University of Southern California Dissertations and Theses

The good, the bad, and the longitudinal: testing dynamic prosocial and toxic behaviors in online commercial games

(USC Thesis Other)

The good, the bad, and the longitudinal: testing dynamic prosocial and toxic behaviors in online commercial games

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content
THE GOOD, THE BAD, AND THE LONGITUDINAL:
TESTING DYNAMIC PROSOCIAL AND TOXIC BEHAVIORS IN
ONLINE COMMERCIAL GAMES

by

Mingxuan Liu

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(COMMUNICATION)

August 2023

Copyright 2023 Mingxuan Liu

ii
ACKNOWLEDGEMENTS
The defense took only a little over an hour on a single day, but it took me 23 years to get
to that point. I grew up in Henan, a major agricultural province and the most populous province
in China. Educational resources were scarce, and the competition was fierce. When I took the
college entrance examination in 2012, there were over 800,000 candidates in Henan. The
acceptance rate for first-tier universities, which meant the probability of entering a relatively
good university, was only 5.83%. Ten years later, in 2022, the number of candidates in Henan
had risen to 1.25 million, the highest in the country, but the acceptance rate for first-tier
universities was the lowest nationwide. From childhood to adulthood, the phrase I heard the most
in school was that the college entrance examination was like a narrow bridge where thousands of
troops and horses crossed, and where a one-point difference in grades can separate thousands or
even tens of thousands of people. I gradually internalized the idea that studying was the only
important thing, and other recreational activities felt like a sin. This sense of guilt towards leisure
activities has not completely faded even now.
When I was little, I skipped kindergarten and went straight to elementary school. There
were more than 60 students in my class, and my grades were always at the bottom. In my fourth
grade, a teacher who was going to leave after teaching our class called me and another classmate
to her office and gave each of us a notebook, saying that she believed I had potential. That was
the first time I felt recognition from a teacher. In junior high school, I still didn't think studying
was a necessary thing for me. My dream at the time was to open a coffee bookstore, which didn't
require good grades. As a result, the teacher in charge of the class didn't like me, and I was not
allowed to participate in extracurricular activities. Sometimes, she would even send me home
during class, asking me to pack my bag in front of the whole class. I wasn't allowed to wear

iii
jeans, keep long hair, or have bangs longer than my eyebrows. These rules and punishments only
applied to girls with poor grades, not boys.
A change happened one day when the teacher said that if there was a typo in the dictation
of an ancient poem the next day, we would have to copy the poem a hundred times. Because I
didn't want to copy the poem a hundred times, I studied the poem diligently, and the next day, I
was the only one in the class who got it completely right. That was the first time I felt the
positive feedback of hard work. Later, my grades improved, and I was admitted to the best high
school in the province. The best high school in the province was indeed different, even though
my initial grades placed me at the bottom again, the teachers patiently guided me. I had a strong
sense of frustration at that time, and my personality changed from outgoing to introverted. I
almost gave up. However, my Chinese teacher did not give up on me. He frequently chatted with
me to understand my dilemmas, tailored his teaching strategies, and gradually restored my
confidence. When I asked him why he paid attention to me back then – each high school teacher
oversaw hundreds of students, he said after reading the articles I wrote, he felt that I shouldn't be
defined solely by my grades and that he believed I had potential. He was the second teacher who
acknowledged my potential.
In Chinese, there is a saying that there are often exceptional horses, but not often
exceptional judges. I am very grateful to these teachers who did not give up on me during my
low periods. East Asian education is said to be "exam-oriented," and from a young age, I
believed that there was only one “correct” path for a fulfilling life. When I was in elementary and
junior high school, the internet was just emerging, and I started writing articles and novels
online, blogging, and interacting with many interesting people. That was probably one of the few
times in my life when I genuinely loved something. However, my father believed it was a

iv
distraction from my studies. One day, he made me announce on my Blog that I was preparing for
the high school entrance examination and wouldn't have time for the internet. I felt ashamed
because I didn't want others to know that I was just a junior high school student, but I still
followed his instructions. After the high school entrance examination, I never picked up the pen
to write a long article online again. Expressing opinions became a luxury reserved for fearless
individuals, but as time passed, I transformed from a fearless person into a self-conscious one.
In college, my goals became clear, and I started to enjoy learning. I knew I wanted to
pursue a Ph.D., and I was on my way to be one. During holidays and special occasions, I would
receive praises from elders like any other children who fit into the mainstream narrative. Every
time Dr. Liu was praised, the childhood blogger Mingxuan and the rebellious teenager Mingxuan
would look at her like floating souls, asking, "Did we die because of her?" It was probably after I
arrived in Los Angeles that I gradually found inner peace again. Los Angeles is vast, with
various kinds of people, including Chinese immigrants who came here through different means.
The Annenberg School is also large, with faculty and staff from different disciplines and
classmates from all over the world. I love this heterogeneous and diverse environment where
there is no uncomfortable attention, no cliques judging outsiders, and those in positions of power
do not look down upon newcomers. Here, I feel like water dissolving into water.
In high school, I used to have conversations with my Chinese teacher during breaks. He
said that he believed there was a special connection between humans and the universe. The
smallest structural units and the orbits of celestial systems are so similar, and the perfection of
circles, how could all this be a mere coincidence? Without any reason, I also came to believe that
the connections between people are inherently close. All humans, like cells, make up a larger
being. The suffering of the world is my own suffering, and the glory of the world is my own

v
glory. So I slowly let go of the obsession with the version of myself that I did not become. If
someone else is doing what I wanted to do, then I feel I have achieved that as well. I am proud of
them as I am proud of possible myself. Los Angeles is such a place. I see people of different skin
colors, genders, and backgrounds, who came here through various paths, doing all kinds of
things. I feel fortunate and privileged to be in this environment. Marshall McLuhan said that
media are extensions of humans. I believe that humans are extension of humans.
I still feel confused about many things. Humans, like yin and yang, are inherently
contradictory. Muscles need to be torn to grow, and the hurts that impact us make us stronger.
However, I am grateful to many people for being there for me when I am confused. In junior
high school, my desk mate Shuang Chen, another girl not allowed to wear jeans in my class, has
been my best friend since I was 12 years old. At that time, I would secretly bring my phone to
school and, upon hearing rumors that the class teacher would search my drawer, I would hide my
phone in her drawer. She is a person who is self-contained and confident, with parents who have
no demands on her. She is the most carefree yet mature and steady person I have ever known.
From junior high to now, she has listened to me the most and given me the greatest support. I
also want to thank my high school friends, Yanrui Huang, Yaojia Wang, and Ke Bai. I remember
when the pressure was too overwhelming, I would run out of school and call Ke Bai, a friend
who received early admission to university and didn't need to be in school. He would accompany
me for a walk and then persuade me to go back to school, saying that even if I could study for a
few minutes, I would gain something worth of a few minutes. Whenever I find it difficult to calm
my mind, I think of his words. He is the person I can turn to when I want to escape from the
world. And special thanks to my friends from college, Zhuangbao Chen, Liyang Shen, and Wen
Sun. Liyang and Wen were my roommates at University of Macau, and we have had our share of

vi
conflicts, but our relationship has become incredibly accommodating because of those conflicts.
I can share my true feelings with them, and this is something that is hard to come by in
interactions with others, even in close relationships. My friends are like family to me, making
sure I am never alone in any situation, always having someone to call, and having someone
unconditionally support me.
As people grow up, it becomes a bit difficult to make new friends. But I am fortunate to
have met many good friends at Annenberg. I want to thank my friend Junyi Lv, who has been
with me throughout these five years. The time we spent together strolling at Westfield, drinking
milk tea, and gossiping has been the highlight of my Ph.D. journey. I am also grateful to my
friend Jack Tang for being a reliable co-author and a candid friend. And to many other wonderful
humans I have met throughout the years—Derek, Qiyao, Kathy, Steffie, Feixue, Lichen,
Sukyoung, Emily, Herbert, Eugene Jang, Eugene Lee, Meiqing, Jingyi, Yiqi, Qiusi, Jade, Grace,
Donna, Yilei, Alex, Becky, Soyun, Hye Min, Yu Xu, Zihan, Youxin, Xiyuan, Zhan, Wenpei,
Jiatao, Xia, Stevie, my cohort friends, my Master’s advisor Narine Yegiyan, my Master’s
committee members Jingwen Zhang and Drew Cingel, as well as professors Emilio Ferrara,
Andrea Hollingshead, Jonathan Gratch, Robin Stevens, Aimei Yang, Su Jung Kim, Tom Valente,
Bo Feng, Cindy Cuihua Shen, Pan Lei, Ying Li, my previous mentors and teachers, staff at UC
Davis and USC: Anne Marie Campian, Sarah Holterman, Stephanie Fallas, and countless
others— for their support. I also want to thank Wargaming and thatgamecompany for providing
access to their anonymized data. I especially wish to thank Tina Lu and Jenova Chen at
thatgamecompany, and Eugene Kislyi and Jeremy Ballenger at Wargaming for their input and
support.

vii
I would also like to express my sincere gratitude to my advisor, Lynn Miller, and my
committee members Dmitri Williams and Lindsay Young. The first time I visited Lynn's
webpage, I felt a strong connection between our research interests in communication science and
psychology. Thank you, Lynn, for meeting with me and choosing me as your advisee. Lynn has
always encouraged me to expand the boundaries of my research and supported my interests and
decisions. She has been like a kind and caring mentor, guiding and taking care of her students.
Under her encouragement, I learned about neuroscience and fMRI and gained a new perspective
on the world that will benefit me for a lifetime. In my first semester at Annenberg, I took
Dmitri's class. He treats students as colleagues and collaborators, and he is so good at writing
fascinating papers. I really enjoy working with Dmitri and feel fortunate to have had the
opportunity to learn from him. In my third year, I met Lindsay and invited her to join my
committee. Lindsay provides her students with care, patience, and love. Having been Lindsay’s
research assistant and teaching assistant, I learned so much from Lindsay, and I truly wish I had
met Lindsay earlier.
Of course, the most deserving of thanks are my parents. Although I have complained
about them viewing academics as the only path to success, which made me feel restricted and
deprived of unconditional love and other possibilities, it is unfair to hold such grievances against
them. They firmly believed in my potential even when there were no signs of academic talent.
They dedicated themselves to nurturing and supporting me. This support was unconditional,
selfless, and luxurious. My father excels both academically and professionally. He is not only a
great scholar but also a remarkable teacher. His passion and dedication to academia are rare and
pure. Moreover, he possesses excellent management and leadership skills. At the same time, he
is an incredibly compassionate person. I have witnessed him supporting elderly patients in the

viii
hospital, assisting an elderly gentleman in the same ward who had no one to care for him when
he needed to use the restroom. This exemplary behavior has had a significant impact on me. My
mother is also exceptionally accomplished. Like many mothers, she often undervalues herself
and claims that she only knows how to handle household chores. For decades, she has
shouldered the majority of the household labor consistently, and I didn't fully realize her value
when I was younger. She is the glue that holds our family together. I hope that both my father
and mother stay healthy and happy.
A significant portion of my research revolves around well-being. Through my studies, I
have discovered a strong connection between psychological well-being and a sense of existence.
We want to be seen, be acknowledged, and be needed, although often we are unaware of our
pursuit of the sense of existence. I am unsure why humans are so fixated on the notion of
existence, but if we are to dig into reasons behind the pursuit of a sense of existence, it is like an
inherent mission, akin to ocean currents propelling water droplets. If finding a sense of existence
is a path to happiness, then I hope my sense of existence stems not from how others perceive me,
but from overcoming fears and pushing beyond my own boundaries. Individuality and
uniqueness hold immense significance in this world, just like the value of diverse cells to human
beings. When I was young and faced skepticism from others, I always wished to grow up
quickly, knowing that I would achieve my goals one day—it was merely a matter of time. Even
now, I maintain this mindset. I haven't yet conducted the research that holds academic and
societal value that meets my own expectations, but I know that I am moving towards that
direction. I cannot wait to be me. Recognizing the support and opportunities I have received, I
am deeply committed to giving back to the community that uplifted me. I aspire to transcend
disciplinary boundaries and address social inequities through digital equity, inclusion, and the

ix
correction of information inequalities. However, I know that there is still much to learn,
accumulate, and digest. I need more time. At 28 years old, I am still growing, eagerly
anticipating the next ten or twenty years and reflecting on my present self.
In conclusion, in this dedicated expression of gratitude, I am thankful for myself, who
continues to live with sincerity and resilience. I am grateful for my parents, relatives, and friends,
for their support. I appreciate my mentors and teachers for their guidance. I am thankful for those
who have shown kindness towards me. I am thankful for my hometown, Zhengzhou, in Henan,
China. I am grateful for the inclusion and diversity of Los Angeles, and for the sunshine, ocean,
and palm trees of California. I appreciate the Chinese chefs in San Gabriel and Rowland Heights,
and I am grateful for Google, YouTube, TikTok, WeChat, and ChatGPT. I am thankful to those
who live authentically and showcase the various possibilities of life.

x
TABLE OF CONTENTS
ACKNOWLEDGEMENTS ............................................................................................................ii
LIST OF TABLES .......................................................................................................................xii
LIST OF FIGURES .....................................................................................................................xiii
ABSTRACT ................................................................................................................................xiv
CHAPTER 1: INTRODUCTION ..................................................................................................1
Multi-Faceted Conceptualizations of Gaming Toxicity ....................................................2
Precedents of Gaming Toxicity: Existing Research ..........................................................3
Precedents of Gaming Toxicity: Missing Pieces ...............................................................4

CHAPTER 2: NETWORK POSITION, POWER, RISK, AND GAMING TOXICITY
(STUDY 1) .........................................................................................................................6
Structural Underpinnings of Toxicity: Power as a Communicated Product in Networks ..8
Network Structure, Power, and Accountability ................................................................10
Network Size, Power, and Risks .......................................................................................11
Network Perspective toward Perpetration-Victimization Overlap ...................................16
Methods .............................................................................................................................18
Research Site .........................................................................................................18
Data .......................................................................................................................19
Measures ...............................................................................................................20
Analysis .................................................................................................................21
Results ...............................................................................................................................22
Discussion .........................................................................................................................25

CHAPTER 3: THE CO-EVOLUTION OF FRIENDSHIP NETWORKS AND TOXIC
BEHAVIORS IN A MULTIPLAYER ONLINE GAME (STUDY 2) .............................34
Friendship Formation: Birds of a Feather .........................................................................36
Friendship Formation: Matthew Effect .............................................................................39
Network Influence on Toxic Behavior .............................................................................40
Coevolution of Behavior and Network .............................................................................43
Methods ............................................................................................................................44
Study Context: World of Tanks ............................................................................44
Data and Sampling ...............................................................................................45
Measures ...............................................................................................................46
Analysis ................................................................................................................49
Results ...............................................................................................................................53
Discussion .........................................................................................................................56

CHAPTER 4: A NATURAL EXPERIMENT OF HOW STAY-AT-HOME ORDERS
UNLEASHED A WAVE OF VIRTUAL ALTRUISM (STUDY 3) ................................63
The Model of Altruism Born of Suffering ........................................................................65

xi
Contextualizing the Model under a Collective Trauma in Virtual Environments ............67
Self-Categorization in the Virtual World ..........................................................................69
Methods ............................................................................................................................72
Study Site ..............................................................................................................72
Data Collection .....................................................................................................73
Measures ...............................................................................................................74
Analysis ................................................................................................................76
Results ...............................................................................................................................78
Discussion .........................................................................................................................80

CHAPTER 5: GENERAL DISCUSSION ....................................................................................88
REFERENCES .............................................................................................................................93

xii
LIST OF TABLES
Table 2.1. Illustration of network positions and accountability cues .............................................8
Table 2.2. A summary of hypotheses testing ...............................................................................18
Table 2.3. Poisson regression using network variables to predict victimization and perpetration
count ..............................................................................................................................................24
Table 2.4. Logistic regression using network variables to predict victims, perpetrators,
aggressive victims, and uninvolved players ..................................................................................25
Table 3.1. Summary of network changes from T1 to T2 .............................................................47
Table 3.2. Summary of behavior changes from T1 to T2 ............................................................48
Table 3.3. Description of the effects included for testing selection and influence processes ......52
Table 3.4. Significance of parameter estimates of the network and toxic behavior submodels ..55
Table 4.1. Average effect of stay-at-home order on CA players’s prosocial behavior ................72

xiii
LIST OF FIGURES
Figure 3.1. Friendship networks of the studied clan at each wave with information about their
toxicity level ..................................................................................................................................54
Figure 3.2. The relative importance of each effect on friendship network formation .................56
Figure 3.3. The relative importance of each effect on toxic behavior evolution .........................56
Figure 4.1. Theoretical framework for the proposed studies .......................................................72
Figure 4.2. Decision tree of ingroup outgroup categorization .....................................................75
Figure 4.3. Average effect of stay-at-home order on CA players’ prosocial behavior toward
unknown players ...........................................................................................................................80
Figure 4.4. Average effect of stay-at-home order on CA players’ prosocial behavior toward
outgroup players ............................................................................................................................80

xiv
ABSTRACT
This dissertation aims to explore how network embeddedness, social interactions, and the
overarching social environment influence toxic and prosocial behavior in online communities.
Using longitudinal data from 69,605 players in a multiplayer online game over two months,
Study 1 analyzed network size and two forms of structural social capital as predictors of gaming
toxicity. The findings showed that players with larger network sizes and network closure were
less likely to engage in incidence-based perpetration and victimization, while those with higher
network brokerage were more likely to be involved in both perpetration and victimization.
Perpetration-victimization overlap was predicted by larger network sizes and lower closure.
Study 2 further examined the coevolution of friendship formation and toxicity diffusion within a
social group in a multiplayer video game. Using stochastic actor-oriented modeling, the study
found that players tend to establish new ties with popular and skillful players. Additionally, the
toxic behavior of group members was influenced by that of their friends, but with considerable
between-person variability. Study 3 examined the effect of the stay-at-home order during the
COVID-19 on players’ prosocial behaviors in a social multiplayer game as a natural experiment.
Results from 306,504 unique daily observations of natural behavior from 13,932 players showed
that the stay-at-home order had a positive effect on players’ prosocial behavior toward unknown
players, but a negative effect on players’ prosocial behavior toward outgroup players. Theoretical
and practical implications of the three studies are discussed.

1
CHAPTER 1: INTRODUCTION
Video games are a form of interactive media that can be played across a wide range of
electronic devices (Bowman et al., 2020). Due to the surge of mobile devices and digital
platforms, the gaming industry has seen substantial growth in recent years. The COVID-19
pandemic has further propelled the industry, with its revenue exceeding that of the global film
industry and North American sports industry combined (Witkowski, 2021). As of 2021, an
estimated 3.24 billion individuals around the globe engage in video gaming, representing
approximately 40% of the global population (Statista, 2021).
The burgeoning of the gaming industry also involves a proliferation of gaming toxicity.
According to a recent report from Unity, in 2021, 72% of multiplayer gamers have witnessed
toxic behavior towards others while playing multiplayer video games, with 68% experiencing it
themselves. Research has shown that players who encounter toxicity early on in their gaming
experience have a 320% higher likelihood of abandoning the game, leading to reduced player
retention, negative reputation, and decreased lifetime value for gaming companies (Grandprey-
Shores et al., 2014). Nevertheless, the impacts of toxicity go beyond economic consequences.
Victims of toxic behavior may suffer from negative psychological effects such as sadness,
isolation, and even suicidal ideation (ADL, 2022). These effects may also extend to loved ones
of the victim and even to the perpetrators themselves (Eisenberg et al., 2015; Marchiondo et al.,
2020). Moreover, as toxicity can spread from one individual to another (Cheng et al., 2017; Shen
et al., 2020), exposure to toxic behavior can raise the likelihood of witnesses engaging in such
behaviors, creating a self-perpetuating cycle of aggression. Toxicity can also have significant
impacts on online and offline social norms, especially for younger generations who have grown
up with digital media (Shen et al., 2020).

2
Multi-Faceted Conceptualizations of Gaming Toxicity
The term "toxicity" encompasses a wide range of deleterious behaviors that have been
further exacerbated by the proliferation of information and communication technology. In the
realm of online gaming, toxicity has been associated with disruptive conduct, including
cyberbullying, griefing, and cheating (Blackburn & Kwak, 2014; Kwak et al., 2015). Given the
harmful effects of toxicity on both players and games, subsequent definitions have emphasized
its outcomes, characterizing it as rule-breaking behavior that leads to anger, frustration, and
negative gaming experiences for others (Neto et al., 2017). To develop a shared language on
gaming toxicity, Kowert (2020) differentiated between dark participation, toxicity, and trolling.
While all deviant verbal and behavioral actions online may fall under the broad heading of dark
participation, only those resulting in negative outcomes on others' health and well-being are
considered toxic. Toxicity is outcome-oriented and culturally defined. Behaviors that are deemed
toxic in one context may not be so in another. For instance, the use of mods to gain an unfair
advantage in gaming is viewed as deviant (i.e., dark participation), but whether it qualifies as
toxic depends on the context and outcome. Trolling is a subcategory of toxicity in which the
perpetrator intentionally seeks to elicit an emotional response or disruptive behaviors from other
players. For example, trash talking, griefing, and doxing are toxic behaviors that are considered
as trolling since they are intended to cause harm to others.
Similarly, definitions of toxicity within the gaming industry vary depending on the game,
with some enforcing rigorous categorization and serious punishment of toxic behavior, while
others take a more lenient approach. To promote a more uniform stance on gaming toxicity,
leading video game companies established the Fair Play Alliance and partnered with the ADL to
combat hate and harassment in video games (ADL, 2020). In their recent Disruption and Harms

3
in Online Gaming Framework (2020), toxicity is seen as an ambiguous term that lumps all
undesirable behaviors together, providing insufficient actionable insights. Similar to the concept
of dark participation, the framework employs “disruptive behavior” as an encompassing term
that highlights behavior that deviates from community norms and impairs player experience.
However, not all disruptive behavior is negative, and behavior that causes significant emotional,
mental, or physical harm to others is defined as “harmful conduct”. The term “disruptive
behavior” is considered useful for game developers as it prompts reflection on what is disrupted
and why, placing more moderating responsibilities on game designers rather than users. In
academic circles, the term “toxicity” aligns more with “harmful conduct” in the gaming industry.
In line with Kowert’s (2020) definition, this paper uses “toxicity” to refer to harmful conduct in
gameplay that deviates from game norms and creates negative gaming experiences for players,
regardless of intentionality.
Precedents of Gaming Toxicity: Existing Research
In recent years, due to the widespread occurrence and detrimental effects of gaming
toxicity, academic research has increasingly focused on investigating the antecedents and
theoretical underpinnings of gaming toxicity from various perspectives (e.g., Kordyaka et al.,
2020; Kwak et al., 2015; Neto et al., 2017). At the platform level, recent research has applied
classic theories from computer-mediated communication, such as the Social Identity Model of
Deindividuation Effects (SIDE) (Postmes et al., 1998) and Online Disinhibition Effect (ODE,
Suler, 2004), to elucidate how the affordances of online games contributes to toxicity
disinhibition by deindividuating other players, especially those in out-groups, in anonymous
virtual spaces. At the game level, studies suggest that zero-sum mechanisms and certain design
features of games may promote or reward aggressive behavior (A. Sparrow et al., 2021). At the

4
individual level, studies have identified socio-demographic characteristics that make players
more susceptible to perpetrating toxic behavior or more vulnerable to being victimized. For
example, LGBTQIA players (Ballard & Welch, 2017), players of color (Gray, 2012; Ortiz,
2019), and female players are disproportionately targeted by gaming toxicity (Fox & Tang, 2017;
McLean & Griffiths, 2019), while young (C. Barlett & Coyne, 2014), male (Fox & Tang, 2017),
heterosexual players (Ballard & Welch, 2017), and players with social dominance orientation,
machiavellianism (Tang et al., 2020), and emotion reactivity (Lemercier-Dugarin et al., 2021)
are more prone to engage in toxic behavior.
Precedents of Gaming Toxicity: Missing Pieces
Although previous research has generated valuable insights into the antecedents of
gaming toxicity, there remains a dearth of literature that explores this phenomenon from a socio-
structural perspective, despite the realization that “a link between bullying and individuals’
position in the social structure is already included in definitions of bullying as characterized by a
power imbalance between perpetrator and target” (Pauksztat & Salin, 2021, p. 455). Gaming
toxicity is fundamentally a social behavior that arises from human interactions and is embedded
in relationships and contexts. On the one hand, network structures confer unique power to
individuals in certain positions, while on the other hand, individuals in different network
positions experience distinct interpersonal and social dynamics, resulting in varying patterns of
behavior. Accordingly, Study 1 investigates how different network positions are associated with
distinct levels and forms of social capital (bridging and bonding) and network centrality that
predict players' roles and involvement in gaming toxicity dynamics. As the results of Study 1 are
correlational in nature and cannot establish causal inferences between network selection and
toxic behavior evolution, Study 2 delves into selection-influence dynamics by modeling the

5
coevolution of network formation and toxic behavior diffusion using stochastic actor-based
modeling. As Study 1 and Study 2 focus on endogenous factors that predict behavior change,
Study 3 employs the exogenous stay-at-home order during the COVID-19 pandemic as a natural
experiment to scrutinize how players' prosocial behavior changes in response to policies and
social contexts towards various types of recipients, including anonymous, ingroup, and outgroup
players.

6
CHAPTER 2: NETWORK POSITION, POWER, RISK, AND GAMING TOXICITY
(STUDY 1)
Multiplayer online video games, as one of the most popular forms of entertainment
media, have provided a wide array of benefits to players, including social connection, a sense of
achievement, and relaxation (Cairns et al., 2021). These factors, along with their accessibility
and connectivity, make video games particularly appealing to marginalized populations such as
minors, underrepresented groups, and individuals with disabilities who may face social isolation
otherwise (Finke et al., 2018). While these vulnerable populations derive benefits from video
games, they are also more susceptible to negative experiences in the virtual realm (Gardner,
2021). According to a recent Anti-Defamation League survey conducted in 2021 (ADL, 2022),
65% of those in marginalized groups, including women, Jewish, Muslim, LGBTQ+, non-white,
or disabled individuals, experienced identity-based harassment in online gaming, compared to
38% in non-marginalized groups. Online incivility greatly harms user engagement (Lu et al.,
2022), yet the deleterious consequences of gaming toxicity extend beyond financial losses and
disruption to the norms of online communities (Grandprey-Shores et al., 2014) to encompass the
health and well-being of those affected, including the perpetrators themselves (ADL, 2022) and
their families (Eisenberg et al., 2015; Marchiondo et al., 2020).
The prevalence and consequences of toxicity in online gaming have spurred numerous
investigations into the contributing factors of such behavior (Lemercier-Dugarin et al., 2021;
Shen et al., 2020) and the identification of players who are susceptible to engaging in or being
targeted by it (Ballard & Welch, 2017; Ortiz, 2019). These studies provide valuable insight into
factors that contribute to gaming toxicity, such as anonymous gaming environments (Kordyaka
et al., 2020), zero-sum game designs (Adinolf & Turkay, 2018), and power imbalances between

7
perpetrators and victims (Lemercier-Dugarin et al., 2021; Tang et al., 2020). However, aside
from a few exceptions (e.g., Yokotani & Takano, 2021), most research has failed to consider the
social networks and relationships in which players are embedded when explaining, identifying,
and predicting gaming toxicity.
Utilizing large-scale, unobtrusively collected longitudinal egocentric network data and
toxic behavior data of the popular multiplayer online video game, World of Tanks (WoT), the
current study integrates the Structural Hole Theory (SHT; Burt, 2001) and the Shadow of the
Future Effect (Axelrod & Hamilton, 1981) to explain influence of players' network positions on
power dynamics, accountability, and behavioral tendencies. On one hand, key network features,
such as network size, network brokerage, and network closure, embody various forms of
informal power and bestow advantages such as potential social capital upon individuals. On the
other hand, individuals occupying distinct network positions may exhibit diverse levels of
cooperation, kindness, and aggression toward others, contingent on the distinct cues of
anonymity, accountability, and "shadow of the future" in their social interactions. From an
evolutionary and game theory standpoint, individuals tend to engage in competition and only
cooperate when they foresee rewards for cooperation and penalties for non-cooperation (Van
Lange et al., 2011). If players consistently interact with familiar counterparts and anticipate
future interactions, they are more inclined to cooperate and regulate their behavior accordingly.
An illustration of network positions, power, and accountability can be found in Table 1.
Theoretically, the findings support a more comprehensive theoretical approach that links
the power and risks associated with various network positions to individuals’ behavioral patterns
in toxicity dynamics. The adoption of an actor-based approach toward gaming toxicity provides
an explanation for the perpetration-victimization overlap phenomenon by identifying the

8
characteristics of aggressive victims. Practically, the findings provide insights for game
developers and community managers that can aid in designing effective interventions and
policies to address toxic behavior and create a more enjoyable gaming experience for all players.
Table 2.1. Illustration of network positions and accountability cues.

Structural Underpinnings of Toxicity: Power as a Communicated Product in Networks
Toxicity refers to behaviors that are disruptive and harmful to the health or well-being of
others, regardless of their intent. Its categorization and consequences depend on the outcome and
context, thus varying across different platforms. Although previous research has generated
valuable insights into the antecedents of gaming toxicity, there remains a dearth of literature that
explores this phenomenon from a socio-structural perspective, despite the realization that “a link
between bullying and individuals’ position in the social structure is already included in
definitions of bullying as characterized by a power imbalance between perpetrator and target”

9
(Pauksztat & Salin, 2021, p. 455). Gaming toxicity is fundamentally a social behavior that arises
from human interactions and is embedded in relationships and contexts. Various network
structures and relational dynamics exert distinct impacts on the perpetration and victimization of
players through two distinct pathways: one stemming from informal power embedded in pivotal
network positions, and the other from anonymity and accountability cues that players experience,
along with the shadow of the future effect in terms of sanctions and loss of the status quo.
Differences in social interactions and network positions lead to differences in power
(Cook & Emerson, 1978), and power differences between individuals are crucial predictors of
toxicity (Aivazpour & Beebe, 2018). Prior studies have found that individuals with formal
power, such as legitimate power in the workplace, are more likely to engage in aggressive
behavior towards those with less power. This is because power differences make it easier for
power holders to commit harmful actions with little consequence (Aquino & Lamertz, 2005).
Similar to the workplace, research on toxicity in large multiplayer online games has revealed that
experienced and skilled players are more inclined to engage in toxic behavior than new players
(Shen et al., 2020). However, research has long recognized that power is not solely an outcome
of formal status, but also a characteristic of one’s position in a network structure, regardless of
whether the individual is cognizant of their position or level of power (Cook & Emerson, 1978).
Therefore, in recent years, there have been mounting calls to study online toxicity from a socio-
structural perspective, using social network analysis to account for individuals’ social ties and
relationships in toxicity dynamics (Laurie-ann et al., 2021; Wegge et al., 2013)
In line with this approach, the Social Structural Model of Social Power and Status Effects
on Victimization (SSM; Lamertz & Aquino, 2004) is one of the pioneering theoretical
frameworks that directly links network structure with power and victimization in organizational

10
settings. The theory posits that different network positions result in an unequal distribution of
status and social power, leading to variations in how social actors experience protection or
vulnerability to harmful behaviors. SSM further distinguishes between formal and informal
power. Formal power is derived from holding a formal position that connects people in
prescribed interactions and thus enjoys certain social resources. Informal power stems from the
possession of informal status and social resources that are valued by others. Unlike formal
power, informal power works as perceived power wherein individuals believe that their actions
have a greater influence on game outcomes compared to others (Zhu et al., 2022).
Similarly, a communication-centric approach to power contends that it is generated
within and by social relations (Barnett & Duvall, 2004), and can only be exercised through
communication (Mumby, 2015). Power is communicatively constructed and exercised through
interdependent relations that shape actors’ capacity to “determine their own circumstances and
fate” (Barnett & Duvall, 2004, p.3). Within the context of a video game, if players are able to
control their own circumstances, they are deemed to have power. This may be accomplished
through formal status, such as possessing superior skills, resulting in the confidence to handle the
consequences of battle, or informal status, such as having a sufficient number of friends to play
with at any time, granting them autonomy in selecting when to play and who to play with. If they
have enough allies during game battles, they may even influence others' circumstances to some
degree. Therefore, the current study considers network structures as the foundation of informal
power, endowing players with the ability to dictate their own circumstances and those of others.
Network Structure, Power, and Accountability
As network structures confer informal power to individuals, with great power comes
great accountability. While some may use it to protect themselves from potential victimization,

11
others may misuse it for their own benefit or become constrained by the contingencies that come
with informal power. According to the Differential Self-Awareness Theory (DSAT; Prentice-
Dunn & Rogers, 1982) and the shadow of the future effect (Axelrod & Hamilton, 1981), if social
interactions signal accountability cues that suggest the degree to which individuals will be held
responsible for their actions, they will reduce deindividuation and increase self-awareness,
resulting in less aggressive behavior. Similarly, if social interactions signal internal attentional
cues that direct an individual's focus on themselves, they will increase self-awareness and
decrease aggressive behavior. In a game setting, individuals who develop more ties in the
network and play with friends may experience a reduction in anonymity and an increase in
internal attentional cues and accountability cues. Because they anticipate having frequent
interactions with their friends in the future, there is a greater "shadow of the future" effect, and
engaging in toxic behavior may result in greater social costs, such as being shunned by friends or
groups, thus losing their status quo or informal power. Compared to individuals playing with
random strangers, those with more network ties are more likely to be held accountable for their
abnormal behavior, which could negatively affect their reputation and power. Thus, while
network structures endow individuals with informal power that shields them from victimization,
whether they use the power for aggressive behavior or not depends on the level of accountability
and attentional cues, as well as the shadow of the future effects.
Network Size, Power and Risks
The communication-centered approach toward power conceptualizes power as the
production of interactions of varying frequencies and symmetries (Olsen et al., 2014). Actors
with more interactions (i.e., larger network size) are endowed with amplified power, as they are
more inclined to shape their own circumstances and those of others (Saffer et al., 2018). Indeed,

12
longitudinal studies conducted in organizational contexts have demonstrated that network
centrality precedes the acquisition of power (Burkhardt & Brass, 1990). From a network
perspective, individuals with more ties in their personal network are perceived as more popular
and wield higher informal authority, thus safeguarding themselves against victimization (Festl &
Quandt, 2013). For example, Mouttapa et al. (2004) examined the social network predictors of
bullying and victimization in a school context and found that the number of friendship
nominations received by students exhibited a negative association with their vulnerability to
victimization. Aside from providing protection against victimization, a larger network size in the
realm of video games carries additional implications for behavior adoption. In the diffusion of
innovation (DOI) (Rogers, 1995, 2010) literature, typically, central members enjoy popularity
and influence, thus facilitating the spread of behaviors through their connections. However, in
circumstances where behaviors are considered risky or incompatible with existing norms, the
diffusion of behaviors may occur in the opposite direction—flowing from peripheral actors to
central actors (Z. Chen et al., 2021). This tendency likely arises from the disincentives for central
members to adopt non-normative and potentially threatening behaviors that could jeopardize
their social status. Conversely, individuals who experience less social pressure, such as those
with fewer social connections or individuals serving as brokers, are more inclined to engage in
initial experimentation of risky and non-normative behaviors (Sgourev, 2013). In addition, when
playing with known others, under the shadow of the future effect, players tend to regulate their
behavior to a greater extent than when engaging with unfamiliar counterparts. This behavior
regulation arises from the desire to avoid being perceived as a toxic player, a reputation that
could hinder future collaborations with friends. Furthermore, their level of deindividuation may

13
diminish, while their public self-awareness may heighten, as they imagine being evaluated
through the lens of their friends. Consequently, it is predicted that:
H1a-b: Controlling for formal social status, individuals who have a larger network size
will (a) engage in fewer toxic behavior and (b) experience less victimization;
Structural Embeddedness, Power and Risks
There are two types of informal power: referent power and expert power (Lamertz &
Aquino, 2004). Referent power is associated with interpersonal identification and attraction,
stemming from a desire to connect with others, whereas expert power is associated with
knowledge and proficiency with information and resources. For example, Litsa and Bekiari
(2022) surveyed high school and university students to examine relationships among network
position, social power, and verbal aggressiveness. They found that students demonstrating
scientific/task attractiveness (i.e., expert power) and social attractiveness (i.e., referent power), as
measured by their positions in advising and friendship networks respectively, have developed
authoritative power and are protected from verbal aggression in schools. Drawing upon the
research on structural social capital (Burt, 2009, 2017), the present study regards network
brokerage as a crucial structural feature that can engender bridging social capital and facilitates
the acquisition of expert power. Similarly, network closure is viewed as a significant network
position that can cultivate bonding social capital and elicit referent power.
Bridging and bonding social capital represent two distinct forms of social capital
(Granovetter, 1973; Putnam, 2000). From the perspective of structural social capital, bridging
social capital is generated through network brokerage, while bonding social capital emerges from
network closure. Network brokerage refers to the social structure where the focal actor connects
structural holes, meaning that the focal actor links to individuals who are otherwise unconnected.

14
Building connections across structural holes enables access to heterogeneous information and
resources, which provide diverse perspectives, information, resources, and opportunities,
ultimately leading to the potential development of bridging social capital (Granovetter, 1973;
Williams, 2006). On the other hand, network closure refers to the social structure where the focal
actor is linked to individuals who are already connected with each other. With everyone
connected to everyone else, members are more likely to share similar views and information and
less likely to engage in deviant behaviors. These redundant strong ties connect closely-knit
networks of individuals or groups that are similar to each other, thus promoting bonding social
capital that provides emotional support, trust, and a sense of belonging (Lee et al., 2018).
To date, many studies have examined the structural view of social capital in virtual
worlds. For example, Ganley and Lampe (2009) explored how network brokerage and closure
generate social capital in a large online community and found that users’ reputation was
positively associated with network closure. Similarly, Burt (2012) investigated network
brokerage and closure in the virtual world, Second Life, and found that brokers tended to
organize more active groups, while network closure was positively linked to the level of trust
within groups. Shen et al. (2014) further examined network social capital in an MMOG and
showed that brokers tended to have higher task performance whereas players embedded in closed
networks were more likely to trust other players.
According to the Power, Approach, and Inhibition Theory (PAI; Keltner et al., 2003),
individuals with elevated power are more likely to perceive others as a means to their own ends
and engage in socially inappropriate behavior to achieve their goals while disregarding the
consequences of their actions. This may be particularly true for individuals with high bridging
social capital and therefore high expert power over others. On the one hand, they have access to

15
more resources and opportunities that can activate their behavioral approach system. On the
other hand, they are embedded in weak ties and their friends are not closely connected with each
other, making them less accountable for their actions. The shadow of the future effect also
applies less to them since losing one connection does not devastate their reputation or informal
power. Consequently, they may be more inclined to engage in aggressive behaviors toward
others, especially in a zero-sum competitive video game where winning is the primary reward.
Although individuals in prominent network positions are generally shielded from
victimization, those with high network bridging social capital may be an exception. As they span
their network across structural holes, they may encounter a more diverse range of players,
manage more conflicting interests and values, and be more vulnerable to attacks or retaliation
from a more diverse range of players. For example, prior research suggests that individuals
having a large network of weak ties were more likely to experience victimization online
(Bastiaensens et al., 2014; Wegge et al., 2014). Moreover, since they tend to perceive themselves
as having higher informal power, they may be more prone to initiating reports when
encountering gaming toxicity. Therefore, it is hypothesized that:
H2a-b: Controlling for formal social status, individuals with higher network bridging
social capital would (a) engage in more toxic behaviors and (b) experience more
victimization;
However, individuals who occupy positions of high power can also exhibit prosocial
behavior, particularly if they are in communal relationships rather than in exchange relationships
where individuals respond to the needs and interests of others solely to receive a benefit in return
(S. Chen et al., 2001). Individuals with high bonding social capital and therefore high referent
power over others may have a stronger sense of identification and trust within the gaming

16
community, which directs their behavior towards attachment-seeking and trust-building and
discourage them from engaging in disruptive behaviors. Their social groups, in turn, provide
them with social support and protection. They may also have a greater sense of security and
safety in their interactions with other players and are less likely to feel threatened to make a
report. In addition to the number of ties, the interrelation among ties also heightens the
accountability cues and attentional cues, creating a stronger shadow of the future effect. As
demonstrated by Burt (2002), network closure enforces the sanctioning of negative behaviors. If
individuals play in a close-knit network, they are more likely to be held accountable for deviant
behaviors. Moreover, the larger shadow of the future effect and the greater consequences for
them if they engage in deviant behaviors make them more vulnerable to negative feedback, even
from a single person. Therefore, we predict the following:
H3a-b: Controlling for formal social status, individuals with higher network bonding
social capital would (a) engage in fewer toxic behaviors and (b) experience less
victimization.
Network Perspective toward Perpetration-Victimization Overlap
In research on offline aggression (e.g., Haynie et al., 2001; Unnever, 2005) and
cyberbullying (e.g., Festl & Quandt, 2013), toxicity interactants have been categorized along the
bully-victim continuum as pure perpetrators (also known as pure bullies), aggressive victims
(also known as bully-victims, provocative victim, etc.), victims, and uninvolved contrasts (also
known as normative contrasts), based on incidents of perpetration and victimization. Each
subgroup is shown to have a unique set of underlying mechanisms that influence their behavior
(Olweus, 2001). Pure perpetrators are those who intentionally use aggression to harm or disturb
relatively powerless individuals (Olweus, 1997). They tend to be deliberate and goal-driven in

17
their decision-making, bullying others for more instrumental purposes, such as gaining power
and resources (Schwartz et al., 2001). Victims, on the other hand, are targets of aggression,
usually due to their power difference from perpetrators. While pure victims are usually
characterized as being submissive and passive, there is a subgroup of victims who are more
prone to aggressive and hostile behavior toward bullies. This subgroup has been referred to as
aggressive victims (Unnever, 2005). They have received great research attention in school
bullying, as they bully others differently than pure bullies, and they are bullied differently than
pure victims. While perpetration from pure perpetrators is goal-driven and instrumental,
perpetration from aggressive victims is more reactive and impulsive. As peers tend to find
impulsivity and emotionality aversive, aggressive victims usually receive greater social rejection
than pure perpetrators, whereas pure perpetrators may even enjoy within-group popularity
(Pellegrini et al., 1999). When victimized, pure victims tend to respond passively and
submissively, while aggressive victims tend to retaliate with greater hostility. As aggressive
victims tend to receive greater social rejection, they are relatively isolated and have less
protection and support from their peers. Consequently, aggressive victims will be victimized
more often than pure victims. Connecting the aforementioned network position and incidence-
based perpetration and victimization to the unique behavioral pattern of the toxicity subgroups, it
is predicted that (for a summary of hypotheses, see Table 2):
H4: Controlling for formal social status, individuals who are at the center of their
network are more likely to be uninvolved in gaming toxicity;
H5: Controlling for formal social status, individuals with higher levels of network
bridging social capital are more likely to be aggressive victims;

18
H6: Controlling for formal social status, individuals with higher levels of network
bonding social capital are more likely to be uninvolved in gaming toxicity;
H7a-c: Controlling for formal social status, individuals with lower informal power (a.
fewer friends, b. lower network bridging social capital, c. lower bonding social capital)
are more likely to be pure victims.
Table 2.2. A summary of hypotheses testing.
Incidence-Based
Perpetration
Incidence-Based
Victimization
Actor-Based Toxicity Subgroup
Informal
Power
Network Centrality –
(H1a)✓
–
(H1b)✓
Uninvolved (+) / Pure Victims (-)
(H4)✓ / (H7a)⤫
Network Bridging SC +
(H2a)✓
+
(H2b)✓
Aggressive Victim (+) / Pure Victims (-)
(H5)⤫ / (H7b)✓
Network Bonding SC –
(H3a)✓
–
(H3b)✓
Uninvolved (+) / Pure Victims (-)
(H6)✓ / (H7c)⤫
Note. SC = Social Capital. H = Hypothesis. “+” indicates positive association and “-” indicates
negative association. “✓” indicates hypothesis supported and “⤫” indicates hypothesis rejected.
Methods
Research Site
The research site was World of Tanks (WoT), a popular vehicle-based Multiplayer Online
Game that has garnered a user base of over 160 million worldwide. WoT features tank combat
between two teams consisting of up to 15 players, with the goal of destroying the enemy’s
central base or eliminating all enemy tanks. Players have the option of joining a random battle at
their convenience, or they can request to join a team with up to two preferred players. Moreover,
players can join clans, a larger and more permanent social structure in WoT, to unlock more
advanced battle types. On average, a battle in WoT lasts from four to 15 minutes. In contrast to
match-based games such as League of Legends, WoT does not match teams based on player

19
skills but by tank types and tiers, resulting in a higher likelihood of teams with disparate skill
levels, which may lead to lopsided results and more emotional swings.
Data
In partnership with the game operator, Wargaming Inc, we acquired and merged three
types of datasets for the WoT North American server, namely player co-play network data, in-
game toxicity report data, and in-game behavioral log data. The co-play network for each player
was an undirected network of WoT players based on their co-play experience in April 2019. A
co-play incidence was recorded any time a player requested to play with another player,
regardless of whether they were friends or not (e.g., players could request to play with their
clanmates). Both players who initiated and accepted the co-play request were included as nodes
in the co-play network. An undirected edge between the two nodes (players) indicates that they
co-played at least once during the data collection period, disregarding the request's initiation.
However, if a player opted to be placed in a random battle, that would not be included in the co-
play instances. Therefore, co-play represents some degree of familiarity and social acceptance.
This study also compiled toxicity report data in May 2019, incorporating player-generated
complaint incidents and toxicity incidents automatically captured by the game system. Although
this study’s design does not enable us to assert causality between key variables, the time lag
between player network embeddedness and their involvement in toxicities puts us in a better
position to speak to alternative explanations than had our data been gathered at a single point in
time. Different types of data were matched using a unique one-way hashed key. The data were
anonymized before reaching the research team.
In WoT, players can report toxic behavior up to 10 times per day, and being reported five
times can result in a permanent game restriction. During a battle, players can report others and

20
specify the type of toxic behavior committed. Four types are provided when making reports:
inappropriate behavior in chat (14% in our dataset), unsportsmanlike conduct (41%), offensive
nickname or clan name (4%), and inaction/bot (41%). Following prior practice (Shen et al.,
2020), we aggregated incidents from all toxicity types for analysis.
Measures
Network size. Network size for each player (M = 98.866, SD =252.785) was measured
by the number of edges a node has in their respective co-play network. It captured a player’s
level of connection with other players. A player who had a high network size was one who has
some familiarity with a lot of other players.
Network brokerage. Adapting the procedure of Burt (1992, 2009), network brokerage,
or structural hole, was measured as the effective size (M = 37.676, SD = 88.918), which
calculates the bridge relations of a focal actor. Consider a focal actor 𝑖’s ego network, a bridge is
said to exist between 𝑖 and alter 𝑗 when 𝑗 has no connection with any other alters that 𝑖 has. It can
also be considered as the non-redundancy, measured by the following formula:
$
!
(1−$
"
𝑝
#"
𝑚
!"
),𝑞 ≠𝑖,𝑗
Where 𝑝
#"
is the proportion of a focal actor 𝑖’s investment in a connection with an alter 𝑞,
and 𝑚
!"
is normalized tie strength of the connection between 𝑗 and 𝑞. In an unweighted and
undirected network (Borgatti et al., 1997) like our co-play network, 𝑚
!"
simply represents the
relation between 𝑗 and 𝑞.
Network closure. Following the procedure of Burt (1992, 2009), this study
operationalized network closure as network constraint (M = 0.285, SD = 0.339), using the
following formula:

21
𝐶
#!
=(𝑝
#!
+$
"
𝑝
#"
𝑝
"!
)
$
,𝑞 ≠𝑖,𝑗
Where 𝑝
#!
is the proportion of a focal actor 𝑖’s investment in a connection with an alter 𝑗.
This constraint measures the interconnectivity between the focal actor 𝑖 and a specific alter 𝑗, in
other words, it measures 𝑖’s dependence on 𝑗, representing the trust, time, and other resources 𝑖
invested into the relation with 𝑗.
Subgroups of players involved in toxicities. Based on individual-battle level
perpetration and victimization, subgroups were categorized as (1) pure perpetrators: players who
have only been reported of committing toxic behaviors but have not reported others (N =
31,179); (2) pure victims: players who have only reported others but not been reported (N =
8,347); and (3) aggressive victims: players who have both reported and been reported (N =
23,980). Players who did not play any battles during the study period were excluded (N = 6,009).
Perpetration count and victimization count. The frequency of toxic behavior
perpetration and victimization were measured by calculating the perpetration count (M = 3.288,
SD = 5.904) and victimization count (M = 3.728, SD = 23.748) for each player, respectively.
This was done by summing the total number of battles in which a player was reported as the
perpetrator or victim of toxic behaviors.
Control Variables. Battle count (M = 298.461, SD = 305.837) was the total number of
battles in which a player participated between January and April 2019. Win rate (M = 43.846, SD
= 17.649) was calculated as the ratio of battles won by a player during this time period to the
total number of battles, indicating player skillfulness. Clan membership was a dummy variable
representing whether a player was part of a clan during the data collection period.

22
Analysis
The data were arranged at the individual level with each row representing an individual
player’s involvement in toxic behaviors in May 2019, and the network measures of individual
players in the co-play network in April 2019. To test H1-H3, Poisson regression models were
utilized with the perpetration count and victimization count serving as dependent variables.
Independent variables included players’ network size, network brokerage, and network closure in
the co-play network. The models also controlled for battle count, win rate, and clan membership.
The output of the Poisson regression models is expressed in parameter estimates of rate ratios,
which represent the expected change in the count of the outcome variable for a unit change in the
predictor variable. This concept is akin to that of an odds ratio in logistic regression, although
rate ratios present the incidence rate of an event occurring rather than its odds. Rate ratio of a
value of 1 indicates no effect, while values greater than or less than 1 represent positive and
negative effects, respectively. Hypotheses H4-H7 were tested using logistic regression models
with the probability of a player being a pure perpetrator, pure victim, aggressive victim, or
uninvolved serving as dependent variables. The same independent variables and control variables
used in the Poisson regression models were included.
Results
H1 investigated whether a player’s popularity in a co-play network is associated with
their involvement in toxic behaviors, namely their engagement in perpetration and the experience
of victimization. Controlling for the indicators of formal social status, specifically, battle count,
win rate, and clan membership, network size (rate ratio = 0.9997, p < 0.001, 𝑅
$
=.1785) was
negatively associated with perpetration count, supporting H1a. Similarly, controlling for the

23
indicators of formal social status, network size (rate ratio = 0.9995, p < 0.001,𝑅
$
=.0775) was
negatively associated with victimization count. H1b was supported.
H2 tested whether a player’s level of brokerage in a co-play network is associated with
their involvement in toxic behaviors. Both H2a and H2b were supported, with the network
brokerage positively related to perpetration count (rate ratio = 1.0015, p < 0.001,𝑅
$
=.1785)
and victimization count (rate ratio = 1.0014, p < 0.001,𝑅
$
=.0775). H3 tested whether a
player’s level of network closure in a co-play network is associated with their involvement in
toxic behaviors. Results showed that network closure was significantly related to perpetration
count (rate ratio = 0.7173, p < 0.001,𝑅
$
=.1785) and victimization count (rate ratio = 0.7188, p
< 0.001,𝑅
$
=.0775), indicating that with one increase in network constraint, there were a
28.27% decrease in perpetration count and 28.12% decrease in victimization count. Both H3a
and H3b were supported.
H4 investigated whether a player’s network size in the co-play network is associated with
the probability of them being uninvolved in gaming toxicity. Controlling for formal social status,
network size was positively related to them being in the uninvolved reference group (odds ratio =
1.0014, p < 0.001, 𝑅
$
=.0532), supporting H4. H5 predicted whether a player’s network
brokerage is associated with the likelihood of them being an aggressive victim. Controlling for
formal social status, the network brokerage was not associated with the aggressive victim group
(odds ratio = 1.0000, p < 0.001, 𝑅
$
=.0614). H5 was not supported. H6 predicted whether a
player’s network closure is associated with the likelihood of them being an uninvolved contrast.
Controlling for formal social status, the network closure was positively related to the uninvolved
group (odds ratio = 1.7675, p < 0.001, 𝑅
$
=.0532), indicating that a one-unit increase in

24
constraint is associated with a 76.75% increase in the odds of a player being in the uninvolved
group, supporting H6.
H7 investigated whether a player's network informal power, specifically, network size
(H7a), network brokerage (H7b), and network closure (H7c), is associated with the probability of
them being a pure victim. Controlling for the formal social status, network size (odds ratio =
1.0006, p < 0.01, 𝑅
$
=.0402) was positively related to the pure victim group, not supporting
H7a; network brokerage (odds ratio = 0.9951, p < 0.001, 𝑅
$
=.0402) was negatively related to
the pure victim group, supporting H7b; and network closure (odds ratio = 1.5253, p < 0.001,
𝑅
$
=.0402) was positively related to the pure victim group, not supporting H7c. All results are
presented in Table 2.3 and Table 2.4.

Table 2.3. Poisson Regression Using Network Variables to Predict Victimization and
Perpetration Count.
Perpetration Count Victimization Count
Rate ratio SE p Rate ratio SE p
Intercept 3.0032 0.0081 *** 4.4432 0.0086 ***
Network Size 0.9997 0.0000 *** 0.9995 0.0000 ***
Network Brokerage 1.0014 0.0001 *** 1.0015 0.0001 ***
Network Closure 0.7173 0.0077 *** 0.7188 0.0075 ***
Clan membership 0.9459 0.0045 *** 0.9739 0.0042
Win rate 1.0010 0.0002 *** 1.0073 0.0002 ***
Battle count 1.0009 0.0000 *** 1.0008 0.0000 ***
Note. ‘***’ p<0.001 ‘**’ p<0.01 ‘*’ p<0.05

25
Table 2.4. Logistic Regression Using Network Variables to Predict Victims, Perpetrators,
Aggressive Victims, and Uninvolved Players.
Perpetrator Aggressive victim Victim Uninvolved

Odds
ratio SE p
Odds
ratio SE p
Odds
ratio SE p
Odds
ratio SE p
Intercept 1.1642 0.0241 *** 0.2520 0.0296 *** 0.1491 0.0349 *** 0.1499 0.0370 ***
Degree 0.9993 0.0001 *** 1.0004 0.0001 * 1.0006 0.0003 * 1.0014 0.0004 ***
Brokerage 1.0014 0.0004 *** 1.0000 0.0004 0.9951 0.0009 *** 0.9939 0.0011 ***
Closure 0.9972 0.0245 0.5201 0.0288 *** 1.5253 0.0353 *** 1.7675 0.0395 ***
Clan 0.8379 0.0161 *** 1.1543 0.0175 *** 1.2256 0.0247 *** 0.8955 0.0287 ***
Win rate 0.9965 0.0005 *** 1.0087 0.0006 *** 1.0037 0.0006 *** 0.9975 0.0007 ***
Battle count 0.9997 0.0000 *** 1.0013 0.0000 *** 0.9982 0.0001 *** 0.9979 0.0001 ***
Note. ‘***’ p<0.001 ‘**’ p<0.01 ‘*’ p<0.05

Discussion
Given the ubiquitous presence of online video games in people’s everyday lives, gaming
toxicity has emerged as a growing concern for game companies, gaming communities,
policymakers, educators, and parents (Neto et al., 2017; Shen et al., 2020). Toxic behaviors not
only affect the psychological well-being of the victims, but also that of the bystanders and
perpetrators, thus compromising the overall user experience (Wegge et al., 2013). Although there
is a growing body of literature investigating the factors that predict toxic behaviors in video
games, most of this research has focused on platform-level factors (Hilvert-Bruce & Neill, 2020;
Kordyaka et al., 2020), game-level factors (Kwak et al., 2015), and personal factors (Tang et al.,
2020), without explicitly examining toxicity dynamics in the context of relationships in social
networks. The few studies that have investigated relational factors predicting gaming toxicity
have indicated that network predictors are more robust in predicting cyberbullying behaviors
than personal factors (Festl, 2016).

26
Theoretically, this study builds upon SHT, DSAT, PAI, and the shadow of the future
effect to investigate the social network positions that contribute both perpetration and
victimization of gaming toxicity. By embedding actors in their social networks and examining
the informal power associated with key network positions, this study further extends the toxicity
subgroup categorization from offline aggression research and cyberbullying to encompass the
realm of gaming toxicity. Meanwhile, this study underscores the distinctiveness of video games
as a space to examine toxicity, given the game affordances and context that shape players' norms
and psychological mechanisms when perpetrating or reporting victimization. Practically, the
results can aid in toxicity prediction and vulnerable player identification, as well as inform the
design of online architectures and interventions that can guard against such viral toxicity.
This study has specifically investigated three crucial network positions and the informal
power associated with them: network size, network brokerage, and network closure. As
predicted, network size is negatively related to perpetration, indicating that players with more co-
play mates are less likely to exhibit toxic behaviors, and vice versa. This result supports the
shadow of the future effect. Having more in-game friends signifies less anonymity and
deindividuation, and higher presence of accountability cues. Thus, players may regulate their
behaviors compared to when they are playing with strangers. Moreover, in the context of video
games, having more friends to play with may indicate that players get to play with their familiar
others, leading to better coordination, enjoyment, and a stronger sense of gaming community.
Furthermore, playing with familiar friends reduces the risk of being paired with strangers who
may not cooperate or who have large skill disparities (Shen et al., 2020). Playing with familiar
friends may also raise the stakes of being toxic. Echoing the person-situation debate in
psychology, it is also possible that individuals who are kinder are more likely to have a higher

27
number of friends who are willing to play with them. From the actor-based approach, individuals
with higher network size are less likely to be pure perpetrators or aggressive victims but are more
likely to be pure victims or uninvolved in toxicity incidents. This suggests that individuals with
more co-play friends are less likely to be reported as toxic. They also do not report others’ toxic
behaviors, if there have been any, thus are more likely to be pure victims.
In addition to network size, this study also examined the potential of network brokerage
and closure in bringing social capital and informal power and predict players’ involvement in
gaming toxicity. Although previous research has shown that differences in social capital can lead
to perpetration and victimization (C. P. Barlett et al., 2022; Evans & Smokowski, 2016) and has
successfully tested the application of structural social capital in video game settings (Shen et al.,
2014), few studies have tested whether structural embeddedness in video games may lead to
perpetration and victimization. As predicted, the study found a positive association between
network brokerage and the count of perpetration and victimization in video games. There are
several possible explanations for this positive association. Firstly, players with greater network
brokerage may have greater access to information about other players and their strategies.
Previous studies have found that higher network brokerage contributes to better task performance
in video games (Shen et al., 2014), which can give players an advantage in the game and increase
their likelihood of engaging in aggressive behaviors. Similarly, research on social network sites
has shown that having many connections on Facebook who are not offline friends increases the
risk of cyberharassment and cyberbullying (Wegge et al., 2015). Perpetrators in online networks
have been found to have a disproportionately higher number of weak ties, indicating that they
may have higher social status than victims. In addition, as the friends of their friends tend not to
know each other, their behavior has less social cost and their level of deindividuation is still

28
relatively high, thus engaing in more perpetration. Alternatively, it is possible that individuals
who engage in toxic behaviors are more likely to be ostracized by their own team or community
and thus seek out connections with different teams.
Network brokerage is also positively correlated with the incidence of victimization.
While brokers spanning across structural holes have a higher potential of possessing greater
bridging social capital, in line with SIDE, regular members of different teams may perceive
brokers as out-group members, thus being more hostile toward them. Research on cyberbullying
on social network sites also suggests that connection with a large number of unfamiliar others is
a risky practice that increases the likelihood of being exposed to negative attention from potential
perpetrators (Dredge et al., 2014). Compared to other subgroups, individuals with higher network
brokerage are less likely to be pure victims or uninvolved players, but more likely to be pure
perpetrators. It is noteworthy that despite the heightened risk of victimization, individuals with
elevated levels of network brokerage tend to refrain from adopting a passive or submissive
stance, potentially due to the influential nature of bridging social capital and the corresponding
power it confers.
Contrary to network brokerage, as predicted, network closure is negatively associated
with the count of perpetration and victimization. This is in line with previous research that
bonding social capital provides emotional support and trust (Lee et al., 2018) that protects
individuals from victimization and prevents individuals from perpetration (Aquino & Lamertz,
2005; Evans & Smokowski, 2016). A study by Ganley and Lampe (2009) examined how
networks of relationships in Slashdot, a popular online community, generate social capital, and
found that network closure positively predicts users’ reputation on the site. For individuals with
higher network bonding social capital (i.e., network closure), their reputation is at stake and there

29
is a larger social cost and shadow of the future effects with them. Burt (2012) and Shen et al.
(2014) also examined network bonding social capital in virtual worlds and found that it was
positively associated with the level of trust within groups. In a video game setting, the level of
trust associated with network bonding social capital may mitigate players’ inclination to blame
others and decrease their chance of being blamed when the feedback of the game was negative
(e.g., losing the game). Interestingly, individuals with higher levels of network closure are less
likely to become victims of aggression, while those with lower levels are more susceptible.
Aggressive victims have been the subject of extensive research due to their unique psychological
profiles and behavioral patterns. The conventional delinquency research has viewed perpetrators
and victims as two distinct populations (Chan & Wong, 2015). However, recent research has
demonstrated that these groups share many commonalities, and many victims are or become
offenders themselves (DeCamp & Newby, 2015). As Reiss (1981) put it, “any theory that
assumes no overlap exists between populations of victims and offenders or that they are distinct
types of persons distorts the empirical research” (p. 711). Overall, the behavioral pattern of
individuals with high network size and network closure is comparable, with the exceptions of
their likelihood of being a pure perpetrator and an aggressive victim. This suggests that the toxic
behavior of individuals with large network sizez typically does not occur and only emerges in
response to their victimization.
In gaming toxicity dynamics, although individuals with higher network size tend to
regulate their toxic behavior, they are more likely to engage in retaliatory behavior or deviant
behavior when victimized. This may imply that while their social group perceives toxic behavior
as disruptive, if their friends are victimized, their network may not impose sanctions on their
aggressive behavior and may even reward out-group antagonism as a way to create a strong in-

30
group solidarity. On the other hand, individuals with high network closure are less likely to be
aggressive victims. This means that even if they encounter gaming toxicity, they are less likely to
engage in toxic behavior. This may be due to the sanction of disruptive behaviors in a close-knit
network. It can also be explained by their stronger sense of security and confidence in their
relationships as a result of their supportive network. They may view the game as a community
and report toxic behavior to improve it, but they will not violate the norms themselves. This
suggests they are less likely to engage in retaliatory behavior, even when provoked.
Researchers have proposed several theories to examine the causes of the perpetration-
victimization overlap. For example, the Routine Activity Theory (RAT; Cohen & Felson, 2010)
posits that individuals who regularly put themselves in dangerous situations without any
supervision are at a higher risk of becoming perpetrators, targets of perpetration, or both. While
RAT has provided valuable insights into the demographics, social activities, and risky behaviors
of the three groups and highlighted the importance of situational factors in determining their
experience, it does not examine how the structural aspect of social networks, such as an
individual's social network position and the quality of their social ties, may affect their likelihood
of becoming a victim, a perpetrator, or an aggressive victim. By examining the role of network
positions and structural social capital, this study provides a more nuanced understanding of the
structural factors that contribute to the perpetration-victimization overlap. Our findings suggest
that individuals with lower levels of network closure are more vulnerable to becoming
aggressive victims, potentially due to their lack of power. On the other hand, building strong
social ties can reduce an individual's vulnerability to victimization. This study thus contributes to
a more comprehensive understanding of the perpetration-victimization overlap and highlights the

31
importance of considering both situational and structural factors when developing effective
prevention and intervention strategies.
The findings of this study sheds light on the power and accountability cues associated
with key network positions and contribute to the growing body of literature calling for a network
perspective in the study of cyberbullying, perpetration-victimization overlap, and online toxicity
(Laurie-ann et al., 2021; Sun & Shen, 2021; Wegge et al., 2013). Although the power associated
with crucial network positions is likely to be comparable in both online and offline contexts, it is
intriguing to consider how accountability cues might differ in the online and offline realms. The
influence of the "shadow of the future" effect is also expected to be applicable in offline
scenarios; however, the cues of anonymity may be absent. These findings also have practical
implications for toxicity identification and intervention. In recent years, researchers have used
players’ in-game behavior and social network analysis to detect and identify key cyberbullies
(Canossa et al., 2021; Choi et al., 2021). Building upon this line of research, the current study
provides further insights into the underlying factors that render network analysis predictive in
proactively identifying vulnerable players through the lens of structural social capital and power.
To reduce gaming toxicity, game companies may contemplate integrating features or
activities that safeguard players rather than endanger them. For instance, penalizing players for
leaving a game or quitting matchmaking due to adverse experiences could be deemed unjust and
indicate that the game's importance takes precedence over the players' health and safety. Instead,
game companies could reconfigure players’ social networks to mitigate gaming toxicity and
empower players. Encouraging co-play with acquaintances rather than strangers, for example,
could shield players from being victimized. Furthermore, helping players cultivate in-game
bonding social capital, such as recommending friends' friends or promoting team-building

32
activities, could mitigate gaming toxicity and retaliation. It is worth noting that the unique
features of video games may result in toxic players exploiting their power and employing
reporting as a means of bullying; therefore, game companies should implement measures to
prevent players from abusing the reporting system. For example, in WoT, players can only
submit a limited number of reports per day, and they have the option to appeal a ticket if they
believe they are innocent. Such measures could minimize the potential for abuse of the reporting
system and establish a more equitable and fair gaming environment. Finally, it would be
beneficial for the game industry to release transparency reports on gaming toxicity and support
game scholars’ and practitioners’ research to understand the risks and benefits of online gaming.
Overall, this study underscores the importance of considering network positions in understanding
and combating gaming toxicity and presents practical recommendations for game companies to
create a safer and healthier gaming environment.
While the results of this study are noteworthy, there are several limitations that warrant
attention. Firstly, the data were collected from one game, and it is unclear whether the results
would be applicable to others, particularly those from diverse genres. While the competitive
nature of the game was appropriate for examining gaming toxicity, the unique mechanisms of
various games may engender distinct forms of toxicity. For instance, toxicity in social games
may manifest differently than in competitive games. Future research may replicate the present
findings across various games and genres to bolster the generalizability of the results. Secondly,
the current classification of subgroups of toxicity may not be comprehensive enough to
encompass all subgroups in a video game context. The potential for toxic players to abuse their
power and utilize reporting as a way of bullying leads to a muddling of the differentiation
between perpetrators and victims. It is plausible that some pure perpetrators are, in fact, victims,

33
and some pure victims are bullies. Future research should develop a more nuanced categorization
of gaming toxicity to cover the full spectrum of player profiles involved in toxicity
dynamics.Thirdly, while this study benefits from a longitudinal design, it did not provide causal
insights into a player’s network position and their toxic behavior. Therefore, Study 2 of the
dissertation models the coevolution of players’ network formation and their toxic behavior
development in order to draw more causal inferences on the self-selection versus social influence
process in codetermining one’s network position and toxic behavior. Lastly, the study's results
only shed light on the endogenous factors that predict player behavior. It is still unclear how
exogenous factors, such as the broader socioeconomic context or policy changes, influence
players’ behavior as a whole. To answer this question, Study 3 of the dissertation leverages the
lockdown order during the COVID-19 pandemic as a natural experiment to examine how the
broader social context influences the helping behavior of those affected.

34
CHAPTER 3: THE CO-EVOLUTION OF FRIENDSHIP NETWORKS AND TOXIC
BEHAVIORS IN A MULTIPLAYER ONLINE GAME (STUDY 2)
Multiplayer online games, more than any other online platform, are subject to a "magic
circle" (Huizinga, 1949) where different rules and norms apply. Based on the reward and
punishment mechanisms in this magic circle, behaviors that would otherwise be considered
inappropriate may become the norm, or even a source of pleasure (Paul et al., 2015), especially
in competitive games and esports (Türkay et al., 2020). Toxicity within online multiplayer video
games is a pervasive issue, with 81% of video game players have experienced some form of
harassment while playing (Anti-Defamation League, 2020).
Although gaming toxicity may be a source of pleasure and normal behavior for some, it
adversely affects the enjoyment and well-being of a significant proportion of players. In the ADL
survey (2020), approximately 25% of players reported that they had quit playing because of
negative experiences. Even more alarming, 11% reported experiencing depressive or suicidal
thoughts due to in-game toxicity. The normalization of toxic behaviors also harms the gaming
community and culture, as it shapes the perceptions of regular players, especially younger and
novice players, regarding what constitute appropriate behavior in games (Shen et al., 2020).
The networks that individuals are embedded in play a crucial role in shaping perceived
norms (Friedkin, 2001; Shepherd, 2017). For example, Shepherd (2017) demonstrated the
association between social network structure and perceptions of descriptive social norms within a
cooperative network. This association arises from reasoning that network positions determine
individuals' exposure to others, the extent of such exposure, and the nature of that exposure. This
aligns with the social influence network theory, which posits that influence is a process whereby
group members' attitudes and opinions on a given matter change in accordance with the weighted

35
averages of influential individuals in the network (Friedkin, 1998). The manner in which
individuals reconcile conflicting opinions depends on the social structures within which this
process occurs, encompassing their initial positions, interpersonal influences, and susceptibilities
to influence. In groups featuring a stable interpersonal communication network, the patterns of
connections indicate channels of interpersonal influence.
Network structure molds descriptive norms, and in turn, norms shape behavior
(Shepherd, 2017). While norms possess considerable influence over behavior, they are not static
phenomena but rather fluid. Social norms “both affect and are affected by human action” (Rimal
& Lapinski, 2015, p. 393). They adapt to human actions and choices (Boyd & Richerson, 1994).
Thus, addressing toxicity in online gaming requires not only identifying mechanisms of behavior
diffusion and norm formation but also clarifying individuals' roles in creating social norms.
However, most prior studies focused solely on the influence of social networks on individuals’
toxic behaviors, without taking individuals’ active network selection processes into account.
More nuanced longitudinal analyses that simultaneously model unobserved micro-changes
underlying the co-evolution of one’s social networks and behaviors are warranted to disentangle
the network selection and social influence processes and provide less biased parameter estimates
(Shalizi & Thomas, 2011).
To disentangle behavior change as a result of social influence and individuals' network
selection preference, stochastic actor-oriented models (SAOM) are often used to estimate the co-
evolution of behaviors and networks. SAOM has been widely adopted in examining health
behaviors embedded in networks, such as physical activities (De La Haye et al., 2011), smoking
(Mercken et al., 2010), and syphilis incidence (Young & Fujimoto, 2021). Despite its
appropriateness, to the best of our knowledge, no studies have applied the model to examine

36
toxic behaviors in gaming networks. Building upon previous studies on network formation and
social influence and following the practices of Lazer (2001) and Lazer et al. (2010), this study
proposes a unified theory of coevolution of toxic behavior and network, where players seek to
accommodate their level of toxicity with that of their friends. This accommodation may be
achieved by choosing friends with similar toxicity level or by adjusting one’s behavior to be in
line with those of their friends, or by some combination of these processes.
Friendship Formation: Birds of a Feather
Homophily theory has been extensively applied to predict and explain network formation
in various contexts and has received wide support. The theory postulates that individuals are
more inclined to interact with those who are similar to them (McPherson et al., 2001). Previous
research has demonstrated the influence of homophily on the formation of ties among
individuals, teams, and organizations based on factors such as age, gender, political orientation,
and geographic location, among others (e.g., Fu, 2019; Lazer et al., 2010; Yuan & Gay, 2006).
For example, people tend to marry others of similar age, date others from the same race
(Newman, 2003), become Facebook friends with others from the same high school (Traud et al.,
2012), recommend books to those of the same gender (Bucur, 2019). Homophily theory finds its
roots in Byrne’s (1971) similarity breeds attraction paradigm and the Self-Categorization Theory
(Turner et al., 1987). Individuals tend to categorize those who resemble them into the same
social category and favor those in the same social category. This is likely because similarity
often presents opportunities for cooperation and the formation of strong relationships, whereas
dissimilarity often leads to competition and weak relationships (Nebus, 2006). Establishing
connections with similar others can also provide informational benefits (Lazer et al., 2010).
When seeking information on unfamiliar topics, it is advisable to consult those who have similar

37
experience and taste. Moreover, as similarity increases predictability, individuals experience less
uncertainty in communicating with similar others, thus are more likely to form relationships with
similar individuals (Yuan & Gay, 2006).
In the context of competitive multiplayer online video games, despite the possibility of
individuals displaying toxic behavior stemming from personal factors or undisclosed matters, it
is plausible that individuals with comparable levels of gaming toxicity within a clan are more
prone to forging friendships. This tendency arises from the shared understanding of clan norms
and acceptable (or unacceptable) conduct, thereby fostering interactions within behaviorally
homogeneous clusters. Consequently, these interactions may entrench the prevalence of
commonly shared behaviors and reinforce behavioral homogeneity within groups (Cho et al.,
2018). Players who display high levels of toxicity may be disliked by clan members who do not
exhibit toxic behavior and may view such behavior as disruptive. Conversely, they may seek to
associate with clan mates who understand their feelings and behaviors and are willing to support
them, especially when confronted with a common adversary (e.g., an opposing team in battles).
For individuals with high toxicity levels, having a friend who displays similar toxicity levels
implies a strong sense of loyalty and camaraderie. Based on the homophily theory, it is predicted
that:
H1: Clan members are more likely to form friendship ties with those who are of similar
level of toxicity.
In addition to attribute-level homophily, assortative mixing is a form of “network
homophily” (Obadimu et al., 2021, p. 10) that refers to the propensity of nodes in a network to
connect with other nodes that share similar properties, such as degree centrality (Aiello et al.,
2012; Newman, 2002). Research on scientific collaboration indicates that scholars with

38
numerous collaborators tend to collaborate with other popular scholars, while scholars with few
collaborators tend to work independently or collaborate with other scholars who have a small
number of collaborators (Newman, 2004). While some networks exhibit assortative mixing in
terms of their degrees, others show disassortative mixing, meaning that nodes with high degrees
are more likely to connect with those with low degrees (Newman, 2002). In general, it is
commonly believed that social networks exhibit assortative mixing, whereas technological
networks (e.g., hyperlink network on World-Wide Web) and biological networks (e.g., protein
network in yeast) show disassortative mixing. However, with the emergence of online social
networks and virtual communities, recent research on network homophily has challenged this
belief and found that online social networks transitioned from a degree assortativity pattern to a
degree disassortativity pattern (Hu & Wang, 2009). This may be due to the fact that online
social networks initially inherit the assortative mixing pattern of the underlying real-world social
network. However, as the online social network develops, low degree users may preferentially
form connections with those with high degrees, resulting in a disassortative pattern.
In the realm of multiplayer online games, previous research on two massively multiplayer
online games, Game X and Travian, has revealed that the message network in Travian displayed
a disassortative mixing pattern, while the attack and trade network in Game X exhibited a
disassortative mixing pattern, indicating that message, attack, and trade activities tend to revolve
around group leaders (Hajibagheri et al., 2015). Based on this line of research on online
communities, it is predicted that in an online game community, players' networks will initially
exhibit an assortative feature that is inherited from real life. However, as time goes on, the
network will display a disassortative pattern where low degree players are more likely to be
connected to high degree players. Therefore, it is predicted that:

39
H2: The friendship network among clan members will show a disassortative pattern
where players with high degree are more likely to form new ties with players with
low degree, and vice versa.
While homophily serves as an exogenous mechanism for tie formation by leveraging
attribute-based similarities, transitivity is endogenous and pertains to the pursuit of relational
closure. Transitivity manifests as the tendency for two nodes to establish connections when they
possess shared connections. In other words, the friend of a friend is more likely to be a friend
(Louch, 2000). Similar to homophily, transitivity also facilitates tie formation by reducing the
search and communication costs associated with forming relationships. Transitivity has been
extensively studied in collaboration networks, where shared connections provide direction for
individuals to seek collaborations, thus reducing the cost of searching. For instance, research on
academic collaboration has shown that the likelihood of scientists forming collaborations
increases with the number of common connections they share (Newman, 2001). Transitivity is
also associated with increased cohesion and trust within groups (Monge et al., 2003).
Relationships between dyads are more stable with the presence of a third node that connects both
members of the dyad (Stark et al., 2020). Drawing upon previous research on transitivity in tie
formation, it is hypothesized that:
H3: Clan members are more likely to form friendship ties with the friends of their
friends.
Friendship Formation: Matthew Effect
Besides the homophily effect, the literature on tie formation also highlights the Matthew
Effect, which refers to the "rich get richer" phenomenon (Merton, 1968). This effect, later known
as preferential attachment in network analysis (Barabási & Albert, 1999), suggests that new

40
nodes in a network tend to connect with well-connected nodes. This rule of preferential
attachment is a self-organizing phenomenon observed in large networks beyond individual
systems. It has been found in scientific collaboration networks (Zhang et al., 2018),
organizational networks (Fu, 2019; Saffer et al., 2022), and social networks for content selection
(Friemel, 2021). At the individual level, as an actor’s degree centrality increases, so does their
visibility, which leads to a higher chance of forming new ties. Furthermore, actors are motivated
to collaborate with popular individuals as it increases their productivity, visibility, and
recognition (Katz, 1994; Zhang et al., 2018). Similarly, organizations tend to affiliate with
popular organizations for their social influence and perceived legitimacy (Lai et al., 2019), which
enhances the focal organization's status and visibility in the network (Shumate & Lipp, 2008). In
social groups within video games, having more in-clan friends increases the visibility of popular
players. Forming friendships with popular players also creates more opportunities to meet new
people and play with known others at will, instead of having to wait in line to be randomly
matched with unknown players. Therefore, players are motivated to form friendships with
popular players. Based on preferential attachment and the study context, it is predicted that:
H4: Clan members are more likely to form ties with those who are popular.
Network Influence on Toxic Behavior
Individuals' social selection drives the formation of networks, while the evolving network
simultaneously affects individuals' attitudes and behaviors. The social influence processes have
been examined across various disciplines, such as political science (Putnam, 1966), sociology
(Festinger et al., 1950), psychology (Friedkin, 1998), and communication (Fulk, 1993). Results
from these streams of research suggest that individuals tend to become more similar to those
around them. Using longitudinal behavioral data in a popular multiplayer online game, World of

41
Tanks (WoT), Shen et al. (2020) found that toxicity is contagious. Exposure to toxic behaviors
significantly increases the likelihood of a player engaging in toxic behaviors themselves.
Moreover, Yokotani and Takano (2021) investigated the effects of online social networks on
cyberbullying from a social contagion perspective. By analyzing the online social networks of
cyberbullying perpetrators and victims, they found that the frequency of perpetrators in one's
network and the focal person's intimacy with perpetrators increase their own risk of being a
perpetrator. This line of research provides valuable insights into the social influence effects on
toxicity and its diffusion mechanisms.
The Social Identity Model of Deindividuation Effects (SIDE; Postmes et al., 1998) and
the general learning model (GLM; Buckley & Anderson, 2006) are two emerging theoretical
underpinnings underlying social influence and toxicity diffusion. SIDE has gained widespread
use in explaining individual behaviors in virtual environments. It posits that individuals entering
online environments may experience uncertainty about how to behave, and therefore look to
others or group norms as references. When entering a game world, the anonymity feature of
computer-mediated environments often leads to deindividuation. Deindividuation causes
individuals to view themselves not as independent beings, but rather as part of a group (Vilanova
et al., 2017). Personal identities become less important than social identities, which can lead to
depersonalization, where individuals categorize themselves as group members and regulate their
behaviors based on group norms. In computer-mediated settings, individuals tend to conform to
group norms, both prosocial and antisocial ones, to guide their behaviors (Chesney et al., 2009;
Kim et al., 2022; Shen et al., 2020). Consequently, game norms play a significant role in shaping
players' behaviors.

42
The GLM has been widely adopted to examine interpersonal learning processes in
behavior adoption and dissemination in both real-world and video game contexts (Barlett &
Anderson, 2012; Greitemeyer et al., 2010; Greitemeyer & Osswald, 2009). GLM proposes that
short-term and long-term media effects are mediated by a variety of learning mechanisms
(Buckley & Anderson, 2006). Grounded in social learning theory and social cognitive theory
(Bandura, 2001), the GLM posits that individuals learn through both direct experience and
indirect observation. For example, in video games, players may observe the behaviors of others
and the consequences of those behaviors, such as whether they are rewarded or punished. Based
on the outcomes of these behaviors, individuals may choose to learn and imitate similar
behaviors, particularly in computer-mediated settings where formal rules are unclear. However,
learning a behavior does not necessarily result in its manifestation. For instance, video game
players may learn toxic behaviors or even that toxicity is a norm in the game, yet they may not
display such behaviors themselves. The enactment of a behavior is dependent on situational
factors and communication partners. According to GLM, personal and situational factors work
together to affect an individual's internal state, which in turn guides how a person behaves in
general (Buckley & Anderson, 2006). The interplay between personal and situational factors
highlights not only the situational influences on individuals' behaviors but also their agency in
actively learning behaviors and selecting networks that fit their behavioral tendencies. Building
on SIDE and GLM, it is predicted that:
H5: The toxicity level of clan members is influenced by the toxicity level of their clan
friends over time.

43
Coevolution of Behavior and Network
Individuals' behaviors are both influenced by and shape their social networks, with
individuals actively selecting and constructing these networks (Lazer, 2001). However, social
influence is not a magic bullet, and understanding the coevolution of individual toxic behavior
and their friendship network has important theoretical and methodological implications. While
the coevolution of network formation and social influence has been examined in various fields,
such as communication, political science, sociology, and psychology (Friemel, 2015; Fu, 2019;
Lazer et al., 2010; Young & Fujimoto, 2021), it has not yet been applied to disentangling the
dynamic interdependencies between toxic behavior diffusion and friendship network evolution in
video games.
Players who repeatedly engage in toxic behavior may choose to be friends with other
perpetrators to engage in otherwise inhibited behaviors within the "magic circle" of the in-group.
Alternatively, they may become friends with other perpetrators due to common friends
(transitivity) or shared in-game experiences and skills. Modeling the selection and influence
processes simultaneously, along with appropriate alternative explanatory variables, can lead to
less biased estimates of parameters (Shalizi & Thomas, 2011). Previous research supports both
the selection and influence processes of gaming toxicity. For the selection process, intimacy with
perpetrators/victims is positively related to one's own tendency to be perpetrators/victims
(Yokotani & Takano, 2021). For the influence process, prior exposure to toxicity increases the
likelihood of adoption (e.g., Neto et al., 2017; Shen et al., 2020). However, it remains unclear
whether either process works when accounting for the other, and how the selection and influence
processes work together over time. Thus, this study seeks to answer the following research
question:

44
RQ1a-b: What is the relative contribution of the aforementioned effects on (a) network
selection and (b) social influence over time?
Methods
Study Context: World of Tanks
With over 160 million users worldwide (Takahashi, 2020), WoT is among the top 10
leading free-to-play multiplayer online games (SuperData, 2021). This highly competitive game
involves player-versus-player team-based shooting, featuring tanks from the mid-20th century.
To participate in WoT, players must collect or purchase tanks of varying types and tiers, and then
form a team of seven to 15 players to engage in battle with their opponents. The objective of the
game is to either destroy all enemy tanks or capture their base. The average battle time ranges
from five to 15 minutes.
WoT offers several game modes to cater to different player preferences. For instance, in
random battles, which is one of the most common play modes, players are randomly assigned to
a team through an automatic matching system. The matchmaking system in WoT has been
criticized since teams are not matched based on skills or experiences but rather on their tank
types and tiers. Skill disparities among teammates can lead to lower win-rates and higher
frustration, as cooperation and coordination are essential for victory in WoT. As a result, players
may opt to play in "platoon" mode, where they can join teams with up to three of their own
friends. Players can also join clans, which are larger and more permanent social structures within
the game. More advanced players may be motivated to join clans to have a more stable pool of
teammates to engage with, allowing for greater autonomy in starting a battle without waiting in
the queue for random matching. Joining a clan also unlocks more advanced game modes, such as
clan-based battles. Upon joining a clan, a player is likely to make new friends with their

45
clanmates and engage in battles with them. Clans serve as meso-ecosystems in the game, with
their own norms. Players select their clans and are influenced by their clanmates and the norms
of their chosen clan.
The highly competitive nature of WoT, coupled with the noticeable skill disparity among
players, often results in toxicity within and between teams (Shen et al., 2020), making the game
an ideal online environment to investigate the process of toxicity contagion. WoT employs two
mechanisms to combat in-game toxicity: a player-driven in-game reporting system and an
automatic toxicity-capturing system. When making complaints, players can choose from four
categories, including inappropriate behavior in chat, unsportsmanlike conduct, offensive
nickname or clan name, and inaction/bot. These categories encompass both text-based and
behavior-based toxicity, with some game-specific subcategories, such as inhibitions of teams and
AFK. To prevent potential abuses of the reporting system, a player can only report players with
whom they have battled. It is important to note that WoT has recently implemented a new tiered
punishment system, effective from March 2022, which takes into account the type, severity, and
frequency of gaming toxicity. It should be noted that this new system differs from the one in
place during the study period.
Data and Sampling
In collaboration with the game operator, Wargaming, this study collected two distinct sets
of data pertaining to friendship networks and toxicity on the North American server in 2019.
Two waves of friendship networks were unobtrusively collected, one in mid-April, 2019 and the
other in mid-July, 2019. Toxicity report data, including both player-initiated reports and the
game auto-captured ones, were collected during a three-month period prior to each wave of the
friendship network. A three-month timeframe was determined to achieve an optimal level of the

46
Jaccard index, a measure indicating the degree of dissimilarity between networks collected at
different time points. For the purpose of ensuring the assumption that the network's evolution
occurs gradually, it is preferable to attain Jaccard values surpassing 0.3 (Snijders et al., 2010).
Utilizing a three-month timeframe resulted in a Jaccard value of 0.5 between the two sets of
network data.
It is worth noting that the number of members in each clan is capped at 100 players, and
each player can only be a member of one clan at a time. To ensure (1) clans had a sufficient
number of players in their networks; (2) clans were active enough to enable considerable
variations in the network to be observed and studied; and (3) clans had enough co-play among
teammates where toxicity could be observed and learned, only clans that had 100 players in April
2019 were considered in the initial selection process. The included clans that had active days
ranging from over 100 days to over 1000 days. In order to ensure that the two waves of networks
had sufficient variability to study network evolution, one of the youngest clans (with around 100
active days) was randomly selected as the sample for this exploratory study. Only players who
were in the clan from April 2019 to July 2019 were included to construct the whole network
within the clan. The final sample included the sociocentric friendship network within the clan,
with 64 players and 2,016 undirected ties among them. The network data were merged with the
toxicity data using a unique, hashed key assigned to each player. All data were anonymized and
any identifying information was removed before reaching the research team. The data collection
process was approved by the Institutional Review Board (IRB).
Measures
Friendship Networks. The friendship lists of players within the selected clans were
collected unobtrusively on a daily basis, from April 2019 to July 2019. The friendship list

47
includes both a player's clan mates and non-clan mates. For the purpose of this study, only
connections between study participants (i.e., clan members) were retained. In WoT, friendships
are always reciprocal. Using each player's friendship list, an unweighted undirected edge list was
constructed and represented as a two-column matrix. Each row in the edge list represents the
presence of a friendship connection between a player in the selected clans (numerically identified
in column 1) and a clanmate (numerically identified in column 2). Subsequently, the friendship
edge list was transformed into a 64 X 64 sociocentric network matrix for further analysis. Two
waves of networks were constructed, one on April 15, 2019 and one on July 15, 2019. For
detailed information on edge changes between observations and its jaccard number, please refer
to Table 3.1. The network density at wave 1 is 0.022 and at wave 2 is 0.044. The average degree
across two waves is 2.062. Though co-play networks might be more influential in toxicity
contagion, friendship networks are preferred for this study to meet the assumption of the SIENA
model, which is suited for studying networks that are relatively enduring and can model gradual
changes in stable relational states (Snijders et al., 2010). Network ties are assumed to be states
(e.g., friendship states) rather than events (e.g., co-play events).
Table 3.1. Summary of network changes from T1 to T2.
Time period 0 → 0 0 → 1 1 → 0 1 → 1 Jaccard coefficient
T1 → T2 1,928 44 0 44 0.500

Toxic Behaviors. This variable was measured as the number of reports a player received
for inappropriate behavior in the game, including both player-filed complaints and those
automatically detected by the game. This data was collected from the game server daily from
January to July 2019, and includes information on the player IDs of both the reporting and

48
reported players, the reason for the report (i.e., inappropriate behavior in chat, unsportsmanlike
conduct, offensive nickname or clan name, and inaction/bot), and the time of the report.
As some players may receive multiple reports from multiple players for various reasons,
different categories of reports were aggregated into a total count of reports received by each
player over two three-month periods (Jan 15 to April 15, and April 16 to July 15, 2019), in order
to match the time stamp of the network data. The distribution of the number of reports was
skewed, with approximately half of the players showing no toxic behavior at both time periods,
followed by those who committed one or two instances of toxic behavior, and a smaller number
of players who committed three or more instances of toxic behavior. In order to meet the
requirements of the SIENA model, which requires nonnegative integer variables ranging from 0
to 10 (Snijders et al., 2010), the dependent behavioral variable was categorized into three levels.
Players who received no reports during the three-month period were given a value of 1 (wave 1:
N = 35, wave 2: N= 39), those who received one or two reports were given a value of 2 (wave 1:
N = 9, wave 2: N = 13), and those who received three or more reports were given a value of 3
(wave 1: N = 20, wave 2: N = 12). It should be noted that these categorizations represent
meaningful patterns of player behavior, as players who receive no reports are considered non-
toxic, while those who receive one or two reports may have been reported accidentally, but those
who receive three or more reports are unlikely to have been reported accidentally. For further
details on changes in player behavior between observations, please refer to Table 3.2.
Table 3.2. Summary of behavior changes from T1 to T2.
Time period Up Down Constant
T1 → T2 11 17 36

49
Covariates. The covariates used in this study were the number of battles played and the
win-rate of players. Previous research has demonstrated that players with higher skill and
experience are more likely to engage in toxic behaviors towards other players (Shen et al., 2020).
To account for these power differences, the number of battles played (M = 322.48, SD = 507.24)
and win-rate (M = 37.28, SD = 14.16), indicative of players’ experience and skill respectively.
These player-level metrics were also unobtrusively retrieved from the game server. To ensure the
analysis was scale-independent, the covariates were scaled before being used in the analysis.
Analysis
Previous research on the diffusion of behavior has often employed lagged regression-
based models to depict the exposure-infection processes (Valente et al., 2019; Wood et al.,
2012). Lagged models are easy to comprehend as they elucidate how prior exposure to a
behavior can result in its future adoption. These models are suitable for examining diffusion
processes in which the network is exogenous to the behavior, and there is no variable that
codetermines the network and the behavior (Valente et al., 2019). Moreover, regression-based
models assume that there are no interdependencies of individuals caused by network structure
(De Vries et al., 2006; Mercken et al., 2010). However, the independence assumption underlying
regression-based models is often violated when investigating human behaviors, particularly
when examining behaviors within groups. Moreover, although regression-based models are
highly valid in modeling social influence processes by including alternative explanatory
variables for the influence processes, they frequently fail to adequately control for network
selection mechanisms and endogenous network effects. Failing to explain or control for
alternative selection mechanisms may result in an overestimation of the strength of the influence
process, although the bias may not be too problematic (see Ragan et al., 2019).

50
Given the limitations of regression-based analyses and the interdependence of clan
members, this study employed SAOM to evaluate the interdependent coevolution of in-clan
friendships and toxic behaviors among clan members. SAOM posits that actors in a network
actively make decisions to change their ties or behaviors (or maintain the status quo) to optimize
their positions in a network (Snijders et al., 2007). At least two waves of longitudinal data are
required to model the co-evolution of networks and behaviors. Changes between waves of
observations are modeled using continuous-time Markov chains to determine the most likely
unobserved micro-changes taken by actors when modifying their network ties or behaviors
between observations. The changes that occur between observations are captured by two
components of the model: the rate function and the evaluation function. The rate function
captures the speed at which the dependent variables (i.e., network or behavior) evolve, whereas
the evaluation function captures rules that motivate these changes (Snijders et al., 2010). These
rules are modeled as “effects” and function as independent variables in the model. Parameter
estimates of these effects provide inferences about which effects are more likely to guide the
unobserved micro-changes that lead to the observed network and behavior (Snijders et al., 2010).
The analyses were performed using the 'RSiena' (Simulation Investigation for Empirical Network
Analysis) package version 1.3.0.1 in R version 4.1.3.
Model Specification. The model aims to analyze the coevolution of in-clan friendship
ties and toxic behaviors by estimating two submodels simultaneously, each with rate and
evaluation functions. The network dynamics submodel simulates the evolution of in-clan
friendship networks, allowing for the study of the network selection process. The behavioral
dynamics submodel simulates the evolution of toxic behaviors for studying the social influence
process. The combined model controls either process for the other one and simulates the

51
selection and influence processes simultaneously. The selection of in-clan friends and the
number of reported toxic behaviors are designated as the dependent variables observed at two
time points. The combined model is summarized in Table 3.3.
The friendship network submodel specifies the preferred direction of network change
through a list of "effects." In this study, a rate function was utilized to capture the pace of change
in the in-group friendship network. Two evaluation effects were employed to assess the impact
of toxic behavior on the likelihood of forming in-group friendships: (1) the effect of a player's
toxic behavior on their inclination to form in-group friendships, and (2) the effects of matches on
toxic behavior between actors and potential friends on the propensity to form in-group
friendships. Additionally, the model accounted for the effects of actor covariates (battle count
and win rate) on changes to in-group friendship ties.
As network dynamics have major endogenous effects, characteristics of the current
network structure are included to account for ways in which in-clan friendship networks are
formed in response to the presence or absence of other ties in the network. These include: (1) a
basic degree effect that models the overall tendency for actors to form in-clan friendship ties, (2)
the tendency for having network closure in in-group friendships (geometrical-weighted edgewise
shared partnerships – gwesp), and (3) an effect that represents a form of preferential attachment,
whereby players with a high number of in-group friends (i.e., high in-group degree) prefer to
connect with other players who also have many in-group friends (degree assortativity).
The toxic behavior evolution submodel accounts for three types of effects (see Table 3.3).
The first two effects can be described as contagion effects, as they account for the impact of prior
instances of toxicity among clan members in the network on a player's likelihood to engage in
toxic behavior. The total exposure effect captures the influence of the total number of contacts

52
engaged in toxic behavior on a player's rate of engaging in toxic behavior themselves. The total
influence effect captures the social influence of an alter's total number of toxic behaviors on an
actor's toxic behavior. Behavior trend effects reflect the pace of change in players' toxic
behaviors. The linear and quadratic terms capture the linear and quadratic shape of toxic
behavior across time points. Effects from covariates, win rate, and battle count, on toxic behavior
are also incorporated in the model.
Table 3.3. Description of the effects included for testing selection and influence processes.
Effect Description
Network Dynamics Submodel
Structural Effects
Degree Tendency to form in-clan friendships
Network transitivity (gwesp) Preference to form friendships with the friends of current friends
Degree Assortativity Preference of high degree actors to form ties with high degree alters
Behavior Effects
Behavior of actor Effect of actor’s toxic behavior on their in-clan friendship formation
Behavior homophily Preference to form in-clan friendships based on same behavior
Covariate Effects
Win rate of actor Effect of actor’s win rate on their in-clan friendship formation
Battle count of actor Effect of actor’s battle count on their in-clan friendship formation
Toxic Behavior Submodel
Contagion Effects
Total exposure to toxic behavior on rate Total number of in-clan friends that are engaged in toxic behavior
Total influence from alter’s behavior Total influence of alther’s toxic behavior on actor’s toxic behavior
Behavior Trend
Rate of period 1 The speed by which toxic behavior changes
Linear shape of toxic behavior The overall linear distribution of toxic behavior
Quadratic shape of toxic behavior The overall non-linear distribution of toxic behavior
Covariate Effects
Effect from win rate Total influence of win rate on actor’s toxic behavior
Effect from battle count Total influence of battle count on actor’s toxic behavior

53
Results
Friendship networks of the studied clan at each wave with information about their
toxicity level can be found in Figure 3.1. The network dynamics (see Table 3.4) submodel
showed that, in general, the formation of friendship ties among clan members from T1 to T2 was
not significantly affected by the level of toxicity exhibited by alter (b = -0.344, p > .05) or
network transitivity (b = 0.299, p > .05). Therefore, H1 and H3 were not supported. On the other
hand, degree (b = 0.206, p < .001) and assortativity mixing (b = -0.586, p < .05) exhibited
significant influences on individuals' formation of friendship ties, suggesting that, in general,
players are more likely to establish new ties with popular players but less likely to do so based on
mutual popularity, indicating preferential attachment and degree disassortative patterns of tie
formation. Consequently, H2 and H4 were supported. Additionally, the control variable, namely
the win rate of focal players, showed a significant influence on the formation of new ties (b =
1.004, p < .001), indicating that players with higher win rates are more likely to form new ties.
The toxic behavior submodel (see Table 3.4) demonstrated that, in general, players'
levels of toxic behavior displayed a decreasing linear pattern (b = -1.037, p < .001). The toxicity
level of clan members was marginally influenced by the toxicity level of their clan friends over
time (b = 0.518, p < .01). However, upon closer inspection of the relative influence of each effect
on each player's network evolution (refer to Figure 3.1) and toxic behavior (refer to Figure 3.2),
it appears that the toxicity level of clan friends played an important role in shaping the toxic
behavior of half of the players over time. In contrast, the other half of players whose toxic
behavior was not influenced by the toxic behavior of friends tended to be marginalized players
with few in-clan ties. Hence, H5 was marginally supported for players who are connected in the
network. To answer RQ1, Table 3.4 summarizes the relative contribution of the effects

54
mentioned above on (a) network selection and (b) social influence over time. Table 3.4 and
Figure 3.2, Figure 3.3 should be interpreted together to obtain a more accurate depiction of the
within-person and between-person variability in the co-evolution of network and behavior
processes.
Figure 3.1. Friendship networks of the studied clan at each wave with information about their
toxicity level.

A) Friendship network of clan members at wave
1
B) Friendship network of clan members at wave
2
Note. This figure illustrates the clan network (n = 64) at wave 1 (April 2019) and wave 2 (July
2019) in the study, highlighting their toxicity levels in the three months preceding the formation
of the network. Each participant is represented by a circle (node), and the circles are color-coded
as follows: green for no recorded instances of toxicity, white for one instance, and pink for two
or more instances. The network was generated using Gephi (version: 0.10.1) with the
Fruchterman Reingold layout algorithm and fixed coordinates.

55
Table 3.4. Significance of parameter estimates of the network and toxic behavior submodels.
Effect PE SE
Network Dynamics Submodel
Structural Effects
Degree 0.206*** 0.049
Network transitivity (gwesp) –0.299 0.542
Degree Assortativity –0.586* 0.244
Behavior Effects
Behavior of actor –0.372 1.648
Behavior homophily –0.344 2.513
Covariate Effects
Win rate of actor 1.004*** 0.302
Battle count of actor –0.171 0.443
Toxic Behavior Submodel
Contagion Effects
Total exposure to toxic behavior on
rate
–0.020 0.068
Total influence from alter’s behavior 0.518† 0.276
Behavior Trend
Rate of period 1 5.044 2.786
Linear shape of toxic behavior –1.037*** 0.297
Quadratic shape of toxic behavior 0.706 0.509
Covariate Effects
Effect from win rate 0.366 0.269
Effect from battle count 0.072 0.189
Note. † p < 0.1; * p < 0.05; ** p < 0.01; *** p < 0.001; all convergence t ratios < 0.1. Overall
maximum convergence ratio 0.24.

56
Figure 3.2. The relative importance of each effect on friendship network formation.

Note. “exp.rel.imp.” = expected relative importance.

Figure 3.3. The relative importance of each effect on toxic behavior evolution.

Note. “exp.rel.imp.” = expected relative importance.
Discussion
The objective of this study was to investigate whether a player's level of toxic behavior
had an impact on their in-clan friendship formation, and if these friendships influenced their
toxic behavior. SAOM was used to test both of these effects simultaneously as coevolving
dynamics. The results reveal the longitudinal influence of one's network on behavior evolution
and the influence of one's behavior on network formation, while taking into account individual
attributes and structural factors. In contrast to what was predicted, clan members' friendship

57
formation was not influenced by homophily in toxicity. This could be because toxicity is a latent
characteristic that is difficult to discern if the two players have not played together or
communicated. It is also possible that since the 64 players were already in the clan at the start of
the observation period, the influence of homophily in toxicity may have already played out and
thus was not captured during the study period. Furthermore, it is plausible that clan members do
not take the alter's toxicity level into account when making tie formation decisions. As the
significant effects of win rate and degree indicate, friendship formation is primarily driven by the
win-rate and popularity of actors. Those popular, active, and competitive players have greater
visibility and attractiveness in the network, thus are more likely to attract potential friends. This
observation may also relate to a ceiling effect wherein a clan can accommodate a maximum of
100 players, leading less popular players to be more inclined to form connections with popular
counterparts. These findings are consistent with the theoretical framework of network evolution
based on node fitness and preferential attachment, where node fitness represents their utility and
importance in the network (Bell et al., 2017). In the context of competitive video games, node
fitness is best captured by their win rate. Therefore, nodes that demonstrate the value of fitness
and popularity become more attractive, and other nodes want to connect to them.
Interestingly, network transitivity showed no significant effect on new tie formation, and
assortativity showed a negative effect on new tie formation. The pattern of non-significant
transitivity is unusual in online communities and is more like an intentional avoidance of
forming friendships with friends’ friends. Rather, popular players are less likely to form
friendships with popular players and more likely to form friendships with less popular ones.
From the network evolution figure (Figure 3.1A and Figure 3.1B), it is also apparent that at T2,
several new nodes emerge as popular nodes and are connected to those isolated nodes in T1, but

58
the popular nodes are not directly connected with each other, making the network more clustered
than T1 and forming three centers. This might be related to the competitive nature of the game
and the characteristics of its male-dominated user base. According to the social exchange theory
(Cook & Emerson, 1978), power is produced through varying interaction frequencies and
symmetries, such as nonreciprocal interaction (Olsen et al., 2014). Network members with more
interactions (degree centrality) tend to have greater power (Flyverbom et al., 2015; Saffer et al.,
2018). In a competitive game like WoT, clan members who have more friends may seek to
connect with less popular clan members to increase their influence, reputation, and power. Clan
members who have fewer friends may seek to connect with more popular clan members to gain
access to more resources, information, and social capital. Consistent with the social exchange
theory, nodes have different preferences for benefits and costs of forming ties and seek to form
ties that maximize their net benefits. Thus, instead of driven to make friends with friends’
friends, players may choose to make friends who are different from them in terms of degree. It is
worth noting how this pattern of individual behaviors may influence the characteristics of the
stability of the clan. Based on previous research, assortative networks are more homogeneous
and robust to removal of high-degree nodes, while dissortive networks are more heterogeneous
and vulnerable (Newman, 2002). Removal of influential nodes from the network may cause the
collapse of the whole clan. Therefore, to increase player retention and clan stability, the game
company may consider keeping the few players with most degree centrality.
The behavior submodel indicated that players' level of toxicity is marginally influenced
by the toxicity level of their clanmates. Upon closer examination at the individual level, the
results reveal that the influence of clanmates' toxic behavior has a greater impact on those who
are more embedded in the network, and less on those who are isolated. These findings support

59
GLM and suggest that players can learn from the behavior of their clanmates, both good and bad,
highlighting the importance of norms. For instance, node 9’s toxic behavior was found to be
largely influenced by their alters. This node’s behavior transitioned from toxic to less toxic,
mainly due to the influence of their in-clan friends. This result highlights the importance of
social contagion and social norms in shaping individuals' behavior, particularly for those who are
well-connected in the network. Moreover, the study suggests that isolates can be influenced
through nudging or incentives that encourage social connections to increase their dependence on
other players, susceptibility to game norms, and loyalty to the game. Game designers and
marketing teams may consider providing incentivized referral codes for both well-connected
players and isolated players to invite their friends to join the game, thus increasing their
connectedness in the game.
Overall, clan members tend to decrease their toxic behavior over time, which may be
related to the increasing intolerance of toxicity within the game and the clan. It is also possible
that clan members learn from their own or vicarious experiences or feedback that toxic behavior
is detrimental to their performance or reputation in the game, and thus may adjust their behavior
accordingly. This finding is consistent with the GLM that suggests players update their behavior
based on learning rewards and punishments from their environment and adopt behavior that
maximizes their expected utility. The decreasing pattern of toxic behavior also supports the
effectiveness of in-game punishments for toxic behaviors. Depending on the frequency of a
player's rule violation, the player will receive different types of punishment, ranging from a chat
ban to a game ban, even to a permanent ban. These measures may explain why players show a
decreasing pattern of toxic behavior over time and why most players had a violation under 2
times during the study period.

60
Meanwhile, the significant difference in the susceptibility of isolated players and
connected players to social influence emphasizes the importance of visualizing and examining
both within-person and between-person differences during data analysis. At the between-person
level, it is crucial to acknowledge individual differences in terms of their proclivities to
cooperate or compete, play solo or in a group, and the motivating factors behind their friendship
formation and toxic behavior. At the within-person level, it is essential to understand what
influences players' perceptions and behaviors over time. By comprehending both within-person
and between-person differences, game designers can implement targeted interventions to
enhance player experience and fulfill player needs. Only by understanding the factors that drive
players' social and unsocial behavior can game designers and scholars design effective
interventions to combat toxicity and foster a positive online environment for everyone.
Theoretically, building on Study 1 which investigates the correlation between network
position and toxic behavior, the current study further tests a unified theory on the coevolution of
network formation and behavior development, elucidating the relative importance of each factor
at the within-person and between-person levels over time. Practically, results of the study
support the efficacy of rule violation policies in a major commercial video game, which can offer
guidance on toxicity management for other game companies and online communities as well as
intervention design for scholars. In general, game designers and scholars should devise games
and rules that penalize toxic behavior and promote prosociality, while simultaneously increasing
connections among network individuals, such that individuals adhering to the rules will exhibit
the highest fitness and attractiveness. As positive norms emerge and individuals become
connected, their behavior will align and toxicity will diminish.

61
Despite the theoretical and practical implications of this study, there are several
limitations that need to be acknowledged. Firstly, the findings are based on data collected from a
single clan within a specific game. While this approach aligns with the research design, it is
recommended that the unified theory is tested using data from other networks in different
contexts, thereby expanding the external validity of the results. Secondly, players are more likely
to observe and learn from each other in co-play, indicating that a co-play network may provide
more insight into the co-evolution of network and behavior. However, as co-play networks are
transient, they exhibit a low Jaccard value across time, making them unsuitable for the SAOM.
Nevertheless, if the results hold for friendship ties, it is plausible that they would be more
prominent in co-play ties. Thirdly, the network evolution modeled in this study pertains only to
clan membership and does not account for friendships outside the clan. At the ego level, some
ties may not be represented in the study. This is a common issue in whole-network constructions
where it is difficult to account for every tie an individual has. The underlying assumption is that
players with close relationships are more likely to join the same clan. However, the study did not
consider players' friendship networks outside the clan, which could also have a significant
impact on their toxic behavior. Future research could address this limitation by studying the
multilevel networks nested within games. Furthermore, while SAOMs allow research to benefit
from a longitudinal design and make causal inferences on the selection-influence problem during
the observational period (Lomi et al., 2011), they do not fully address the causal inquiries.
Causal interpretations must be complemented by results from experiments and qualitative studies
to provide more robust evidence. While Study 1 and Study 2 offer valuable insights into the
emergence and diffusion of toxicity in gaming, they do not explore external factors that may
influence collective behavior in social systems. To enhance the current understanding of the

62
antecedents of variation in toxic and prosocial behavior, Study 3 incorporates exogenous factors,
specifically the stay-at-home order during the COVID-19 pandemic in the United States, as a
natural experiment to draw causal inferences about whether social isolation during a period of
collective trauma leads to changes in helping behaviors.

63
CHAPTER 4: A NATURAL EXPERIMENT OF HOW STAY-AT-HOME ORDERS
UNLEASHED A WAVE OF VIRTUAL ALTRUISM (STUDY 3)
The magnitude and impact of the COVID-19 pandemic are unparalleled, having brought
nearly the entire world to a halt (Amin et al., 2020). To mitigate the spread of COVID-19,
various public health measures, such as quarantine, lockdown, and social distancing, were
implemented globally. These measures have caused significant disruptions to regular work,
education, and social activities, prompting individuals to remain at home to flatten the curve.
Consequently, many individuals have shifted their social and entertainment activities from
offline to online, particularly toward video games. During the week of March 23-29, 2020, when
the stay-at-home order went into effect worldwide, Nielsen reported a 45% surge in video game
usage in the United States (Shanley, 2020a), while Verizon reported a 75% increase in video
game usage during peak hours (Shanley, 2020b).
With the rise of video game usage amidst the pandemic, many researchers have explored
the advantages and potential drawbacks of using them to cope with the social fallout of COVID-
19. While numerous studies have provided valuable insights into the helpful (e.g., Pahayahay &
Khalili-Mahani, 2020; Viana & de Lira, 2020) and harmful effects of video games (e.g., Amin et
al., 2020), most have relied on retrospective surveys to investigate video game use and player
experiences. However, these surveys may be distorted by memory or self-serving biases
(Vollhardt, 2009; Zoellner & Maercker, 2006). No study, to our knowledge, has utilized
unobtrusively collected behavioral data to examine actual user behavior during collective trauma.
Leveraging access to a cooperative multiplayer social game, Sky: Children of the Light, this
study views the stay-at-home order amidst the COVID-19 pandemic as a natural experiment to

64
explore players’ prosocial behavior variation before and after the state-wide lockdown order, and
whether the kindness extends toward outgroup members.
Players' in-game behaviors are critical to consider, especially during the pandemic, as
engagement in or exposure to prosocial behaviors has been shown to positively predict well-
being (e.g., Halbrook et al., 2019). Conversely, the involvement in or exposure to toxic behaviors
may harm players’ enjoyment and well-being (e.g., Türkay et al., 2020). According to the Model
of Altruism Born of Suffering (ABS; Vollhardt, 2009), individuals who have undergone
adversity are more inclined to engage in prosocial behaviors not only despite their difficulties but
because of them. The ABS model outlines several contextual factors that may shape the
manifestation of prosocial behavior and urge further investigation to delineate the circumstances
under which acts of kindness extend beyond the boundaries of one's own social group to
encompass members of other groups (Staub, 2005). While offline studies have consistently
supported ABS (see a review in Vollhardt, 2009), the examination of its applicability to online
behavior remains deficient. Furthermore, prior studies have been constrained by their focus on
post-event behaviors, a limitation that may arise from inadequate access to historical data
capturing human behavior during the occurrence of events. Additionally, longitudinal studies
incorporating behavioral measures have been scarce, posing challenges in establishing causality.
These limitations were also acknowledged by the founder of ABS as challenging yet
essential aspects that necessitate future research (Vollhardt, 2009). Fortunately, virtual
environments present opportunities to address some of these challenges by expanding the
repertoire of behaviors explicable within the ABS model and by applying it to online contexts. In
this natural experiment, the treatment is the lockdown order imposed in California, the first state
in the U.S. to adopt a lockdown order, and the control group is the absence of such an order in

65
Oregon. The difference-in-differences (DiD) approach was employed to compare changes in
prosocial behavior between the treatment and control groups before and after the intervention.
By using DiD, the study can better control for other factors that may influence the outcome, such
as pre-existing disparities between the treatment and control groups. To further explore the
inclusiveness and scope of prosocial behavior toward others, this study expands the ABS by
examining whether changes in prosocial behavior are contingent upon the in-group and out-
group membership of other players. Theoretically, this study broadens the scope of prosocial
behavior and the inclusiveness of recipients within the ABS model, while also extending its
applicability to virtual environments. Practically, this study sheds light on the potential variations
in human behavior during times of collective trauma and provides recommendations for game
companies to enhance player experiences and well-being, particularly in challenging times.
The Model of Altruism Born of Suffering
The literature on negative psychological effects and violent behaviors in response to
adverse experiences (e.g., Bonanno & Jost, 2006; Moya, 2018; Skitka et al., 2004), as well as
positive psychological effects and prosocial behaviors in response to positive stimuli (e.g., Aknin
et al., 2018; Penner et al., 2005), is well-documented. However, this linear perspective may not
fully encompass human resilience, post-traumatic growth, and the potential for thriving in
response to adversity. ABS proposed that some negative experiences may give rise to more
prosocial behaviors from those who endured those negative experiences. The proposition is
based on a large body of empirical evidence on individual behaviors during and after adverse or
traumatic life events (e.g., Staub, 2005; Staub & Ervin, 2003; Staub & Vollhardt, 2008),
and research from clinical psychology and social psychology explaining the motivational
processes underlying victims' choices. The model provides a detailed explanation of

66
circumstances where individuals choose to engage in prosocial behaviors rather than toxic
behaviors (e.g., revenge), their underlying psychological mechanisms, and some moderators that
either strengthen or weaken their motivations to help.
Specifically, ABS differentiates whether the suffering is experienced individually (e.g.,
losing close others) or collectively (e.g., natural disasters), and whether the suffering is caused
intentionally (e.g., terrorism) or unintentionally (e.g., accident), or even without human agency
(e.g., illness). The social impact of suffering depends on whether it is experienced by individuals
or as a collective. The intention behind the harm inflicted shapes the subsequent behaviors.
Intentional harm frequently increases the chances of violence and revenge, whereas unintentional
harm tends to foster prosocial behavior (Vollhardt, 2009). Moreover, the timing and beneficiaries
of prosocial behaviors span various levels of inclusivity and scope, encompassing factors such as
the immediacy of such behaviors in relation to the suffering experienced, whether the prosocial
acts are directed at an interpersonal or collective level, whether the prosocial behavior is targeted
towards members of one's own group or individuals outside of it, and whether the prosocial acts
extend to those who share a similar fate.
The model draws from clinical psychology and social psychology to explain the
motivational processes underlying altruism born of suffering. From the clinical perspective,
helping is an effective mechanism to cope with posttraumatic stress and can facilitate
posttraumatic growth (Midlarsky, 1994). From a social psychological perspective, altruism born
of suffering results from the interaction of personal and situational factors. At the personal level,
individuals who have experienced suffering will show more empathy and perspective-taking
toward other people’s suffering, especially when the perceived similarity between the help
provider and the help recipient is high (Epley et al., 2004). In addition to perceived similarity,

67
perceived common fate also leads to perceived shared group membership (i.e., common ingroup
membership) and subsequent helping behaviors. For example, helping behaviors among EU
citizens increased after natural disasters in other EU countries. The EU citizen identity becomes
more salient than the national identity after natural disasters (Levine & Thompson, 2004). At the
situational level, individuals are more likely to help others in contexts with a reciprocity norm of
helping (Southwick et al., 2005).
As previously mentioned, it is important to note that suffering does not always produce
altruism, as suffering can also result in cycles of violence. This underscores the significance of
identifying what factors facilitate (and inhibit) the link from suffering to altruism. In ABS, the
focus was on the presence and absence of facilitating factors, given the primary interest in
studying prosocial behaviors. Five facilitating factors were identified, including selective
attention (i.e., whether victims demonstrate a heightened awareness of others' suffering or remain
focused on their own difficulties), encoding control (i.e., whether victims recognize situational
cues that require help, such as perceived shared fate), emotion control (i.e., whether victims can
manage their negative emotions resulting from their suffering), motivational control (i.e.,
whether the motivation to help takes precedence over competing goals), and environmental
control (i.e., whether the environment permits or discourages one's helping behaviors). In this
study, the focus will be on encoding control (i.e., whether perceived shared fate exists or not) and
environmental control (i.e., the affordances and norms present in the environment).
Contextualizing the Model under a Collective Trauma in Virtual Environments
Expanding on ABS, this research conceptualizes the COVID-19 pandemic as a collective
trauma that was not caused by human agency and was experienced by society as a whole.
Collective trauma refers to “the psychological reactions to a traumatic event that affect an entire

68
society” (Hirschberger, 2018, p.1). In accordance with ABS, individuals influenced by the stay-
at-home order during the pandemic are more likely to exhibit prosocial behaviors rather than
violent behaviors due to its classification as a collective trauma that was caused unintentionally.
Therefore, the primary objective of this study is to investigate any potential increase in prosocial
behavior during the collective trauma, particularly following the enforcement of the lockdown
order. This study views the state-level lockdown order as a natural experiment since it indicates
the severity of the trauma and limits social interactions to primarily online spaces.
Previous studies have suggested that virtual environments, where millions of people
interact in a somewhat lifelike manner, have the potential to serve as research sites for examining
social and behavioral phenomena (Bainbridge, 2007). Similarly, Williams (2010) posited that by
building upon the parallel trends between online and offline worlds, human actions in one might
inform our understanding of human behaviors in the other. Since offline behavior is largely
limited and unobservable following the lockdown order, this study is situated in a virtual world
Sky and explores whether the ABS model still applies to the virtual realm in explaining players’
behavioral changes during the collective trauma.
Sky presents distinct game mechanisms and norms, where situational factors can act as
both mediators and moderators. Situational factors that "require" individuals to assist others due
to explicit requests or social norms of helping serve as mediators, while the affordances of
situations that support or hinder one's helping behaviors serve as moderators (Southwick et al.,
2005). Recognizing the importance of situational factors in shaping behaviors, this study
considers situational factors as given and examines the variability of prosocial behavior in Sky, a
game that encourages prosocial behaviors through its game design (Kim et al., 2022).
Specifically, the study features Sky players on the international server and sets the treatment time

69
to be the time of the first state-level lockdown order in the U.S. As the first state that adopted the
stay-at-home order was California, based on ABS, it is predicted that:
H1: Overall, CA players will exhibit more prosocial behaviors after the lockdown than
before the lockdown;
Self-Categorization in the Virtual World
In line with ABS, Self-Categorization Theory (SCT; Turner et al., 1987) predicted that
the salience of shared group membership gives rise to prosocial behavior that favors ingroup
members over outgroup members (Stürmer et al., 2005). Shared group memberships are
determined by two key factors: similarity and common fate (Castano et al., 2003). Individuals
categorize others as ingroup members when they perceive a high degree of resemblance or
shared fate. According to SCT, this ingroup-outgroup categorization shapes individuals' attitudes
towards others, particularly in virtual environments where nonverbal cues are limited. In such
contexts, individuals tend to display ingroup favoritism and outgroup antagonism, driven by
perceptions of limited resources and competition posed by outgroup members. Ingroup
favoritism may motivate players to exhibit greater prosocial behavior towards fellow ingroup
members, while outgroup antagonism may discourage the display of prosocial behavior towards
outgroup members. As the international server of Sky accommodates players from various
regions, players' national identities become salient following socialization activities, such as
chatting. Specifically, using CA players as the treatment group, other English-speaking players,
especially those from North America (NA), are more likely to be perceived as ingroup members
due to their shared characteristics with CA players and perceived common fate in the context of
the COVID-19 pandemic. Therefore, the hypothesis is formulated as follows:

70
H2: During the lockdown, CA players will exhibit more prosocial behaviors toward
ingroup members than before the lockdown;
The ABS model delineates a range of prosocial behaviors that span from interpersonal
assistance towards ingroup members in situations of immediate suffering, to collective aid
towards outgroup members in the aftermath. Nevertheless, it does not present specific
hypotheses regarding the conditions that prompt altruistic behavior towards outgroup members
(Vollhardt, 2009). Previous investigations indicate that this is contingent upon individuals' level
of abstraction of the concept of "common fate." If individuals perceive shared destiny in a
manner that surpasses overt characteristics such as nationality, ethnicity, or gender,
encompassing anyone affected by the disaster, they are more prone to encompass outgroup
members within a common ingroup and demonstrate benevolence towards them (Gaertner et al.,
2000). Although previous studies predicted that natural disasters lead to increased identification
among victims and that social barriers temporarily disappear (Eränen & Liebkind, 1993),
empirical research has found that structurally disadvantaged groups, such as ethnic minorities,
the elderly, and the less educated, are still less likely to receive help (Kaniasty & Norris, 1995;
Norris et al., 2005). This might be due to the importance of initial perceptions of similarity in
extending prosocial behavior to outgroup members (Flippen et al., 1996). This could explain why
outgroup members are not always included in the increased helping behavior after natural
disasters (Kaniasty & Norris, 1995; Norris et al., 2002), and why negative phenomena, such as
stereotyping and victim blaming, may arise (Napier et al., 2006). Interestingly, a review article
found that intentional harm caused by other human beings is often followed by altruism towards
other disadvantaged and outgroup members, whereas altruism following non-intentional harm
(such as natural disasters or illness) was more often limited to close others or those suffering

71
from the same event (Vollhardt, 2009). As the current study focuses specifically on interpersonal
prosocial behavior toward both ingroup and outgroup members in situations of immediate
suffering without human agency (see Figure 1 for the theoretical model of the study), based on
this line of research, it is predicted that:
H3: During the lockdown, CA players will exhibit fewer prosocial behaviors toward
outgroup members after the lockdown;
However, to date, there is no research that has examined whether prosocial behavior
extends to individuals whose group identities are unknown. Given that all players initially appear
as anonymous silhouettes before forming connections and engaging in conversation, it is
plausible that players in the Sky may perceive all other players as ingroup members based on a
shared identity as gamers, leading them to be more inclined to assist all Sky players in general.
Alternatively, it is conceivable that the unknown identities of other players lie somewhere along
the continuum between ingroup and outgroup. In this scenario, players may exhibit a greater
propensity to help ingroup members compared to strangers and outgroup members. Additionally,
it is also plausible that showing kindness to individuals about whom one knows nothing could
represent the utmost manifestation of altruism that players may engage in during the lockdown
period. Therefore, the following research question is posed:
RQ1: During the lockdown, will CA players exhibit more or fewer prosocial behaviors
toward members with unknown group membership?

72
Figure 4.1. Theoretical framework for the proposed studies.

Methods
Study Site
Sky: Children of the Light. Sky, a free-to-play mobile title, has garnered over 50 million
installations worldwide since its launch in July 2019 (Kerr, 2020). Following its release, it has
gained recognition as the top shared iOS game, the fifth most discussed iOS game, and has been
awarded Apple's iPhone Game of the Year for 2019, placing it as the twelfth-best iOS game of
that year (Metacritic, 2019). Additionally, the game has been honored with accolades for its
remarkable design and innovation (Apple Newsroom, 2020), and has received the esteemed
People's Choice Award from Games for Change (2021). Sky features a captivating and
beautifully animated virtual world that players can explore without boundaries. Unlike other
multiplayer online games, Sky has a tranquil pace and emphasizes cooperation rather than
competition. The game centers on traveling to new locations and collecting light, which players
can use to upgrade their capes for extended flights and to create candles for unlocking social
features and personalizing avatars. The architecture of Sky fosters help, trust, and social support,
making it a promising platform for studying prosocial behavior during collective trauma.

73
Another feature that makes Sky ideal for researching prosocial behavior is its social
mechanics. Building friendships is a core aspect of the game, and while it is possible to play
alone, forming connections enhances the experience and provides numerous benefits. Holding
hands and flying with friends enables players to soar higher and further, facilitating the
exploration of new areas or overcoming levels that require collaborative efforts. Upon entering
Sky, players encounter others from the global community, and everyone appears as a dark
silhouette. In other words, all players look the same when first met, and no identifiable
information is disclosed until players become friends. Typically, befriending other players in the
game requires exchanging one candle with each other. However, players can establish
friendships without offering a candle if they were previously acquainted and added each other
via a personalized QR code. To perform more advanced social actions like high fives, hugs, and
piggybacks, more candles are required, which can be obtained through candle farming (also
known as wax farming) or monetary purchases. Additionally, candles can be exchanged for
hearts, which players can then use as gifts for their friends. The allure of acquiring in-game
cosmetics through hearts makes this feature particularly appealing to players. The act of offering
candles is the targeted prosocial behavior for this study.
Data Collection
In collaboration with thatgamecompany (TGC), the developer of the game, this study
accessed the international server of Sky to collect data for the study. Data were anonymized by
TGC before reaching the research team. To investigate the impact of the lockdown order on
player behavior during the COVID-19 pandemic, this study designated the first state to
implement the lockdown order, California (CA), as the treatment group. Further, to apply the
difference-in-differences method to make causal inferences in this natural experiment, a control

74
group with a similar trend of prosocial behavior as CA before the treatment is required. This
study selected players from Oregon (OR) as the control group based on the following criteria: (1)
geographical proximity to CA; (2) it is the only nearby state that passed the parallel trend
assumption in DiD; (3) implemented the lockdown order a few days after California (rather than
right after or before CA), which enables a time window for a comparative analysis of the
differential effects of the lockdown order on players.
Data collection took place between March 1, 2020 and March 22, 2020, with the time
period preceding March 19, 2020 (the first day of the CA lockdown order) serving as the
baseline to observe parallel behavioral trends between players from CA and OR. The treatment
window spanned from March 19 to March 22, ending on the latter date as OR also implemented
the lockdown order beginning March 23. Players have to play for at least 30 min in March 2023
to be classified as active players that can be included in the analysis. The analytic dataset
includes 306,504 unique daily observations of 13,932 players, with 12,664 players from CA (N =
306,504 unique daily observations) and 944 players from OR (N = 20,768 unique daily
observations). The study was approved by the Institutional Review Board (IRB).
Measures
Prosocial Behavior. Prosocial behavior is defined as the act of offering candles to other
players and is categorized into three groups: unknown group membership (M = 0.15, SD = 0.93),
ingroup (M = 0.01, SD = 0.18), and outgroup (M = 0.01, SD = 0.21) based on the recipient's login
region, device language, and previous acquaintance. In the case of friendship formed via QR
code invitation or those who have unlocked the chat function, the group membership is assumed
to be salient. Otherwise, it is considered to be unknown. In the case where the group membership
is salient, ingroup membership is defined by a U.S. login regionand an English device language.

75
Otherwise, the group membership is determined to be outgroup. The decision tree of
categorizing the three groups can be seens in Figure 2.
Figure 4.2. Decision tree of ingroup outgroup categorization.

Control Variables. Players' pre-March 1, 2020 play time (M = 566.76, SD = 1977.95)
was included as a covariate. This includes their time spent in minutes from the start of their first
play to the end of Febuary 2020, accounting for their level of experience and engagement within
the game. Additionally, the number of wax units farmed from their first play to the end of
Febuary 2020 (M = 2504.37, SD = 9017.69) was incorporated as a control variable to signify the
amount of candles available to each player. Finally, the number of friends a player had from their
first play to the end of Febuary 2020 (M = 12.75, SD = 40.33) was included as a covariate,
accounting for the size of their social networks.

76
Analysis
To examine the hypotheses, this study employed the difference-in-differences (DiD)
analysis, a widely utilized approach across various disciplines, including the social sciences and
business. DiD has been effectively applied to investigate the impacts of COVID-19 lockdown
orders on individual well-being (e.g., Auffhammer & Kellogg, 2011; Brodeur et al., 2021). By
incorporating a control group, DiD is more robust against time-dependent confounding factors. It
effectively addresses the non-random assignment issue inherent in natural experiments by
introducing a comparison group that undergoes a similar trend as the treatment group but
remains untreated. The canonical DiD design encompasses two distinctions: (1) the post-
treatment difference relative to the pre-treatment period within the exposed group (expressed as
B2 - B1); (2) the post-treatment difference relative to the pre-treatment period within the
unexposed group (expressed as A2 - A1). The change exclusively attributable to the treatment,
beyond the background trends, can be estimated through the difference-in-differences analysis,
simplistically represented as (B2 - B1) - (A2 - A1). When there exists no association between the
treatment and the change in prosocial behavior, the expression (B2 - B1) - (A2 - A1) equates to
zero. Established DiD practices in regression analysis explore the interaction between the pre-
post variable and the exposed-unexposed variable (Dimick & Ryan, 2014). If the parameter of
the interaction term significantly deviates from zero, it indicates a substantial relationship
between the treatment and the change in prosocial behavior. In the current analysis, daily
prosocial behavior data were aggregated into three time spans: from March 01 to March 10 (pre-
treatment), from March 11 to March 18 (pre-treatment), and from March 19 to March 22 (during
treatment). This aggregation was done for two main reasons: By aggregating the data into larger
time spans, I aimed to reduce the large fluctuation in player activity observed between pre-

77
treatment and during-treatment periods at the daily level. This approach allowed for a more
stable comparison in the DiD analysis, as behavior during the treatment was compared with the
average activity across the respective pre-treatment time span. Another reason for aggregating
the data was to ensure that each time span had a relatively equal distribution of dates. This helps
to minimize any potential bias or confounding effects caused by uneven distribution of
observations within each time span.
The primary focus of this study centers on assessing the "Average Treatment Effect on
the Treated" (ATT), which quantifies the impact of the treatment on individuals who have
undergone the stay-at-home order. Leveraging the advancements in DiD methodology, the
primary analyses were performed using doubly robust difference-in-differences (DR DID)
estimators, implemented through the "did" package in R version 4.1.3 (Callaway & Sant’Anna,
2021; Sant’Anna & Zhao, 2020). DR DID possesses several advantages over alternative
approaches. Firstly, in contrast to regression models and other DiD models, the DR DID method
employs two distinct models to estimate the treatment effect: a propensity score model to gauge
the likelihood of receiving the treatment and an outcome regression model to estimate the
relationship between the treatment and the outcome. By jointly employing these models, the DR
DID method generates estimates that are less sensitive to model misspecification compared to
employing either model individually. Secondly, the DR DID method accommodates
nonparametric data by utilizing a nonparametric estimator for the propensity score. This
characteristic renders the method more flexible and robust against deviations from parametric
assumptions, such as normality or linearity. Thirdly, the DR DID method can be effectively
employed with high-dimensional data, where the number of covariates is substantial relative to

78
the sample size. In summary, the DR DID method is designed to yield reliable estimates of the
treatment effect even in the presence of model misspecification and violations of assumptions.
Results
The results of the DiD analysis are presented in Table 1. Parallel trends assumption
between CA and OR was met for all dependent variables (all p > .05), indicating that CA and OR
players demonstrated similar trend in the dependent variables before the stay-at-home order.
Control variables in DiD are used to reinforce the assumptions of parallel trends and have no
direct causal interpretation in model estimates (Keele et al., 2020), thus are not reported. This
study did not find a significant effect of the stay-at-home order on players' prosocial behavior
toward all players or in-group players. The ATT for candles offered to all players increased from
0.018 (SE = 0.133, 95% CI = [-0.276, 0.312]) before the treatment to 0.249 (SE = 0.126, 95% CI
= [-0.029, 0.527]) after the treatment, but this increase was not statistically significant. Similarly,
the ATT for candles offered to in-group players decreased slightly from 0.001 (SE = 0.022, 95%
CI = [-0.044, 0.045]) before the treatment to -0.004 (SE = 0.021, 95% CI = [-0.046, 0.039]) after
the treatment, but this decrease was also not statistically significant. Therefore, H1 and H2 were
not supported.
Interestingly, the stay-at-home order had a positive effect on players' prosocial behavior
toward unknown players (see Figure 3). The ATT increased from -0.016 (SE = 0.128, 95% CI =
[-0.284, 0.253]) before the treatment to 0.284 (SE = 0.114, 95% CI = [0.045, 0.523]) after the
treatment. This represents a percentage change of 1,875%. There was also a significant negative
effect of the stay-at-home order on players' prosocial behavior toward out-group players (see
Figure 4). The ATT for candles offered to out-group players decreased from 0.033 (SE = 0.022,
95% CI = [-0.016, 0.082]) before the treatment to -0.031 (SE = 0.013, 95% CI = [-0.060, -0.002])

79
after the treatment. This represents a percentage change of -93.9%. Taken together, H3 was
supported and RQ1 was answered.
Table 4.1. Average effect of stay-at-home order on CA players’s prosocial behavior.
Time ATT SE 95% CI
Candles Offered to
All Players
Pre-Treatment 0.018 0.133 [-0.276, 0.312]
Post-Treatment 0.249 0.126 [-0.029, 0.527]
Candles Offered to
Unknown Players
Pre-Treatment -0.016 0.128 [-0.284, 0.253]
Post-Treatment 0.284* 0.114 [0.045, 0.523]
Candles Offered to
In-Group Players
Pre-Treatment 0.001 0.022 [-0.044, 0.045]
Post-Treatment -0.004 0.021 [-0.046, 0.039]
Candles Offered to
Out-Group Players
Pre-Treatment 0.033 0.022 [-0.016, 0.082]
Post-Treatment -0.031* 0.013 [-0.060, -0.002]
Note: ATT = average treatment effect; SE = standard error; CI = confidence interval.

80
Figure 4.3. Average effect of stay-at-home order on CA players’ prosocial behavior toward
unknown players.

Figure 4.4. Average effect of stay-at-home order on CA players’ prosocial behavior toward
outgroup players.

Discussion
The distinctive design and characteristics of Sky establish it as an ideal and controlled
research environment for conducting a natural experiment. Upon entering the game, players are
immediately exposed to a diverse global community, although specific group membership

81
information may not be initially shared until players establish friendships and engage in further
interactions. As social interaction represents the primary objective of the game, players may,
albeit not guaranteed, discern cues related to group membership once they unlock the chat
function and become acquainted with the language and geographic cues provided by their
friends. This facilitates the observation of the effects of the stay-at-home order on players'
prosocial behavior towards various groups. The findings of the study unveiled a mixed impact of
the stay-at-home order on players' prosocial behavior towards different groups within the realm
of the online social video game. The order did not yield a significant effect on players' prosocial
behavior towards all other players, possibly due to the counterbalancing effect of increased
prosocial behavior towards unknown players and decreased prosocial behavior towards outgroup
players.
While players did not universally extend their prosocial behavior to all other players,
there was indeed an observed increase in prosocial behavior towards unknown players,
potentially attributed to the influence of initial perceived similarity. When the group identity of
other players remains unknown, they are perceived as part of the general player population, thus
being viewed in a similar manner to the focal player. This perceived similarity is rooted in the
shared membership of being a Sky player, rather than being influenced by collective trauma.
Consequently, players are more inclined to exhibit prosocial behavior towards unknown players
based on their common player membership. Alternatively, players may be driven by the inherent
need for socialization, prompting them to engage in prosocial behavior and offer candles to
unfamiliar individuals. During periods of social isolation, when offline social interactions are
limited, players may experience an increased desire for connectedness and turn to the virtual

82
realm to expand their social networks, communicate with others, and experience a sense of
belonging.
Given the constrained availability of candles, if players opt to assign a higher proportion
of candles to unknown players, a sustained allocation of candles to ingroup players signifies a
distinct prioritization of candles towards the ingroup. This inclination to offer candles to ingroup
players remains steadfast even amid fluctuations, due to the game's innovative design and
affordances that promotes reciprocity. It is noteworthy that three candles can be exchanged for a
heart, with gifting among friends serving as the primary avenue for obtaining hearts. Hearts, in
turn, can be utilized for the acquisition of diverse cosmetics, such as hairstyles and outfits. The
daily constraint of gifting only one heart to a specific friend serves as an impetus for players to
engage in gifting across a broader spectrum of individuals, rather than concentrating exclusively
on a particular friend. Consequently, players may strategically allocate a greater number of
candles to unfamiliar players as a means to expand their network circle and thereby heighten
their prospects of receiving additional hearts. In their quest for reciprocal interactions, players
often choose to preserve their candles for exchanging with familiar players, given the greater
degree of trust established within such relationships. Conversely, the allotment of candles to
unknown players may be restricted, particularly during periods of collective trauma, owing to a
lack of trust in reciprocal exchanges.
As predicted, the implementation of the stay-at-home order led to a decrease in prosocial
behavior towards members of outgroups. This is consistent with previous research under ABS, as
well as the propositions in SCT. SCT distinguishes between social identity and personal identity,
suggesting that both can operate simultaneously. While social identity is influenced by group
membership, personal identity is relatively more independent of group affiliations (Trepte &

83
Loy, 2017). In situations where all players are represented by dark silhouettes, devoid of
identifiable information that could differentiate their group membership, individuals' personal
identity as human beings may assume greater prominence. Consequently, players are more likely
to engage in prosocial behavior towards others based on a shared sense of humanity, rather than
their respective group memberships.
Nevertheless, when players establish friendships and engage in chatting, this mediated
form of communication can render their social identity more salient due to language and
geolocational cues. Players from countries implementing the lockdown order may perceive
players from countries without the order as outgroup members, perceiving a lack of common
fate, and consequently exhibit a decreased likelihood of displaying prosocial behavior towards
them. Based on SCT, it is also plausible that once group categorization becomes salient,
individuals may engage in outgroup derogation, leading to negative evaluations of the outgroup
and reduced inclination to exhibit kindness towards them. Additionally, when group membership
is salient, individuals are more prone to engaging in social comparisons and strive to be
evaluated more favorably than outgroups. One strategy individuals may employ to achieve a
more positive evaluation compared to outgroups is social competition. In the context of the
game, where candles hold an equivalent value to currency, individuals may be more inclined to
engage in a social competition process rather than giving away their own in-game currency, as
giving away candles would have an adverse impact on their own resources.
This finding provides a nuanced understanding of the ABS model by elucidating the
circumstances that influence the likelihood of individuals displaying prosocial behavior. It is
important to acknowledge that the current observation focuses on the immediate situation,
specifically the suffering and distress experienced by players affected by the stay-at-home order.

84
Previous research within the framework of ABS has revealed that the most inclusive caring,
which encompasses outgroup members, tends to occur in the aftermath rather than during the
immediate situation. This is because individuals in distress require temporal distance from the
events (Trope & Liberman, 2003) and time to restore their psychological resilience and attain
closure (Skitka et al., 2004), enabling them to engage in more inclusive forms of altruism arising
from suffering. Hence, in the present study with a relatively narrow observational window,
players may lack the mental distance and resources necessary to exhibit prosocial behavior
towards others, particularly outgroup members. However, if provided with a more extended
observational window, players may expand the inclusiveness and breadth of their kindness.
Although the overall prosocial behavior of CA players increased following the
implementation of the stay-at-home order, the comparison with the control group and the pre-
lockdown period did not yield statistical significance. This lack of significance may be attributed
to CA players being the first group in the U.S. to experience the stay-at-home order.
Consequently, they may not have perceived all other U.S. players as sharing a common fate,
leading to a reduced inclination to engage in prosocial behavior towards them. Therefore, a
viable next step would be to develop a more nuanced categorization of ingroup and outgroup
membership. For instance, CA players might only consider other CA players as ingroup
members during the lockdown, or they may view players from states with similar political
ideologies as ingroup members. Furthermore, it is worth considering the indirect psychological
impact of the stay-at-home order on players in Oregon, who did not have a formal order in place
during the observational period. Given their awareness of the situation in CA and the anticipation
of a potential lockdown, Oregon players may have been psychologically affected by the events in
CA, which could have influenced their behavior. In light of this observation, an interesting

85
avenue for future research would be to investigate whether the lockdown order lowered players'
inhibitions in seeking help, particularly by examining whether players actively seek or ask for
candles through interpersonal communication. This would provide valuable insights into how
players navigate and utilize social interactions during challenging times, shedding light on the
mechanisms underlying prosocial behavior in online communities.
This study provides empirical support for ABS and SCT and expands their applications in
the virtual world. Previous studies have consistently supported the phenomenon in offline
settings (e.g., Brown et al., 2008; Hannah & Midlarsky, 2005; Norris et al., 2005). The present
findings reveal a similar trend within an online community, albeit with additional nuances in
behavioral variations in response to the group membership of recipients. ABS-related
phenomena have predominantly been explored in the fields of clinical psychology and
psychiatry, with limited systematic investigation into how adverse life experiences can motivate
altruism and prosocial behavior in the virtual world, especially towards outgroup members
(Vollhardt, 2009). Understanding the conditions that facilitate prosocial behavior, particularly
regarding the inclusion or exclusion of outgroups, holds significant theoretical and practical
implications for comprehending human behavior during challenging times and intergroup
relations. The findings highlight the potential of online communities to empower individuals
experiencing hardship by recognizing and harnessing their capacity to help others, thereby
enhancing their self-efficacy and well-being. Consequently, this study recommends that the
game industry leverage the social affordances of video games to promote and incentivize
prosocial behaviors among players, particularly in the context of collective trauma. For instance,
within Sky, players tend to prioritize their social interactions with familiar individuals, driven by
a sense of trust in reciprocal heart gifting, which in turn can be utilized to acquire in-game

86
cosmetics. While this aspect encourages social interactions within one's familiar social circle,
redirecting players' focus from outfit competition towards features that facilitate collaboration
with less familiar friends could cultivate a more supportive environment and activate a broader
range of social activities among players. Furthermore, game designers may consider prioritizing
cues that foster a shared gamer identity over those that perpetuate divisions based on observable
characteristics. Such an approach would help mitigate the negative impacts experienced by
individuals undergoing personal struggles and strengthen the resilience of game communities at
the collective level.
Despite the implications of this work, there are several limitations. Firstly, the data are
drawn from one single game. Although the game has unique and controlled affordances that are
particularly suited for the natural experiment, the results should be interpreted within context and
with caution when generalized to other settings. It is important to consider that not all games
possess the same level of prosocial and cooperative elements as Sky. Conducting future research
within games that exhibit a more neutral position on the altruism-toxicity spectrum could provide
a stronger natural experiment and offer insights into the salience of the results. Second, despite
CA being the first state in the U.S. that adopted the stay-at-home order, players in OR may
expect their state to implement a similar order. Thus, the behaviors of OR players may assimilate
the behaviors of CA players after the CA lockdown order. In fact, OR adopted the stay-at-home
order a few days after the CA order, thus leaving the study a limited time window to observe the
unfold of behavior dynamics. Lastly, although the study employed doubly robust estimates to
make causal inferences, it was not an experiment, and therefore the causal inferences should be
interpreted with caution. By combining the findings of Study 1 and Study 2, Study 3 further
elucidates the external factors that influence collective player behavior as a whole. The studies

87
reveal that players' behavior is impacted by the network structures they are embedded in, the
social group they belong to, and policy changes that significantly alter their social environment.
These findings highlight the importance of adopting an ecological perspective in the study of
behavior for future research. However, obtaining psychometric information at the player level
from large-scale behavioral data is challenging. Thus, future research may consider using
surveys or interviews to gain insights into specific individual differences, such as their
personality and attachment style, which influence their propensity to engage in toxic or prosocial
behavior.

88
General Discussion
Motivated by a pressing need in academia and the gaming industry to understand the
origins and diffusion of digital exclusion and gaming toxicity, as well as an interest in the
variation of human behavior arising from individual characteristics and social influence, this
dissertation endeavors to explore how network structures, friendship selection, social influence,
and policy change impact players’ toxic and prosocial behavior variation. The results from a
series of three studies reveal that, overall, the network positions that players occupy confer both
power and responsibility within a social system, shaping their inclination towards perpetration
and susceptibility to victimization. While individuals primarily choose friends based on their
fitness within the social system, their toxic behavior is also influenced by their friends' toxic
behavior. Furthermore, amidst the social isolation and collective trauma associated with the
COVID-19 pandemic, players are more likely to display altruistic behavior towards anonymous
players. However, once the salience of national identity is introduced, players display a
decreased propensity to exhibit such behavior towards out-group members compared to the pre-
isolation period.
Study 1 integrates several theoretical perspectives, including the Structural Hole Theory,
Differential Self-Awareness Theory, Power, Approach, and Inhibition Theory, and the Shadow
of the Future Effect, to explore how network structures confer power and accountability to
players in virtual environments. Specifically, the study examines how network size, brokerage,
and closure are related to players' roles in gaming toxicity dynamics. The results indicate that
different network positions may lead to distinct patterns of toxic behavior and demonstrate the
potential of network analysis in identifying toxic and vulnerable players. However, it remains
unclear whether players actively choose to occupy specific network positions or if such positions

89
present opportunities and risks that shape players’ behavior. Therefore, Study 2 employs a
stochastic actor-based modeling approach to investigate the selection-influence dynamics and the
co-evolution of network and behavior using longitudinal digital footprint data.
Study 2 utilizes a combination of homophily theory, the Matthew Effect, and preferential
attachment to predict network evolution, and draws on social influence theories such as the
Social Identity Model of Deindividuation Effects and the General Learning Model to predict
social influence. The findings demonstrate support for the Matthew Effect and preferential
attachment in network evolution, revealing that individuals with more friends and higher win
rates are more likely to form new ties. While individuals do not consider homophily in toxic
behavior in forming new relationships, their toxic behavior can be influenced by their friends if
they are well-connected in a network. These results underscore the importance of promoting
positive norms to encourage positive social influence and foster a positive online environment.
Furthermore, the study employs stochastic actor-based modeling to establish causal relationships
between network and behavior coevolution over the observational period. Nonetheless, the study
does not consider the impact of external influences on players' behaviors, nor does it explore
whether and how behaviors extend to ingroup and outgroup members.
To investigate the impact of exogenous factors on collective behavior in online
multiplayer games, Study 3 utilizes the stay-at-home order during the COVID-19 pandemic as a
natural experiment to explore causal relationships between players' prosocial behavior towards
different types of players (anonymous, ingroup, outgroup) before and after the order. The results
demonstrate that players in the state that implemented the order exhibited more helping behavior
towards unknown players, but less helping behavior towards players from a different country or
speaking a different language, compared to before the lockdown and an adjacent state that did

90
not implement the order. These findings not only extend the Altruism Born of Suffering model to
the online world, but also link it with the Self-Categorization Theory to explain the degree to
which altruistic behavior applies to others in the virtual world.
The three studies presented in this dissertation offer valuable insights into the complex
interplay between network structures, individual differences, and contextual factors that shape
toxic and prosocial behaviors in online communities. The studies provide evidence supporting a
number of theories that help understand how network structures confer power and accountability
cues to individuals, how network structures co-evolve with behavior, and how behavior is
influenced by norms within a network. Additionally, the studies demonstrate how online user
behaviors are susceptible to sociocultural context and policy changes. Taken together, these
findings suggest an overarching theory that explains and predicts individuals’ prosocial and toxic
behavior in online communities based on their positions in social networks, social interactions,
norms in social networks, contextual factors, as well as individual differences.
However, the studies discussed have limitations that should be taken into account when
interpreting the results. Several dilemmas arose, and trade-offs were made in study execution.
Firstly, the generalizability of the findings was weighed against data availability. While testing
the hypotheses and research questions across various contexts and platforms would increase the
generalizability of the findings, obtaining access to natural user behavior from commercial
games can be challenging. Thus, the study relied on available data from two specific games and
their players, and the results may not fully represent the diversity and complexity of online
gaming as a whole. Future research should collect data from multiple games and platforms to
examine the robustness and variability of the network structure and toxicity relationship across
different contexts and settings.

91
Another trade-off was between internal and external validity. As the study relied on
naturalistic observation to make causal inferences on player behavior, the causal link can only be
inferred during the observational period. Since there was no real control group or random
assignment, there is a trade-off between the internal and external validity of the study. Future
research should complement observational studies with real experimental designs that can test
the causal effects of network structure and social influence on toxicity, as well as their
interaction effects of contextual factors and individual differences.
A third trade-off was the balance between utilizing large-scale behavioral data and
obtaining individual-level, in-depth understanding of players' psychometric information. Future
research may benefit from examining the influence of individual differences on players'
willingness to help or harm others. The General Aggression Model (Allen & Anderson, 2017)
suggests that various personal and situational factors interact to predict bullying behaviors. For
instance, attachment style, personality traits, and the nature of the communication medium may
all play a role. According to attachment theory (Bowlby, 1982), early attachment-related
experiences shape individuals' needs, expectations, and behaviors. Anxiously attached
individuals may be more likely to engage in aggression, whereas avoidantly attached individuals
may withhold their emotions in response to conflict (Miga et al., 2010). Individuals with insecure
attachment tend to be a bully, a victim, or an aggressive victim (Varghese & Pistole, 2017).
Research in the Big-Five personality traits paradigm has also found that high extraversion, low
conscientiousness, and low agreeableness are related to cyber victimization (Kokkinos &
Saripanidis, 2017). Furthermore, some individuals may generally be helpful, but have a lower
threshold for feeling threatened, which can result in a failure to exhibit helping behavior even
when experiencing similar suffering.

92
In addition to dispositional differences, instantaneous affect also influences players'
behaviors. For instance, according to the attribution-affect model (Roseman, 1996), anger is a
prevalent emotional response to unjust treatment. The model has been employed in video game
research elucidate players' misconduct caused by anger triggered by disruptive gaming behaviors
such as account theft (Teng et al., 2012). Moreover, according to Roseman (Roseman, 2001),
emotions are initiated by appraisals of events rather than the events themselves. Disruptive
actions that impede an individual's goals and are attributable to others tend to trigger anger.
Therefore, players who encounter toxicity in games might feel resentful towards the responsible
parties, obstructing their goal attainment and feeling motivated to rationalize their retaliatory
behavior as appropriate, thereby becoming an aggressive victim. Similarly, the power that an
individual holds relative to the target may also influence their readiness for action and
engagement in aggressive conduct, even if they feel anger. Thus, appraisal and affect wield
substantial influence over an individual's toxic behavior and behavioral response to toxicity.
Future research should utilize survey and qualitative research methodologies to obtain an in-
depth understanding of the psychological mechanisms underlying individuals' behavior. Based
on results of the current studies and future research on individual differences, a unified theory
can be proposed to explain and predict the occurrence, frequency, and recipient of toxic and
prosocial behaviors.

93
References
Aiello, L. M., Barrat, A., Schifanella, R., Cattuto, C., Markines, B., & Menczer, F. (2012).
Friendship prediction and homophily in social media. ACM Transactions on the Web
(TWEB), 6(2), 1–33.
A. Sparrow, L., Gibbs, M., & Arnold, M. (2021). The ethics of multiplayer game design and
community management: Industry perspectives and challenges. Proceedings of the 2021
CHI Conference on Human Factors in Computing Systems, 1–13.
Adinolf, S., & Turkay, S. (2018). Toxic behaviors in Esports games: Player perceptions and
coping strategies. Proceedings of the 2018 Annual Symposium on Computer-Human
Interaction in Play Companion Extended Abstracts, 365–372.
Aivazpour, Z., & Beebe, N. (2018). Cyberbullying: Investigating the Roles of Power and
Communication Medium. AMCIS.
Aknin, L. B., Van de Vondervoort, J. W., & Hamlin, J. K. (2018). Positive feelings reward and
promote prosocial behavior. Current Opinion in Psychology, 20, 55–59.
Allen, J. J., & Anderson, C. A. (2017). General aggression model. The International
Encyclopedia of Media Effects, 1–15.
Amin, K. P., Griffiths, M. D., & Dsouza, D. D. (2020). Online gaming during the COVID-19
pandemic in India: Strategies for work-life balance. International Journal of Mental
Health and Addiction, 1–7.
Anti-Defamation League. (2020, December). Disruption and Harms: Online Gaming
Framework. Retrieved from https://www.adl.org/resources/report/disruption-and-harms-
online-gaming-framework
Anti-Defamation League (ADL). (2020, December). Fair Play Alliance and ADL announce
partnership to rally industry and combat hate. Retrieved from
https://www.adl.org/resources/press-release/fair-play-alliance-and-adl-announce-
partnership-rally-industry-combat-hate

94
Anti-Defamation League. (2022, June). Online Hate and Harassment Survey 2022. Retrieved
from https://www.adl.org/sites/default/files/pdfs/2022-06/Online-Hate-and-Harassment-
Survey-2022-v7.pdf
Apple Newsroom. (2020, June). Apple Honors Eight Developers with Annual Apple Design
Awards. Retrieved from https://www.apple.com/newsroom/2020/06/apple-honors-eight-
developers-with-annual-apple-design-awards/
Aquino, K., & Lamertz, K. (2005). A Relational Model of Workplace Victimization: Social
Roles and Patterns of Victimization in Dyadic Relationships. The Journal of Applied
Psychology, 89, 1023–1034. https://doi.org/10.1037/0021-9010.89.6.1023
Auffhammer, M., & Kellogg, R. (2011). Clearing the air? The effects of gasoline content
regulation on air quality. American Economic Review, 101(6), 2687–2722.
Axelrod, R., & Hamilton, W. D. (1981). The evolution of cooperation. Science, 211(4489),
1390–1396.
Bainbridge, W. S. (2007). The scientific research potential of virtual worlds. Science, 317(5837),
472–476.
Ballard, M. E., & Welch, K. M. (2017). Virtual warfare: Cyberbullying and cyber-victimization
in MMOG play. Games and Culture, 12(5), 466–491.
Bandura, A. (2001). Social cognitive theory of mass communication. Media Psychology, 3(3),
265–299.
Barabási, A.-L., & Albert, R. (1999). Emergence of scaling in random networks. Science,
286(5439), 509–512.
Barlett, C. P., & Anderson, C. A. (2012). Examining media effects: The general aggression and
general learning models. The International Encyclopedia of Media Studies.
Barlett, C., & Coyne, S. M. (2014). A meta-analysis of sex differences in cyber-bullying
behavior: The moderating role of age. Aggressive Behavior, 40(5), 474–488.

95
Barlett, C. P., Roth, B. R., & Rinker, A. M. (2022). Muscles, popularity, social capital, and
computer skills: Examining “power” in cyberbullying. Aggressive Behavior, 48(6), 608–
615.
Barnett, M., & Duvall, R. (2004). Power in global governance (Vol. 98). Cambridge University
Press.
Bastiaensens, S., Vandebosch, H., Poels, K., Van Cleemput, K., DeSmet, A., & De
Bourdeaudhuij, I. (2014). Cyberbullying on social network sites. An experimental study
into bystanders’ behavioural intentions to help the victim or reinforce the bully.
Computers in Human Behavior, 31, 259–271.
Bell, M., Perera, S., Piraveenan, M., Bliemer, M., Latty, T., & Reid, C. (2017). Network growth
models: A behavioural basis for attachment proportional to fitness. Scientific Reports,
7(1), 42431.
Blackburn, J., & Kwak, H. (2014). STFU NOOB! predicting crowdsourced decisions on toxic
behavior in online games. Proceedings of the 23rd International Conference on World
Wide Web, 877–888.
Bonanno, G. A., & Jost, J. T. (2006). Conservative shift among high-exposure survivors of the
September 11th terrorist attacks. Basic and Applied Social Psychology, 28(4), 311–323.
Borgatti, S. P., Everett, M. G., & Freeman, L. C. (1997). UCINET 5.0 for Windows. Columbia,
SC: Analytic Technologies. Inc.
Bowlby, J. (1982). Attachment and loss: Retrospect and prospect. American Journal of
Orthopsychiatry, 52(4), 664.
Bowman, N. D., Ahn, S. J., & Mercer Kollar, L. M. (2020). The paradox of interactive media:
The potential for video games and virtual reality as tools for violence prevention.
Frontiers in Communication, 104.
Boyd, R., & Richerson, P. J. (1994). The evolution of norms: An anthropological view. Journal
of Institutional and Theoretical Economics (JITE)/Zeitschrift Für Die Gesamte
Staatswissenschaft, 72–87.

96
Brodeur, A., Clark, A. E., Fleche, S., & Powdthavee, N. (2021). COVID-19, lockdowns and
well-being: Evidence from Google Trends. Journal of Public Economics, 193, 104346.
Brown, S. L., Brown, R. M., House, J. S., & Smith, D. M. (2008). Coping with spousal loss:
Potential buffering effects of self-reported helping behavior. Personality and Social
Psychology Bulletin, 34(6), 849–861.
Buckley, K. E., & Anderson, C. A. (2006). A theoretical model of the effects and consequences
of playing video games. Playing Video Games: Motives, Responses, and Consequences,
363–378.
Bucur, D. (2019). Gender homophily in online book networks. Information Sciences, 481, 229–
243.
Burkhardt, M. E., & Brass, D. J. (1990). Changing patterns or patterns of change: The effects of
a change in technology on social network structure and power. Administrative Science
Quarterly, 104–127.
Burt, R. S. (1992). Structure holes. Harvard University Press, Cambridge, MA.
Burt, R. S. (2001). Closure as social capital. Social Capital: Theory and Research, 31–55.
Burt, R. S. (2009). Structural holes: The social structure of competition. Harvard university
press.
Burt, R. S. (2012). Network-related personality and the agency question: Multirole evidence
from a virtual world. American Journal of Sociology, 118(3), 543–591.
Burt, R. S. (2017). Structural holes versus network closure as social capital. Social Capital, 31–
56.
Byrne, D. E. (1971). The attraction paradigm (Vol. 462). Academic press.
Cairns, P., Power, C., Barlet, M., Haynes, G., Kaufman, C., & Beeston, J. (2021). Enabled
players: The value of accessible digital games. Games and Culture, 16(2), 262–282.

97
Callaway, B., & Sant’Anna, P. H. (2021). Difference-in-differences with multiple time
periods. Journal of Econometrics, 225(2), 200-230.
Canossa, A., Salimov, D., Azadvar, A., Harteveld, C., & Yannakakis, G. (2021). For Honor, for
Toxicity: Detecting Toxic Behavior through Gameplay. Proceedings of the ACM on
Human-Computer Interaction, 5(CHI PLAY), 1–29.
Castano, E., Yzerbyt, V., & Bourguignon, D. (2003). We are one and I like it: The impact of
ingroup entitativity on ingroup identification. European Journal of Social Psychology,
33(6), 735–754.
Chan, H. C. O., & Wong, D. S. (2015). The overlap between school bullying perpetration and
victimization: Assessing the psychological, familial, and school factors of Chinese
adolescents in Hong Kong. Journal of Child and Family Studies, 24, 3224–3234.
Chen, S., Lee-Chai, A. Y., & Bargh, J. A. (2001). Relationship orientation as a moderator of the
effects of social power. Journal of Personality and Social Psychology, 80(2), 173.
Chen, Z., Oh, P., & Chen, A. (2021). The role of online media in mobilizing large-scale
collective action. Social Media+ Society, 7(3), 20563051211033810.
Cheng, J., Bernstein, M., Danescu-Niculescu-Mizil, C., & Leskovec, J. (2017). Anyone can
become a troll: Causes of trolling behavior in online discussions. Proceedings of the 2017
ACM Conference on Computer Supported Cooperative Work and Social Computing,
1217–1230.
Chesney, T., Coyne, I., Logan, B., & Madden, N. (2009). Griefing in virtual worlds: Causes,
casualties and coping strategies. Information Systems Journal, 19(6), 525–548.
Cho, J., Ahmed, S., Keum, H., Choi, Y. J., & Lee, J. H. (2018). Influencing myself: Self-
reinforcement through online political expression. Communication Research, 45(1), 83–
111.
Choi, Y.-J., Jeon, B.-J., & Kim, H.-W. (2021). Identification of key cyberbullies: A text mining
and social network analysis approach. Telematics and Informatics, 56, 101504.

98
Cohen, L. E., & Felson, M. (2010). Social change and crime rate trends: A routine activity
approach (1979). In Classics in environmental criminology (pp. 203–232). Routledge.
Cook, K. S., & Emerson, R. M. (1978). Power, Equity and Commitment in Exchange Networks.
American Sociological Review, 43(5), 721–739.
DeCamp, W., & Newby, B. (2015). From bullied to deviant: The victim–offender overlap among
bullying victims. Youth Violence and Juvenile Justice, 13(1), 3–17.
De La Haye, K., Robins, G., Mohr, P., & Wilson, C. (2011). How physical activity shapes, and is
shaped by, adolescent friendships. Social Science & Medicine, 73(5), 719–728.
De Vries, H., Candel, M., Engels, R., & Mercken, L. (2006). Challenges to the peer influence
paradigm: Results for 12–13 year olds from six European countries from the European
Smoking Prevention Framework Approach study. Tobacco Control, 15(2), 83–89.
Dibbell, J. (1994). A rape in cyberspace or how an evil clown, a Haitian trickster spirit, two
wizards, and a cast of dozens turned a database into a society. Ann. Surv. Am. L., 471.
Dimick, J. B., & Ryan, A. M. (2014). Methods for evaluating changes in health care policy: The
difference-in-differences approach. Jama, 312(22), 2401–2402.
Dredge, R., Gleeson, J., & De la Piedad Garcia, X. (2014). Presentation on Facebook and risk of
cyberbullying victimisation. Computers in Human Behavior, 40, 16–22.
Entertainment Software Association. (2022). Essential facts about the video game industry.
Retrieved from https://www.theesa.com/resource/2022-essential-facts-about-the-video-
game-industry/
Eisenberg, M. E., Gower, A. L., McMorris, B. J., & Bucchianeri, M. M. (2015). Vulnerable
bullies: Perpetration of peer harassment among youths across sexual orientation, weight,
and disability status. American Journal of Public Health, 105(9), 1784–1791.
Ellwardt, L., Labianca, G. J., & Wittek, R. (2012). Who are the objects of positive and negative
gossip at work?: A social network perspective on workplace gossip. Social Networks,
34(2), 193–205.

99
Epley, N., Keysar, B., Van Boven, L., & Gilovich, T. (2004). Perspective taking as egocentric
anchoring and adjustment. Journal of Personality and Social Psychology, 87(3), 327.
Eränen, L., & Liebkind, K. (1993). Coping with disaster: The helping behavior of communities
and individuals. International handbook of traumatic stress syndromes, 957-964.
Escortell, R., Aparisi, D., Martínez-Monteagudo, M. C., & Delgado, B. (2020). Personality traits
and aggression as explanatory variables of cyberbullying in Spanish preadolescents.
International Journal of Environmental Research and Public Health, 17(16), 5705.
Evans, C. B., & Smokowski, P. R. (2016). Theoretical explanations for bullying in school: How
ecological processes propagate perpetration and victimization. Child and Adolescent
Social Work Journal, 33, 365–375.
Festinger, L., Schachter, S., & Back, K. (1950). Social pressures in informal groups; a study of
human factors in housing.
Festl, R. (2016). Perpetrators on the internet: Analyzing individual and structural explanation
factors of cyberbullying in school context. Computers in Human Behavior, 59, 237–248.
https://doi.org/10.1016/j.chb.2016.02.017
Festl, R., & Quandt, T. (2013). Social relations and cyberbullying: The influence of individual
and structural attributes on victimization and perpetration via the internet. Human
Communication Research, 39(1), 101–126.
Finke, E. H., Hickerson, B. D., & Kremkow, J. M. (2018). “To be quite honest, if it wasn’t for
videogames I wouldn’t have a social life at all”: Motivations of young adults with autism
spectrum disorder for playing videogames as leisure. American Journal of Speech-
Language Pathology, 27(2), 672–689.
Flippen, A. R., Hornstein, H. A., Siegal, W. E., & Weitzman, E. A. (1996). A comparison of
similarity and interdependence as triggers for in-group formation. Personality and Social
Psychology Bulletin, 22(9), 882-893.
Flyverbom, M., Christensen, L. T., & Hansen, H. K. (2015). The transparency–power nexus:
Observational and regularizing control. Management Communication Quarterly, 29(3),
385–410.

100
Fox, J., & Tang, W. Y. (2017). Women’s experiences with general and sexual harassment in
online video games: Rumination, organizational responsiveness, withdrawal, and coping
strategies. New Media & Society, 19(8), 1290–1307.
Friedkin, N. E. (1998). A structural theory of social influence. Cambridge University Press.
Friemel, T. N. (2015). Influence versus selection: A network perspective on opinion leadership.
International Journal of Communication, 9, 1002–1022.
Friedkin, N. E. (2001). Norm formation in social influence networks. Social Networks, 23(3),
167–189.
Friemel, T. N. (2021). Co-orientation of media use: Studying selection and influence processes
in social networks to link micro behavior of TV and YouTube use to meso-level
structures. Communication Methods and Measures, 15(4), 312–331.
Fu, J. S. (2019). Unpacking the influence of informational, organizational, and structural factors
on the longitudinal change of the NPO follower-followee network on Twitter.
International Journal of Communication, 13, 24.
Fulk, J. (1993). Social construction of communication technology. Academy of Management
Journal, 36(5), 921–950.
Gaertner, S. L., Dovidio, J. F., Banker, B. S., Houlette, M., Johnson, K. M., & McGlynn, E. A.
(2000). Reducing intergroup conflict: From superordinate goals to decategorization,
recategorization, and mutual differentiation. Group Dynamics: Theory, Research, and
Practice, 4(1), 98.
Games for Change. (2020, July). Past Winners. Retrieved from https://g4cawards.secure-
platform.com/a/page/past_winners
Ganley, D., & Lampe, C. (2009). The ties that bind: Social network principles in online
communities. Decision Support Systems, 47(3), 266–274.

101
Gardner, M. (2021, December 3). Why Games Are Still Letting Down People With Disabilities.
Forbes. Retrieved from https://www.forbes.com/sites/mattgardner1/2021/12/03/why-
games-are-still-letting-down-people-with-disabilities/?sh=14cc7f707be0
Grandprey-Shores, K., He, Y., Swanenburg, K. L., Kraut, R., & Riedl, J. (2014). The
identification of deviance and its impact on retention in a multiplayer game. Proceedings
of the 17th ACM Conference on Computer Supported Cooperative Work & Social
Computing, 1356–1365.
Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6),
1360–1380.
Gray, K. L. (2012). Deviant bodies, stigmatized identities, and racist acts: Examining the
experiences of African-American gamers in Xbox Live. New Review of Hypermedia and
Multimedia, 18(4), 261–276.
Greitemeyer, T., & Osswald, S. (2009). Prosocial video games reduce aggressive cognitions.
Journal of Experimental Social Psychology, 45(4), 896–900.
Greitemeyer, T., Osswald, S., & Brauer, M. (2010). Playing prosocial video games increases
empathy and decreases schadenfreude. Emotion, 10(6), 796.
Hajibagheri, A., Lakkaraju, K., Sukthankar, G., Wigand, R. T., & Agarwal, N. (2015). Conflict
and communication in massively-multiplayer online games. Social Computing,
Behavioral-Cultural Modeling, and Prediction: 8th International Conference, SBP 2015,
Washington, DC, USA, March 31-April 3, 2015. Proceedings 8, 65–74.
Halbrook, Y. J., O’Donnell, A. T., & Msetfi, R. M. (2019). When and How Video Games Can
Be Good: A Review of the Positive Effects of Video Games on Well-Being. Perspectives
on Psychological Science, 14(6), 1096–1104.
Hannah, M. E., & Midlarsky, E. (2005). Helping by siblings of children with mental retardation.
American Journal on Mental Retardation, 110(2), 87–99.
Haynie, D. L., Nansel, T., Eitel, P., Crump, A. D., Saylor, K., Yu, K., & Simons-Morton, B.
(2001). Bullies, victims, and bully/victims: Distinct groups of at-risk youth. The Journal
of Early Adolescence, 21(1), 29–49.

102
Hilvert-Bruce, Z., & Neill, J. T. (2020). I’m just trolling: The role of normative beliefs in
aggressive behaviour in online gaming. Computers in Human Behavior, 102, 303–311.
Hirschberger, G. (2018). Collective trauma and the social construction of meaning. Frontiers in
Psychology, 9, 1441.
Hu, H.-B., & Wang, X.-F. (2009). Disassortative mixing in online social networks. Europhysics
Letters, 86(1), 18003.
Huizinga, J. (1949). Homo ludens: A study of the play element in culture London. Citado en
Nobert Elias & Eric Dunning: op. cit.
Kaniasty, K., & Norris, F. H. (1995). In search of altruistic community: Patterns of social support
mobilization following Hurricane Hugo. American journal of community
psychology, 23(4), 447-477.
Katz, J. (1994). Geographical proximity and scientific collaboration. Scientometrics, 31(1), 31–
43.
Keele, L., Stevenson, R. T., & Elwert, F. (2020). The causal interpretation of estimated
associations in regression models. Political Science Research and Methods, 8(1), 1-13.
Keltner, D., Gruenfeld, D. H., & Anderson, C. (2003). Power, approach, and inhibition.
Psychological Review, 110(2), 265.
Kerr, C. (2020, October). Sky: Children of the Light has topped 50 million installs worldwide.
Retrieved from https://www.gamedeveloper.com/business/-i-sky-children-of-the-light-i-
has-topped-50-million-installs-worldwide
Kim, S. S., Huang-Isherwood, K. M., Zheng, W., & Williams, D. (2022). The art of being
together: How group play can increase reciprocity, social capital, and social status in a
multiplayer online game. Computers in Human Behavior, 133, 107291.
Kokkinos, C. M., & Saripanidis, I. (2017). A lifestyle exposure perspective of victimization
through Facebook among university students. Do individual differences matter?
Computers in Human Behavior, 74, 235–245.

103
Kordyaka, B., Jahn, K., & Niehaves, B. (2020). Towards a unified theory of toxic behavior in
video games. Internet Research.
Kowert, R. (2020). Dark participation in games. Frontiers in Psychology, 11, 2969.
Kwak, H., Blackburn, J., & Han, S. (2015). Exploring cyberbullying and other toxic behavior in
team competition online games. Proceedings of the 33rd Annual ACM Conference on
Human Factors in Computing Systems, 3739–3748.
Lai, C.-H., She, B., & Ye, X. (2019). Unpacking the network processes and outcomes of online
and offline humanitarian collaboration. Communication Research, 46(1), 88–116.
Lamertz, K., & Aquino, K. (2004). Social Power, Social Status and Perceptual Similarity of
Workplace Victimization: A Social Network Analysis of Stratification. Human Relations,
57(7), 795–822.
Laurie-ann, M. H., Crespi, I., Hendry, B., & Fermani, A. (2021). Extending the Current
Theorization on Cyberbullying: Importance of Including Socio-Psychological
Perspectives. Italian Journal of Sociology of Education, 13(3).
Lazer, D. (2001). The co-evolution of individual and network. Journal of Mathematical
Sociology, 25(1), 69–108.
Lazer, D., Rubineau, B., Chetkovich, C., Katz, N., & Neblo, M. (2010). The coevolution of
networks and political attitudes. Political Communication, 27(3), 248–274.
Lee, S., Chung, J. E., & Park, N. (2018). Network environments and well-being: An examination
of personal network structure, social capital, and perceived social support. Health
Communication, 33(1), 22–31.
Lemercier-Dugarin, M., Romo, L., Tijus, C., & Zerhouni, O. (2021). “Who Are the Cyka Blyat?”
How Empathy, Impulsivity, and Motivations to Play Predict Aggressive Behaviors in
Multiplayer Online Games. Cyberpsychology, Behavior, and Social Networking, 24(1),
63–69.

104
Levine, M., & Thompson, K. (2004). Identity, place, and bystander intervention: Social
categories and helping after natural disasters. The Journal of Social Psychology, 144(3),
229–245.
Litsa, M., & Bekiari, A. (2022). The powerful, the powerless, and the empowered: Visualizations
of power in high school and university through social network analysis. Power and
Education, 14(3), 262–281.
Lomi, A., Snijders, T. A., Steglich, C. E., & Torló, V. J. (2011). Why are some more peer than
others? Evidence from a longitudinal study of social networks and individual academic
performance. Social Science Research, 40(6), 1506–1520.
Louch, H. (2000). Personal network integration: Transitivity and homophily in strong-tie
relations. Social Networks, 22(1), 45–64.
Marchiondo, L. A., Fisher, G. G., Cortina, L. M., & Matthews, R. A. (2020). Disrespect at Work,
Distress at Home: A Longitudinal Investigation of Incivility Spillover and Crossover
Among Older Workers. Work, Aging and Retirement.
McLean, L., & Griffiths, M. D. (2019). Female gamers’ experience of online harassment and
social support in online gaming: A qualitative study. International Journal of Mental
Health and Addiction, 17(4), 970–994.
McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a feather: Homophily in social
networks. Annual Review of Sociology, 27(1), 415–444.
Mercken, L., Snijders, T. A., Steglich, C., Vartiainen, E., & De Vries, H. (2010). Dynamics of
adolescent friendship networks and smoking behavior. Social Networks, 32(1), 72–81.
Merton, R. K. (1968). The Matthew effect in science: The reward and communication systems of
science are considered. Science, 159(3810), 56–63.
Metacritic. (2019, July). Sky: Children of the Light. Retrieved from
https://www.metacritic.com/game/ios/sky-children-of-the-light

105
Midlarsky, E. (1994). Altruism through the life course. Family Caregiving across the Lifespan,
69–95.
Miga, E. M., Hare, A., Allen, J. P., & Manning, N. (2010). The relation of insecure attachment
states of mind and romantic attachment styles to adolescent aggression in romantic
relationships. Attachment & Human Development, 12(5), 463–481.
Monge, P. R., Contractor, N. S., & Contractor, P. S. (2003). Theories of communication
networks. Oxford University Press, USA.
Mouttapa, M., Valente, T., Gallaher, P., Rohrbach, L. A., & Unger, J. B. (2004). Social network
predictors of bullying and victimization. Adolescence, 39(154), 315.
Moya, A. (2018). Violence, psychological trauma, and risk attitudes: Evidence from victims of
violence in Colombia. Journal of Development Economics, 131, 15–27.
Mumby, D. K. (2015). Organizing power. Review of Communication, 15(1), 19–38.
Napier, J. L., Mandisodza, A. N., Andersen, S. M., & Jost, J. T. (2006). System justification in
responding to the poor and displaced in the aftermath of Hurricane Katrina. Analyses of
Social Issues and Public Policy, 6(1), 57-73.
Nebus, J. (2006). Building collegial information networks: A theory of advice network
generation. Academy of Management Review, 31(3), 615–637.
Neto, J. A., Yokoyama, K. M., & Becker, K. (2017). Studying toxic behavior influence and
player chat in an online video game. Proceedings of the International Conference on Web
Intelligence, 26–33.
Newman, M. E. (2001). Clustering and preferential attachment in growing networks. Physical
Review E, 64(2), 025102.
Newman, M. E. (2002). Assortative mixing in networks. Physical Review Letters, 89(20),
208701.

106
Newman, M. E. (2003). Mixing patterns in networks. Physical Review E, 67(2), 026126.
Newman, M. E. (2004). Coauthorship networks and patterns of scientific collaboration.
Proceedings of the National Academy of Sciences, 101(suppl_1), 5200–5205.
Norris, F. H., Baker, C. K., Murphy, A. D., & Kaniasty, K. (2005). Social support mobilization
and deterioration after Mexico’s 1999 flood: Effects of context, gender, and time.
American Journal of Community Psychology, 36(1), 15–28.
Norris, F. H., Friedman, M. J., Watson, P. J., Byrne, C. M., Diaz, E., & Kaniasty, K. (2002).
60,000 disaster victims speak: Part I. An empirical review of the empirical literature,
1981—2001. Psychiatry, 65(3), 207-239.
Obadimu, A., Khaund, T., Mead, E., Marcoux, T., & Agarwal, N. (2021). Developing a socio-
computational approach to examine toxicity propagation and regulation in COVID-19
discourse on YouTube. Information Processing & Management, 58(5), 102660.
Olsen, P. I., Prenkert, F., Hoholm, T., & Harrison, D. (2014). The dynamics of networked power
in a concentrated business network. Journal of Business Research, 67(12), 2579–2589.
Olweus, D. (1997). Tackling peer victimization with a school-based intervention program.
Cultural Variation in Conflict Resolution: Alternatives to Violence, 215–231.
Olweus, D. (2001). Peer harassment: A critical analysis and some important issues. Peer
Harassment in School: The Plight of the Vulnerable and Victimized, 3–20.
Ortiz, S. M. (2019). “You can say I got desensitized to it”: How men of color cope with everyday
racism in online gaming. Sociological Perspectives, 62(4), 572–588.
Pahayahay, A., & Khalili-Mahani, N. (2020). What Media Helps, What Media Hurts: A Mixed
Methods Survey Study of Coping with COVID-19 Using the Media Repertoire
Framework and the Appraisal Theory of Stress. Journal of Medical Internet Research,
22(8), e20186.
Paul, H. L., Bowman, N. D., & Banks, J. (2015). The enjoyment of griefing in online games.
Journal of Gaming & Virtual Worlds, 7(3), 243–258.

107
Pauksztat, B., & Salin, D. (2021). Workplace bullying: A social network perspective. Concepts,
Approaches and Methods, 449–484.
Pellegrini, A. D., Bartini, M., & Brooks, F. (1999). School bullies, victims, and aggressive
victims: Factors relating to group affiliation and victimization in early adolescence.
Journal of Educational Psychology, 91(2), 216.
Peluchette, J. V., Karl, K., Wood, C., & Williams, J. (2015). Cyberbullying victimization: Do
victims’ personality and risky social network behaviors contribute to the problem?
Computers in Human Behavior, 52, 424–435.
Penner, L. A., Dovidio, J. F., Piliavin, J. A., & Schroeder, D. A. (2005). Prosocial behavior:
Multilevel perspectives. Annu. Rev. Psychol., 56, 365–392.
Postmes, T., Spears, R., & Lea, M. (1998). Breaching or building social boundaries? SIDE-
effects of computer-mediated communication. Communication Research, 25(6), 689–
715.
Prentice-Dunn, S., & Rogers, R. W. (1982). Effects of public and private self-awareness on
deindividuation and aggression. Journal of Personality and Social Psychology, 43(3),
503.
Putnam, R. D. (1966). Political attitudes and the local community. American Political Science
Review, 60(3), 640-654.
Putnam, R. D. (2000). Bowling alone: The collapse and revival of American community. Simon
and schuster.
Ragan, D. T., Osgood, D. W., Ramirez, N. G., Moody, J., & Gest, S. D. (2019). A comparison of
peer influence estimates from SIENA stochastic actor–based models and from
conventional regression approaches. Sociological Methods & Research,
0049124119852369.
Reiss Jr, A. J. (1981). Towards a revitalization of theory and research on victimization by crime.
J. Crim. L. & Criminology, 72, 704.

108
Rimal, R. N., & Lapinski, M. K. (2015). A re-explication of social norms, ten years later.
Communication Theory, 25(4), 393–409.
Rogers, E. M. (1995). Diffusion of Innovations: Modifications of a model for
telecommunications. In Die diffusion von innovationen in der telekommunikation (pp.
25–38). Springer.
Rogers, E. M. (2010). Diffusion of innovations. Simon and Schuster.
Roseman, I. J. (1996). Appraisal determinants of emotions: Constructing a more accurate and
comprehensive theory. Cognition & Emotion, 10(3), 241–278.
Roseman, I. J. (2001). A model of appraisal in the emotion system. Appraisal Processes in
Emotion: Theory, Methods, Research, 68–91.
Saffer, A. J., Pilny, A., & Sommerfeldt, E. J. (2022). What influences relationship formation in a
global civil society network? An examination of valued multiplex relations.
Communication Research, 49(5), 703–732.
Saffer, A. J., Yang, A., & Taylor, M. (2018). Reconsidering power in multistakeholder
relationship management. Management Communication Quarterly, 32(1), 121–139.
Sant’Anna, P. H., & Zhao, J. (2020). Doubly robust difference-in-differences estimators. Journal
of Econometrics, 219(1), 101-122.
Schwartz, D., Proctor, L. J., & Chien, D. H. (2001). The aggressive victim of bullying. Peer
Harassment in School: The Plight of the Vulnerable and Victimized, 147–174.
Sgourev, S. V. (2013). How Paris gave rise to Cubism (and Picasso): Ambiguity and
fragmentation in radical innovation. Organization Science, 24(6), 1601–1617.
Shalizi, C. R., & Thomas, A. C. (2011). Homophily and contagion are generically confounded in
observational social network studies. Sociological Methods & Research, 40(2), 211–239.

109
Shanley P. (2020a, April). U.S. Video Game Usage Up 45 Percent, Nielsen Reports. Retrieved
from https://www.hollywoodreporter.com/business/digital/us-video-game-usage-up-45-
percent-nielsen-reports-1288738/
Shanley P. (2020b, March). Gaming Usage Up 75 Percent Amid Coronavirus Outbreak, Verizon
Reports. Retrieved from https://www.hollywoodreporter.com/news/general-
news/gaming-usage-up-75-percent-coronavirus-outbreak-verizon-reports-1285140/
Shen, C., Monge, P., & Williams, D. (2014). Virtual brokerage and closure: Network structure
and social capital in a massively multiplayer online game. Communication Research,
41(4), 459–480.
Shen, C., Sun, Q., Kim, T., Wolff, G., Ratan, R., & Williams, D. (2020). Viral vitriol: Predictors
and contagion of online toxicity in World of Tanks. Computers in Human Behavior, 108,
106343.
Shepherd, H. R. (2017). The structure of perception: How networks shape ideas of norms.
Sociological Forum, 32(1), 72–93.
Skitka, L. J., Bauman, C. W., & Mullen, E. (2004). Political tolerance and coming to
psychological closure following the September 11, 2001, terrorist attacks: An integrative
approach. Personality and Social Psychology Bulletin, 30(6), 743–756.
Statista. (2021). Number of video gamers worldwide from 2014 to 2025 (in billions). Retrieved
from https://www.statista.com/statistics/293304/number-video-
gamers/#:~:text=In%202021%2C%20there%20were%20almost,billion%20gamers%20ac
ross%20the%20globe.
Shumate, M., & Lipp, J. (2008). Connective collective action online: An examination of the
hyperlink network structure of an NGO issue network. Journal of Computer-Mediated
Communication, 14(1), 178–201.
Snijders, T. A., Steglich, C. E., Schweinberger, M., & Huisman, M. (2007). Manual for SIENA
version 3.1. University of Groningen: ICS/Department of Sociology.
Snijders, T. A., Van de Bunt, G. G., & Steglich, C. E. (2010). Introduction to stochastic actor-
based models for network dynamics. Social Networks, 32(1), 44–60.

110
Southwick, S. M., Vythilingam, M., & Charney, D. S. (2005). The psychobiology of depression
and resilience to stress: Implications for prevention and treatment. Annu. Rev. Clin.
Psychol., 1, 255–291.
Stark, T. H., Rambaran, J. A., & McFarland, D. A. (2020). The meeting of minds: Forging social
and intellectual networks within universities. Sociological Science, 7, 433–464.
Staub, E. (2005). The roots of goodness: The fulfillment of basic human needs and the
development of caring, helping and nonaggression, inclusive caring, moral courage,
active bystandership, and altruism born of suffering.
Staub, E., & Ervin, S. (2003). The psychology of good and evil: Why children, adults, and groups
help and harm others. Cambridge University Press.
Staub, E., & Vollhardt, J. (2008). Altruism born of suffering: The roots of caring and helping
after victimization and other trauma. American Journal of Orthopsychiatry, 78(3), 267–
280.
Stürmer, S., Snyder, M., & Omoto, A. M. (2005). Prosocial emotions and helping: The
moderating role of group membership. Journal of Personality and Social Psychology,
88(3), 532.
Suler, J. (2004). The online disinhibition effect. Cyberpsychology & behavior, 7(3), 321-326.
Sun, Q., & Shen, C. (2021). Who would respond to A troll? A social network analysis of
reactions to trolls in online communities. Computers in Human Behavior, 121, 106786.
SuperData. (2020). Worldwide digital games market: November 2020. Retrieved from
https://www.superdataresearch.com/blog/worldwide-digital-games-market
Takahashi, D. (2020). How Wargaming enlisted 160 million World of Tanks players over a
decade. Retrieved from https://venturebeat.com/2020/08/15/how-wargaming-racked-up-
200-million-players-over-a-decade-of-world-of-tanks/

111
Tang, W. Y., Reer, F., & Quandt, T. (2020). Investigating sexual harassment in online video
games: How personality and context factors are related to toxic sexual behaviors against
fellow players. Aggressive Behavior, 46(1), 127–135.
Teng, C.-I., Tseng, F.-C., Chen, Y.-S., & Wu, S. (2012). Online gaming misbehaviours and their
adverse impact on other gamers. Online Information Review, 36(3), 342–358.
Traud, A. L., Mucha, P. J., & Porter, M. A. (2012). Social structure of facebook networks.
Physica A: Statistical Mechanics and Its Applications, 391(16), 4165–4180.
Trepte, S., & Loy, L. S. (2017). Social identity theory and self‐categorization theory. The
International Encyclopedia of Media Effects, 1-13.
Trope, Y., & Liberman, N. (2003). Temporal construal. Psychological Review, 110(3), 403.
Türkay, S., Formosa, J., Adinolf, S., Cuthbert, R., & Altizer, R. (2020). See No Evil, Hear No
Evil, Speak No Evil: How Collegiate Players Define, Experience and Cope with Toxicity.
In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems
(pp. 1–13). Association for Computing Machinery.
Turner, J. C., Hogg, M. A., Oakes, P. J., Reicher, S. D., & Wetherell, M. S. (1987).
Rediscovering the social group: A self-categorization theory. basil Blackwell.
Unnever, J. D. (2005). Bullies, aggressive victims, and victims: Are they distinct groups?
Aggressive Behavior: Official Journal of the International Society for Research on
Aggression, 31(2), 153–171.
Valente, T. W. (1996). Social network thresholds in the diffusion of innovations. Social
Networks, 18(1), 69–89.
Valente, T. W. (2012). Network interventions. Science, 337(6090), 49–53.
Valente, T. W., Pitts, S., Wipfli, H., & Yon, G. G. V. (2019). Network influences on policy
implementation: Evidence from a global health treaty. Social Science & Medicine, 222,
188–197.

112
Van Lange, P. A., Klapwijk, A., & Van Munster, L. M. (2011). How the shadow of the future
might promote cooperation. Group Processes & Intergroup Relations, 14(6), 857–870.
Varghese, M. E., & Pistole, M. C. (2017). College student cyberbullying: Self-esteem,
depression, loneliness, and attachment. Journal of College Counseling, 20(1), 7–21.
Viana, R. B., & de Lira, C. A. B. (2020). Exergames as coping strategies for anxiety disorders
during the COVID-19 quarantine period. Games for Health Journal, 9(3), 147–149.
Vilanova, F., Beria, F. M., Costa, Â. B., & Koller, S. H. (2017). Deindividuation: From Le Bon
to the social identity model of deindividuation effects. Cogent Psychology, 4(1),
1308104.
Vollhardt, J. R. (2009). Altruism born of suffering and prosocial behavior following adverse life
events: A review and conceptualization. Social Justice Research, 22(1), 53–97.
Wegge, D., Van Cleemput, K., Vandebosch, H., & Eggermont, S. (2013). Setting an agenda for
future research into cyberbullying using social network analysis.
Wegge, D., Vandebosch, H., & Eggermont, S. (2014). Who bullies whom online: A social
network analysis of cyberbullying in a school context. Communications, 39(4), 415–433.
Wegge, D., Vandebosch, H., Eggermont, S., & Walrave, M. (2015). The strong, the weak, and
the unbalanced: The link between tie strength and cyberaggression on a social network
site. Social Science Computer Review, 33(3), 315–342.
Wejnert, B. (2002). Integrating models of diffusion of innovations: A conceptual framework.
Annual Review of Sociology, 28(1), 297–326.
Williams, D. (2006). On and off the’Net: Scales for social capital in an online era. Journal of
Computer-Mediated Communication, 11(2), 593–628.
Williams, D. (2010). The mapping principle, and a research framework for virtual worlds.
Communication Theory, 20(4), 451–470.

113
Witkowski, W. (2021, April 4). Videogames are a bigger industry than movies and North
American sports combined, thanks to the pandemic. MarketWatch.
https://www.marketwatch.com/story/videogames-are-a-bigger-industry-than-sports-and-
movies-combined-thanks-to-the-pandemic-11608654990
Wood, J. J., Lynne-Landsman, S. D., Langer, D. A., Wood, P. A., Clark, S. L., Mark Eddy, J., &
Ialongo, N. (2012). School attendance problems and youth psychopathology: Structural
cross-lagged regression models in three longitudinal data sets. Child Development, 83(1),
351–366.
Yokotani, K., & Takano, M. (2021). Social contagion of cyberbullying via online perpetrator and
victim networks. Computers in Human Behavior, 119, 106719.
Young, L. E., & Fujimoto, K. (2021). The Co-evolution of online social networks and syphilis
incidence among young black men who have sex with men. Social Science & Medicine,
272, 113764.
Yuan, Y. C., & Gay, G. (2006). Homophily of network ties and bonding and bridging social
capital in computer-mediated distributed teams. Journal of Computer-Mediated
Communication, 11(4), 1062–1084.
Zhang, C., Bu, Y., Ding, Y., & Xu, J. (2018). Understanding scientific collaboration:
Homophily, transitivity, and preferential attachment. Journal of the Association for
Information Science and Technology, 69(1), 72–86.
Zoellner, T., & Maercker, A. (2006). Posttraumatic growth in clinical psychology—A critical
review and introduction of a two component model. Clinical Psychology Review, 26(5),
626–653.

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Crisis and stasis: understanding change in online communities

PDF

The evolution of multilevel organizational networks in an online gaming community

PDF

The formation and influence of online health social networks on social support, self-tracking behavior and weight loss outcomes

PDF

Gamification + HCI + CMC: effects of persuasive video games on consumers’ mental and physical health

PDF

Modeling social and cognitive aspects of user behavior in social media

PDF

Civic games with 'local fit': embedding with real‐world neighborhoods and place‐based networks

PDF

We're all in this (game) together: transactive memory systems, social presence, and social information processing in video game teams

PDF

The evolution of knowledge creation online: what is driving the dynamics of Wikipedia networks

PDF

Relationship formation and information sharing to promote risky health behavior on social media

PDF

Cooperation in wireless networks with selfish users

PDF

Contagious: social norms about health in work group networks

PDF

Ecology and network evolution in online innovation contest crowdsourcing

PDF

Network influences in health initiatives: multimedia games for youth in Peru

PDF

The patterns, effects and evolution of player social networks in online gaming communities

PDF

Game recognize game: performative archives and alternate reality games

PDF

Working through death and grief with a video game: the design and development of Where the sea meets the sky

PDF

Gathering understanding: players, systems and developers in Magic: the gathering

PDF

Identity, trust, and credibility online: evaluating contradictory user-generated information via the warranting principle

PDF

Social capital and community philanthropy: the impact of social trust and social networks on individual charitable behavior and community foundation development

PDF

The power of social media narratives in raising mental health awareness for anti-stigma campaigns

Asset Metadata

Creator Liu, Mingxuan (author)

Core Title The good, the bad, and the longitudinal: testing dynamic prosocial and toxic behaviors in online commercial games

School Annenberg School for Communication

Degree Doctor of Philosophy

Degree Program Communication

Degree Conferral Date 2023-08

Publication Date 07/10/2023

Defense Date 05/03/2023

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag behavior diffusion,gaming toxicity,inferential network analysis,natural experiment,OAI-PMH Harvest,user behavior

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Miller, Lynn (committee chair)

Creator Email lgmx721@hotmail.com,mliu@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113262904

Unique identifier UC113262904

Identifier etd-LiuMingxua-12038.pdf (filename)

Legacy Identifier etd-LiuMingxua-12038

Document Type Dissertation

Format theses (aat)

Rights Liu, Mingxuan

Internet Media Type application/pdf

Type texts

Source 20230710-usctheses-batch-1063 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu