual metric,
or which metrics are likely to best predict elite response. If we consider social move-
ments as issue publics (Krosnick, 1990), we might conjecture that they would exercise
the most power in conversations on that issue. However, strong interest in an issue does
not guarantee strength—if movement opponents have greater access to the mass media
or politicians, for example, they may be able to overwhelm even highly enthusiastic
activists. It is also conceivable that MN outlets could draw large numbers of united
onlookers at times when major stories break. The phrasing of the following research
questions reflects these uncertainties:

•• RQ1. How do the three communities compare on each of the three power metrics,
and how do these comparisons change over time?
•• RQ2. How well does each community predict elite responses?
•• RQ3. How well does each metric predict elite responses?
•• RQ4. How often does each community’s distinct users and hashtags appear in elite
tweets?

Data and methods

This study analyzes Twitter data pertaining to police brutality. We purchased from Twitter
all public tweets posted during the yearlong period between 1 June 2014 and 31 May
2015, containing at least one of 45 keywords related to BLM and police killings of Black
people that some perceived as unjustified (see Table 1).5 The keywords consist mostly of
the full and hashtagged names of 20 Black individuals killed by police in 2014 and 2015.
The resulting dataset contains 40,815,975 tweets posted by 4,435,217 unique users.
The names in Table 1 were collated from two sources: a series of tweets posted by
the National Association for the Advancement of Colored People (NAACP) Legal
Defense Fund’s Twitter account (@naacp_ldf) on 3 December 2014 containing the
names of unarmed Black people killed by police between 1999 and 2014, and (2) a 1
May 2015 Buzzfeed article listing a number of unarmed Black males killed by police
in 2014 and 2015 (Quah and David, 2015).6 Neither of these lists is necessarily com-
plete, but they were the most comprehensive we could find. From the NAACP list we
extracted all of the 2014 names, and from the Buzzfeed list we extracted all names
except two, resulting in a combined total of 20 names.7 To these keywords, we added
the hashtags #blacklivesmatter and #ferguson (the birthplace of the movement) and the
phrase “black lives matter.”
We analyzed these tweets using Python and R scripts written by the first author. We
included only tweets posted between 8 August 2014 (the day before Michael Brown was
Freelon et al. 997

Table 1.  Twitter keywords.

Keyword
#ferguson
“michael brown”/“mike brown”/#michaelbrown/#mikebrown
#blacklivesmatter
“eric garner”/#ericgarner
“freddie gray”/#freddiegray
“walter scott”/#walterscott
“tamir rice”/#tamirrice
“black lives matter”
“john crawford”/#johncrawford
“tony robinson”/#tonyrobinson
“eric harris”/#ericharris
“ezell ford”/#ezellford
“akai gurley”/#akaigurley
“kajieme powell”/#kajiemepowell
“tanisha anderson”/#tanishaanderson
“victor white”/#victorwhite
“jordan baker”/#jordanbaker
“jerame reid”/#jeramereid
“yvette smith”/#yvettesmith
“phillip white”/#philipwhite
“dante parker”/#danteparker
“mckenzie cochran”/#mckenziecochran
“tyree woodson”/#tyreewoodson

killed) and 31 May 2015 (the end of our data collection period) because many of the
tweets posted before this period were false positives (e.g. referencing other individuals
named Michael Brown). This 297-day period includes 99.4% of all tweets and 99.1% of
all unique users in the full dataset (40,563,224 tweets; 4,393,926 users).
Next came the task of identifying the like-minded communities on which our analysis
is based. While small-scale studies have identified social media communities manually
(Adamic and Glance, 2005; Hargittai et al., 2008), unsupervised network community
detection algorithms are more effective for larger datasets (Aragón et al., 2013; Conover
et al., 2011). However, most of these methods only generate cross-sectional communi-
ties. Freelon et al. (2015) describe a method of tracking network communities over a
period of months, but it is not effective for smaller time units. Hence, we introduce a
novel method of identifying and tracking social media communities that is equally effec-
tive for all time units.
We began by creating a distinct retweet-based network for each of the 42 weeks of our
dataset, as retweets have been observed to signify ideological affinity in politically ori-
ented Twitter networks (Aragón et al., 2013; Bode et al., 2015; Conover et al., 2011).
Next, we used the Louvain community detection algorithm (Blondel et al., 2008) to sepa-
rate each network into a set of communities characterized by dense retweeting patterns.
998 new media & society 20(3)

When applied to large Twitter networks, Louvain creates a small number of large com-
munities and a large number of small communities, many of which consist of one or two
users retweeting one another. Within each week, we retained the 10 largest communities,
as these are the ones most likely to represent politically consequential constituencies.
This resulted in 420 retweet-based network communities, 10 for each week.
The next major step was to separate the communities into categories based on mem-
bership similarity. To do this, we used Latent Dirichlet allocation (LDA), a popular
method of unsupervised machine classification (also known as topic modeling). We cre-
ated a document-by-term matrix to serve as the input in which the documents were com-
munities and the terms were users. Each user was weighted by network in-degree so that
users who were retweeted more often were considered proportionally more important in
the topic-generation process. Based on these input data, LDA created a series of topics or
collections of network communities with similar memberships. Because LDA requires
researchers to set the number of topics (k) manually and because there are no universal
rules for choosing k, we ran LDA on our data 10 separate times using k values ranging
from 4 to 13. Next, we qualitatively identified three clusters of topics with similar sets of
prominent participants across the 10 LDA runs: one representing BLM (present in all 10
runs), one representing MN (present in eight runs), and one representing PC (present in
seven runs). These were by far the most frequently recurring topics we could identify.
These clusters still needed some winnowing down, in part because certain participants
appeared in more than one cluster. We discarded all communities that appeared in fewer
than half of each cluster’s topics to ensure that only communities that were consistently
placed together in the same topic were retained. We then placed participants appearing in
multiple clusters into the cluster in which they appeared most often, discarding all those
that appeared in at least two clusters equally often.
This entire process yielded three persistent communities—one representing BLM,
one representing MN, and one representing PC—whose participants were consistently
grouped together. As Table 2 shows, the PC community is by far the smallest, while
BLM is only slightly larger than MN. The 10 most retweeted users in each community
demonstrate the face validity of our method: all those in MN are institutional accounts
for MN outlets (including @blackvoices, which is operated by the Huffington Post).
Most of the top PC users are conservative journalists and pundits, while BLM is domi-
nated by anti-brutality activists, most of whom are Black. The three communities overlap
a great deal in terms of hashtag use; #ferguson is the most commonly used hashtag across
all three, and two other hashtags are also present in each community’s top 5 (#blacklives-
matter and #mikebrown).
Our main predictor variables are unity, numbers, and commitment measured on a per-
day, per-community basis. To measure our main outcome variable, elite response, we
manually compiled a list of the Twitter screen names (where available) of the following
elected and appointed government officials:

•• The US President, the First Lady, and all Cabinet members;

•• The official accounts of the Cabinet-level federal agencies (Justice, Labor, State,
etc.);
•• All House and Senate members of the 113th and 114th Congresses;
Freelon et al. 999

Table 2.  Summary statistics for three persistent communities.

BLM (Black Lives MN (Mainstream PC (Political

Matter) News) Conservatives)
Total users 470,655 422,651 260,111
  Rank  
Most 1 @deray @cnnbrk @chris_1791
retweeted 2 @shaunking @cnn @waynedupreeshow
users 3 @bipartisanism @nbcnews @foxnews
4 @antoniofrench @abc @breaking911
5 @nettaaaaaaaa @bbcbreaking @larryelder
6 @michaelskolnik @nytimes @fox2now
7 @ryanjreilly @rt_com @ksdknews
8 @talbertswan @huffingtonpost @amymek
9 @bassem_masri @ac360 @patdollard
10 @youranonglobal @blackvoices @michaelcalhoun
Most 1 #ferguson #ferguson #ferguson
frequently 2 #blacklivesmatter #ericgarner #tcot
used hashtags 3 #mikebrown #blacklivesmatter #blacklivesmatter
4 #ericgarner #mikebrown #mikebrown
5 #freddiegray #freddiegray #michaelbrown
6 #icantbreathe #michaelbrown #ericgarner
7 #michaelbrown #walterscott #freddiegray
8 #tamirrice #icantbreathe #darrenwilson
9 #walterscott #baltimore #baltimore
10 #baltimore #fergusondecision #pjnet

•• All US governors in office during the dataset time span;

•• The lieutenant governors, attorneys general, and members of the state legislatures
of Maryland, Missouri, New York, and Ohio, the states in which five of the six
most discussed incidents occurred;8
•• The mayors and top local prosecutors for the cities in which the above five inci-
dents occurred.

This list contains 1498 screen names, of which 298 (20%) tweeted at least once in the
data. These 298 users contributed 2524 total tweets. In total, 169 names appeared in one
or another of the three persistent communities; these were removed from the communi-
ties prior to analysis.

Results
RQ1 calls for a comparison between the three communities in terms of the three metrics.
We begin with numbers, the most easily interpreted metric. Figure 1 displays the number
of unique users from each community per day. While participation from each community
spikes at the same times, BLM is nearly always the largest. Interestingly, PC is usually
1000 new media & society 20(3)

Figure 1.  Numbers over time.

more active than MN on non-peak days, but MN tends to surpass it when attention
focuses on a major event such as a killing or a major legal decision. Participation from
all three communities spike around peak periods, but this effect is stronger proportion-
ally for MN than for PC.
For the unity metric, which we operationalize as the Gini coefficient of hashtags
used by each community, Figure 2 reveals substantial differences. BLM is consistently
more unified than PC, which is more unified than MN. In other words, BLM’s hashtag
use was more concentrated among a smaller number of hashtags than were the other two
communities. MN’s unity values also fluctuate far more than PC’s or BLM’s: the vari-
ance of its daily Ginis is 0.0072, while PC’s variance is 0.0021 and BLM’s is 0.0014.
These numbers quantify the disparities in longitudinal variation that can be seen clearly
in Figure 2.
Figure 3 shows the longitudinal changes in each community’s repeat participation
rates (i.e. commitment), which are simply the proportions of unique users on any
given day that post at least once during the following 3 days. Those paying attention
primarily to MN are the least committed, with rates that usually fall below 0.25. BLM
and PC are both higher than MN during non-peak periods, with BLM usually slightly
higher than PC.
To summarize briefly before proceeding, BLM definitively exceeds the other two
communities on all three power metrics most of the time. PC generally comes in second
and MN third. Spikes in attention seem to result in sharp increases of all three metrics for
all three communities.
To answer RQ2 and RQ3, we estimate Granger causalities between the nine commu-
nity/metric variables and the daily number of elite tweets. Extended discussions of this
Freelon et al. 1001

Figure 2.  Unity over time.

technique’s logic and value for communication research are available elsewhere (Bastos
et al., 2015; Neuman et al., 2014), so we will not repeat them here. Instead, we offer a
highly condensed description: variable X Granger-causes variable Y if past values of X
enable more accurate predictions of Y than past values of Y alone. (This should not be
confused with commonsense notions of causality.) Granger causality can be estimated by
computing one vector autoregression (VAR) model in which prior values of outcome vari-
able Y are the sole predictors and a second model in which prior values of an independent
predictor X are added to the first model. If the ratio of the variance of the first model’s
error term to that of the second is sufficiently greater than 1, we conclude that X Granger-
causes Y. After examining models with lags of 1–5 days using Breusch–Godfrey tests, we
chose a 4-day lag for all models because it yielded the lowest levels of autocorrelation.
Despite this, we were unable to completely eliminate autocorrelation in some models.
Our Granger analysis proceeds in two steps. First, we estimate bidirectional Granger
causalities between daily elite tweet (DET) counts and each of the nine community/met-
ric variables.9 We call these direct Granger causes because there are no intermediate vari-
ables between them and the outcome. Second, we examine the extent to which each of
these nine variables Granger-causes one another. We call these indirect Granger causes.
Our results are summarized in Table 3, which requires some explanation. The coeffi-
cients in the second column from the right and the second column from the left are
F-statistics giving the ratio described in the preceding paragraph, which indicate the
magnitude of the reduction in error term variance occasioned by the corresponding vari-
able. Each F-statistic is one of a pair: for the second column from the right, arrows point-
ing right indicate metric-to-DET Granger causality, while those pointing left indicate
DET-to-metric Granger causality (i.e. reverse Granger causality). The F-statistic with the
1002 new media & society 20(3)

Figure 3.  Commitment over time.

greater value in each pair is indicated in bold. The variables in the middle column are the
direct Granger causes of DET, while the leftmost column contains all statistically signifi-
cant indirect Granger causes and their corresponding reverse Granger causes.
The first important finding Table 3 reveals is that elites are clearly following the cues of
the communities, as opposed to the reverse. The magnitudes of the metric-to-DET F-statistics
for all nine direct Granger causes are much greater than those of their DET-to-metric coun-
terparts. This is clear evidence that direct Granger causality overwhelmingly runs in one
direction. In answer to RQ2, comparing the direct variables, MN is the clear leader in elicit-
ing elite responses as well as the direct cause least affected by autocorrelation. MN commit-
ment metrics are the first-, fourth-, and sixth-strongest direct Granger causes of DET, while
BLM’s are the second-, fifth-, and eighth-strongest direct Granger causes of DET. PC exerts
the weakest direct influence, coming in at third, seventh, and ninth places.
Examining the significant forward and reverse indirect Granger causes generally sup-
ports this story. One concurring finding is that the significant indirect causes of MN
commitment are fairly modest, with BLM clearly stronger than PC. But MN commit-
ment is a much stronger cause of BLM numbers than the opposite, which may indicate
the power of the media to draw users to BLM during periods of high attention. It is a
weaker, though still relatively strong, cause of PC numbers. BLM commitment seems to
exert some influence on MN numbers, but the equation is autocorrelated, reducing con-
fidence in its coefficient.
Turning to RQ3, which concerns the metrics of the greatest predictive capacity, com-
mitment emerges as the leader. The commitment metrics are the top 3 direct Granger
causes and 14 of the 23 significant indirect ones. Also, some of the strongest indirect
Granger causes by F-statistic magnitude are commitment metrics. As for the other power
metrics, numbers is a stronger direct Granger cause than unity, as it appears twice (in the
Freelon et al. 1003

Table 3.  Direct and indirect Granger causes of daily elite tweets (4-day lag).

Indirect Granger cause Indirect F Direct Granger cause Direct F  

BLM numbers → 4.32 MN commitment → 24.52  
← 71.1 ← 0.98
BLM commitment → 3.79 DET
← 2.71 (Daily elite
PC numbers → 3.02   tweets)
← 42.7
MN numbers† → 7.32 BLM commitment† → 17.56
← 83.9 ← 0.69
MN unity → 7.34
← 47.1
MN commitment → 2.71  
← 3.79
MN numbers → 2.87 PC commitment† → 10.04
← 28.3 ← 1.04
BLM commitment† → 83.9 MN numbers → 9.80
← 7.33 ← 1.02
PC commitment → 28.3
← 2.87
BLM numbers† → 3.34  
← 1.46
BLM unity† → 2.41  
← 0.44
MN commitment → 71.1 BLM numbers → 8.19
← 4.33 ← 0.57
PC commitment → 22.0
← 1.90
BLM commitment → 47.14 MN unity → 6.50
← 7.34 ← 0.67
PC commitment → 10.93
← 1.10
BLM unity† → 3.14
← 0.26
BLM numbers → 2.59
← 0.32
MN commitment → 42.66 PC numbers → 5.01
← 3.02 ← 1.35
BLM commitment → 32.76
← 1.83
MN commitment† → 16.34 BLM unity → 4.74
← 1.51 ← 0.89
PC commitment → 7.98
← 2.09
BLM commitment → 15.04 PC unity† → 2.97
← 0.62 ← 1.40
MN commitment → 10.75
← 0.72

MN: mainstream news; BLM: Black Lives Matter; PC: Political Conservatives.
F values above 2.4 = p < .05; above 3.35 = p < .01; above 4.73 = p < .001. Daggers indicate autocorrelated equa-
tions (Breusch–Godfrey p < .05). To reduce repetition, only the significant (p < .05) indirect causes of each
direct cause are listed.
1004 new media & society 20(3)

Figure 4.  Total and unique users mentioned by political elites.

fourth and fifth spots) before unity appears once. Numbers surpasses unity as an indirect
cause, appearing twice as often and generally with slightly higher F-statistics.
RQ4, which concerns how often each community’s users and hashtags appear in elite
tweets, can be answered using basic computational techniques. For users, we simply
counted the numbers of unique and total screen names mentioned by elites that belonged
to each persistent community. Figure 4 shows that BLM users are mentioned more often
by elites whether unique or total users are considered. MN users are mentioned slightly
more often than PC in each case.
The hashtag analysis is more complicated because the most popular hashtags were
used extensively by all three communities (see Table 2). Therefore, we created a list of
hashtags used disproportionately more often by each community compared to the other
two. We call these each community’s distinctive hashtags. To ensure that our results were
not specific to the choice of a single disproportion constant, we used two, examining
hashtags used by a given community’s participants in proportions at least 1.5 times and
two times greater than the other two. Figure 5 shows how often elites used each com-
munity’s distinctive hashtags at both the 1.5× and 2× levels. In both cases, BLM achieves
only a slight advantage over the next-ranked community. But at the 1.5× level, PC comes
in second, while MN occupies that position at the 2× level. Figure 5 suggests that
although all three communities are sensitive to the choice of disproportion constant,
BLM’s presence is felt most consistently.

Discussion
This study presents convergent, highly suggestive evidence of power as projected through
social media by a connective social movement and two competing communities. It intro-
duces three theoretically derived, movement-relevant metrics of social media power;
Freelon et al. 1005

Figure 5.  Distinct hashtags mentioned by political elites.

measures them longitudinally over the course of nearly 300 days; and estimates the extent
to which they Granger-cause elite responses. Our results indicate that unaligned news
outlets and their audiences are more successful than the other two communities in provok-
ing elite responses. We also find modest but convergent evidence that BLM helped to
generate the media attention in the first place (we present further such evidence in Freelon
et al. [2016]).
These results contribute a novel answer to a central question in the literature on digi-
tally enabled social movements: How, if at all, does social media use contribute to move-
ment goals? We demonstrate for the first time that social movements can attract elite
attention via social media as their concerns are broadcast through news outlets. This
finding is consistent with evidence that offline activism can influence elites through
media coverage (Vliegenthart and Walgrave, 2012; Walgrave and Vliegenthart, 2012).
Of our three power metrics, commitment is by far the strongest. Its overall effects are
stronger than either numbers (as measured by the total number of individuals tweeting on
any given day) or unity (as measured through hashtag use).
While Granger causality is not “true” causality, our method definitively fulfills two of
the three criteria for causal inference and substantially, although incompletely, addresses
the third. Causal inference is widely considered to be valid when three criteria are
obtained: correlation, time precedence, and the elimination of alternative explanations
(Babbie, 2012: 93–94; Vogt and Johnson, 2015: 55). Granger causality demonstrates cor-
relation through the VAR models on which it is based and time precedence through its
use of lagged predictors. And while it cannot eliminate all potential rival explanations, it
can account for some of the most obvious ones. First, reversing the Granger causal order
of each pair of variables tests for the presence of reverse and bidirectional causation.
While this occurs to some degree in our results, in most instances Granger causality is
1006 new media & society 20(3)

much stronger in one direction than in the other. Second, we test the possibility that nine
different variables may directly Granger cause DET, some of which turn out to be much
more consequential than others. Third, we examine indirect causes to account for the
possibility of a multistep causal process. These measures add additional support, though
not definitive proof, of a probabilistic causal interpretation.
That said, the absence of non-Twitter variables is this article’s chief limitation and
may have caused some of the autocorrelation in the VAR models. It is likely that the
political elites were motivated to speak out on this issue through a number of channels,
with Twitter being only one. Other media channels, letters and phone calls from con-
stituents, conversations with colleagues, and events occurring in one’s district are a
few of the plausible possibilities. But the fact that news outlets sourced much of their
reporting on police killings in 2014 and 2015 from social media (Freelon et al., 2016)
supports our multistep model of online protest power. It is also impossible to com-
pletely separate the influence of offline protests from protest tweets, given that they
spiked around the same times. However, elites’ extensive use of relevant hashtags and
mentions of movement-associated participants and media outlets support the notion
that the tweets had some impact. Further research that includes additional variables
may well discover new causes.
The metrics of social media power we have introduced may exhibit predictive power
in other studies, but they are interesting in and of themselves. Although unity (as opera-
tionalized through hashtags) proved to be the least powerful metric in our Granger analy-
sis, it has the potential to contribute to the voluminous literature on collective action
frames (Corrigall-Brown and Wilkes, 2011; Sanfilippo et al., 2008; Snow et al., 1986).
Given its long-standing status as a key concern in studies of offline protest (McCarthy
et al., 1996; Soule and Earl, 2005), numbers will likely remain so in social media con-
texts. And among its other potential uses, commitment in the form of repeat participation
rates is a new method of examining “serial activism” in social media (Bastos et al., 2013;
Bastos and Mercea, 2015).
We also contribute a computationally tractable method of identifying and tracking
distinct Twitter communities over time. While community detection is relatively straight-
forward for cross-sectional research, it is far less so for longitudinal studies. As a result,
the predominant cross-sectional approaches typically used in network studies have been
unable to analyze much of theoretical interest in social media, which generate inherently
longitudinal data. Our method creates persistent communities whose variables (unity,
numbers, commitment, etc.) can be measured at any desired level of time granularity. Its
utility is not limited to the study of social media power: it can be applied to any large-N
Twitter conversation in which multiple distinct communities participate.
This study’s two main contributions go hand in hand: a falsifiable model of social
media power as exercised by social movements and others interested in a given issue,
and an innovative methodology for measuring it. Future studies may use our methods to
investigate the extent to which the model applies to other social movements. We might
expect that connective movements with similar characteristics to BLM—situated within
an advanced democracy, led by marginalized but tech-savvy youth, and eager for policy
change—may use Twitter to similar effect. But it may or may not apply equally well to
other platforms or types of movements. Nevertheless, the finding that social movements
Freelon et al. 1007

can, under certain circumstances, further policy-relevant goals directly through tweeting
is one with powerful theoretical and practical implications.

Funding
The author(s) disclosed receipt of the following financial support for the research, authorship and/
or publication of this article: The authors gratefully acknowledge funding support from the Spencer
Foundation, grant #201600019.

Notes
1. Counter-movements “make competing claims on the state on matters of policy and politics
and vie for attention from the mass media and the broader public” (Meyer and Staggenborg,
1996: 1632) and, when opposing left-wing movements (as does the one analyzed here), “seek
to maintain the currently dominant field frame and thus maintain the status quo by opposing,
or countering, the efforts of movements seeking change” (Brulle, 2014: 683).
2. In addition to these methodological considerations, Black Lives Matter (BLM) would likely
condemn this conception of worthiness as counterproductive “respectability politics” (Smith,
2014).
3. See also Klandermans (1997), who concurs that “the more committed to a movement some-
one is, the more likely it is that he or she will continue to participate” (p. 29).
4. We chose a 3-day period to strike a balance between a week, which we felt would be too
liberal, and 1 day, which would be too conservative.
5. Data purchased from Twitter includes all public tweets matching the buyer’s search criteria,
which is not guaranteed when collecting data from the platform’s Application Programming
Interfaces (Jackson and Welles, 2016; LeFebvre and Armstrong, 2018).
6. The first tweet in this series is here: [Link]
7. Two names (Dontre Hamilton and Rumain Brisbon) were omitted from our final list due to a
clerical error.
8. We originally included the states where the five most discussed killings occurred, but since
the fifth and sixth most discussed killings (Tamir Rice and John Crawford, respectively) both
took place in Ohio, we decided to include it instead of South Carolina, where the fourth most
discussed killing (Walter Scott) occurred.
9. Daily elite tweets (DET) and the three numbers metrics were transformed prior to analy-
sis using the inverse hyperbolic sine function (Burbidge et al., 1988) to satisfy the Granger
method’s assumptions of normality and stationarity. The unity and commitment metrics were
not transformed because they are already normalized.

References
Adamic LA and Glance N (2005) The political blogosphere and the 2004 US election: divided
they blog. In: Proceedings of the 3rd International Workshop on Link Discovery, Chicago,
IL, 21–24 August, pp. 36–43. New York: ACM.
Agbaria AK and Mustafa M (2012) Two states for three peoples: the “Palestinian-Israeli” in the
Future Vision Documents of the Palestinians in Israel. Ethnic and Racial Studies 35(4):
718–736.
Anderson M and Hitlin P (2016) Social Media Conversations About Race: How Social Media
Users See, Share and Discuss Race and the Rise of Hashtags Like #BlackLivesMatter. Pew
Research Center. Available at: [Link]
Race-and-Social-Media_FINAL.pdf
1008 new media & society 20(3)

Aouragh M (2012) Social media, mediation and the Arab revolutions. TripleC: Communication,
Capitalism & Critique 10(2): 518–536.
Aragón P, Kappler KE, Kaltenbrunner A, et al. (2013) Communication dynamics in Twitter during
political campaigns: the case of the 2011 Spanish national election. Policy & Internet 5(2):
183–206.
Babbie E (2012) The Practice of Social Research. Belmont, CA: Cengage Learning.
Barnhurst KG (2005) News ideology in the twentieth century. In: Pöttker H and Høyer S (eds)
Diffusion of the News Paradigm 1850–2000. Gothenburg: Nordicom, pp. 239–262.
Bastos MT and Mercea D (2015) Serial activists: political Twitter beyond influentials and the twit-
tertariat. New Media & Society 18(10): 2359–2378.
Bastos MT, Mercea D and Charpentier A (2015) Tents, tweets, and events: the interplay between
ongoing protests and social media. Journal of Communication 65(2): 320–350.
Bastos MT, Puschmann C and Travitzki R (2013) Tweeting across hashtags: overlapping users and
the importance of language, topics, and politics. In: Proceedings of the 24th ACM Conference
on Hypertext and Social Media (HT ’13), Paris, 1–3 May, pp. 164–168. New York: ACM.
Available at: [Link] (accessed 25 June 2015).
Benkler Y, Roberts H, Faris R, et al. (2015) Social mobilization and the networked public sphere:
mapping the SOPA-PIPA debate. Political Communication 32(4): 594–624.
Bennett WL (2003) New media power. In: Couldry N and Curran J (eds) Contesting Media Power.
Lanham, MD: Rowman & Littlefield, pp. 17–37.
Bennett WL and Segerberg A (2012) The logic of connective action: Digital media and the per-
sonalization of contentious politics. Information, Communication & Society 15(5): 739–768.
Bennett WL and Segerberg A (2013) The Logic of Connective Action: digital Media and the
Personalization of Contentious Politics. Cambridge: Cambridge University Press.
Blondel VD, Guillaume J-L, Lambiotte R, et al. (2008) Fast unfolding of communities in large
networks. Journal of Statistical Mechanics: Theory and Experiment 2008(10): P10008.
Bode L, Hanna A, Yang J, et al. (2015) Candidate networks, citizen clusters, and political expres-
sion strategic hashtag use in the 2010 midterms. Annals of the American Academy of Political
and Social Science 659(1): 149–165.
Bonilla Y and Rosa J (2015) #Ferguson: digital protest, hashtag ethnography, and the racial poli-
tics of social media in the United States. American Ethnologist 42(1): 4–17.
Brulle RJ (2014) Institutionalizing delay: foundation funding and the creation of U.S. climate
change counter-movement organizations. Climatic Change 122(4): 681–694.
Burbidge JB, Magee L and Robb AL (1988) Alternative transformations to handle extreme
values of the dependent variable. Journal of the American Statistical Association 83(401):
123–127.
Button JW (1989) Blacks and Social Change: Impact of the Civil Rights Movement in Southern
Communities. Princeton, NJ: Princeton University Press.
Castells M (2012) Networks of Outrage and Hope: Social Movements in the Internet Age. Malden,
MA: Polity Press.
Conover MD, Ratkiewicz J, Francisco M, et al. (2011) Political polarization on Twitter. In:
Proceedings of the 5th International Conference on Weblogs and Social Media, Barcelona,
17–21 July, pp. 89–96. Palo Alto, CA: AAAI.
Corrigall-Brown C and Wilkes R (2011) Picturing protest: the visual framing of collective action
by first nations in Canada. American Behavioral Scientist 56(2): 223–243.
Couldry N and Curran J (2003) The paradox of media power. In: Couldry N and Curran J (eds)
Contesting Media Power: Alternative Media in a Networked World. Lanham, MD: Rowman
& Littlefield, pp. 3–15.
Freelon et al. 1009

Croeser S and Highfield T (2014) Occupy Oakland and #oo: uses of Twitter within the occupy
movement. First Monday 19(3). Available at: [Link]
view/4827 (accessed 16 June 2016).
De Choudhury M, Jhaver S, Sugar B, et al. (2016) Social media participation in an activist move-
ment for racial equality. In: Proceedings of the Tenth International AAAI Conference on Web
and Social Media, Cologne, 17–20 May 2016. Available at: [Link]
BLM_ICWSM16.pdf (accessed 15 June 2016).
Diani M (1992) The concept of social movement. Sociological Review 40(1): 1–25.
Earl J and Kimport K (2011) Digitally Enabled Social Change: Activism in the Internet Age.
Cambridge, MA: The MIT Press.
Freelon D (2014) On the interpretation of digital trace data in communication and social comput-
ing research. Journal of Broadcasting & Electronic Media 58(1): 59–75.
Freelon D, Lynch M and Aday S (2015) Online fragmentation in wartime: a longitudinal analysis
of tweets about Syria, 2011–2013. Annals of the American Academy of Political and Social
Science 659(1): 166–179.
Freelon D, McIlwain CD and Clark MD (2016) Beyond the hashtags: #Ferguson, #Blacklivesmatter,
and the online struggle for offline justice. Center for Media and Social Impact, American
University. Available at [Link]
Gallagher RJ, Reagan AJ, Danforth CM, et al. (2016) Divergent discourse between protests and
counter-protests: #BlackLivesMatter and #AllLivesMatter. Available at: [Link]
abs/1606.06820 (accessed 18 September 2016).
Giddens A (1987) A Contemporary Critique of Historical Materialism: The Nation-State and
Violence. Berkeley, CA: University of California Press.
González-Bailón S, Borge-Holthoefer J and Moreno Y (2013) Broadcasters and hidden influen-
tials in online protest diffusion. American Behavioral Scientist 57(7): 943–965.
Hargittai E, Gallo J and Kane M (2008) Cross-ideological discussions among conservative and
liberal bloggers. Public Choice 134(1): 67–86.
Jackson SJ and Welles BF (2015) Hijacking #myNYPD: social media dissent and networked coun-
terpublics. Journal of Communication 65(6): 932–952.
Jackson SJ and Welles BF (2016) #Ferguson is everywhere: initiators in emerging counterpublic
networks. Information, Communication & Society 19(3): 397–418.
Jensen MJ and Bang HP (2013) Occupy wall street: a new political form of movement and com-
munity? Journal of Information Technology & Politics 10(4): 444–461.
Kahn R and Kellner D (2004) New media and internet activism: from the “Battle of Seattle” to
blogging. New Media & Society 6(1): 87–95.
Kang JC (2015) ‘Our demand is simple: stop killing us’. The New York Times, 4 May. Available
at: [Link]
html (accessed 22 January 2016).
Kelley RDG (2015) Beyond ‘Black Lives Matter’. Kalfou 2(2). Available at: [Link]
[Link]/[Link]/kalfou/article/view/71 (accessed 4 February 2016).
Klandermans B (1997) The Social Psychology of Protest. 1st ed. Oxford; Cambridge, MA: Wiley-
Blackwell.
Krosnick JA (1990) Government policy and citizen passion: a study of issue publics in contempo-
rary America. Political Behavior 12(1): 59–92.
LeFebvre RK and Armstrong C (2018) Grievance-based social movement mobilization in the
#Ferguson Twitter storm. New Media & Society 20(1): 8–28.
McCarthy JD and Zald MN (1977) Resource mobilization and social movements: a partial theory.
American Journal of Sociology 82(6): 1212–1241.
1010 new media & society 20(3)

McCarthy JD, McPhail C and Smith J (1996) Images of protest: dimensions of selection bias
in media coverage of Washington demonstrations, 1982 and 1991. American Sociological
Review 61(3): 478–499.
Mann L (1974) Counting the crowd: effects of editorial policy on estimates. Journalism & Mass
Communication Quarterly 51(2): 278–285.
Meyer DS and Staggenborg S (1996) Movements, countermovements, and the structure of politi-
cal opportunity. American Journal of Sociology 101(6): 1628–1660.
Moodie-Mills D (2015) OpEd: are we all just one bullet away from becoming a hashtag? NBC
News. Available at: [Link]
all-just-one-bullet-away-becoming-hashtag-n338111 (accessed 27 February 2016).
Movement for Black Lives (2016) A vision for black lives: policy demands for black power, free-
dom & justice. Available at: [Link]
Neuman WR, Guggenheim L, Jang SM, et al. (2014) The dynamics of public attention: agenda-
setting theory meets big data. Journal of Communication 64(2): 193–214.
Olteanu A, Weber I and Gatica-Perez D (2015) Characterizing the demographics behind the
#BlackLivesMatter Movement. Available at: [Link] (accessed 19
January 2016).
Pew Research Center (2016) On Views of Race and Inequality, Backs and Whites are Worlds
Apart. Pew Research Center. Available at: [Link]
ST_2016.06.27_Race-[Link]
Quah N and David LE (2015) Here’s a Timeline of Unarmed Black People Killed By Police Over
Past Year. BuzzFeed. Available at: [Link]
line-of-unarmed-black-men-killed-by-police-over (accessed 22 January 2016).
Reese SD (1990) The news paradigm and the ideology of objectivity: a socialist at the Wall Street
Journal. Critical Studies in Media Communication 7(4): 390–409.
Sanfilippo A, Franklin L, Tratz S, et al. (2008) Automating frame analysis. In: Liu H, Salerno JJ
and Young MJ (eds) Social Computing, Behavioral Modeling, and Prediction. New York:
Springer, pp. 239–248.
Shirky C (2011) The political power of social media: technology, the public sphere, and political
change. Foreign Affairs 90: 28.
Smith M (2014) Affect and respectability politics. Theory & Event 17(3). Available at: https://
[Link]/article/559376 (accessed 18 September 2016).
Snow DA, Rochford EB Jr, Worden SK, et al. (1986) Frame alignment processes, micromobiliza-
tion, and movement participation. American Sociological Review 51(4): 464–481.
Soule S and Earl J (2005) A movement society evaluated: collective protest in The United States,
1960–1986. Mobilization 10(3): 345–364.
Stephen B (2015) How Black Lives Matter Uses Social Media to Fight the Power. WIRED.
Available at: [Link]
fight-the-power/ (accessed 27 February 2016).
Tarrow S (1998) “The very excess of democracy”: state building and contentious politics in
America. In: Costain AN and McFarland AS (eds) Social Movements and American Political
Institutions. Lanham, MD: Rowman & Littlefield, pp. 20–38.
Theocharis Y (2013) The wealth of (occupation) networks? Communication patterns and informa-
tion distribution in a Twitter protest network. Journal of Information Technology & Politics
10(1): 35–56.
Tilly C (1999) Conclusion: from interactions to outcomes in social movements. In: Giugni M,
McAdam D and Tilly C (eds) How Social Movements Matter. Minneapolis, MN: University
of Minnesota Press, pp. 253–270.
Tilly C and Wood LJ (2013) Social Movements 1768–2012. New York: Routledge.
Freelon et al. 1011

Tufekci Z and Wilson C (2012) Social media and the decision to participate in political protest:
observations from Tahrir Square. Journal of Communication 62(2): 363–379.
Valenzuela S, Arriagada A and Scherman A (2014) Facebook, Twitter, and youth engagement:
a quasi-experimental study of social media use and protest behavior using propensity score
matching. International Journal of Communication 8: 25.
Vliegenthart R and Walgrave S (2012) The interdependency of mass media and social movements.
In: Semetko H and Scammell M (eds) The Sage Handbook of Political Communication.
Thousand Oaks, CA: SAGE, pp. 387–398.
Vogt WP and Johnson RB (2015) The SAGE Dictionary of Statistics & Methodology: a
Nontechnical Guide for the Social Sciences. Thousand Oaks, CA: SAGE.
Walgrave S and Vliegenthart R (2012) The complex agenda-setting power of protest: demonstra-
tions, media, parliament, government, and legislation in Belgium, 1993–2000. Mobilization
17(2): 129–156.

Author biographies
Deen Freelon is an associate professor of communication at American University.
Charlton McIlwain is an associate professor of media, culture and communication and Associate
Dean for Faculty Development and Diversity at New York University.
Meredith Clark is an assistant professor of digital and print news at the University of North Texas.