Public Relations Review 44 (2018) 549–561

Contents lists available at ScienceDirect

Public Relations Review

journal homepage: www.elsevier.com/locate/pubrev

Disentangling social media influence in crises: Testing a four-factor

T
model of social media influence with large data
⁎
Xinyan Zhaoa, , Mengqi Zhanb, Brooke F. Liuc
a
Hong Kong Baptist University, Hong Kong
b
University of Texas at Arlington, United States
c
University of Maryland, United States

A R T IC LE I N F O ABS TRA CT

Keywords: Social media empower publics by providing a platform for their voices during crises. Digital-
Crisis communication enabled platforms allow individuals to become influentials by sharing their insights and expertise
Social media influence with others. Confronted with the fast-paced and complex dynamics of crises, we lack a systematic
Measurement conceptualization and a valid measure of social media influence in the crisis context. By in-
Influencer
tegrating diverse perspectives on influence, we propose a new framework that theorizes different
dimensions of social media influence based on publics’ communicative behaviors during crises.
This integrated framework offers a refined conceptualization and measurement of social media
influence in crises by incorporating the network perspective. We tested the framework with large-
scale Twitter data from four crises. Results from multigroup CFA on Twitter influencers suggest
that social media influence is composed of four factors: output, reactive outtake, proactive
outtake, and network positioning. Each factor is associated with a distinct set of users’ behavioral
indicators (e.g., retweet). Implications for crisis communication and public relations are dis-
cussed.

1. Theoretical framework

Social media have reshaped how public relationspractitioners and researchers understand publics, especially during organiza-
tional crises. Today’s media landscape includes the continued decline of media as gatekeepers, the waning of organizations’ dom-
inance in their crisis response discourse, and the empowerment of publics online during crises (e.g., Frandsen & Johansen, 2016).
Digital media platforms like Twitter, Instagram, and Snapchat allow individuals to establish their roles as influentials by “sharing
opinions, insights, experiences and perspectives with others” (Marken, 2007, p. 10). Social media platforms have emerged as integral
communities for publics to voice their opinions toward organizations in crises.
Paralleling the surge of social media in crises, there has been a growing body of research examining influence on social media
(e.g., Freberg, Graham, McGaughey, & Freberg, 2011; Jin & Liu, 2010; Sommerfeldt & Kent, 2015), particularly engagement (e.g.,
Men & Tsai, 2013; Smith & Taylor, 2017). For example, crisis communication theories such as the social-mediated crisis commu-
nication theory (SMCC) discuss the influencers’ attributes of involvement and motivation (e.g., Jin & Liu, 2010). This growing body of
literature has provided valuable insights to our understanding of social media influence. Yet, we know much less about how social
media influence unfolds in an ever-changing and decentralized environment such as crises on social media. The field would benefit
greatly from a systematic conceptualization and valid measurement of social media influence in the crisis communication context.

⁎
Corresponding author at: School of Communication, Hong Kong Baptist University, Hong Kong.
E-mail address: zhaoxy@terpmail.umd.edu (X. Zhao).

https://doi.org/10.1016/j.pubrev.2018.08.002
Received 1 May 2018; Received in revised form 27 June 2018; Accepted 1 August 2018
Available online 23 August 2018
0363-8111/ © 2018 Elsevier Inc. All rights reserved.
X. Zhao et al. Public Relations Review 44 (2018) 549–561

The fast-paced and complex dynamics of social media crises call for a new model of social media influence. By integrating diverse
perspectives on influence, we propose a new framework that conceptualizes different dimensions of social media influence during
crises (i.e., output, reactive outtake, proactive outtake, and network positioning). This framework is based on publics’ communicative
behaviors on social media (e.g., liking a post). Namely, influentials are assumed to generate more output, receive more commu-
nicative responses from their followers, and occupy more central positions in a network than non-influentials.
Our study features three contributions. First, setting a common ground for crisis communication theory building, we offer an
integrated framework of social media influence based on literature on social psychology, mass communication, crisis communication,
and public relations. By examining different dimensions of social media influence, we delineate a more refined conceptual picture of
influence on social media in crises. Second, existing theories on social media influence in crises emphasize characteristics of in-
formation source, such as knowledge and involvement (e.g., Jin & Liu, 2010; Hallahan, 2000). However, stakeholders interact with
each other and form communicative networks on social media in crises (e.g., Zhao, Zhan, & Wong, 2018). As such, the network
perspective provides a complementary view to the literature by considering the role of social media influence in shaping networks
(Yang & Taylor, 2014). Third, our model is more applicable than prior research to the real-time and short-term social media en-
vironment because we tested our model with complete Twitter datasets on crisis events. A multitude of influencers exists on social
media (Li, 2016; McCorkindale & DiStaso, 2014). But it is still unknown whether and how these influencers share certain char-
acteristics giving rise to influence on social media during crises. Unobtrusive real crisis data from social media can capture the
immediate behavioral indicators of social media influence and thus provide more valid and reliable measures.
In sum, our study aims to examine different dimensions of social media influence in crises and reveal new opportunities for theory
building. To disentangle the dimensions, we conducted a multigroup confirmatory factor analysis on Twitter influencers in four
crises. These influencers were retrieved from an automation program for processing large-scale Twitter data. The four studied crises
consist of two corporation-related crises and two government-related crises for higher generalizability. Our theory-grounding
measurement of social media influence with Twitter data offers a comprehensive picture of influence on social media, thereby
informing how public relations practitioners can identify and interact with influential publics during crises.

1.1. Different perspectives to influence on social media

Social impact is conceptualized as any influence on individual attitude, behavior, and feelings that occur as a result of the real,
implied, or imagined actions of others (Latané, 1981; Nowak, Szamrej, & Latané, 1990). In social psychologist Latané’s seminal work
(1981), he argued that an important characteristic of social influence was salience of influencers. Specifically, the number, strength,
and immediacy of a source contribute positively to the amount of its influence. Relatedly, the structural theory of social influence
prescribed that influence was established by receivers’ repeated responses (e.g., comments) to the influential source (Friedkin, 1998).
Based on Cialdini and Goldstein’s (2004) review, to obtain social influence, a source needs to activate targets’ three central moti-
vations: to be accurate, to affiliate, and to maintain a positive self-concept. Namely, social influence entails sources providing ac-
curate descriptions of social situations, offer opportunities to maintain meaningful social relationships with their publics, and en-
hance publics’ positive self-concept. Moreover, influence can be considered as a product of information cascade (Bikhchandani,
Hirshleifer, & Welch, 1992), which occurs when individuals follow their peers’ behaviors without considering their own information.
The concept of influencer originated from mass communication (i.e., opinion leadership; Lazarsfeld, Berelson, & Gaudet, 1944;
Rogers & Cartano, 1962) and has been extensively developed in various communication areas (e.g., Jin & Liu, 2010; Jiang, Luo, &
Kulemeka, 2016; Sommerfeldt & Kent, 2015). Investigating the diffusion of news messages within communities, Lazarsfeld et al.
(1944) identified a group of individuals called opinion leaders. The opinion leaders give heed to an issue, discuss the issue frequently,
and consider themselves highly convincing to persuade others on issue-related matters. According to the two step-flow of information
model, opinion leaders pass on the information from mass media to general publics (Katz & Lazarsfeld, 1955). More recently,
Huffaker (2010) considered opinion leaders as those who diffuse information and potentially shape public opinions by joining
discussions in various online forums (e.g., Google group). These opinion leaders “trigger feedback, spark conversations within the
community, or even shape the way that other members of a group ‘talk’ about a topic” (Huffaker, 2010, p. 594).
In the realm of crisis communication and public relations, the concept of social influence and influencers also occupies a central
position. Echoing the opinion leadership literature, crisis communication scholars conceptualize influence based on the individuals’
characteristics, such as self-involvement (e.g., Jin & Liu, 2010). Distinct from previous studies, public relations scholars conceptualize
influence from the perspectives of engagement (e.g., Jiang et al., 2016) and network (e.g., Sommerfeldt & Kent, 2015). Below we
synthesize three streams of crisis communication and public relations research highlighting specific aspects of influence in the social
media context.

1.1.1. Influence and social-mediated crises

Identifying influential social media users in crises is an essential step for effective organizational crisis communication. Yet, how
to measure organizational crisis communication effectiveness via social media remains an open area for research (Cheng & Cameron,
2017). Social media can serve as effective crisis communication tools, due to their real-time and engaging nature (Fraustino, Liu, &
Jin, 2017). Communication through social media resulted in higher organizational reputation and less secondary crisis reactions such
as boycotting (Schultz, Utz, & Göritz, 2011; Utz, Schultz, & Glocka, 2013).
Furthermore, successful identification of powerful influentials on social media help organizations prioritize their use of limited
resources (e.g., personnel, time, money) to prevent and mitigate organizational crisis, and subsequent potential reputation damage
(Jin & Liu, 2010). If organizations use social media effectively, they also can efficiently reach and connect with enormous numbers of

550
X. Zhao et al. Public Relations Review 44 (2018) 549–561

people during crises (Lachlan, Spence, & Lin, 2017).

Scholars have conceptualized influence on new media platforms during crises. Park, Jeong, and Han, (2008) proposed a model to
identify influential bloggers by the level of issue involvement and communication activities in terms of message production and
consumption. Following Park et al. (2008) reasoning, Jin and Liu (2010) proposed that influential bloggers engage themselves in
crisis-related blogging due to issue-involvement and self-involvement. Specifically, they argued that influential bloggers are those
who are motivated and able to talk about the issue (i.e., issue-involvement) and have the need to build up authority and leadership by
talking about the issue (i.e., self-involvement).
According to the social-mediated crisis communication (SMCC) model (Austin, Liu, & Jin, 2012; Jin & Liu, 2010), influentials,
such as bloggers and/or organizations, affect followers directly and indirectly. On the one hand, influentials influence their followers
directly through their communicative behaviors, such as selectively publicizing some first-hand information, namely gatewatching
(for details, see Bruns, 2005). On the other hand, influentials exert their influence on followers indirectly through interpreting the
news media content from a more relevant and tailored perspective for their audience (Zhao et al., 2018). Similarly, the interactive
crisis communication model on social media (Cheng, 2018) suggests that social media influentials may be identified by the levels of
knowledge and involvement.
Based on the literature on crisis influentials, we argue that social media influence during crises may emerge when users are
concerned about the crisis, have expertise on the crisis, and talk a lot about the crisis. However, one limitation of this line of research
is that social media influence is conceptualized by influencers’ certain characteristics (e.g., ability, motivation). In social media crises,
besides the influencers’ characteristics, social influence also encompasses publics’ response to the influential source (Cialdini &
Goldstein, 2004). In the following section, we discuss engagement as a dimension of social media influence.

1.1.2. Influence and engagement

Social media engagement is generated by a meaningful dialogue among social media influentials and their publics (Jiang et al.,
2016). Couched in social exchange theory, we argue that social media users may reciprocate positive behaviors (e.g., social media
engagement behaviors) toward a social media influential when they receive a benefit from the relationship (Hollebeek, 2011). Thus,
social media engagement can be considered as one type of behavioral manifestations of influence. Moreover, regulatory engagement
theory prescribes that value is a motivational force that corresponds to experiencing attraction (i.e., something positive; Higgins &
Scholer, 2009). As the motivational force varies in intensity, social media users’ level of engagement as behavioral manifestations are
differentiated (Hollebeek, 2011). Besides passively receiving and processing messages, social media users can like, share messages,
and comment on the posts. They can also create their own messages to connect with the trendy topics.
Although consensus has yet to be reached for the meaning of engagement in the public relations literature, scholars have generally
agreed that engagement has a behavioral component (e.g., Jiang et al., 2016; Smith & Gallicano, 2015; Men & Tsai, 2013; Taylor &
Kent, 2014). Public relations scholars have started to use social media engagement as a barometer to assess publics’ experiences; yet
most prior researchers adopted an organization-centric view and measured organizational activities as a proxy for publics’ social
media engagement (e.g., Agostino, 2013; Jiang et al., 2016; Lovejoy, Waters, & Saxton, 2012). For example, Lovejoy et al. (2012)
measured frequency of tweets, hyperlinks, public messages (messages that start with “@”), retweets, and hashtags of nonprofits’
official Twitter accounts to examine the extent to which nonprofits engage publics on Twitter.
Different types of engagement behaviors can signal different involvement with the crisis or information source. Namely, public
engagement behaviors on social media can be understood as a hierarchical behavioral construct with varied activity levels, from
passive message consumption to active content contributing (Men & Tsai, 2013; Paine, 2011; Tsai & Men, 2013). This is because
active content contribution may indicate deeper crisis involvement to the organization in crisis or the influencer. Indeed, Cho,
Schweickart, and Haase (2014) found that social media likes, sharing, and comments represented different levels of engagement (i.e.,
from low to moderate to high) for publics on Facebook.

1.1.3. Influence and social network

The social network framework is one of the most useful conceptualizations to map influence (e.g., Wasserman & Faust, 1994).
Social networks are structures of interactions created by social actors, such as individuals, partners, and organizations (Himelboim,
Golan, Moon, & Suto, 2014; Sommerfeldt & Kent, 2015). As a budding theoretical and methodological paradigm, social network
analysis (SNA) focuses on social entities’ patterns of relationships in systems (Freeman, 2004; Sommerfeldt & Kent, 2015), which
makes SNA an ideal candidate for studying public relations (Botan, 1992). Relying on network theories such as structural hole theory
and resource dependency theory, public relations scholars started using SNA to study relationships among social entities more than a
decade ago (e.g., Taylor & Doerfel, 2005; Kent, Sommerfeldt, & Saffer, 2016). For example, using SNA to study relationships among
non-governmental organizations, Sommerfeldt and Kent (2015) found that organizations with high degree centrality are more likely
to be invested in civil society actors. SNA has also been used to study relationships among Twitter accounts (e.g., Himelboim et al.,
2014).
SNA provided valuable insights for influence. As Yang and Taylor (2014) pointed out, from a network ecology perspective,
networks can be shaped by communicative behaviors (e.g., replies). Social media network positions may be an indicator of influence.
For example, Brass (1984) found that communication network centrality was strongly correlated with perception of influence.
Sparrowe and Liden (2005) found that employee centrality was positively related to employee influence in an advice-giving network.
Moreover, Dubois and Gaffney (2014) found that network positioning measures helped to identify influentials in Twitter political
discussion communities.

551
X. Zhao et al. Public Relations Review 44 (2018) 549–561

1.2. Unpacking influence with large-scale data

The concept of large-scale data is ubiquitous in today’s society, but is just beginning to be applied to public relations research
(e.g., van der Meer, 2016; Zhao et al., 2018). Equipped with a data mining perspective, data science researchers have made tre-
mendous efforts in translating various features into a social media influence scoring system (e.g., Rao, Spasojevic, & DSouza, 2015;
Song, Chi, Hino, & Tseng, 2007). These researchers have offered valuable practical tools for predicting social media influence in
public relations and strategic communication with large-scale data (Cha, Haddadi, Benevenuto, & Gummadi, 2010; Rao, Spasojevic,
Li, & DSouza, 2015; Song et al., 2007; Ye & Wu, 2010).
However, this line of work has limited contribution to the communication discipline in term of theory and methodology. First,
social media influence is a latent construct that can be manifested by different behavioral indicators. Use of a single surrogate
measure such as retweets constrains the development of valid and reliable measurement of social media influence. Second, the latent
construct of influence can disperse different dimensions. Without conducting factor analysis and dimensionality identification, the
treatment of influence as an overall score may be problematic. Third, these scoring algorithms bear limited relevance for crisis
communication research. Social media influence during crises may be issue specific, but prior research has not tested this possibility
due to not examining whether influences changes by crisis type.

1.3. The multidimensional structure of social media influence

Cutting across the multidisciplinary perspectives of social media influence, we propose a theory-grounded model of social media
influence during crises. Following Paine’s (2007) influence framework in public relations, we argue that social media influence
should be multifaceted. As social media crises feature ever-changing publics and complex dynamics between stakeholders, we
complement the existing multidimensional influence framework with the network perspective (Taylor & Kent, 2014). Namely, social
media influence can be factorized into four dimensions: output, proactive outtake, reactive outtake, and network positioning, all
relevant and important for establishing influence in the rapidly changing social media environment. In the following section, we
provide the rationales for the four factors in the social media crisis context.

1.3.1. Output
Outputs are usually the immediate communication product of public relations efforts (Lindenmann, 2003; Paine, 2007). Stacks
and Bowen (2013) defined output as “what is generated as a result of a public relations program or campaign that may be received
and processed by members of a target audience, and may have cognitive impact on outtakes” (p. 21). Outputs include the number of
communication products generated and number of times the communication products are distributed, which usually indicates the
amount of exposure.
In the context of social-mediated crises, output can be a dimension of social media influence. Lazarsfeld et al. (1944) prescribed
that opinion leadership emerges when someone articulates frequently about an issue. Due to the abundance of crisis-related in-
formation on social media, users who strive to gain influence, such as organizations or bloggers, usually publish crisis information
consistently and routinely to win audience attention and engagement. Indeed, opinion leaders on Twitter were found to tweet often
(Park, 2013).
Consistently and routinely discussing a crisis requires one’s knowledge and motivation (Austin et al., 2012; Jin & Liu, 2010).
Social media followers retweet, reply, and mention communication outputs from bloggers, organizations in crisis, and the media, all
of which are relevant resources to learn about crises. So the availability of large amount of outputs or relevant resources allows for
the possibility of high influence on the followers. Thus, we concluded that a social media user who has a high level of social media
influence is one who frequently communicates about a crisis (i.e., measured by the total number of posts sent) and who is highly
engaged during all crisis stages (i.e., measured by the longevity of engagement).

1.3.2. Outtake: reactive vs. Proactive

Outtakes measure the extent to which audiences received, paid attention, comprehended, and retained the messages directed to
them (Lindenmann, 2003). According to Stacks and Bowen (2013), outtake is a measurement of audiences’ responses to commu-
nication products, such as further information seeking or reaction of favorability toward the communication product. As social media
engagement is a behavioral process resulting from meaningful dialogues that emerge among social media users (Jiang et al., 2016;
Taylor & Kent, 2014), posting comment on Twitter and liking a Tweet can both be considered as behavioral indicators of dialogic
engagement (Taylor & Kent, 2014).
There are two types of outtakes that differ in the level of publics’ social media engagement. The first type of outtake is reactive
outtake. We argue that reactive outtakes act like signatures of audience message receipt. When someone likes a tweet, it requires
minimal cognitive processing of the message and this action may carry less commitment toward the source (cf., Kim, 2016). However,
reactive outtakes can serve as amplifiers and endorsers of the original social media crisis-related messages. This distribution power of
reactive outtakes increases the levels of the reach and social approvals of the messages exponentially. Thus, reactive outtakes are
useful for maintaining the presence on social media.
The second type of outtakes is proactive outtakes. Proactive outtakes resemble two-way communication that encourages more
interactions than reactive outtakes. Proactive outtakes require a higher level of cognitive, behavioral, and emotional resource al-
location than reactive outtakes. For instance, replying to a tweet requires cognitive processing, behavioral reaction, and message
production. Proactive outtakes are strong social endorsements, which should exert influence on the endorser’s social media followers.

552
X. Zhao et al. Public Relations Review 44 (2018) 549–561

In fact, social endorsements were found to be much stronger heuristic cues than information sources when followers select in-
formation on social media (Messing & Westwood, 2014; Muchnik, Aral, & Taylor, 2013).). In addition, publics’ sentiment should be
considered in proactive outtakes (Jiang et al., 2016; Jin & Liu, 2010; Kang, 2014). For example, Jiang et al. (2016) pointed out the
importance of understanding and attending to the emotional component within publics’ conversations and actions.
In the context of social media and crises, proactive outtakes measure a different dimension of influence than reactive outtakes.
Reactive outtakes are manifested by reactive behaviors including sharing (e.g., retweets), consumption (e.g., favorites), and reach
(e.g., size of followers). Proactive outtakes are manifested by proactive behaviors including conversation initiation (e.g., mentions),
conversation attending (e.g., replies), and positive emotion (e.g., positive references).

1.3.3. Network positioning

Going beyond the content and conversation levels, network positioning examines influence by considering a user’s position in the
network structure. In a social media crisis communication network, the nodes are social media users and the edges (i.e., the inter-
active relationship that links the nodes) are any type of communicative acts, such as a reply, mention, and retweet. In such a social
network, network positioning represents a social media user’s access to other users’ attention or others’ dependence on them (Tsai,
2001). Salancik and Pfeffer (1977) posited that power, or the inverse of dependence, comes from control of relevant resources. A
relevant resource has high demand (Emerson, 1962) and thus is the source of influence.
Based on SMCC model, social media followers are motivated to fulfill their high demand for crisis-related informational and
emotional needs (Austin et al., 2012; Jin & Liu, 2010). By interacting with others, social media users potentially fulfill their in-
formational and emotional needs during crises. In fact, social media influentials were found to be more likely to be sought out for
advice and reassurance about brands (Freberg et al., 2011). More importantly, in the same process, the followers create dependency
on influentials, which renders those social media accounts influence during crises. Therefore, these networks represent co-con-
structed communities that feature conversations and content forwarded in multiple directions, thereby increasing the influence of
influential (Jiang et al., 2016).
Network positioning measures a user’s salience in the network in terms of three aspects. First, “traffic” through a user, which is the
number of communicative actions that a user sends out and receives in the network (i.e., degree centrality). Second, a user’s control
and gatekeeping role in the network, namely whether the user stands between different groups (i.e., betweenness centrality). Third,
the quality of a user’s connections, whether the user is connected with “important” others (i.e., PageRank).
In summary, we propose that the multidimensional construct of social media influence consists of output, proactive outtakes,
reactive outtakes, and network positioning. As such, we propose a four-factor structure of social media influence:
H1. Social media influence can be measured by four factors, including output (number of posts and longevity), reactive outtake
(retweets, favorites, and size of followers), proactive outtake (mentions, replies, positive references), and network positioning (degree
centrality, betweenness centrality, and PageRank).
Based on the literature, we also considered three plausible alternative models. First, some scholars consider outtakes as audiences’
responses to a communication product without differentiating proactive outtake from proactive outtake (e.g., Paine, 2007; Stacks &
Bowen, 2013). Thus, social media influence may be measured by three factors, including output, outtake, and network positioning.
Second, social media influence may be conceptualized by two aspects: one related to content, the other related to network structure.
As such, social media influence may be measured by two factors, a content-related factor (i.e., output and outtake as one factor) and a
network-related factor (i.e., network positioning). Last, there may be a one-factor model in which all indicators loaded on the latent
construct of social media influence. Based on our framework, the four-factor model should be a better measurement of influence than
the three alternative models.
H2. The four-factor model better measures social media influence than the three alternative models (i.e., three-factor model, two-
factor model, and one-factor mode).

2. Method

2.1. Data collection

Among the diverse social media platforms, we focused on Twitter because of its popularity for social media strategies and
relevance for crisis communication. First, Twitter has a total of 320 million users (Twitter, 2016); 49% Twitter users follow brands or
companies (Sproutsocial, 2015). Additionally, Twitter users tend to communicate directly with an organization on Twitter as
compared with users on the other platforms (Smith, Fischer, & Chen, 2012). Therefore, Twitter allows public relation practitioners to
build influencer relationships more effectively (Meltwater, 2015) and allows for publics to develop influencer relationships with
other publics. Furthermore, Twitter is an ideal platform for crisis communication because it features real-time content. Several studies
have showed that publics often exhibit active presence on Twitter during crises (e.g., Lachlan et al., 2017; Veil, Buehner, & Palenchar,
2011). Last, Twitter data are public such that researchers can access any individual’s profiles and posts, unlike other popular social
media sites like Facebook. With Twitter data, we can develop a comprehensive list of indicators for social media influence.
We chose four organizational crises to study to increase the generalizability of the results. Two of the crises were corporate crises:
the Target data leak in 2013 (Forbes, 2014) and Samsung GalaxyNote7 explosion in 2016 (Time, 2016). The other two crises were
government crises: the National Security Agency data leak in 2017 (Wall Street Journal, 2017) and the Ariana Grande concert

553
X. Zhao et al. Public Relations Review 44 (2018) 549–561

bombing in 2017 (PR Week, 2017). The four crises were comparable to some extent in that two involved explosions and two involved
data leaks. These crises occurred in the last four years and generated a good amount of media exposure as well as discussion on social
media based on analyses from Google Trend. Thus, they allow for us to look at social media influence over time.
Twitter only provides open access to a sample of tweets in the past week through the Search Application Program Interface (API).
For capturing a full set of crisis-related tweets, we built a data collection pipeline through three steps. First, we developed a complete
list of key words, terms, and hashtags for each crisis. For example, twelve keywords for the Ariana Grande concert bombing were used
(e.g., #PraryForManchester, Ariana concert attack). Second, the time frame was from the first day after the crisis broke out to the last
day people still talked about the crisis on Twitter. For each week, we archived all the keyword search results returned by twitter.com
in Web ARChive (WARC) files, using an open source web archiving service. WARC is a file format for storing web crawls with
metadata attached to each record. Third, we parsed the web archiving files into JavaScript Object Notation (JSON) format data
frames with text and metadata in Python.
For each crisis, all the data frames were aggregated, with redundant tweets removed. This resulted in 245,588 crisis-related
tweets: Target data leak (N = 55,612), Samsung Note7 explosion (N = 90,901), NSA data leak (N = 14,876), and Ariana Grande
concert bombing (N = 84,199).

2.2. Data preprocessing

The unit of analysis in the raw data was each tweet. Our goal was to examine the indicators of social media influence of a Twitter
user. To achieve this goal, we needed to change the unit of analysis to a Twitter user and compute all the indicators for each user.
Thus, we created an automation program in Python. For each crisis dataset, the program automatically finds all the unique users,
computes all our predefined influence indicators for each user, visits the API for retrieving additional personal information for each
user, and outputs Comma-separated Values (CSV) files for subsequent statistical analyses. The indicators’ measures are detailed in the
following section.
After the automatic data preprocessing, our unit of analysis became Twitter user: Target (N = 25,904), Samsung (N = 46,702),
NSA (N = 11,315), and Ariana concert (N=63,950). In the dataset, each row represents a user and each column represents an item of
social media influence.
An overwhelming number of users in the dataset were followers who just posted one tweet in the crisis (see Zhao et al., 2018 for a
discussion of this problem). These users with one tweet and zeros on the remaining measures give rise to a large quantity of zeros in
the data matrix (i.e., sparse matrix). Matrix sparsity usually introduces convergence problems in factor analysis. Given our goal of
examining social media influence, we decided to focus on users who possess relatively more influence than other users. Operationally,
we filtered our sample with the following criteria: The sampled user should at least post one tweet, receive one like and retweet, and
have a degree centrality larger than zero. Applying these criteria, we ended up with a total sample of 5030 users who show some
variation on influence (N=951 for Target, N=1676 for Samsung, N=910 for NSA, and N=1493 for Ariana Grande concert).

2.3. Measures

We computed the indicators of social media influence with the automation program designed for this study. We detail how these
indicators were measured below.

2.3.1. Transformation and scaling

We followed the standard transformation procedures suggested by Fink (2009). For variables measured by count, the distribution
of scores were nonnormal and positively skewed (e.g., the skewness of number of post was 34.90). Thus, the scores of these variables
were transformed by exponentiation or logarithm (e.g., we added 1 to the score and raised the score to a power of a value of 0.2). In
addition, the variables were measured by different scales. For example, the number of followers ranged from 0 to 52,827,416 whereas
the number of posts ranged from 1 to 887. Thus, we standardized these variables by subtracting the mean from scores and then
dividing that figure by the standard deviation. After transformation and standardization, these variables appeared relatively normal
(i.e., the skewness or kurtosis of the distribution of scores of a variable ranges between -3 and 3). Our analyses were conducted on the
transformed and standardized scores of variables. Below, we report the original scores for all measures except for betweenness
centrality and PageRank. For betweenness centrality and PageRank, standardized scores (i.e., mean is zero and standard deviation is
1) are reported.

2.3.2. Number of tweets

The number of tweets measures the total number of tweets that a user posted during the crises. The tweet could be a post, reply, or
mention from a particular user. The median number of posts was 2 (M = 4.88, SD=17.66) and posts ranged from 1 to 887.

2.3.3. Longevity
Longevity measures the number of days during which a user remained active. A user was considered active if he or she posted a
tweet anytime in a crisis. Longevity was computed by subtracting the last day of activity from the first day of activity for each user.
The median longevity was 2 (M = 17.91, SD=46.87) and longevity ranged from 0 to 1493.

554
X. Zhao et al. Public Relations Review 44 (2018) 549–561

2.3.4. Number of followers

The number of followers measures how many people follow an individual Twitter account. The median number of follower was
5161 (M = 228,109.55, SD=1,645,050.66) and the number of followers ranged from 0 to 52,827,416.

2.3.5. Number of retweets

A retweet is a re-posting of a tweet from another Twitter account. Retweet indicates users’ consumption of information from a
source and their subsequent sharing of the information. Operationally, retweets were identified with particular strings (i.e., RT, via or
by followed by a username) using regular expressions (i.e., a pattern describing and matching algorithm; see Brin, 1998 for details).
The number of retweets measures the total number of retweets that a user receives in a crisis. The median number of retweets was 4
(M = 83.17, SD=775.56) and the number of retweets ranged from 1 to 28,986.

2.3.6. Number of favorites

One can like a tweet by clicking the heart icon. The number of favorites measures the total number of likes that a user receives in a
crisis. The median number of favorites was 4 (M = 87.19, SD=893.61) and favorites ranged from 1 to 34,082.

2.3.7. Number of mentions

A mention tweet contains another user’s name anywhere in the body of a tweet. The number of mentions measures the total
number of mentions that a user receives in a crisis. The median number of mentions was 0 (M = 3.08, SD=70.35) and the number of
mentions ranged from 1 to 4378.

2.3.8. Replies
A reply is a response to another person’s tweet. The number of replies measures the total number of replies that a user receives in a
crisis. The median number of replies was 0 (M = 0.26, SD=4.22) and replies ranged from 0 to 245.

2.3.9. Positive references

The number of positive references of a user was calculated by summing the number of positive mentions and positive replies that
the user receives. Positivity was labeled by the VADER (i.e., Valence Aware Dictionary for Sentiment Reasoning; see Hutto & Gilbert,
2014). VADER is a rule-based lexicon for sentiment analysis of social media data. The reliability of VADER sentiment analysis on
social media data has been supported by several studies (e.g., γ = 0.88 between the classifier and multiple human coders in Hutto &
Gilbert, 2014). The median number of positive references was 0 (M = 1.31, SD=50.83) and positive references ranged from 0 to
3555.

2.3.10. Degree centrality

Incorporating the NetworkX Python package (Hagberg, Swart, & Schult, 2008), our program computes various network centrality
statistics. In the network, the node is a user and the edge can be any type of communicative relations, including reply, mention,
retweet, etc. The number of connections between nodes were represented by weight. The centrality measures were computed from
the weighted directional communicative network. Degree centrality indicates the number of connections one has with other nodes in
the network. The median degree centrality was 2 (M = 9.26, SD=93.07) and ranged from 1 to 5489. This means that a half of users
in our data has more than 2 communicative connections with other users (e.g., either send or receive a reply) and the other half has
fewer than 2 connections.

2.3.11. Betweenness centrality

Betweenness centrality represents the degree of which a node stands between each other. A node with a higher betweenness
centrality is assumed to have more control over the network than other nodes because much information needs to pass through the
gate to travel in the network. As such, those with high betweenness centrality are considered as information brokers in the network
(Yang & Taylor, 2014). Operationally, betweenness centrality of a node was computed by the sum of the fraction of all-pairs shortest
paths that pass through the node with the algorithm in NetworkX. The standardized median betweenness centrality was - 0.55 (M =
0, SD=1).

2.3.12. PageRank
PageRank is a variant of eigenvector centrality for directed network (Newman, 2010). PageRank takes the quality and number of
connections that a node has into consideration. A node with a connection with a more “important” node should be considered
“important” too. PageRank was originally designed by Google as an algorithm to rank web pages (Page & Brin, 1998). The stan-
dardized median PageRank centrality was - 0.35 (M = 0, SD=1).

3. Results

We conducted a multigroup confirmatory analysis (CFA) with the R “Lavaan” package (Rosseel, 2011). Four groups comprised
Twitter users with relatively more influence than other users from four crises. Operationally a sampled user should at least post one
tweet, receive one like and retweet, and have a degree centrality larger than zero. Based on our framework, we specified a four-factor
structure (see Fig. 1). Latent factors were output, reactive outtake, proactive outtake, and network positioning. The factors were

555
X. Zhao et al. Public Relations Review 44 (2018) 549–561

Fig. 1. Multigroup Confirmatory Factor Analysis for Social Media Influence.

Note. N = 5030. Maximum Likelihood estimation (MLR). Standardized coefficients were reported. All factor loadings were significant and thus were
only marked before these loadings. Two corporation-related crises were shown in black and two government-related crises were shown in purple.
(1) The variance of number of tweets and longevity explained by output was [.51, .27, .45, .29] and [.43, .23, .06, .01] for four crises. (2) The
variance of number of retweets, favorites and followers explained by reactive outtake was [.71, .42, .66, .56], [.72, .78, .66, .55], and [.55, .33, .47,
.48] for four crises. (3) The variance of number of mentions, replies, and positive references explained by proactive outtake was [.88, .60, .88, .81],
[.47, .45, .46, .67], and [.77, .58, .95, .73] for four crises. (4) The variance of degree centrality, betweenness centrality, and PageRank explained by
network positioning was [.86, .99, .67, .30], [.22, .49, .48, .42], and [.88, .91, .76, .79] for four crises.

allowed to covary because they represent different dimensions of social media influence. In Fig. 1, the number of tweets and longevity
loaded on Output. The number of retweets, favorites, and followers loaded on reactive outtake. The number of mentions, replies, and
positive references loaded on proactive outtake. Degree centrality, betweenness centrality, and PageRank loaded on network posi-
tioning. Maximum likelihood estimation with robust standard errors (MLR) was employed for parameter estimation. The factor
loadings of the first indicator of a factor were fixed for estimation. Standardized coefficients were reported in the following sections.
The correlation matrix of variables is in Table 1.

3.1. Factors of social media influence

The overall model-data fit was good, SRMR=0.04, RMSEA=0.070 (90%CI = [0.067, 0.072]), and CFI=0.94. The SRMR was less
than 0.08, consistent with recommendations by Hu and Bentler (1999). The RMSEA was smaller than 0.08, as recommend by
MacCallum, Browne, and Sugawara, (1996). Both indices suggest that the overall discrepancy between the observed and model-
implied covariance matrix was not significant. The incremental fit index, CFI, was close to 0.95. This suggests that the proposed
model fit the data better than the baseline null model. Different chi-squared values were reported for the four groups: χ 2 (144,
N=951)=250.33, p<.001 for Target, χ 2 (144, N=1676)=232.02, p<.001 for Samsung, χ 2 (144, N=910)=343.813, p<.001 for
NSA, and χ 2 (144, N=1493)=245.38, p<.001 for Ariana. Note that the chi-squared value is always significant for a model with a
sample size of 400 or more. In summary, considering the large sample size, the model-data fit was quite good.
Because of our large sample size, the factor loadings can be always significant. Therefore, we interpreted the factor loadings in
terms of its magnitude. As shown in Fig. 1, the number of tweets loaded strongly on output (β = .71, 0.52, .67, .54 or four crises).
Longevity loaded moderately on output. In addition, the latent factor of output can be differentiated because it was weakly associated

556
X. Zhao et al. Public Relations Review 44 (2018) 549–561

Table 1
Correlation Matrix for the Variables in the Multigroup CFA.
Pearson Correlations

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)

(1) Number of Tweets 1.00

(2) Longevity .28*** 1.00
(3) # Followers .02 .15*** 1.00
(4) # Retweets .27*** .08** .42*** 1.00
(5) # Favorites .25*** .01 .39*** .80*** 1.00
(6) # Replies .09** .05* .35*** .30*** .32*** 1.00
(7) # Mentions .10** .12** .47*** .39*** .34*** .53*** 1.00
(8) # Positive References .08** .08** .34*** .35*** .33*** .64*** .69*** 1.00
(9) Degree Centrality .12*** .21*** .44*** .41*** .38*** .39*** .53*** .49*** 1.00
(10) Between Centrality .16*** .10** .31*** .30*** .22*** .31*** .45*** .37*** .40*** 1.00
(11) PageRank −.01 .19*** .47*** .37*** .37*** .37*** .49*** .48*** .88*** .31*** 1.00

Note. N = 5030. The scores of the variables were standardized. Thus, for each variable, mean is 0 and standard deviation is 1. So the mean and
standard deviation was not shown in the table for simplicity.
*p < .05. **p < .01. ***p < .01.

with reactive outtake (γ = .21, 0.13, .21, .20 for four crises), proactive outtake (γ = .23, 0.08, .07, .05), and network positioning
(γ = .28, 0.18, 0.14, 0.05).
Regarding reactive outtake, the number of retweets, favorites, and followers loaded significantly and strongly on reactive outtake
across the four crises (see Fig. 1). These factor loadings across crises varied from 0.58 to 0.88 for each crisis, suggesting strong
relationships. For the indicators of outtake, the variance explained by the model fluctuated between 0.334 to 0.777 for the different
crises, suggesting that the latent factor of outtake explained some variance of its indicators. On the other hand, the number of
mentions, replies, and positive references significantly loaded on proactive outtake. These factor loadings for each crisis varied from
0.67 to .97, suggesting very strong relationships. For indicators of proactive outtake, the variance explained by the model fluctuated
between 0.454 to 0.952 for different crises, suggesting that the latent factor of proactive outtake explained a good amount of variance
of its indicators. Furthermore, reactive and proactive outtake covary with each other moderately as expected (γ = .50, 0.40, 0.40, 33
for four crises). Together, these results suggest that reactive and proactive outtake were distinct factors of social media influence.
Last, we can see that all freely estimated factor loadings on the construct of network positioning were significant in Fig. 1.
Betweenness centrality loaded moderately on network positioning (β = .47, 0.70, .69, .65 for four crises), PageRank loaded strongly
on network positioning (β = .94, 0.96, .87, .89), and degree centrality loaded strongly on network positioning (β = .93, 1.00, 0.82,
.55). In addition, the covariance between network positioning and the other three factors was moderate to small. Therefore, it was
concluded that network positioning was a dimension of social media influence. Therefore, H1 was supported.

3.2. Model comparison

To test H2 and provide additional evidence for the previous hypotheses, we constructed alternative models and compared them
with the four-factor model. Namely, we constructed a three-factor model, two-factor model, and one-factor model (see Table 2 for
details of model specification).
We compared the performance of different models by conducting ANOVA on these models’ chi squares. A model was considered
significantly better than another model if its chi square was significantly smaller than that of the alternative model. A smaller chi
square indicates a smaller model-data discrepancy and a higher model-data fit. As shown in Table 2, the four-factor model fits the
data significantly better than the three-factor model (Δχ 2 = 99.40, p < .001), the two-factor model (Δχ 2 = 1012.43, p < .001),
and the one-factor model (Δχ 2 = 702.03, p < .001). In addition, the four-factor model also had much lower Akaike Information
Criterion (AIC) and Bayesian Information Criterion (BIC) scores, as compared with the remaining models. AIC and BIC are important
comparative indices of fit when there are two or more estimated models. Therefore, H2 was supported: The four-factor model better

Table 2
Multigroup CFA Model Comparison.
DF AIC BIC Chi-square Chi-square diff & sig

The 4-factor model 148 100710 101754 2764.2

The 3-factor model 156 101806 102797 3875.5 99.40***
The 2-factor model 168 103848 104761 5941.7 1012.43***
The 1-factor model 172 104211 105098 6312.7 702.03***

Note. The 4-factor model has output, reactive outtake, proactive outtake, and network positioning as four factors. The 3-factor model has output,
outtake (reactive + proactive outtake), and network positioning as three factors. The 2-factor model has output, outtake (reactive + proactive
outtake), and network positioning as three factors. In the 2-factor model, output and outtake were considered as one latent factor, along with
network. In the 1-factor model, all the items were loaded on one latent construct of social media influence.

557
X. Zhao et al. Public Relations Review 44 (2018) 549–561

measures social media influence than the alternative models (Henson, Reise, & Kim, 2007). Therefore, H2 was supported: The four-
factor model better measures social media influence than the three specified alternative models.

4. Discussion

Drawing from the social media influence literature from crisis communication, social network, and data science research, we
established an integrated framework of social media influence by explicating its different components. The framework was tested by
complete Twitter datasets on crisis events, preprocessed with our program that automatically samples influencers and computes the
influence indicators. The use of Twitter data from four organizational crises with different organizational entities (corporation vs.
government) increases the generalizability of our results. Multigroup CFA on influencers from these crises support our four-factor
influence structure. There are four dimensions of social media influence: Output, reactive outtake, proactive outtake, and network
positioning. And each dimension can be represented by a distinct set of users’ behavioral traces on social media.
First, output is a primary aspect of social media influence. Our analyses supported that output could be represented by two
indicators: (1) the total number of posts from a user and (2) the number of days during which a user posts any content. Previous
studies argued that the actual distribution of public relations programs or campaigns determined the possible exposure to the public
relations message among publics, which was a necessary condition of social media influence (Paine, 2007; Stacks & Bowen, 2013).
Our results are consistent with this line of research. Namely, across different types of organizational crises, the number of posts was a
strong indicator of output. Given that Twitter features real-time content, the frequency of posts also should be an indicator of output.
Indeed, in the two corporate crises that we studied, longevity strongly loaded on output. However, longevity was a weak indicator of
output influence for the two government-related crises that we studied. One reason may be that more consumers and publics were
affected by the corporate crises of the Target data leak and Samsung Note7 exploding battery. The two corporations published
multiple press releases to address the potential consequence from using the defective product or hacked service. In the NSA data leak
and Ariana concert bombing, relatively fewer people were directly involved and the crises were more like “news” for most publics.
Thus, in the latter cases, an influential does not need to frequently post crisis-related content to establish impact.
Second, the results from both CFA and model comparisons supported the differentiation of reactive outtake and proactive outtake
as distinct factors of social media influence. Specifically, across the four organizational crises, the number of retweets, favorites, and
followers were strong indicators of reactive outtake whereas the number of mentions, replies, and positive references were strong
indicators of proactive outtake. These results support the literature that discuss the varying level of social media engagement (Cho,
Schweickart, & Haase, 2014; Jiang et al., 2016; Men & Tsai, 2013; Tsai & Men, 2013) and further crystalize how these levels of social
media engagement can be reciprocated into different dimensions of social media influence. Complementing the literature, certain
indicators appear to carry more weight toward the proactive outtakes than other indicators. For example, mentioning (vs. replying)
was a stronger indicator of proactive engagement. Furthermore, positive references loaded on proactive outtake, consistent with prior
research (Jiang et al., 2016; Jin & Liu, 2010). This confirms the importance of including an emotional component in the measurement
of social media influence. Positive references, together with mentions and replies, bring about proactive outtakes, which may be
particularly useful for establishing social media influence during crises.
Last, we found that network positioning was a distinct factor of social media influence, which has not been proposed in prior crisis
communication research. Across four organizational crises, degree centrality, betweenness centrality, and PageRank emerged as
important indicators of network positioning. Literature on network analysis has recorded how one’s relationship with others in the
system can affect one’s behaviors and the implications of these behaviors (Himelboim et al., 2014; Kent et al., 2016; Sommerfeldt &
Kent, 2015). Our study contextualized this line of thought in social media influence and incorporated the network into a more
comprehensive influence framework. Interestingly, our results suggest that degree centrality and PageRank in a communicative
network are stronger indicators of network positioning, as compared with betweenness centrality. One reason may be that both the
number (i.e., degree centrality) and quality (i.e., PageRank) of connections determine social media influence to a large extent.
Meanwhile, whether one stands between different communities and plays the role of information broker speaks less to the social
media influence. This finding can be explained by the diverse and fragmented communities on social media in crises (Zhao, Jie, &
Zhan, 2017). As such, there is not a single gatekeeper that a social media user has to bypass to obtain unique, valuable information, as
social media user can access and consume content through word-of-mouth flowing from multiple directions (e.g., Jin & Liu, 2010).
For organizations, the findings imply that social media influencers exist in dispersed communities, which may have different in-
formation needs during crises.

4.1. Theoretical implications

Our framework offers a holistic perspective to social media influence by integrating different perspectives in social media crises.
As our results revealed the multidimensional structure of influence, future research in crisis communication should consider different
types of influence.
First, for crisis communication research, our study extends the SMCC model (Jin & Liu, 2010) by being the first study to refine and
measure how publics become influential online during crises. Influentials may demonstrate impact on one or more of our four
dimensions and a topology of influentials may be developed in future research. Furthermore, our results support varying level of
engagement with real crisis data. Future research is needed to determine the associations between organization-publics relationship
and different types of outtakes, namely proactive and relative outtakes in their crisis responses. Our study provides the first validated
measures for such important future research. Last, regarding social network research, our study suggests that researchers should go

558
X. Zhao et al. Public Relations Review 44 (2018) 549–561

beyond the number of connections and emphasize quality of connections in social media networks. For example, PageRank and
related network statistics such as Eigenvector centrality can measure quality of connections.
In sum, our framework brings coherence to the disparate research and reveals new opportunities. We believe that theory-
grounding large data analytics, aided by automation and statistical modeling, can provide a more comprehensive picture of social
media influence.

4.2. Practical implications

In today’s media environment, organizations compete for influence during crises along with the traditional media and prominent
social media users. Public relations practitioners must be able to assess to what extent their organizations’ crisis messages are: (1)
reaching their online publics, (2) meeting organizational goals, (3) penetrating through the online clutter, and (4) competing with
other messages on social media. This study’s four-factor model of social media influence during crises serves as a starting point for
strategically guiding how public relations practitioners assess online crisis influence.
First, reactive outtakes (i.e., retweets and favorites) amplify organizational crisis messages whereas proactive outtakes (i.e.,
mentions, replies, and positive emotion) can bring about the dialogue that is essential for relationship and community building. It
may be that early in a crisis the focus should be on reactive outtakes to spread important life-saving information whereas later in the
crisis proactive outtakes may cement positive-organizational relationships. Namely, mentioning was a stronger indicator of proactive
engagement than replying. This finding indicates that to build influence online, organizations may want to mention supportive
influentials in their crisis tweets rather than replying directly to influentials.
Second, this study reveals the importance of optimizing social media influence by weaving positive emotional components into
social media crisis communication. We know that publics go online for emotional coping during crises (Jin & Liu, 2010). This study
adds that meeting these coping needs in a positive manner appears to have the benefit of also increasing one’s influence online. Last
but not least, organizations may need to tailor crisis information for different communities on social media, because publics in
dispersed communities may have different information needs.

4.3. Limitations and future directions

This study is limited by several factors. First, there may be additional indicators of social media influence during crises. For
example, the quality of the content or the credibility of the linked website in the post may be important indicators of output. Future
research should examine the message features associated with social media influence. Second, our model was developed in the
context of organizational crises on Twitter. Future research should test the factor structure of social media influence in additional
social media contexts. Third, another component of social media influence may be how publics cognitively assess messages in terms
of beliefs, attitudes, and behaviors (see Smith & Taylor, 2017). Yet, publics’ perceptions were not testable with behavioral traces
extracted from Twitter data. Future research should include the perceptual component in a social media influence measurement
model. Despite the study’s limitations, we provide validated measures for social media influence during crises. We unobtrusively
tested the measures in real crises, thereby accurately capturing how publics and organizations can exert influence in social media
during crises.

Author note

Dr. Xinyan Zhao is with the School of Communication, Hong Kong Baptist University. Dr. Mengqi Zhan is with Department of
Communication, University of Texas at Arlington. Dr. Brooke F. Liu is with the Department of Communication, University of
Maryland, College Park.
We thank Dr. Chau-Wai Wong for suggesting constructive comments during revisions.

References

Agostino, D. (2013). Using social media to engage citizens: A study of Italian municipalities. Public Relations Review, 39, 232–234. https://doi.org/10.1016/j.pubrev.
2013.02.009.
Austin, L., Liu, B. F., & Jin, Y. (2012). How audiences seek out crisis information: Exploring the social-mediated crisis communication model. Journal of Applied
Communication Research, 40, 188–207. https://doi.org/10.1080/00909882.2012.654498.
Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of fads, fashion, custom, and cultural change as informational cascades. The Journal of Political Economy,
100(5), 992–1026.
Botan, C. (1992). International public relations: Critique and reformulation. Public Relations Review, 18, 149–159. https://doi.org/10.1016/0363-8111(92)90006-k.
Brass, D. J. (1984). Being in the right place: A structural analysis of individual influence in an organization. Administrative Science Quarterly, 29, 518–539. https://doi.
org/10.2307/2392937.
Brin, S. (1998). Extracting patterns and relations from the world wide web. MarchInternational workshop on the world wide web and databases. Berlin, Heidelberg:
Springer172–183.
Bruns, A. (2005). Gatewatching: Collaborative Online News Production. New York: Peter Lang.
Cha, M., Haddadi, H., Benevenuto, F., & Gummadi, P. K. (2010). Measuring user influence in Twitter: The million follower fallacy. ICWSM, 10, 10–17.
Cheng, Y. (2018). How social media is changing crisis communication strategies: Evidence from the updated literature. Journal of Contingencies and Crisis Management,
26(1), 58–68.
Cheng, Y., & Cameron, G. (2017). The status of social-mediated crisis communication (SMCC) research: An analysis of published articles 2002-2014. In L. Austin, & Y.
Jin (Eds.). Social media and crisis communication (pp. 9–20). New York: Routledge.
Cho, M., Schweickart, T., & Haase, A. (2014). Public engagement with nonprofit organizations on Facebook. Public Relations Review, 40, 565–567. https://doi.org/10.

559
X. Zhao et al. Public Relations Review 44 (2018) 549–561

1016/j.pubrev.2014.01.008.
Cialdini, R. B., & Goldstein, N. J. (2004). Social influence: Compliance and conformity. Annual Review of Psychology, 55, 591–621. https://doi.org/10.1146/annurev.
psych.55.090902.142015.
Dubois, E., & Gaffney, D. (2014). The multiple facets of influence: Identifying political influentials and opinion leaders on Twitter. The American Behavioral Scientist, 58,
1260–1277. https://doi.org/10.1177/0002764214527088.
Emerson, R. M. (1962). Power-dependence relations. American sociological review, 27, 31–41. https://doi.org/10.2307/2089716.
Fink, E. L. (2009). The FAQs on data transformation. Communication Monographs, 76, 379–397. https://doi.org/10.1080/03637750903310352.
Forbes (2014). Target data breach spilled information on as many as 70 million customers. Retrieved from https://www.forbes.com/sites/maggiemcgrath/2014/01/10/
target-data-breach-spilled-info-on-as-many-as-70-million-customers on Oct 25, 2017.
Frandsen, F., & Johansen, W. (2016). Organizational crisis communication: A multivocal approach. London, UK: Sage.
Fraustino, J. D., Liu, B. F., & Jin, Y. (2017). Social media use during disasters: A research synthesis and road map. In L. Austin, & Y. Jin (Eds.). Social media and crisis
communication (pp. 283–295). New York: Routledge.
Freberg, K., Graham, K., McGaughey, K., & Freberg, L. A. (2011). Who are the social media influencers? A study of public perceptions of personality. Public Relations
Review, 37, 90–92. https://doi.org/10.1016/j.pubrev.2010.11.001.
Freeman, L. C. (2004). The development of social network analysis: A study in the sociology of science. North Charleston, SC: BookSurge, LLC.
Friedkin, N. E. (1998). A structural theory of social influence. Cambridge, MA: Cambridge University Press.
Hagberg, A., Swart, P., & Schult, D. (2008). Exploring network structure, dynamics, and function using NetworkX (No. LA-UR-08-05495; LA-UR-08-5495). Los Alamos
National Laboratory (LANL).
Hallahan, K. (2000). Inactive publics: The forgotten publics in public relations. Public Relations Review, 26, 499–515. https://doi.org/10.1016/S0363-8111(00)
00061-8.
Henson, J. M., Reise, S. P., & Kim, K. H. (2007). Detecting mixtures from structural model differences using latent variable mixture modeling: A comparison of relative
model fit statistics. Structural Equation Modeling A Multidisciplinary Journal, 14(2), 202–226. https://doi.org/10.1080/10705510709336744.
Higgins, E. T., & Scholer, A. A. (2009). Engaging the consumer: The science and art of the value creation process. Journal of Consumer Psychology, 19, 100–114. https://
doi.org/10.1016/j.jcps.2009.02.002.
Himelboim, I., Golan, G. J., Moon, B. B., & Suto, R. J. (2014). A social networks approach to public relations on Twitter: Social mediators and mediated public
relations. Journal of Public Relations Research, 26(4), 359–379. https://doi.org/10.1080/1062726x.2014.908724.
Hollebeek, L. (2011). Exploring customer brand engagement: Definition and themes. Journal of Strategic Marketing, 19, 555–573. https://doi.org/10.1080/0965254x.
2011.599493.
Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation
Modeling A Multidisciplinary Journal, 6, 1–55. https://doi.org/10.1080/10705519909540118.
Huffaker, D. (2010). Dimensions of leadership and social influence in online communities. Human Communication Research, 36(4), 593–617. https://doi.org/10.1111/
j.1468-2958.2010.01390.x.
Hutto, C. J., & Gilbert, E. (2014). Vader: A parsimonious rule-based model for sentiment analysis of social media text. MayEighth international AAAI conference on weblogs and
social media.
Jiang, H., Luo, Y., & Kulemeka, O. (2016). Social media engagement as an evaluation barometer: Insights from communication executives. Public Relations Review, 42,
679–691. https://doi.org/10.1016/j.pubrev.2015.12.004.
Jin, Y., & Liu, B. F. (2010). The blog-mediated crisis communication model: Recommendations for responding to influential external blogs. Journal of Public Relations
Research, 22, 429–455. https://doi.org/10.1080/10627261003801420.
Kang, M. (2014). Understanding public engagement: Conceptualizing and measuring its influence on supportive behavioral intentions. Journal of Public Relations
Research, 26, 399–416. https://doi.org/10.1080/1062726X.2014.956107.
Katz, E., & Lazarsfeld, P. F. (1955). Personal influence: The part played by people in the flow of mass communications. New York: Routledge.
Kent, M. L., Sommerfeldt, E. J., & Saffer, A. J. (2016). Social networks, power, and public relations: Tertius Iungens as a cocreational approach to studying relationship
networks. Public Relations Review, 42, 91–100. https://doi.org/10.1016/j.pubrev.2015.08.002.
Kim, C. M. (2016). Social Media Campaigns: Strategies for Public Relations and Marketing. New York: Routledge.
Lachlan, K. A., Spence, P., & Lin, X. (2017). Natural disasters, twitter, and stakeholder communication: What we know and directions for future inquiry. In L. Austin, &
Y. Jin (Eds.). Social media and crisis communication (pp. 296–305). New York: Routledge.
Latané, B. (1981). The psychology of social impact. The American Psychologist, 36, 343–356. https://doi.org/10.1037/0003-066X.36.4.343.
Lazarsfeld, P., Berelson, B., & Gaudet, H. (1944). The people’s choice. New York: Columbia University Press.
Li, Z. (2016). Psychological empowerment on social media: Who are the empowered users? Public Relations Review, 42, 49–59. https://doi.org/10.1016/j.pubrev.2015.
09.001.
Lindenmann, W. K. (2003). Guidelines for measuring the effectiveness of PR programs and activities. Institute for public relations. Retrieved from http://www.instituteforpr.
org/research.
Lovejoy, K., Waters, R. D., & Saxton, G. D. (2012). Engaging stakeholders through twitter: How nonprofit organizations are getting more out of 140 characters or less.
Public Relations Review, 38, 313–318. https://doi.org/10.1016/j.pubrev.2012.01.005.
MacCallum, R. C., Browne, M. W., & Sugawara, H. M. (1996). Power analysis and determination of sample size for covariance structure modeling. Psychological
Methods, 1, 130–149. https://doi.org/10.1037//1082-989x.1.2.130.
Marken, G. A. (2007). Social media…The hunted can become the hunter. Public Relations Quarterly, 52(4), 9–12. https://doi.org/10.1037/e380852004-038.
McCorkindale, T., & DiStaso, M. W. (2014). The state of social media research: Where are we now, where we were and what it means for public relations. Research
Journal of the Institute for Public Relations, 1, 1–17.
Meltwater (2015). Facebook vs. Twitter: Social media strategy differences. Retrieved from https://www.meltwater.com/blog/facebook-vs-twitter-social-media-strategy-
differences on Oct 20, 2017.
Men, L. R., & Tsai, W. S. (2013). Towards an integrated model of public engagement on corporate social network sites: Antecedents, the process, and relational
outcomes. International Journal of Strategic Communication, 7, 257–273. https://doi.org/10.1080/1553118x.2013.822373.
Messing, S., & Westwood, S. J. (2014). Selective exposure in the age of social media: Endorsements trump partisan source affiliation when selecting news online.
Communication Research, 41(8), 1042–1063. https://doi.org/10.1177/0093650212466406.
Muchnik, L., Aral, S., & Taylor, S. J. (2013). Social influence bias: A randomized experiment. Science, 341(6146), 647–651.
Newman, M. E. J. (2010). Networks: An introduction. New York: Oxford..
Nowak, A., Szamrej, J., & Latané, B. (1990). From private attitude to public opinion: A dynamic theory of social impact. Psychological Review, 97(3), 362–376. https://
doi.org/10.1037//0033-295x.97.3.362.
Page, L., & Brin, S. (1998). The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems, 30, 107–117.
Paine, K. D. (2007). How to measure social media relations: The more things change, the more they remain the same. Institute for public relations. Retrieved from http://www.
instituteforpr.org/research.
Paine, K. D. (2011). Measure what matters: Online tools for understanding customers, social media, engagement, and key relationships. New Jersey, NY: John Wiley & Sons.
Park, C. S. (2013). Does Twitter motivate involvement in politics? Tweeting, opinion leadership, and political engagement. Computers in Human Behavior, 29(4),
1641–1648.
Park, N., Jeong, J., & Han, J. H. (2008). Who are the power bloggers as potential target public in PR?: Public issue involvement-production of messages model.
MarchProceedings of 11th International Public Relations Research Conference571–582.
PR Week (2017). Case study: The challenges of launching one love Manchester, less than two weeks after the attack. Retrieved from https://www.prweek.com/article/
1436284/case-study-challenges-launching-one-love-manchester-less-two-weeks-attack on Oct 25, 2017.

560
X. Zhao et al. Public Relations Review 44 (2018) 549–561

Rao, A., Spasojevic, N., Li, Z., & DSouza, T. (2015). Klout score: Measuring influence across multiple social networks. OctoberBig Data (Big Data), 2015 IEEE International
Conference2282–2289.
Rogers, E. M., & Cartano, D. G. (1962). Methods of measuring opinion leadership. Public Opinion Quarterly, 435–441. https://doi.org/10.1086/267118.
Rosseel, Y. (2011). Lavaan: An R package for structural equation modeling and more (Version 0.4-9 beta) [Computer software]. Retrieved fromhttp://byrneslab.net/classes/
lavaan_materials/lavaanIntroduction4-9.pdf.
Salancik, G. R., & Pfeffer, J. (1977). Who gets power—And how they hold on to it: A strategic-contingency model of power. Organizational Dynamics, 5(3), 3–21.
https://doi.org/10.1016/0090-2616(77)90028-6.
Schultz, F., Utz, S., & Göritz, A. (2011). Is the medium the message? Perceptions of and reactions to crisis communication via twitter, blogs and traditional media.
Public Relations Review, 37, 20–27. https://doi.org/10.1016/j.pubrev.2010.12.001.
Smith, B. G., & Gallicano, T. D. (2015). Terms of engagement: Analyzing public engagement with organizations through social media. Computers in Human Behavior, 53,
82–90. https://doi.org/10.1016/j.chb.2015.05.060.
Smith, B. G., & Taylor, M. (2017). Empowering engagement: Understanding social media user sense of influence. International Journal of Strategic Communication, 11,
148–164. https://doi.org/10.1080/1553118x.2017.1284072.
Smith, A. N., Fischer, E., & Chen, Y. (2012). How does brand-related user-generated content differ across YouTube, Facebook, and Twitter? Journal of Interactive
Marketing, 26, 102–113. https://doi.org/10.1016/j.intmar.2012.01.002.
Sommerfeldt, E. J., & Kent, M. L. (2015). Civil society, networks, and relationship management: Beyond the organization–Public dyad. International Journal of Strategic
Communication, 9, 235–252. https://doi.org/10.1080/1553118x.2015.1025405.
Song, X., Chi, Y., Hino, K., & Tseng, B. (2007). Identifying opinion leaders in the blogosphere. November Proceedings of the Sixteenth ACM Conference on Conference on
Information and Knowledge Management, 971–974.
Sparrowe, R. T., & Liden, R. C. (2005). Two routes to influence: Integrating leader-member exchange and social network perspectives. Administrative Science Quarterly,
50(4), 505–535. https://doi.org/10.2189/asqu.50.4.505.
Sproutsocial (2015). Facebook vs. Twitter: Which is best for your brand. Retrieved fromhttps://sproutsocial.com/insights/facebook-vs-twitter.
Stacks, D., & Bowen, S. (2013). Dictionary of public relations measurement and researchRetrieve from(3rd ed.). Gainsville, FL: Institute for Public Relations. http://www.
instituteforpr.org/dictionary-public-relations-measurement-research-third-edition/.
Taylor, M., & Doerfel, M. L. (2005). Another dimension to explicating relationships: Measuring inter-organizational linkages. Public Relations Review, 31, 121–129.
https://doi.org/10.1016/j.pubrev.2004.11.013.
Taylor, M., & Kent, M. L. (2014). Dialogic engagement: Clarifying foundational concepts. Journal of Public Relations Research, 26(5), 384–398. https://doi.org/10.
1080/1062726x.2014.956106.
Time (2016). A brief history of Samsung’s troubled Galaxy Note 7 smartphone. Retrieved from http://time.com/4526350/samsung-galaxy-note-7-recall-problems-
overheating-fire/ on Oct 25, 2017.
Tsai, W. (2001). Knowledge transfer in intraorganizational networks: Effects of network position and absorptive capacity on business unit innovation and performance.
The Academy of Management Journal, 44, 996–1004. https://doi.org/10.2307/3069443.
Tsai, W. S., & Men, L. R. (2013). Motivations and antecedents of consumer engagement with brand pages on social networking sites. Journal of Interactive Advertising,
13, 76–87. https://doi.org/10.1080/15252019.2013.826549.
Twitter (2016). Twitter Usage/Company facts. Retrieved fromhttps://about.twitter.com/company/.
Utz, S., Schultz, F., & Glocka, S. (2013). Crisis communication online: How medium, crisis type and emotions affected public reactions in the Fukushima Daiichi
nuclear disaster. Public Relations Review, 39, 40–46. https://doi.org/10.1016/j.pubrev.2012.09.010.
van der Meer, T. G. L. A. (2016). Public frame building: The role of source usage in times of crisis. Communication Research, 1, 1–26. https://doi.org/10.1177/
0093650216644027.
Veil, S. R., Buehner, T., & Palenchar, M. J. (2011). A work-in-process literature review: Incorporating social media in risk and crisis communication. Journal of
Contingencies and Crisis Management, 19(2), 110–122. https://doi.org/10.1111/j.1468-5973.2011.00639.x.
Wall Street Journal (2017). Russian hackers stole NSA data on U.S. Cyber defense. Retrived from https://www.wsj.com/articles/russian-hackers-stole-nsa-data-on-u-s-
cyber-defense-1507222108 on Oct 25, 2017.
Wasserman, S., & Faust, K. (1994). Social network analysis: Methods and applications. Structural analysis in the social sciences. New York: Cambridge University Press.
Yang, A., & Taylor, M. (2014). Looking over, looking out, and moving forward: Positioning public relations in theorizing organizational network ecologies.
Communication Theory, 25, 91–115. https://doi.org/10.1111/comt.12049.
Ye, S., & Wu, S. F. (2010). Measuring message propagation and social influence on twitter.com. SocInfo, 10, 216–231. https://doi.org/10.1007/978-3-642-16567-2_16.
Zhao, X., Jie, C., & Zhan, M. (2017). Toward a social-mediated crisis theory (SCARE): Modeling topics from big data for chipotle E. coli crisis. November Paper
Presented at 103th Annual Conference of National Communication Association (Public Relations Division).
Zhao, X., Zhan, M., & Wong, C.-W. (2018). Segmenting and understanding publics in a social media information sharing network: An interactional and dynamic
approach. International Journal of Strategic Communication, 12, 25–45. https://doi.org/10.1080/1553118X.2017.1379013.

561