PUBLIC PERCEPTIONS OF ADOLESCENCE ON SOCIAL MEDIA: A BIG DATA
ANALYSIS OF TWITTER DATA

By
Mingzhang Chen

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Human Development and Family Studies – Doctor of Philosophy
2023



ABSTRACT
Despite the crucial role of adolescence in human development, little research has
explored how the general public perceives this stage of life. Existing research has primarily
examined individual perceptions of adolescence and focused on the negative and stereotyped
perceptions based on the storm-and-stress view of adolescence. The current study employed both
theory-driven and data-driven approaches enabled by machine learning and big data analytics to
understand public perceptions of adolescence on Twitter. I extracted millions of publicly
available tweets that discussed adolescence in 2019 on Twitter, and employed word embeddings,
a neural network algorithm, to analyze the latent meanings of different words in the data as
reflecting public perceptions. I then analyzed the outputs of word embeddings to identify public
perceptions of adolescence for different terms of adolescence as well as different subgroups of
adolescents based on race/ethnicity, gender, and sexual orientation. A unique mixed-methods
analytic approach was developed, in which I first leveraged the quantitative analysis of word
embeddings to generate descriptors that were closely associated with adolescence, and then
conducted qualitative content analysis of the descriptors while referencing the original tweets.
Results suggested that adolescence is a complex and comprehensive construct that cannot be
captured by a single item. While certain terms of adolescence such as teen was perceived as
negative and stereotypical, the overall construct of adolescence was perceived as neutral or
positive. Both similarities and differences existed in the perceptions of different subgroups of
adolescents, with certain issues being identified such as the lack of voices in anti-racism
movements for Asian adolescents. My dissertation can also inform future practices to promote
positive framing of adolescence and create a better digital environment for adolescent
development.



Copyright by
MINGZHANG CHEN
2023



ACKNOWLEDGEMENTS
I would like to thank the following people or organizations for helping me undertake the
dissertation study:
Dr. Wang, my advisor and chair of my committee, for introducing me to the field of
research on adolescence, providing me with essential guidance and support in program training
and research, and helping me finish my dissertation study;
Dr. Johnson, Dr. Qin, and Dr. Peng, my committee members, for providing knowledge
and expertise to help me design, conduct, and refine my dissertation study;
College of Social Science, Department of Human Development of Family Studies,
Graduate School, Michigan State University, for providing me with funding support for the study;
Dr. Vallotton and Dr. Tobe, my program directors, and everyone in the Department of
Human Development of Family Studies, Michigan State University, for supporting and helping
me in the program;
My mom, for supporting me to study abroad and providing me with love and funding.

iv



PREFACE
Acknowledgement of Bias and Harmful Content
This dissertation incorporates content from public posts on Twitter collected through the
Twitter API during the study period. It is important to note that some of the content may contains
bias, offensive language, or harmful depictions, such as stereotyped, racist, sexist language
related to adolescents. I have included these materials with the sole purpose of illustrating the
perceptions and depictions of adolescence on Twitter in 2019. It is crucial to clarify that these
contents do not align with, represent, or reflect my personal perceptions or opinions regarding
adolescence. My intent is to analyze and discuss the external perspectives found on Twitter
during the specified time frame.

v



TABLE OF CONTENTS
CHAPTER 1: INTRODUCTION ................................................................................................... 1
CHAPTER 2: THEORETICAL FRAMEWORKS ......................................................................... 6
CHAPTER 3: LITERATURE REVIEW .......................................................................................16
CHAPTER 4: METHODOLOGY .................................................................................................31
CHAPTER 5: RESULTS ...............................................................................................................44
CHAPTER 6: DISCUSSION ........................................................................................................ 61
REFERENCES ..............................................................................................................................88
APPENDIX 1: TABLES ............................................................................................................. 100
APPENDIX 2: FIGURES ........................................................................................................... 108

vi



CHAPTER 1: INTRODUCTION
Contemporary research generally acknowledges adolescence as a distinct developmental
period that occurs between childhood and adulthood. In the field of developmental science,
adolescence has become a crucial area of study, prompting extensive investigations into its
defining characteristics (see review by Crosnoe & Johnson, 2011). However, the way in which
research views adolescence is not constant but rather evolving (National Academies of Sciences,
Engineering, and Medicine, 2019). The pioneering psychologist credited with initiating the
scientific study of adolescence, Stanley Hall (1904) described adolescence as a period of storm
and stress, seeing conflicts and abnormality as universal and immutable during adolescence. In
contrast, later generations of researchers have revisited this view and proposed more positive
perspectives to understand adolescence (e.g., revisited storm-and stress view that negative
characteristics such as risky behavior is more prevalent but not universal during adolescence,
Arnett, 1999; positive youth development that highlights positive characteristics such as
responsibility and creativity, Damon, 2004). Alongside this revolution in academia, multiple
discernible youth cultures began to emerge in the public fields since the 1950s (e.g., the Mods;
Austin & Willard, 1998), leading to diversified portraits of adolescence capturing the lives of
young people from various sociodemographic backgrounds (e.g., region, race/ethnicity, religion,
etc.; Jeffrey & Dyson, 2008).
Despite the crucial role of adolescence in human development, little research has
explored how the general public perceives this stage of life. Specifically, there is a dearth of
studies that examine the collective attitudes and beliefs about adolescence held by a group of
people, also known as public perceptions of adolescence (Dowler et al., 2006). Informed by
theories highlighting the importance of sociocultural influences on development (e.g.,

1



Bronfenbrenner, 1979), values and beliefs that are pervasive in the larger social contexts (e.g.,
social norm) have a strong impact on adolescent development (Benner et al., 2022; Wang et al.,
2019). Similarly, public perceptions of adolescence could be another meaningful factor in the
sociocultural context that influences adolescent development. Negative and stereotyped
perceptions of adolescence may pose stereotype threat (Steele & Aronson, 1995) to adolescents
and hinder their development, whereas positive perceptions may empower adolescents to
promote their positive development (Benson, 1997).
Although research has not directly examined public perceptions of adolescence, a handful
of studies has investigated individual perceptions of adolescence held by parents and adolescents
themselves as well as their influences on development (e.g., Buchanan & Holmbeck, 1998; Qu et
al., 2016). This research shows that youth whose parents or who themselves hold negative and
stereotyped perceptions of adolescence (e.g., expecting rebelliousness and alienation during
adolescence) tend to have poor developmental outcomes such as increased externalizing
behaviors and parent-child conflicts (Buchanan, 2003; Buchanan & Hughes, 2009). Although
these studies provide initial evidence for the value of investigating perceptions of adolescence,
they have used small community samples to examine individual perceptions that cannot be
immediately generalized to public perceptions. Investigating public perceptions of adolescence
will deepen the scholarly understanding of the broader sociocultural context within which young
people develop. Moreover, by gaining a better understanding of how the public perceives
adolescence, we can take the first step towards shaping public discourse and promoting positive
youth development through effective framing of adolescence in public spaces (Kendall-Taylor,
2020).

2



Social media, an important context for contemporary adolescent development (Uhls et al.,
2017), can serve as a valuable source of information to understand public perceptions of
adolescence. Social media provides the general public with a platform to share information,
experiences, and perspectives around various issues (Weinberg, 2009). The content and discourse
on social media, as well as the values and beliefs associated with such information, have been
found to influence adolescent beliefs and behaviors around sexual development (e.g., Bleakley et
al, 2009) and emotion regulation (e.g., Marino et al., 2020). Among various social media
platforms, Twitter is a popular site where millions of active users post millions of small bursts of
information per day (Twitter, 2022a). National data show that U.S. adolescents between 13 to 17
years old have a high level of exposure to social media platforms, with 23% of them indicating
having used Twitter ever (Vogels et al., 2022). An emerging body of literature has examined
public perceptions and opinions around a variety of topics such as presidential election (Karami
et al., 2018) and mental health (e.g., Pavlova & Berkers, 2022) on Twitter. Similarly, studying
perceptions of adolescence on Twitter can provide valuable insights into the social media context,
laying the groundwork for understanding the developmental implications of these perceptions in
future research.
Beyond the lack of attention to public perceptions, previous research is also limited by
the focus on negative perceptions of the adolescence (e.g., risk-taking and rebelliousness) that
are heavily informed by earlier developmental theories, such as the storm and stress view
(Buchanan & Holmbeck,1998). This narrow focus on negative perceptions ignores the positive
and neutral perceptions of adolescence that have been emphasized by more recent adolescent
theories (e.g., adolescents have unique potentials and strengths; Damon, 2004), potentially
precluding the understanding of a full range of views that may exist in the public. Furthermore,

3



existing research has exclusively investigated individual perceptions of the general adolescent
population, without considering the diversity of adolescent subgroups (e.g., racial/ethnic and
sexual minorities) who may have unique developmental contexts and characteristics (McLoyd &
Steinberg, 1998). This omission may obscure potential differences in how adolescents from
various sociodemographic backgrounds are viewed by the public. Understanding these
differences is crucial to challenge the potential negative perceptions associated with adolescent
subgroups and to promote positive development among diverse youth.
To address these limitations, the current study employed both theory-driven and datadriven approaches enabled by machine learning and big data analytics to understand public
perceptions of adolescence on Twitter. I also conducted an initial exploration on how these public
perceptions may differ for adolescents from various sociodemographic backgrounds (i.e., by
gender, race/ethnicity, and sexual orientation). Specifically, I extracted information from big
textual data (i.e., text entries obtained from Twitter) using unsupervised learning algorithms in
word embedding models. This approach represented each word that appeared in the dataset as
high-dimensional vectors, including terms of adolescence (e.g., “adolescent,” “teen,” “youth”),
terms that indicated membership in adolescent subgroups based on race/ethnicity, gender, and
sexual orientation (e.g., “boy,” “girl,” “lesbian”), and descriptors that may reflected Twitter users’
perceptions of adolescence (e.g., “risk taking,” “civic engagement,” “creativity”). The
associations between different terms (e.g., “adolescent” and “risk taking”) can be quantitatively
assessed, reflecting how people associate meanings of adolescence with different characteristics,
which can be interpreted as public perceptions of adolescence on Twitter (e.g., Miklov et al.,
2013). For the theory-driven approach, I tested how terms of adolescence were quantitative
associated with pre-defined lists of descriptors (i.e., stereotypical vs. neutral descriptors, positive

4



vs. negative descriptors) compared to the terms of adult and people. For the data-driven approach,
I developed an innovative method that leveraged both quantitative results and qualitative analysis.
I first used the word embedding model to generate the top 60 descriptors (e.g., “creativity”) that
were found to be most closely associated with a certain term of adolescence (e.g., “adolescent”).
I then recruit coders to conduct qualitative content analysis of the descriptors while referencing
example tweets that contained the descriptors to identify themes regarding public perceptions of
adolescence. My dissertation revealed a full range of characteristics and themes that people
associate with adolescence, including both positive, neutral, and negative perceptions, for diverse
populations. The results provided insights into how the public perceives adolescence and can
inform future efforts to promote positive framing of adolescence in public discourse.

5



CHAPTER 2: THEORETICAL FRAMEWORKS
Understanding how researchers define and characterize the term of adolescence can be
the first step of unpacking public perceptions of adolescence, specifically about who is in the
stage of adolescence and what are the characteristics of the stage. Although few theoretical
works have directly discussed the importance of public perceptions of adolescence,
developmental theories such as the ecological systems theory (Bronfenbrenner, 1979, 1995) and
positive youth development (Damon, 2004) can inform the current study by highlighting how
social values and beliefs can impact adolescent development. In this section, I first reviewed the
trends of definition of adolescence as a developmental stage. I then summarized theoretical
views of adolescence since Hall’s (1904) initial conception. I also discussed theoretical
frameworks that may inform the importance of studying public perceptions of adolescence.
Adolescence as A Developmental Stage
From a historical perspective, the understanding and definition of adolescence continued
to evolve over time. The word “adolescence” was first used in English in the 15th century, which
meant “to grow up or to grow into maturity (Lerner & Steinberg, 2009, p. 3),” yet it did not refer
to a specific developmental period at that time. Thirteen-year-olds are considered adolescents
and immature in the current society, but they would be considered adults in western countries in
the 16th century (Staff, 1995). Scholars generally acknowledge Stanley Hall as discovering and
defining adolescence in its modern sense (Henig, 2010) and initiating the scientific study of
adolescence (Arnett, 1999). Hall (1904) defined adolescence as a new developmental period
between 12 and 25 years of age, characterized by storm and stress, seeing adolescence as the
recapitulation of a turbulent transitional time in the evolutionary history of human race.

6



The modern concept of adolescence continues to evolve after Stanley Hall (1904). As of
now, adolescence is widely accepted by both scholars and lay people as a developmental stage
between childhood and adulthood. Current researchers generally acknowledge adolescence as a
distinct phase with both biological maturation and major social role transitions (Sawyer et al.,
2018). The World Health Organization has formally defined adolescence as the period between
10 and 19 years of age (Kuruvilla et al., 2016).
Several other terms are often used to refer to a roughly similar developmental stage. The
term “teenage” has been used interchangeably with “adolescence” in academia (e.g., Altikulac et
al., 2019). The term “puberty/pubescence” refers to the process of sexual maturation
accompanied by significant changes in body, brain, and behavior (Mendle et al., 2019), which
has been acknowledged as a key element in the physical and cognitive development during
adolescence (Sawyer et al., 2018). The term “youth” became popular after the United Nations’
first International Youth Year in 1985 (Sawyer et al., 2018). Youth was formally defined as a
social category without age boundaries (United Nations Educational, Scientific and Cultural
Organization; UNESCO, 1985), yet the United Nations usually include people between 15 and
24 years of age when reporting statistics, which overlaps with the age definition of adolescence
but also includes young adulthood or emerging adulthood. Indeed, researchers now consider the
period between 18 and 24 years of age as emerging adulthood, of which the characteristics such
as identity exploration and instability resemble adolescence more compared to adulthood (Arnett,
2000).
Theoretical Views of Adolescence
While adolescence as a developmental period can be neutral, the valence of its
characteristics – whether positive or negative – can differ across theoretical frameworks. Stanley

7



Hall’s (1904) initial definition of adolescence was closely associated with his storm-and-stress
view of adolescence. Specifically, Hall (1916) extended Darwin’s concept of biological evolution
into a psychological theory of recapitulation, suggesting that different developmental periods can
be seen as recapitulations corresponding to different stages in the evolutionary history of human
race. The period of adolescence in Hall’s (1916) framework corresponds to the turbulent
transitional stage in human history, therefore characterized with high levels of emotional
disturbance and stress. The so-called storm-and-stress view of adolescence was also adopted by
psychoanalysts such as Anna Freud (1946, 1958), who argued that conflicts and abnormality is
universal and immutable during adolescence. Hall’s storm-and-stress view of adolescence
corresponds to a major trend of contemporary research that focuses on internalizing and
externalizing problems in adolescents (e.g., Memmott-Elison et al., 2020). While such research
does not necessarily reflect any negative perceptions of adolescence within each study itself, the
large proportion of adolescent studies with a deficit perspective may indicate an overall negative
characterization of adolescence in academics (Arnett, 1999; Rich, 2003).
Nevertheless, later generations of researchers have shifted their perspectives of
adolescence from Hall’s (1904) storm-and-stress view. On the one side, researchers have
revisited the storm-and-stress view, reaching to the conclusion that although the levels of storm
and stress such as risky behavior and emotional conflicts are more likely to increase during
adolescence, the overall degree is moderate, and there are individual differences as well as
cultural variations (Arnett, 1999). On the other side, more positive theoretical frameworks have
been proposed to capture the whole picture of adolescence. For instance, the positive youth
development (PYD) framework suggested that previous research on adolescence focused more
on the negative aspects, such as the deficits of adolescents, and underestimated the true

8



capacities of them, therefore proposing to focus more on the positive aspects (e.g., positive
attitudes such as self-efficacy, positive competence such as empathy, positive behaviors such as
social engagement) and capture the full potential of adolescents (e.g., Damon, 2004; Larson,
2000). The PYD framework has widely been adopted to guide the directions of research on
adolescence as well as to facilitate practices to promote adolescent development, such as to
understand and promote youth civic engagement (Youniss et al., 2002), suggesting an
increasingly positive characterization of adolescence in the research field.
Another central theme in adolescent research, identity development, also reflects the
complexity of crisis and opportunities during this developmental period. Erikson’s (1968) theory
posits that the primary psychosocial task of adolescence is identity formation with both
opportunities and crisis for adolescent development. On the one side, increased cognitive and
physical abilities as well as independence and autonomy enable adolescents to explore topics
such as vocations, ideologies, and relationships through the interactions with schools,
communities, and neighborhoods, which may help adolescents form their identity. On the other
side, changing physical, cognitive, and environmental factors lead adolescents to experience
some form of role confusion, in which adolescents doubt the meaning and purpose of their
existence. Therefore, Erikson (1968) suggested that adolescents must form their identity to solve
the crisis of role confusion. While Erikson (1968) mainly discussed the theme of identity
development from the perspective of solving developmental crisis and conflict, later research has
emphasized more on the opportunities of identity development during adolescence, such as
adolescence as a period when individuals actively explore the characteristics of the self (Meeus,
2011) and develop self-concepts such as self-esteem (Harter, 1998). Researchers have also
explored adolescents’ identity development in specific social categories such as ethnic and racial

9



identity development, in which ethnic/racial minority adolescents may face challenges such as
stereotypes (Way et al., 2013) but also benefit from it by navigating through racism and
oppression (Phinney & Kohatsu, 1997).
Moreover, later generations of researchers have increasingly realized the importance of
sociocultural backgrounds for understanding adolescent development. For example, when
examining adolescence in non-western cultures, Margaret Mead (1928) found that Samoa people
did not have a clear conception of adolescence and most children there had a smooth transition
between childhood and adulthood with a minimum level of storm and stress. This line of
research has led contemporary researchers to emphasize cultural and individual differences in the
understanding of adolescent development (e.g., Arnett, 1999) and to investigate diverse
populations (McLoyd & Steinberg, 1998), leading to diversified characterizations of adolescence
rather than assuming adolescent characteristics as universal. For instance, while identity
development has been found to be a key feature of adolescence across cultures, the patterns of
identity development and its key factors (e.g., cultural values) may vary across different
adolescent subgroups (e.g., ethnic/racial groups, Smith, 1991; sexual-minority youth, SavinWilliams, 2011).
In line with this trend, contemporary research on adolescence has emphasized contextual
factors in adolescent development. By reviewing literature, Dornbusch (1989) concluded that
research on adolescence has shifted from studying individual adolescents carrying out their
developmental tasks (e.g., cognitive development, identity development) to understanding the
intersections between adolescents and their environments. Such a trend has been strengthened in
later works, with more and more research adopting theories such as the ecological systems
theory (Bronfenbrenner & Morris, 1998) and the life course paradigm (Elder, 1998) to

10



investigate contextual factors in adolescence (Crosnoe & Johnson, 2011). Adolescence is now
seen as embedded in multiple layers of social systems (Crosnoe & Johnson, 2011).
Contextual factors such as peer relationships have been seen as one of the most important
characteristics (e.g., adolescents are oriented to and impacted by their peers for both positive
outcomes such as prosocial behavior and negative outcomes such as substance use) of
adolescence (Brown & Larson, 2009). In addition to peer relationships, other contexts such as
family, school, and neighborhood have also been highlighted in contemporary research.
Specifically, adolescents’ relationship with their parents remains to be impacting their
development, including positive relationship such as closeness and warmth, positive relationship
such as conflicts and disagreements, and other themes such as autonomy seeking/support,
communication, conflict resolution (Smetana & Rote, 2019). Adolescents spend more time in
school than other settings and engage in important activities such as learning and extracurricular
activities (Waters et al., 2009). Researchers have also highlighted adolescent-teacher relationship,
peer interaction, school culture, school safety in the school context for adolescence (Eccles &
Roeser, 2011). Meanwhile, neighborhood has been considered as an important context for
adolescent development in terms of externalizing and internalizing problems, academic
achievement, health, and cultural development, especially for adolescents from underrepresented
or minoritized backgrounds (White et al., 2021).
Theoretical Views on Public Perceptions of Adolescence
Although there is no widely accepted definition of public perceptions of adolescence in
the literature, I proposed the following definition based on previous discussions of public
perceptions of other topics, such as food, risk, and health (Dowler et al., 2006). Public
perceptions of adolescence can be defined as the collective beliefs, attitudes, understanding,

11



impression, or opinions of the general public towards adolescence as a developmental period and
towards adolescents as a group. This definition is also informed by Buchanan and Holmbeck’s
(1998) work on individual perceptions of adolescents, which refers to the beliefs and attitudes
that individuals (e.g., parents, young people themselves) hold towards adolescents as a social
category. By contrast, public perceptions of adolescence capture the broader views of society as a
whole, reflecting the collective attitudes and opinions of the general public towards this critical
period of development. To date, there are few theoretical works that specifically discuss public
perceptions of adolescence. However, the importance of this topic can be inferred from several
classical theories related to adolescent development.
Bronfenbrenner’s (1979, 1995) ecological systems theory suggested that adolescent
development is embedded in multiple ecological systems. In this theory, public perceptions of
adolescence are perhaps most relevant in the macrosystem. Macrosystem refers to the culture or
society including its values and ideologies that shapes the fundamental background of
development and impacts the structures and relationships among the other systems (e.g., the
exosystem consisting of multiple social structures that indirectly influence the developing
individual such as neighborhood and parents’ workplaces). Public perceptions of adolescence can
be an important factor in this system because it can be seen as an element of sociocultural values,
therefore fundamentally impacting other systems such as the perceptions of adolescents held by
neighbors, parents, and teachers as well as their attitudes towards and interactions with the
adolescents based on their perceptions. While Bronfenbrenner (1979) discussed more on social
categories such as socioeconomic status and ethnicity in the macrosystem, adolescence can also
be seen as an important social category that has profound developmental implications.

12



Specifically, public perceptions of adolescence can also be incorporated as an important
contextual factor into Bronfenbrenner’s (1995, 2005) process-person-context-time (PPCT) model,
in which he emphasized the interrelatedness between the developing person and the
developmental contexts (i.e., the ecological systems such as the macrosystem of the culture or
society) via reciprocal interactional processes over time. Public perceptions of adolescence in the
macrosystem can impact other developmental contexts that are more proximal to the developing
individual. For instance, public perceptions of adolescence can be adopted by parents and
teachers in the microsystem (i.e., the immediate settings in which the developing individual can
interact with such as family and school) as individual perceptions of adolescence or as social
norms about adolescent development if certain perceptions become pervasive in the settings.
Public perceptions of adolescence can also impact the developing person by shifting how one
perceive and expect one’s self-development in the current developmental stage. Then, individual
perceptions of adolescence as held by individuals (e.g., parents, teachers, and the developing
person) and social norms about adolescent development can impact the developing person
through reciprocal processes such as parent-child and teacher-student interactions. For instance,
parental beliefs on development may impact their parenting attitudes and behavior (Sigel, 1985).
Adolescents’ own perceptions about their social categories (e.g., gender) can also impact their
own developmental outcomes (e.g., gender stereotypes, Starr & Simpkins, 2021).
The positive youth development (PYD) framework (e.g., Benson, 1997; Lerner et al.,
2005) also speaks to the importance of understanding public perceptions of adolescence. In
Benson’s (1997) framework of developmental assets, he highlighted the importance of multiple
internal and external assets to facilitate adolescent development, many of which are related to
perceptions of adolescence. For instance, Benson (1997) advocated positive values (e.g., caring,

13



honesty) and identity (e.g., personal power, positive view of personal future) as internal assets,
which can only be achieved when the adolescents have a positive perception of their own age
group. Benson (1997) also advocated empowerment from multiple contexts (e.g., community,
school, and family) such as acknowledging the value of youth in the community, which requires
positive perceptions of adolescence to be prevalent in these contexts. Furthermore, Lerner and
colleagues’ (2005) 6Cs models of PYD proposed competence, confidence, connection, character,
caring/compassion, and contribution as six important indicators of PYD. The second indicator of
confidence suggested that the individual should establish a global positive view of the self,
which requires the developing adolescents to possess a positive perception of their own age
group. From this perspective, positive perceptions of adolescence should be established in
multiple contexts (e.g., community, school, and family) as well as the developing adolescents
themselves, all of which are impacted by public perceptions of adolescence as illustrated in the
prior discussion on the PPCT model. Although contemporary researchers have revisited Hall’s
storm-and-stress view of adolescence, adolescent research and practices are still influenced by
Hall’s perspective, focusing more on themes related to the storm-and-stress view, such as risky or
problem behaviors and mood disruptions (Arnett, 1999; Rich, 2003), which may strengthen the
negative perceptions of adolescence held by the public and further hinder the positive
development of adolescents. The PYD framework has the potential to break such a chain of
negative influences by advocating a positive framework of understanding adolescence.
In addition to developmental theories, several basic psychological and communication
theories also help explain the importance of how we understand or perceive adolescence. For
instance, the theory of self-fulfilling prophecies suggests that public perceptions or beliefs have a
strong impact on individual beliefs, which will in turn impact individual behavior based on the

14



prophecies of self-fulfilling (i.e., the phenomenon whereby the expectations about an individual
or entity result in the individual or entity acting in the ways that confirm the expectations; Jussim,
1986, 1991). The theory of self-fulfilling prophecies is also relevant to the theory of stereotype
threat, which suggests that stereotyped perceptions of social groups may cause the group
members to confirm such perceptions by engaging in particular activities (Spencer et al., 2016).
From this perspective, public perceptions of adolescence may impact adolescent development
through self-fulling, in which adolescents may tend to behave in ways that confirm public
perceptions of adolescence. In addition, communication research on framing effects suggests that
perceptions or beliefs can influence how we frame or present certain issues in communication
messages, which will in turn influence how the audience perceive the issues from the messages
and their later behaviors (e.g., Entman, 1993; Vliegenthart & van Zoonen, 2011).

15



CHAPTER 3: LITERATURE REVIEW
While extensive research has investigated the characteristics of adolescence, less
attention has been paid to people’s perceptions of adolescence. A handful of studies have
examined and measured individual perceptions of adolescence (e.g., Buchanan & Holmbeck,
1998; Gross & Hardin, 2007; Hines & Paulson, 2006) as well as the impact of such perceptions
on adolescent development (e.g., Buchanan, 2003; Qu et al., 2016; Qu et al., 2018). However,
these study all surveyed small samples of participants and focused on the school and family
contexts. They also employed a theory-driven approach that tended to focus on negative
perceptions of adolescence. In this section, I presented a literature review of existing research
that measured individual perceptions of adolescence or examined the impact of individual
perceptions of adolescence on adolescent development. I then discussed the limitations of
existing research.
Individual Perceptions of Adolescence
A growing body of literature suggests that people with different identities (e.g., parents,
teachers, college students, and adolescents) and country of origin (e.g., the U.S., Netherlands,
and China) have had stereotyped perceptions of adolescents since the 1960s (e.g., Friedman,
1975; Gross & Hardin, 2007; Qu et al., 2016). For instance, in a survey conducted in 1965,
American teachers agreed that adolescence is a period with great emotional disturbance and
complete personality change (Friedman, 1975). In the 1980s, 192 American college students
(approximately 75% female, 73% White) generally agreed with the storm-and-stress view of
adolescence (Holmbeck & Hill, 1988), and 90 elementary school teachers and 636 parent pairs of
early adolescents (97% identified their children as White) in the U.S. endorsed stereotyped
beliefs that adolescence is a difficult period (Buchanan et al., 1990). In the 1990s, 361 college

16



students (53% female, 89% White) and 112 parents of adolescents (69% mothers) in the U.S.
indicated more negative perceptions of adolescence such as risk-taking and rebellious compared
to neutral or positive perceptions such as friendly and conforming (Buchanan & Holmbeck,
1998). In the 2000s, 106 American college students in the (79% female, 44% White) exhibited
both explicit and implicit stereotyping of adolescents compared to older adults (Gross & Hardin,
2007), while 1112 teachers and social workers attending training programs in Hong Kong had
much more negative comments of adolescence (e.g., “problematic self-conceptions,”
“irresponsible,” “materialistic”) compared to positive comments (e.g., “creative,” “energetic”;
Shek & Chan, 2011). In the 2010s, 203 American adolescents (46% female, 73% White) were
more likely to perceive adolescence as a time of decreased family responsibility and increased
school disengagement compared to 194 Chinese adolescents (54% female, 99% Han decent; Qu
et al., 2016).
Among these studies, two of them specifically explored the contents and structures of
individual perceptions of adolescence in a systematic manner. Buchanan and Holmbeck (1998)
first asked 62 college students and 61 parents (demographic information was not reported) to
answer open-ended questions about adolescent personality characteristics, they selected 48
descriptors from participants’ answers and also added 17 descriptors based on their
understanding of academic and popular literature. Then, they developed the descriptors into
scales and administered the questionnaires to 361 college students (53% female, 89% White) and
112 parents (69% mothers, the majority had a college degree or beyond). Principle components
analysis resulted in 9 factors with positive, negative, and neutral perceptions, including risktaking/rebellious, friendly, problem behaviors, conforming, etc. They also did a series of
measurement analysis to establish reliability and validity. This scale (“Stereotypes of

17



Adolescents”) is often used by studies to examine perceptions of adolescence their influence on
adolescent development. However, these studies mainly focused on the subscales with negative
perceptions (e.g., risk-taking/rebelliousness, problem behaviors, and internalizing problems;
Buchanan & Hughes, 2009; Hines & Paulson, 2006).
Using a similar approach, Qu et al. (2016) asked 203 U.S. adolescents (46% female, 73%
White) and 194 Chinese adolescents (54% female, 99% Han decent) to write descriptions or
characteristics of youth their age using an open-ended measure. The responses were coded into
one of the four categories, including individuation, responsibility, social relationships, and
emotionality. The categories were developed based on Buchanan and Holmbeck’s work (1998)
while also ensuring interrater reliability. Qu et al. (2016) also developed a closed-ended measure
that was administered to the same participants after the open-ended measure. This measure also
included four categories based on developmental literature (i.e., individuation, family
responsibility, school engagement, and peer orientation), and items were developed by the
research team or modified from other relevant scales (e.g., the family responsibility scale,
Fuligni et al., 1999). They used reliability and confirmative factor analysis to evaluate the
measure. Both results from the open-ended measure and closed-ended measure were compared
between U.S. adolescents and Chinese adolescents. For the open-ended measure, American
adolescents perceived adolescence as more associated with identity exploration, bullying, being
disrespectful, and being oriented towards peers as well as less associated with being responsible,
experiencing positive emotions. For the closed-ended measure, American adolescents were more
likely to characterize adolescence as a time of individuation from parents, heightened
disengagement from school, and not being responsible to the family.

18



In addition to academic research, several public reports have also investigated public
perceptions of adolescence using survey data or media materials (e.g., Aubrun & Grady, 2000;
Gilliam & Bales, 2001; Bostrom, 2004). For instance, Bostrom (2001) reviewed a series of
public polls that surveyed peoples’ perceptions of adolescence nationally in the U.S. (e.g., the
CBS News Poll, the ABC News Poll, and polls conducted by the Gallup Organization, each
surveyed between 800 and 3,000 adults that were randomly selected using certain sampling
strategies) and found that American adults tend to have a negative perception of adolescence and
such perception is becoming more and more negative. Several reports from the Frameworks
Institute analyzed the media contents about adolescents (e.g., prime-time entertainment
programming, Aubrun & Grady, 2000) and found that the media tends to show a negative and
stereotyped image of the general adolescent population, which may in turn strengthen the
negative perceptions of adolescents held by the public.
The Impact of Individual Perceptions of Adolescence
There is an emerging line of research examining the impact of individual perceptions of
adolescence on adolescent development (e.g., Buchanan, 2003; Qu et al., 2016; Qu et al., 2018).
Focusing on parents’ perceptions of adolescence, several studies show that negative perceptions
of adolescence are associated with negative parental attitudes and behaviors, leading to less
desirable developmental outcomes of their children. For example, one study found that among 75
mothers (89% White), those holding more positive beliefs about adolescents as a group tended to
have more positive expectations for their own children in adolescence; they also expected a
closer relationship with their children (Buchanan, 2003). Another study found that parents
holding a storm-and-stress view of adolescence exhibited more authoritarian parenting, less
supportive of their adolescents’ autonomy (Holmbeck, 1996). Moreover, among 1193 parents

19



(62% mothers), those with stereotyped perceptions of adolescence tended to have children with
lower self-rated of social ability and having more deviant friends (Jacobs et al., 2005).
A newer body of literature has investigated youth’s own understanding of adolescence
and shown that stereotyped perceptions of adolescence held by youth themselves may hinder
their positive development, yet interventions may be implemented to shift their perceptions and
promote positive development (e.g., Qu et al., 2016; Qu et al., 2018; Qu et al., 2020). For
example, among 270 adolescents (52% female, 64% White), those holding a storm-and-stress
view of adolescence also showed parallel behavior and attributes one year later (e.g.,
externalizing behavior and internalizing symptoms, Buchanan & Hughes, 2009). Among 203 U.S.
adolescents (46% female, 73% White) and 194 Chinese adolescents (54% female, 99% Han
decent), American youth’s understanding of adolescence as associated with less family
responsibility and more school disengagement explained the observed differences that American
youth were less engaged in school compared to Chinese youth (Qu et al., 2016). Among 23
adolescents (43% female, 64% White), those holding stereotyped perceptions of adolescence in
middle school showed altered prefrontal cortex activation during cognitive control during the
transition to high school, which was in turn associated with more risk taking (Qu et al., 2018).
An intervention countering youth’s stereotypes of adolescence promoted youth’s academic
engagement and performance as well as decreased their risk taking among middle school
students in China (Study 1: 125 youths, 48% female, 99% Han decent; Study 2: 319 youths, 50%
female; Qu et al., 2020).
Limitations of Existing Research
Although previous research has examined individual perceptions of adolescence and
tested their impact on adolescent development, there are several limitations associated with this

20



body of work. First, existing research mainly examined perceptions of adolescence from the
perspectives of parents, teachers, and youth, focusing on the school and family contexts. Few
studies have investigated this topic from the perspective of the general public (Heintz-Knowles,
2000; Nichols & Good, 2004). The ecological systems theory suggests that youth development is
embedded in multiple systems and the transactions among systems (Bronfenbrenner, 1979).
Public perceptions of adolescence reflect the societal value on the social category of adolescence
and therefore serve as an important contextual factor in the macrosystem. They can impact
individual perceptions of adolescence in other systems (e.g., teachers in school, parents in
family), which will in turn influence the developing adolescents through reciprocal processes
such as teacher-student and parent-child interactions. In fact, both Buchanan and Holmbeck
(1998) and Qu et al. (2016) suggested that perceptions of adolescence held by parents and youth
are influenced by societal stereotypes or the society’s mainstream ideas. Therefore, it is essential
to examine public perceptions of adolescence from a broader perspective as the underlying
context for individual perceptions.
Furthermore, existing research has primarily employed a theory-driven approach and
focused more on the negative perceptions of adolescence as informed by Hall’s (1904) stormand-stress view. For instance, both Buchanan & Holmbeck (1998) and Qu et al. (2016) coded
participants open-ended responses about their perceptions of adolescence into a limited number
of categories based on developmental theories. Their follow-up studies focused more on the
negative perceptions of adolescence that reflect the storm-and-stress view rather than other
neutral or positive themes that were identified in their initial work (e.g., Buchanan & Hughes,
2009; Qu et al., 2018; Qu et al., 2020). The focus on negative perceptions also overlooks positive
or neutral views of adolescence that are emphasized to a greater degree by later theoretical work

21



(e.g., the PYD framework). A data-driven approach that examines a full range of public
perceptions may better address this limitation and facilitate future theoretical development.
Moreover, existing research focuses on perceptions of the general adolescent population
without paying attention to various subgroups by race/ethnicity, gender, and sexual orientation.
Research has identified various stereotypes being held towards adolescents based on their social
categories (e.g., gender stereotypes; Ellemers, 2018). It remains unknown how public
perceptions of adolescence may differ by social categories (e.g., gender, race, ethnicity, sexual
orientation). For instance, adolescent boys may be perceived as more associated with rebellion
and risk taking compared to adolescent girls (Abbott-Chapman et al., 2008). Racial/ethnic
minority adolescents may be perceived as more engaged with risky health behaviors compared to
the general adolescent population (Wallace et al., 2011). On the other hand, social media posts
about adolescent subgroups may not reflect stereotypes but convey messages about the actual
developmental disparities such as gender disparities in STEM (Beede et al., 2011). Such
messages may also be captured as reflecting public perceptions that we need to address these
disparities. Either for the stereotypes or the disparities associated with different subgroups of
adolescence, little is known about how they unfold in public perceptions of adolescence.
Investigating the differential perceptions associated with adolescent subgroups may be the first
step towards addressing inequality among adolescent subgroups in broader contexts via potential
prevention and intervention programs that seeks to promote the positive framing of adolescent
subgroups in public perceptions.
Finally, the definition and characteristics of adolescence have kept evolving in
accordance with sociocultural changes, and public perceptions of adolescence may also be
constantly shifting. For instance, as the society has increasingly advocated for the importance of

22



youth engagement in facilitating positive social change (e.g., MacKay et al., 2020), public
perceptions of adolescence may become more associated with positive themes such as
contributing to the society. Furthermore, Bronfenbrenner’s (1995, 2005) suggested in his PPCT
model that the macrotime (i.e., the macro-level time factor that impacts development) involves
the changing expectations in the larger society, which requires capturing potential changes in
public perceptions of adolescence. Therefore, an update of knowledge on the perceptions of
adolescence is needed.
My dissertation sought to address the aforementioned limitations with two major aims: 1)
to provide an updated, comprehensive, and data-driven description of current public perceptions
of adolescence, with a focus on the social media platform of Twitter, which can expand the
previous work from the individual level to a broader context; 2) to conduct initial exploration of
how public perceptions of adolescence differ by various adolescent subgroups (e.g., racial/ethnic,
gender, and sexual minorities).
Twitter as a Unique Window
The society has entered into the era of Web 2.0, with Internet users actively creating
contents and interacting with each other in online communities (Cormode & Krishnamurthy,
2008). As of 2018, 45% of American adolescents reported that they were online almost
constantly, and 44% reported being online several times a day (Anderson & Jiang, 2018). Social
media platforms, such as Twitter, is becoming the major and most important online communities
and attracting more and more Internet users. Twitter is a free microblogging social network
platform that allows users to post small bursts of information called tweets within 140 characters
of length. As of 2018, Twitter has more than 330 million active users monthly, and these users
produced more than 500 million tweets per day; around 32% of American adolescents were using

23



Twitter (Anderson & Jiang, 2018), a much higher rate than that of adults (24%; Perrin &
Anderson, 2019).
Twitter has become one of the most important platforms for users to express their
opinions across various issues on the Internet (Java et al., 2007). An extensive body of research
has examined political expression on Twitter (e.g., Vaccari et al., 2015), and to a lesser extent
sexual assault (PettyJohn et al., 2018) and public health (Hays & Daker-White, 2015). A recent
study examined perceptions of older adults and COVID-19 on twitter (Jimenez-Sotomayor et al.,
2020). Similarly, Twitter can be a unique window to explore people’s perceptions of adolescence.
The information sharing and opinion expression nature of the platform, as well as the large
number of users and posts, may enable generalization of users’ perceptions of adolescence.
Moreover, tweets reflect a more spontaneous form of opinion expression (Christensen, 2011),
rather than responses given during intentional inquires such as surveys or interviews. This
approach enables us to explore public perceptions of adolescence in a more naturalistic and
realistic way, avoiding potential biases that may arise from asking participants to report their
views in a more controlled setting and, thus, enhancing the study’s ecological validity.
Using a Big Data and Machine Learning Approach
I employed a big data and machine learning approach to examine public perceptions of
adolescence on Twitter. Big data and machine learning have recently been introduced into social
sciences (e.g., Jacobucci & Grimm, 2020; McFarland et al., 2016), yet the paradigms and
standards of their applications have yet to be established. Grimmer et al. (2021) suggested that
machine learning is better to be applied to discover new concepts, measure the prevalence of the
concepts, assess causal effects, and make predictions. These applications may facilitate social

24



sciences to move from a deductive approach, where we make predictions based on theories and
test them, to an inductive approach, where we learn from observations.
The focus of my dissertation can be seen as discovering and measuring the public
perceptions of adolescence in big data, which is suitable for using machine learning with an
agnostic and inductive approach. Although the concept of adolescence is not new and has already
been formally defined, there is no clear definition or theoretical account on public perceptions of
adolescence. Public perceptions of adolescence can be best captured by describing how the
public perceive adolescence, rather than identify them by judging if the data meet a certain
theoretical account of adolescence. Therefore, it is appropriate to use the big data/machine
learning approach for my dissertation, and my research questions also align more with the
agnostic approach because of my research focus and methodology.
Unpacking the Meanings of Adolescence Using Word Embeddings
Word embedding models are a powerful machine learning tool that has widely been
employed in the field of artificial intelligence and natural language processing. By feeding
textual data from large corpora to a word embedding model, the model will extract information
from the data based on the cooccurrence of words and then represent the meanings of words as
high-dimensional vectors (i.e., word embeddings). The resulting word embeddings can be
thought of as a way of measuring a construct with multiple items or variables. While
conventional social science research typically measures a construct using a small set of items
(e.g., 2-10 items), word embeddings represent the meanings of each word using hundreds of
dimensions (e.g., 300 dimensions; Grag et al., 2018; Kozlowski, 2019). This allows for a much
more nuanced characterization of word meanings. Indeed, word embeddings have been found to
show compositionality, an interesting attribute suggesting that word embeddings can capture the

25



latent meanings of words (Mikolov et al., 2013). For instance, the word embeddings of “Vietnam”
plus the word embeddings of “capital” was found to be close to the word embeddings of “Hanoi”,
which is indeed the capital of Vietnam.
If word embeddings can capture the latent meanings of words in large corpora, they can
be interpreted as reflecting perceptions of the words held by the people writing the texts, as the
use of words are subjective and reflect individual’s understanding about the words. Indeed, an
emerging body of literature has shown that word embeddings may be used in this way. Inspired
by the Implicit Association Test (IAT) in psychology, Caliskan et al. (2017) developed the WordEmbedding Association Test (WEAT) based on word embeddings trained with big textual data
collected from the Internet. Whereas IAT assessed human participants’ biases with experiments,
WEAT assessed how different words was associated with each other in the space of word
embeddings as reflecting biases (e.g., if the word embeddings of the word of “flower” is closer to
positive terms compared to the word embeddings of “insect”). The two tests are comparable as
they can assess the same set of biases. Their results showed that many findings from IAT can be
replicated in WEAT, suggesting that human biases in big data can be identified with word
embeddings as indicating potential bias in people’s language.
Grag et al. (2018) trained a word embedding model with 100 years of English literature
data and focused specifically on gender and ethnic stereotypes. Specifically, they measured
gender and ethnic stereotypes as the relative distance between a term for gender/ethnicity (e.g.
“woman”) and a neutral term (e.g., “researcher”) compared to the opposite term for
gender/ethnicity (e.g. “man”). In this example, the measure corresponds to gender stereotype
towards the occupation of researcher. Moreover, Grag et al. (2018) took a historical perspective
and analyzed 100 years of textual data to identify the patterns of change. The identified

26



trajectories of gender and ethnical stereotypes were found to be meaningful, such that historical
trends (e.g., how “woman” is close to “researcher” compared to “man”) were also closely
associated with demographic and occupation shifts (e.g., the actual prevalence of females in
academia). This study provides another example of how word embeddings can capture the latent
meanings of terms in big data and reflect stereotypes associated with gender and ethnicity.
Kozlowski et al. (2019) presented another example of constructing the meanings of terms
using word embeddings trained with English literature data. They identified six dimensions of
class based on literature (i.e., affluence, cultivation, employment, education, status, morality) and
explored the meanings of these dimensions by examining how the word embeddings of these
dimensions (e.g., “rich” and “poor” as capturing the dimension of affluence) in relation to the
word embeddings of other sociological constructs of interests (e.g., gender, race/ethnicity,
occupation, etc.). Kozlowski et al. (2019) specifically made the claim that word embeddings can
be used to analyze the cultural meanings of important constructs in theory.
The aforementioned three studies have validated the basic analytic framework of training
word embeddings and analyzing the associations between word embeddings of different terms to
interpret the meanings of terms. The meanings of terms as interpreted based on word embeddings
are not objective but rather subjective to the data, capturing latent nuances underlying the data.
The current study will train word embeddings using Twitter data about adolescence. Because
Twitter data are user-generated contents that are related to users’ opinion expression, interpreting
the latent meanings of terms related to adolescence may reflect public perceptions of adolescence
on Twitter.
I think this analytic approach has several advantages. First, the word embeddings are
high-dimensional, usually with 300 dimensions. Although it is difficult to interpret the meanings

27



of each dimension, the 300 dimensions as a whole may better capture the complexity and
nuances compared to other conventional quantitative approaches such as factor analysis. Second,
this analytic approach is data-driven and unsupervised. No prior knowledge or assumptions need
to be fed into the algorithm, it is thus more likely to capture more meanings and perceptions of
adolescence, whereas previous studies have mostly relied on theory-driven approaches and
focused on the storm-and-stress view.
The Current Study
The current study sought to move from previous work that measured individual
perceptions of adolescence to assessing public perceptions of adolescence in a broader context.
To achieve this, I leveraged a large dataset consisting of millions of tweets about adolescence
from Twitter that is extracted throughout 2019 before the COVID-19 pandemic. The use of big
data may enhance the generalizability of the current study to capture public perceptions on social
media. The dataset is relatively new and can provide an updated description of perceptions of
adolescence, which may be reflective of the change in sociocultural backgrounds in recent years.
Empowered by word embedding models, an innovative and powerful machine learning tool that
can extract information from textual data and construct the meanings of concepts with hundreds
of dimensions, the current study also has the potential to transcend previous work by not only
focusing on negative perceptions but also explore and evaluate other forms of perceptions.
Furthermore, the current study also explored public perceptions of adolescence for different
adolescent subgroups (e.g., racial/ethnic and sexual minorities), which may fill in the scarce in
literature as there are few studies investigating this issue.
My dissertation has two aims that are descriptive and exploratory in nature. As such, I
presented some general hypotheses informed by prior work and theories of adolescence.

28



Aim 1. To describe public perceptions of adolescence on Twitter. I hypothesize that,
overall, there is a negative perception (e.g., the storm-and-stress view) of adolescence on Twitter,
which may be indicated by the observation that terms of adolescence (e.g., “teenage”) are more
closely associated with negative or stereotypical descriptors (e.g., “rebellious”) compared to
positive or neutral descriptors (e.g., “patience”) based on prior research on individual perceptions
(Buchanan & Holmbeck, 1998). However, I also expect to identify positive and neutral
perceptions (e.g., youth civic engagement) of adolescence on Twitter, as indicated by the
observation that terms of adolescence are also closely associated with neutral and positive
descriptors.
Furthermore, themes of public perceptions of adolescence can be identified from Twitter
data, as achieved by conducting qualitative thematic analysis to assign descriptors that are
closely associated with adolescence into different themes. Specifically, analyzing word
embeddings can generate a list of relevant terms (e.g., “friend”) associated with the terms of
adolescence, which resemble existing approaches in which participants were asked to assign
descriptors to adolescence (Buchanan & Holmbeck, 1998; Qu et al., 2016). I can then code the
list of relevant terms into different themes to facilitate the understanding of perceptions of
adolescence.
Aim 2. To explore potential variations in public perceptions of adolescents from various
subgroups (based on race/ethnicity, gender, and sexual orientation). This will be done by using
quantitative comparisons of how different terms of adolescent subgroups (e.g., “female
adolescent”) differ in their associations with descriptors and qualitative comparisons of how
different themes can be identified with different adolescence subgroups (based on the
aforementioned steps of qualitative thematic analysis). For instance, I hypothesize that

29



racial/ethnic minority adolescents will be more closely associated with descriptors and themes
related to risky behavior and less associated with competence compared to the general adolescent
population. Male adolescents will be more closely associated with descriptors and themes related
to risky behavior, autonomy, competence compared to female adolescents. Adolescents
belonging to the LGBTQ+ groups may be uniquely associated with descriptors and themes
related to their sexual orientation.

30



CHAPTER 4: METHODOLOGY
Data
The Twitter data were extracted using the Historical PowerTrack API (application
programming interface; Twitter, 2022b) provided and authorized by the Developer Platform of
Twitter, Inc, which is a common data extraction approach in research using Twitter data (e.g.,
Jimenez-Sotomayor et al., 2020). The API is a commercial enterprise API that provides access to
the entire historical archive of public Twitter data with certain rules of search since the initiation
of the Twitter platform. The data extracted from the API were all publicly available (i.e., they
were searchable and could be seen by Twitter users on Twitter) at the time of extraction.
To identify tweets that are relevant for adolescence, I searched a set of keywords related
with adolescence (i.e., adolescent, adolescents, adolescence, adolescency, teen, teens, teenage,
teenagers, youth, youths, juvenile, juveniles, juvenility, juvenescence, puberty, pubertal,
pubescent, pubescents, pubescence, youngster, youngsters, tween, tweens). The keywords were
selected based on English dictionaries, literature on adolescence, and consultation with
researchers studying adolescent development. Based on the keywords, I extracted a 10% random
sample (the highest ratio of sampling that Twitter can offer) of all tweets that were written in
English, posted in 2019 (between December 31, 2018 and January 1, 2020), and contained any of
the aforementioned keywords (n = 8,094,988) by accessing the Historical PowerTrack API via
python. The data were extracted in 2020 with the support of funding from the Research Scholars
Award issued by the College of Social Sciences at Michigan State University. The time period of
2019 was chosen due to the consideration of avoiding potential influences of the COVID-19
pandemic on Twitter environment.

31



The extracted data were recorded in the format of JavaScript Object Notation (JSON),
containing both the text part of the tweets and the metadata (e.g., user profiles). For the purpose
of the current study, I focused on the text part of the tweets, whereas the metadata may be
analyzed for future research to contextualize the tweets.
The extracted data can be categorized into original tweets (n = 2,455,227; 30.3%) and
retweets (n = 5,639,761; 69.7%). Whereas original tweets were solely written and posted by the
users, retweets were based on original tweets but were reposted by users. Some retweets also
contained comments of the users who retweeted the original tweets. Because both original tweets
and retweets can reflect users’ perceptions of adolescence, I kept both types for data analysis. For
retweets, I merged the text of the original tweets that the users reposted with the comments, as
both parts may reflect the users’ perceptions of adolescence.
There were 4,367,889 tweets (54.0% of all tweets) that were duplicates in terms of the
text part, consisting of 1.9% of the original tweets and 76.6% of the retweets, suggesting that
duplicate tweets were prevalent in retweets but not original tweets. Because duplicate tweets may
result in potential biases in the word embedding model by overweighing these samples compared
to other tweets, I excluded the duplicate tweets. As sensitivity analysis, I also trained another
word embedding model that included the duplicate tweets. The results were less meaningful
compared to the model without duplicate tweets. For instance, the word embedding of “teenager”
was found to be most closely associated with “think” in sensitivity analysis, because a tweet
containing text of “my teenager thinks so” had many duplicates in the dataset. After removing
duplicates, the results were more meaningful as the associations among different terms were
more reflective of their meanings rather than simply co-occurrence.

32



The final analytic sample included 3,727,099 tweets, consisting of 2,407,476 original
tweets (64.6%) and 1,319,623 retweets (35.4%).
Data Analysis
Data Cleaning
Following previous research analyzing Twitter data (e.g., Eichstaedt et al., 2015), I
applied a series of steps to clean the textual data, with the goals of removing meaningless texts,
standardizing words in different forms, and facilitating the training of the word embedding
model, including: 1) lowercasing all the letters; 2) removing textual patterns that were
meaningless, including the initial parts of retweets (e.g., “rt @author:”), the parts when users
were mentioning other uses (e.g., “@author”), the hashtag symbol (“#”; a hashtag is a word or
phrase that classifies or categorizes the accompanying text, usually used to facilitate the
discussion of the same topic; I only removed the “#” symbol but kept the word or phrase),
website links (e.g., “www.google.com”); 3) removing punctuations that were not in the middle of
a word (e.g., “whether you’re a new parent, grand parent, first-grader” was cleaned as “whether
you’re a new parent grand parent first-grader”); 4) removing numbers that were not tied to letters
(e.g., “4themany” was kept but “19 01 11” was removed); 5) tokenizing the tweets (i.e.,
separating a tweet into a list of tokens; tokens are individual linguistic units; e.g., “you make me
a teenager” was tokenized as “you,” “make,” “me,” “a,” “teenager”); 6) lemmatizing the tokens
(i.e., shifting the inflected or variant forms of the same word into the original form; e.g., “a
teenager goes to school” was cleaned as “a teenager go to school”).
To facilitate the training and interpretation of the word embedding model, I manually
shifted different keywords of adolescence to more consistent forms. Specifically, keywords of
“adolescents,” “adolescence,” and “adolescency” were shifted to “adolescent”. Keywords of

33



“teens,” “teenager,” “teenagers,” and “teenage” were shifted to “teen”. Keywords of “juveniles,”
“juvenility,” and “juvenescence” were shifted to “juvenile”. Keywords of “pubertal,” “pubescent,”
“pubescents,” and “pubescence” were shifted to “puberty”. Keywords of “youngsters” were
shifted to “youngster”. Keywords of “tweens” were shifted to “tween”. Therefore, there were
seven terms of adolescence left, including “adolescent,” “teen,” “youth,” “juvenile,” “puberty,”
“youngster,” and “tween”.
Sentiment Analysis
To better understand the descriptives of the data, I conducted sentiment analysis of the
data using the VADER (Valence Aware Dictionary and sEntiment Reasoner; Hutto & Gilbert,
2014) algorithm. VADER is an algorithm that was specifically designed to distinguish sentiments
expressed in social media. VADER evaluates the sentiments of tweets by identifying sentiment
intensity lexicons (i.e., words that indicate sentiment intensity based on existing research) as well
as integrating rules (e.g., the exclamation point “!” increases the sentiment intensity without
modifying the semantic orientation; for this part of the analysis I did not remove punctuations
from the data). Previous research suggested that this algorithm outperformed other lexicon-based
models, machine learning models, and even human raters in analyzing the sentiment of tweets
(Hutto & Gilbert, 2014). VADER can automatically analyze the tweets and provide scores for
positive sentiment, negative sentiment, neutral sentiment, and the overall sentiment. For positive,
negative, and neutral sentiment, the scores range from 0 to 1, with higher scores indicating
higher levels of sentiment. For overall sentiment, the scores range from -1 to 1, with scores
higher than .05 indicating positive sentiment, scores lower than -.05 indicating negative
sentiment, and scores between -.05 and .05 indicating neutral sentiment. I fed all the tweets into
the VADER algorithm and generated sentiment scores for each tweet. I then compared the

34



average sentiment scores among tweets containing different terms of adolescence (i.e.,
“adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”) using ANOVA.
Word Embedding Model
I used the word2vec algorithm (Miklov et al., 2013), an unsupervised neural network
model that only require the input of large corpora data and then generate word embeddings (i.e.,
representations of words) for all the different words in the data as vectors with multiple
dimensions (i.e., word embeddings). This approach has been employed by recent studies to
analyze the meanings and perceptions of sociological concepts such as gender, ethnicity, and
social class (Grag et al., 2018; Kozlowski et al., 2019). While the algorithm is complex and its
internal possessing mechanisms remain to be a black box, its output of word embeddings as
high-dimensional vectors can be simply interpreted as constructing the latent meanings of words
using multiple dimensions, which are learned from the data. Each word in the data will be
assessed using the same set of dimensions. Therefore, their associations can be quantitatively
assessed using cosine similarity (a mathematical formula to calculate the distance of two vectors).
Such associations may be used for theoretical interpretation, such as why a word of interest (e.g.,
“adolescent”) is more closely associated with another word (e.g., “risk-taking”) compared to
other words (e.g., “driving”). In the context of the current study, since the data were usergenerated contents around adolescence on Twitter, the meanings of various words of adolescence
(e.g., “teen”) can be interpreted as reflecting public perceptions of adolescence on Twitter.
After data cleaning, the tweets were fed into the word2vec model. For each tweet, the
model identified all the words in the tweet and analyzed words with their surrounding words,
representing each word as vectors with multiple dimensions. Following previous research using
the word2vec algorithm to analyze the meanings of words (e.g., Kozlowski et al., 2019), I set the

35



size of the context window (i.e., the number of words that are identified as surroundings) as 10,
the number of vector dimensions as 300, and the minimum requirement for word occurrence as
20 (i.e., words that appeared less than 20 times in the dataset would not be represented as
vectors).
Word2vec is an unsupervised machine learning model and there is no golden standard in
evaluating model performance. While some of previous studies validated model outputs by
testing if the word embeddings can be applied to pass syntactic and semantic analogy tasks (e.g.,
if the word embedding of “worse” minus the word embedding of “bad” resembles the word
embedding of “higher” minus the word embedding of “high”; e.g., Grag et al., 2018 ), they used
larger datasets with billions of words and the datasets were domain-general, whereas the dataset
used in the current study was smaller (i.e., hundreds of tweets) and domain-specific (all related to
adolescence on Twitter). Nevertheless, even with a smaller dataset, word2vec can still be applied
to synthesis information from the data and represent words into meaningful word embeddings,
given that its mechanism is to factorize a word-context matrix and embed words into the contexts
of data (Levy & Goldberg, 2013). Because the focus of the current study is on adolescence, I
validated the model outputs by examining the word embeddings of terms of adolescence (i.e.,
“adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”). I tested if the
word embeddings of these terms were close to each other (i.e., the cosine similarity scores
among them) compared to other words (i.e., the cosine similarity scores between any of the word
embedding of terms of adolescence and any of 1,000 random word embeddings in the model). I
noted that no prior information or knowledge (e.g., “adolescent” is similar to “teenager”) was fed
into the model. Therefore, if the model can automatically build closer associations among terms

36



of adolescents compared to other words, it may suggest that the concept of adolescence has been
formed in the model and can be used for further analysis.
A conceptual illustration of representing words as word embeddings is shown in Figure 1.
[Figure 1]
Public Perceptions of Adolescence
Word embedding models can capture the latent meanings of words in contexts, therefore
the model used in the current study captured the latent meanings of words in tweets, which
reflected public perceptions on Twitter. To analyze the public perceptions of adolescence on
Twitter, I focused on the word embeddings of the aforementioned seven terms (i.e., “adolescent,”
“teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”). In addition, I also created a
new word embedding as the centroid of the seven word embeddings (i.e., the center point of the
seven terms in the word embedding model, which should capture their shared meanings) as a
composite measure of the construct of adolescence. Therefore, there were altogether eight word
embeddings of adolescence as the focus of analysis, including “adolescent,” “teen,” “youth,”
“juvenile,” “puberty,” “youngster,” “tween,” and the composite measure.
For each word embedding of adolescence, I computed the cosine similarity scores (range:
0 and 1; Miklov et al., 2013) between it and the word embedding of any other word (i.e., the
descriptor, as it is not a term of adolescence but can be used to describe perceptions of
adolescence) that was ever used in the Twitter data (e.g., “rebellious,” “responsibility”). These
similarity scores capture the closeness of association between a pair of words (e.g., “teen” and
“rebellious”) in terms of their latent meanings in the data.
I employed two approaches to analyze public perceptions associated with these word
embeddings of adolescence. The first one is a theory-driven approach testing the associations

37



between terms of adolescence and other pre-defined descriptors. The second approach is datadriven without prior assumptions. Both approaches are described in the following sections.
Quantitative Analysis of Valence of Public Perceptions. Informed by previous work
that examined the stereotypes of adolescence (Gross & Hardin, 2007), which tested how human
participants linked the construct of adolescence (i.e., “teenager”) with a set of stereotypical
descriptors(e.g., “stubborn”) and another set of neutral descriptors (e.g., “chair”) compared to the
construct of adulthood (i.e., “adult”), I compared how each word embedding of adolescence (i.e.,
“adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” “tween,” and the composite
measure) as well as the word embedding of “adult” and “people” (“adult” was selected based on
Gross & Hardin’s (2007) work; “people” was selected as a more neutral term; I did not select
“child” or “children” as the comparison terms, as they sometimes refer to adolescents) was
linked with a set of stereotypical descriptors used in previous work (i.e., “rebellious,” “reckless,”
“risk-taking,” “stubborn,” “rude,” “impulsive,” “selfish,” “crazy”; adapted from Buchanan &
Holmbeck, 1998; Gross & Hardin, 2007) and another set of neutral descriptors (i.e., “get,” “say,”
“still,” “also,” “thing,” “would,” “as,” “go”; identified from the data). For each word embedding
of adolescence (e.g., “adolescent”) and the word embedding of adulthood (i.e., “adult”), I
computed the cosine similarity scores between it and any of the aforementioned stereotypical
descriptors (e.g., “rebellious”) and neutral descriptors (e.g., “get”). I then compared the average
cosine similarity scores for stereotypical descriptors and neutral descriptors using paired sample
t-tests. I further compared different terms using two-way ANOVA examining the main effects of
term and descriptor type (stereotypical vs. neutral) as well as their interaction effect on the cosine
similarity scores.

38



Informed by previous work that examined the biases of gender and ethnicity terms in
word embedding models (Caliskan, 2017; Grag et al., 2018), which tested how the word
embeddings of gender and ethnicity terms were linked with positive descriptors (e.g., “honest”)
and negative descriptors (e.g., “evil”), I compared how each word embedding of adolescence
(i.e., “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” “tween,” and the
composite measure) as well as the word embedding of adult and people was linked with a set of
positive descriptors (i.e., “agreeable,” “bright,” “diligent,” “gentle,” “helpful,” “honest,” “kind,”
“nice”; identified from the data) and a set of negative descriptors (i.e., “aggressive,” “careless,”
“dishonest,” “foolish,” “rude,” “lazy,” “selfish,” “silly”; identified from the data). For each term
of adolescence (e.g., “adolescent”) and the term of adulthood (i.e., “adult”), I computed the
cosine similarity scores between it and any of the aforementioned positive descriptors (e.g.,
“agreeable”) and negative descriptors (e.g., “aggressive”). I then compared the average cosine
similarity scores for positive descriptors and negative descriptors using paired sample t-tests. I
further compared different terms using two-way ANOVA examining the main effects of term and
descriptor type (positive vs. negative) as well as their interaction effect on the cosine similarity
scores.
A Mixed Methods Analysis of Themes of Public Perceptions. The second approach is
data-driven without prior assumptions. I extracted a list of 60 descriptors (similar to prior work
that analyzed around 60 descriptors to identity individual perceptions of adolescence, Buchanan
& Holmbeck, 1998) that are most closely associated with each word embedding of adolescence,
excluding other terms of adolescence (since these terms will be closely related to each other) and
meaningless words (e.g., “people,” “say,” “think,” etc.). I first presented the lists of descriptors in
word cloud figures, in which descriptors that were closer to the word embeddings of adolescence

39



were of larger sizes. I then conducted qualitative content analyses (Vaismoradi et al., 2016) of
these descriptors to describe public perceptions of adolescence associated with these word
embeddings of adolescence. I recruited three coders who were graduate students studying human
development. They read the descriptors and example tweets (i.e., tweets that contained a certain
term of adolescence and a descriptor), coded the descriptors into different themes, and
interpreted the themes in a qualitative manner. The themes were interpreted in terms of how they
reflect Twitter users’ perceptions of adolescence.
Since there were eight word embeddings of adolescence (i.e., “adolescent,” “teen,”
“youth,” “juvenile,” “puberty,” “youngster,” “tween,” and the composite measure), I built eight
lists with 480 descriptors (some of the descriptors may be the same across different terms of
adolescence) and 1,440 example tweets (three randomly selected tweets for each descriptors; all
coders agreed that three tweets were enough to help them identify how the descriptors can reflect
certain perceptions of adolescence, but I acknowledged that other perceptions of adolescence
may be reflected in tweets that we did not select as examples; I did not select more than three
example tweets because extracting and coding more tweets were time-consuming). I first
randomly selected 50% of the 480 descriptors as the pilot data, had the three coders code the
pilot data independently (i.e., identifying themes and assigning the descriptors into the themes),
compared and discussed their coding results to establish consensus, and then developed a coding
scheme of different themes and their coding criteria. I then had the three coders code the rest of
the descriptors based on the coding scheme. I checked interrater reliability based on Cohen’s
(1960) kappa and the interrater reliability was satisfactory (Cohen’s kappa = .86). The few
inconsistent coding results were further resolved by the team during coding meetings. This
coding approach and procedures are consistent with Qu et al. (2016).

40



A conceptual illustration of data analysis in this part is shown in Figure 2.
[Figure 2]
During the coding process, the coders and I found that the word embedding model
identified descriptors and example tweets that reflected online sexual abuse of adolescents that
are outside of the scope of the current study. I believed this was a critical issue to be addressed to
create a safe social media space for youth. However, these contents do not necessarily reflect
public perceptions of the nature of adolescence or adolescent development. As such, I removed
these descriptors in the word cloud figures and did not code descriptors or tweets related to
online sexual abuse as themes for the perceptions of adolescence. The removed descriptors and
example tweets are available upon request from me. I discussed the potential implications of
these contents in the discussion section.
Public Perceptions of Subgroups of Adolescents
The outputs of word embedding models have been found to have additive
compositionality, such that the word embeddings of different words can be combined to create a
meaningful new word embedding (Miklov et al., 2013). For instance, Miklov et al. (2013)
demonstrated that, in their word embedding model, combining the word embeddings of
“Vietnam” and “capital” generated a new word embedding that was close to the word embedding
of “Hanoi”, which was indeed the capital of Vietnam. This property of word embeddings can be
utilized to analyze public perceptions of subgroups of adolescents. Specifically, I combined the
word embedding of the composite measure of adolescence (i.e., the centroid of the word
embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”)
with other terms indicating subgroups of adolescents (i.e., race/ethnicity: “Black,” “Latinx,”
“Asian,” “White”; gender: “boy,” “girl”; sexual orientation: “gay,” “lesbian,” “LGBT”; I

41



identified these terms from the data as they were frequently used to indicate group identity; I did
not include other subgroups because they were less frequently mentioned in the dataset) to
created new word embeddings. For each word embedding of adolescent subgroups, I applied the
same set of analyses as described in the last section, including testing stereotypes in subgroups of
adolescents and qualitative content analysis based on the results from the word embedding model.
During qualitative content analysis, the coders first coded the descriptors and identified the
themes for each subgroup separately. The coders then divided the word embeddings of subgroups
of adolescents into the categories of race/ethnicity (Black adolescents, Latinx adolescents, Asian
adolescents, White adolescents), gender (male adolescents and female adolescents), and sexual
orientation (gay adolescents, lesbian adolescents, LGBT adolescents; the three groups are not
mutually exclusive; I selected these three groups because they were frequently mentioned in the
data compared to other groups) and further refined, compared, and interpreted the themes within
each category separately to facilitate the identification of themes and comparison.
Similar to the above investigation, during the coding process, the coders and I identified
descriptors and tweets that captured online sexual abuse of adolescents. I removed these
descriptors in the word cloud figures and did not code them as themes for the perceptions of
adolescence. The removed descriptors and example tweets are available upon request from me. I
discussed the potential implications of these contents in the discussion section.
Positionality Statement
As the author of the dissertation, I acknowledge my standpoint as a researcher studying
adolescent development, an educated Chinese man growing up in China, and a young adult who
has not entered the job market. As a researcher studying adolescent development, I may tend to
interpret the word embeddings in association with existing research on adolescence, which may

42



limit my ability to identify public perceptions of adolesence that have not been addressed by
literature. As an educated Chinese man growing up in China, my interpretation of the word
embeddings may be biased due to my group identity and culture. As a young adult who has not
entered the job market, my perspectives on adolescence may be constrained by my own age
group and work condition. Meanwhile, the data I analyzed were tweets on Twitter. I have a
Twitter account but I am not an active Twitter user, therefore I may be unfamiliar with certain
idioms of the tweets and may not understand the tweets well. I acknowledge my positionality
influenced the current study to some extent. I tried to avoid potential bias of data interpretation
by consulting people from diverse standpoints (e.g., active Twitter users growing up in the U.S.)
about the meanings of tweets.
The coders I recruited for qualitative analysis generally acknowledge their standpoints as
researchers or students studying human development. They may also tend to associate the word
embeddings with existing research rather than identify public perceptions of adolescence that
have not been addressed by literature.

43



CHAPTER 5: RESULTS
Descriptives
Table 1 presents the number of tweets containing the terms of adolescence (i.e.,
“adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”) with an example
tweet for each term. Table 2 presents the mean and standard deviation (SD) for four types of
sentiment scores (i.e., positive, negative, neutral, and overall) for tweets containing different
terms of adolescence. Except the term of “juvenile,” tweets containing other terms of
adolescence were overall slightly or moderately positive, whereas tweets containing the term of
“juvenile” tended to be neutral.
[Table 1]
[Table 2]
MANOVA suggested that tweets containing different terms of adolescence differed
significantly in the four types of sentiment scores, F(4, 3639482) = 8672.70, p < .001, partical η2
= .01. Post-hoc analyses using Tukey tests suggested that all pair-wise comparisons were
significant due to large sample size. From the descriptives as presented in Table 2, tweets
containing terms of youth and youngster tended to be more positive compared to other tweets,
indicating more positive public perceptions on Twitter. Tweets containing terms of juvenile and
teen tended to be more negative compared to other tweets, indicating more negative public
perceptions on Twitter.
Word Embeddings
The average cosine similarity score for all the pairs among the word embeddings of
adolescence (i.e., “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”)
was .29 (SD = .09), whereas the average cosine similarity score for any of the word embedding

44



of adolescence with any of 1,000 random word embeddings was .12 (SD = .07), suggesting that
the model successfully distinguish the concept of adolescence by associating different terms of
adolescence closer compared to other random words. Figure 3 presents a plot describing the
relative location of different terms of adolescence as well as some other words that were close to
them in the word embedding model in a two-dimensional space (i.e., the word embeddings had
300 dimensions and I used t-distributed Stochastic Neighbor Embedding to reduce the
dimensions to 2 for visualization).
[Figure 3]
Public Perceptions of Adolescence
Quantitative Analysis of Valence of Public Perceptions
Table 3 presents the average cosine similarity scores for different word embeddings of
adolescence, “adult,” and “people” in their associations with stereotypical descriptors (i.e.,
“rebellious,” “reckless,” “risk-taking,” “stubborn,” “rude,” “impulsive,” “selfish,” “crazy”) and
neutral descriptors (i.e., “get,” “say,” “still,” “also,” “thing,” “would,” “as,” “go”) as well as
results of the paired sample t-tests.
[Table 3]
All the word embeddings of adolescence and adulthood were significantly less associated
with stereotypical descriptors compared to neutral descriptors (because the neutral terms I
selected were common words that frequently occurred with other words in the data), yet such
degree differed to some degree. The word embedding of “people” tended to be more closely
associated neutral descriptors compared to stereotypical descriptors, followed by “teen,” the
composite measure of adolescence, “youth,” “adult,” and “youngster,” whereas the word
embedding of “juvenile” tended to be more closely associated with stereotypical descriptors vs.

45



neutral descriptors compared to other word embeddings, followed by “adolescent,” “puberty,”
and “tween”.
Two-way ANOVA test suggested that the main effect of word embedding was significant,
F(9,140) = 14.71, p < .001, partical η2 = .49. The main effect of descriptor type (stereotypical vs.
neutral) was significant, F(1,140) = 321.45, p < .001, partical η2 = .70. The interaction effect
between word embedding and descriptor type was also significant, F(9,140) = 5.94, p < .001,
partical η2 = .28, suggesting that different word embeddings did differ in their relative
association with stereotypical vs. neutral descriptors.
Table 4 presents the average cosine similarity scores for different embeddings of
adolescence, adult, and people in their associations with positive descriptors (i.e., “agreeable,”
“bright,” “diligent,” “gentle,” “helpful,” “honest,” “kind,” “nice”) and negative descriptors (i.e.,
“aggressive,” “careless,” “dishonest,” “foolish,” “rude,” “lazy,” “selfish,” “silly”) as well as
results of the paired sample t-tests.
[Table 4]
The word embeddings of “youth” and “youngster” were significantly more closely
associated with positive descriptors compared to negative descriptors, whereas the word
embedding of “juvenile” was significantly more closely associated with negative descriptors
compared to positive descriptors. Other terms did not differ in their associations with positive vs.
negative descriptors, yet their associations with positive descriptors tended to be slightly stronger
(except people).
Two-way ANOVA test suggested that the main effect of word embedding was significant,
F(9,140) = 3.79, p < .001, partical η2 = .20. The main effect of descriptor type (positive vs.
negative) was not significant, F(1,140) = 2.37, p = .126, partical η2 = .02. The interaction effect

46



between word embedding and descriptor type was marginally significant, F(9,140) = 1.79, p
= .075, partical η2 = .10, suggesting that different word embeddings did differ in their relative
association with positive vs. negative descriptors.
A Mixed Methods Analysis of Themes of Public Perceptions
I identified distinct patterns of themes for different word embeddings of adolescence.
Table 5 presents the summary of themes associated with different word embeddings (the
composite measure of adolescence was not included as it was comprised of themes associated
with the other terms).
[Table 5]
“Adolescent”. Figure 4 presents the descriptors associated with the word embedding of
“adolescent,” with larger font sizes indicating higher levels of closeness.
[Figure 4]
Two themes emerged for the term of “adolescent,” including:
1) Research on adolescence to be advocated: identified from descriptors that were about
research or frequently used in academic language.
“Important paper that robustly challenges claims for high incidence of bipolar disorder
in children & adolescents made by some US studies. In fact suggests much research shows
prevalence rate of 0% below 18 years of age.”
Most of the descriptors that were found to be closely associated with the term of
“adolescence” were the ones that are frequently used in academic research, with many of the
example tweets introducing research on adolescence or discussing relevant issues. The top five
descriptors that were identified into this theme were “meta-analysis,” “systematic review,”
“springerlink,” “comorbidity,” “psychosocial”.

47



2) Concern for adolescent health: identified descriptors that were related to physical or
mental health issues of adolescents.
“Sometimes children and adolescents suffer from depressive symptoms or are even have a
#depression.”
Many descriptors were related to physical and mental health issues, with some of them
also identified in the previous theme of advocating research on adolescence. The top five
descriptors that were identified into this theme were “mum-to-be,” “nutrition,” “comorbidity,”
“health,” “over-nutrition”.
“Teen”. Figure 5 presents the descriptors associated with the word embedding of “teen”.
I identified descriptors that indicated online sexual abuse of adolescents. These descriptors were
removed in the figure and not coded for public perceptions of adolescence. The removed
descriptors are available upon request from the author.
[Figure 5]
One theme emerged for the term of “teen,” including:
1) Hate and bias for teenagers. Identified from descriptors that expressed or indicated
negative attitudes towards teenagers.
“Fuck teens in your area.”
“A fucking loser that takes to social media more than a teenage girl trying to one up
another.”
The descriptors of “fuck” and “fucking” were found to be most closely associated with
the term of “teen,” with the example tweets reflecting users’ hate and bias for adolescence. Such
bias can be implicit, with several example tweets implicitly but not explicitly linked the term of

48



“teen” with undesirable characteristics. The top five descriptors that were identified into this
theme were “fuck,” “fucking,” “horny,” “dirty,” “worthless”.
“Youth”. Figure 6 presents the descriptors associated with the word embedding of
“youth”.
[Figure 6]
Three themes emerged for the term of “youth,” including:
1) Youth as an important element of society. Identified from descriptors related to social
units (e.g., “country,” “community”) and descriptors that indicated the importance of youth.
“Leaders like Hong Hoang, who mobilized a youth-led movement to create a greener
world after becoming the first Vietnamese person to visit Antarctica.”
Descriptors and example tweets in this theme highlighted the importance of youth in
society, such as leading social movements and being regarded as the future of society. The top
five descriptors that were identified into this theme were “support,” “country,” “nation,” “future,”
“community”.
2) Youth with positive characteristics. Identified from descriptors related to positive
characteristics of youth.
“Youths contribute to economy and the growth of Nigeria.”
Many descriptors and example tweets indicated the positive characteristics, such as
leadership, kindness, and prosocial attitudes. The top five descriptors that were identified into
this theme were “leader,” “great,” “young,” “good,” “well”.
3) Youth to be empowered and supported. Identified from descriptors related to help and
support.

49



“End 2018 being the change you wish to see. help us spread kindness & empower youth.
together, we can prevent #bullying.”
Some descriptors and example tweets called for empowering or supporting youth. Some
also expressed concern for youth status and the solutions to improve their status, such as youth
status in the job market. The top five descriptors that were identified into this theme were
“support,” “hope,” “help,” “opportunity,” “work”.
“Juvenile”. Figure 7 presents the descriptors associated with the word embedding of
“juvenile”.
[Figure 7]
One theme emerged for the term of “juvenile,” including:
1) Juvenile justice and delinquency. Identified from descriptors related to justice and
delinquency.
“Earlier this month, congress passed the juvenile justice and delinquency prevention act
that now awaits presidential signature.”
Most of the descriptors and example tweets were related to juvenile justice and
delinquency affairs. The top five descriptors that were identified into this theme were
“delinquent,” “delinquency,” “justice reform,” “justice,” “peer justice”.
“Puberty”. Figure 8 presents the descriptors associated with the word embedding of
“puberty”.
[Figure 8]
One theme emerged for the term of “puberty,” including:

50



1) Puberty as a period of physical development and sexual maturation. Identified from
descriptors related to physical development and sexual maturation, many of which were
physiological terms.
“Sophia began identifying as a boy at 11 years old...to halt the effects of female puberty.”
Most descriptors were physiological terms with the example tweets introducing pubertal
phenomenon or discussing relevant issues. Some also expressed concern for health issues in
puberty. The top five descriptors that were identified into this theme were “blocker,” “hormone,”
“hrt,” “estrogen,” “testosterone”.
“Youngster”. Figure 9 presents the descriptors associated with the word embedding of
“youngster”.
[Figure 9]
One theme emerged for the term of “youngster,” including:
1) Youngster sports players getting attention. Identified from descriptors related to sports.
“A mixture of first team players and youngsters will all get their chance to impress ahead
of 2019!”
Most of the descriptors were the names of young sports players or related to sports. The
example tweets expressed interests and support for youngster sports players. The top five
descriptors that were identified into this theme were “Arsche,” “player,” “Sanju Samson,”
“Minamino,” “best”.
“Tween”. Figure 10 presents the descriptors associated with the word embedding of
“tween”.
[Figure 10]
One theme emerged for the term of “tween,” including:

51



1) Tweens’ everyday life. Identified from descriptors related to tweens’ life and activities.
“What do you buy? gift ideas for teens who love dogs #giftfinder #teensandtweens”
Descriptors and example tweets were about tweens’ life in many aspects, such as reading,
friendship, birthday, etc. Many of the tweets seemed to be posted by parents of tweens, such as
finding the books or gifts for their children. The top five descriptors that were identified into this
theme were “kids book,” “childrens book,” “gift finder,” “birthday gifts,” “school library”.
Composite measure. Figure 11 presents the descriptors associated with the word
embedding of the composite measure of adolescence (i.e., the centroid of the word embeddings
of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”). I identified
descriptors that indicated online sexual abuse of adolescents. These descriptors were removed
from the figure and not coded for public perceptions of adolescence. The removed descriptors are
available upon request from the author.
[Figure 11]
Descriptors did not indicate any new themes but recalled many the themes identified
above, including youth with positive characteristics (e.g., “good,” “better”), youth to be
empowered and supported (e.g., “help,” “support”), tweens’ activities and life (e.g., “life,”
“family”), concern for adolescent health (e.g., “mum-to-be,” “dosing”), youth as an important
element of society (e.g., “today,” “change”), puberty as a period of physical development and
sexual maturation (e.g., “biological”). The coders agreed that themes associated with the
composite measure of adolescence were overall positive whereas few negative or stereotypical
descriptors were found to be associated with the composite measure.

52



Public Perceptions of Subgroups of Adolescents
Leveraging the property of addictive compositionality of word embeddings, I generated
new word embeddings that can represent subgroups of adolescents by combining certain
indicators of group identity (e.g., “boy”) and the composite measure of adolescence (i.e., the
centroid of the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,”
“youngster,” and “tween”), including male adolescents (“boy”), female adolescents (“girl”),
Black adolescents (“Black”), Latinx adolescents (“Latinx”), Asian adolescents (“Asian”), White
adolescents (“White”), LGBTQ+ adolescents (“gay,” “lesbian,” “LGBT”).
Quantitative Analysis of Valence of Public Perceptions
Table 6 presents the average cosine similarity scores for different word embeddings of
subgroups of adolescents in their associations with stereotypical descriptors (i.e., “rebellious,”
“reckless,” “risk-taking,” “stubborn,” “rude,” “impulsive,” “selfish,” “crazy”) and neutral
descriptors (i.e., “get,” “say,” “still,” “also,” “thing,” “would,” “as,” “go”) as well as results of
the paired sample t-tests. As comparisons, I also added the word embeddings of the composite
measure (i.e., the centroid of the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,”
“puberty,” “youngster,” and “tween”), adulthood (“adult”), and a neutral term (“people”).
[Table 6]
All the word embeddings were significantly less associated with stereotypical descriptors
compared to neutral descriptors (because the neutral terms I selected were common words that
frequently occurred with other words in the data), yet such degree differed by word embeddings.
The word embeddings of both male and female adolescents showed tendency towards neutral
descriptors rather than stereotypical descriptors, compared with other word embeddings. The

53



word embedding of Latinx adolescents were more closely associated with stereotypical
descriptors rather than neutral descriptors, compared with other word embeddings.
Two-way ANOVA test suggested that the main effect of word embedding was significant,
F(11,168) = 12.05, p < .001, partical η2 = .44. The main effect of descriptor type (stereotypical vs.
neutral) was significant, F(1,168) = 645.08, p < .001, partical η2 = .79. The interaction effect
between word embedding and descriptor type was also significant, F(11,168) = 5.29, p < .001,
partical η2 = .26, suggesting that different word embeddings did differ in their relative
association with stereotypical vs. neutral descriptors.
Table 7 presents the average cosine similarity scores for word embeddings of subgroups
of adolescents in their associations with positive descriptors (i.e., “agreeable,” “bright,”
“diligent,” “gentle,” “helpful,” “honest,” “kind,” “nice”) and negative descriptors (i.e.,
“aggressive,” “careless,” “dishonest,” “foolish,” “rude,” “lazy,” “selfish,” “silly”) as well as
results of the paired sample t-tests. As comparisons, I also added the word embeddings of the
composite measure (i.e., the centroid of the word embeddings of “adolescent,” “teen,” “youth,”
“juvenile,” “puberty,” “youngster,” and “tween”), adulthood (“adult”), and a neutral term
(“people”).
[Table 7]
All the word embeddings did not differ significantly in their associations with positive vs.
Neutral descriptors, yet most of them had slightly stronger association with positive descriptors
compared to negative descriptors, except gay adolescents and people.
Two-way ANOVA test suggested that the main effect of word embedding was significant,
F(11,168) = 2.52, p < .01, partical η2 = .14. The main effect of descriptor type (positive vs.
negative) was significant, F(1,168) = 2.02, p = .157, partical η2 = .01. The interaction effect

54



between word embedding and descriptor type was not significant, F(11,168) = 0.45, p = .931,
partical η2 = .03, suggesting that different word embeddings did not differ in their relative
association with positive vs. negative descriptors.
A Mixed Methods Analysis of Themes of Public Perceptions
I identified distinct patterns of themes for different word embeddings of subgroups of
adolescence. Table 8 presents the summary of themes associated with different word embeddings.
[Table 8]
Black adolescents. Figure 12 the descriptors associated with the word embedding of
Black adolescents, with larger font sizes indicating higher levels of closeness. I identified
descriptors that indicated online sexual abuse of adolescents. These descriptors were removed
from the figure and not coded for public perceptions of adolescence. The removed descriptors are
available upon request from the author.
[Figure 12]
Two themes emerged for Black adolescents, including:
1) Black adolescents engaging in social justice movements. Identified from descriptors
related to social movements.
“Stop telling the Black youth half a bread is better than nothing, we want the whole
bakery.”
Many descriptors and example tweets were related to Black movements, showing support
and proud of the Black identity, while some also expressed hate for White people. The top five
descriptors that were identified into this theme were “bbconly,” “say no to white boys,” “history,”
“black twitter movement,” “black lives matter, black is better”.

55



2) Concern for Black adolescents’ life. Identified from descriptors related to adolescents’
life and activities.
“Almost half of Black youth in the US grow up in the bottom 20% of the income
distribution.”
Similar to what I identified for the word embedding of “tween”, some descriptors were
about Black adolescents’ life in many aspects, such as fashion, school, and family. However, the
example tweets were not just about Black adolescents’ life. Many example tweets discussed
problems in or concern for Black adolescents’ life. The top five descriptors that were identified
into this theme were “poncho_free,” “melaninpoppin,” “school,” “family,” “group”.
Latinx adolescents. Figure 13 the descriptors associated with the word embedding of
Latinx adolescents, with larger font sizes indicating higher levels of closeness.
[Figure 13]
One theme emerged for Latinx adolescents, including:
1) Increasing racial equity for Latinx adolescents. Identified from descriptors related to
racial equity.
“New reports confirm #soda companies disproportionately target #latino & #black youth
with their #caffeine laced.”
Descriptors and example tweets associated with Latinx adolescents were primarily related
to racial equity, with example tweets highlighting unequal conditions of Latinx adolescents or
advocating racial equity for them. The top five descriptors that were identified into this theme
were “representationmatters,” “racialjustice,” “hispanicheritagemonth,” “educolor,”
“minoritymentalhealthmonth”.

56



Asian adolescents. Most of the descriptors were identified as related to online sexual
abuse of adolescents. Therefore, I did not present the descriptor or code them as perceptions of
adolescence. The descriptors are available upon request from the author.
White adolescents. Figure 14 the descriptors associated with the word embedding of
White adolescents, with larger font sizes indicating higher levels of closeness. I identified
descriptors that indicated online sexual abuse of adolescents. These descriptors were removed
from the figure and not coded for public perceptions of adolescence. The removed descriptors are
available upon request from the author.
[Figure 14]
Two themes emerged for White adolescents, including:
1) Hate and Bias for White adolescents. Identified from descriptors that expressed or
indicated negative attitudes towards White adolescents.
“Stupid white girls.”
“What’s happening in the home that’s making these White teenage boys shoot up schools?”
Similar to what I identified with the word embedding of “teen,” many descriptors and
example tweets associated with White adolescents expressed hate and bias for White adolescents.
Such attitudes can be held implicitly but not explicitly expressed, such as the aforementioned
example in which the users associated White teenage boys with school shooters. The top five
descriptors that were identified into this theme were “fuck,” “fucking,” “wrong,” “bad,” “weird”.
2) White adolescents involved in discussion about race and racism. Identified from
descriptors related to race and racism.

57



“If you are a white female teen and have a black boyfriend, don't have fear, at least one of
your friends will follow your example soon! #irmovement #womengoingblack #interraciallife
#junglefever #blackisbetter #blackguysonly #saynotowhiteboys2019”
Some descriptors and example tweets criticized White adolescents for being a racist or
supremacist, yet some advised them not to do so. The top five descriptors that were identified
into this theme were “racist,” “supremacist,” “supremacy,” “white nationalism,” “white
supremacist terrorism”.
Comparison between subgroups by race. I identified a theme related to racism and
racial movements for both Black, Latinx, and White adolescents, yet such theme was missing for
Asian adolescents. White adolescents had a theme of hate and bias for White adolescents, yet this
theme was not identified for other subgroups, suggesting that the negative and stereotypical
perception of adolescence may be mainly associated with White adolescents.
Male adolescents. Figure 15 presents the descriptors associated with the word
embedding of male adolescents, with larger font sizes indicating higher levels of closeness. I
identified descriptors that indicated online sexual abuse of adolescents. These descriptors were
removed from the figure and not coded for public perceptions of adolescence. The removed
descriptors are available upon request from the author.
[Figure 15]
One themes emerged for male adolescents, including:
1) Male adolescents’ everyday life. Identified from descriptors related to male
adolescents’ life and activities.
“A teenage boy on holiday”

58



Many descriptors and example tweets were about male adolescents’ life in many aspects,
such as school, family, and friendship. The top five descriptors that were identified into this
theme were “school,” “family,” “son,” “group,” “home”.
Female adolescents. Figure 16 presents the descriptors associated with the word
embedding of female adolescents, with larger font sizes indicating higher levels of closeness. I
identified descriptors that indicated online sexual abuse of adolescents. These descriptors were
removed from the figure and not coded for public perceptions of adolescence. The removed
descriptors are available upon request from the author.
[Figure 16]
One themes emerged for female adolescents, including:
1) Female adolescents’ everyday life. Identified from descriptors related to female
adolescents’ life and activities.
“The retention rate for girls playing youth sports can be so much better.”
Some descriptors and example tweets were about female adolescents’ life in many
aspacts, such as school, family, friendship, sports, and entertainment. The top five descriptors
that were identified into this theme were “family,” “daughter,” “school,” “friend,” “life”.
Comparison between subgroups by gender. All the coders agreed that although both
male and female adolescents were identified with the theme about their everyday life. Such
theme was more evident for male adolescents compared to female adolescents (i.e., more
descriptors were identified into the theme for male adolescents compared to female adolescents).
Gay adolescents. Most of the descriptors were identified as related to online sexual
abuse of adolescents. Therefore, I did not present the descriptor or code them as perceptions of
adolescence. The descriptors are available upon request from the author.

59



Lesbian adolescents. Most of the descriptors were identified as related to online sexual
abuse of adolescents. Therefore, I did not present the descriptor or code them as perceptions of
adolescence. The descriptors are available upon request from the author.
LGBT adolescents. Figure 17 presents the descriptors associated with the word
embedding of LGBT adolescents, with larger font sizes indicating higher levels of closeness.
[Figure 17]
One theme emerged for gay adolescents, including:
1) Increasing awareness and supporting LGBT adolescents. Identified from descriptors
related to LGBTQ+ movements.
“#hope #loveislove one man's advice for friends and relatives of lgbtq youth.”
Most of the descriptors associated with LGBT adolescents were about LGBTQ+
movements, with example tweets expressing support of these movements or concern for the
unfair condition of LGBTQ+ adolescents. The top five descriptors that were identified into this
theme were “pridemonth,” “homophobia,” “loveislove,” “lgbtqoftwitter,” “nooutsiders”.
Comparison between LGBTQ+ subgroups. All the coders agreed that the descriptors
associated with LGBT adolescents were different from the other two subgroups (i.e., gay
adolescents and lesbian adolescents) as the theme was mainly about supporting LGBTQ+
adolescents.

60



CHAPTER 6: DISCUSSION
Social media is an influential developmental context that young people have contact with
on a daily basis. However, we currently know little about how social media portraits adolescents,
which may have important implications for young people’s everyday well-being and how young
people think about themselves. My dissertation addressed this critical gap in the literature by
exploring public perceptions of adolescence on Twitter using a big data, machine learning
approach. I first employed a theory-driven and quantitative approach to test how various terms of
adolescence (i.e., “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”)
was associated with stereotypical (vs. neutral) or positive (vs. negative) descriptors compared to
the terms of adult and people. Results showed that the overall construct of adolescence was not
perceived as negative or stereotypical but relatively neutral and unstereotypical. Moreover,
perceptions varied by different terms of adolescence (e.g., “youth” vs. “juvenile”), suggesting
that adolescence is a comprehensive construct that may not be captured by a single term. I then
developed a data-driven and mixed-methods approach in which I generated lists of 60 descriptors
that were most closely associated with terms of adolescence and then conducted qualitative
content analysis on the descriptors while reference example tweets containing the descriptors.
For each term of adolescence, a variety of themes were identified including not only negative
themes but also neutral and positive themes. These themes highlighted public perceptions of
adolescence in various aspects and many of them were meaningful in terms of promoting
adolescent development. I also explored variations in perceptions of adolescents by race/ethnicity,
gender, and sexual orientation, highlighting both risk and resilience processes within each
adolescent subgroup. In this section, I discussed the details of my findings as well as their
interpretation and implications for future research and practices. I also discussed the

61



methodology of my dissertation in terms of its strengths, implications for future research, and
limitations.
Understanding Adolescence Using the Word Embedding Model
The word embedding model was built by feeding cleaned tweets into the algorithm. The
model then automatically learned the associations of different words in the data and represented
each word in the dataset as a word embedding with 300 dimensions. Although there has been no
absolute measure to evaluate model performance, I found evidence suggesting that the word
embedding model I built successfully distinguished the concept of adolescence. Without any
prior information or knowledge, the model associated terms of adolescence (i.e., “adolescent,”
“teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”) closer to each other compared
to other random words, indicating that the model captured the concept of adolescence based on
these terms. Furthermore, other words that were found to be closely associated with the terms of
adolescence appeared to be meaningful and capture public perceptions of adolescence, such as
words related to teenagers’ life and environment (e.g., “parent,” “friend,” “school”), youth’s role
in society (e.g., “country,” “world,” “change”).
While existing literature has employed word embedding models to extract and analyze
implicit meanings of constructs such as social class (Kozlowski et al., 2019) and race/ethnicity
(Grag, 2018) in large corpora, the data they used were relatively large (i.e., billions of words
were fed into the model), domain-general (i.e., the content of the data were not limited to a
specific topic), and formal (i.e., the data were from published news or books). My dissertation
further validated the use of word embedding models to explore the implicit meanings of
adolescence in a dataset that was relatively small (i.e., millions of tweets, with total word count
less than one billion), domain-specific (i.e., the data were tweets related to the discussion of

62



adolescence), and informal (i.e., the language used in tweets was somewhat informal). Therefore,
my dissertation may inform future work to apply word embedding models to smaller datasets to
analyze the meanings of words in specific contexts.
This approach may contribute to existing literature in two ways. First, existing studies
analyzing people’s perceptions or opinions of certain constructs usually recruited a small sample
of participants and intentionally collected their responses using structured questions (e.g.,
parental opinion of school sexuality education; Heller & Johnson, 2012). It is possible to collect
a textual dataset consisting of people’s expression of opinions around certain topic and then use
word embeddings to identify people’s opinions associated with relevant terms. This approach
may complement existing studies by leveraging bigger data (because it may be easier to access
larger textual datasets compared to recruiting participants) and identifying opinions in a datadriven manner (rather than theory-driven).
Second, existing studies analyzing big textual data usually employed a series of
traditional linguistic modeling to analyze the data, such as latent Dirichlet allocation (LDA, Blei
et al., 2003) and Linguistic Inquiry and Word Count (LIWC, Pennebaker et al., 2001), these
modeling techniques have been shown to be less accurate compared to word embeddings (e.g.,
Rudkowsky et al., 2018). It is possible to update such analyses using word embeddings to better
extract information from the data. For instance, Kern et al. (2013) analyzed over 70,000
Facebook users’ posts to examine the aging positivity effect (i.e., older people are happier than
young people) with traditional modeling techniques, whereas it is now possible to update their
analysis by feeding data into a word embedding model and test how the term of “old” is
associated with terms indicating positive emotions compared to the term of “young”.

63



Public Perceptions of Adolescence on Twitter
Quantitative Analysis of Valence of Public Perceptions
I used two different approaches using word embeddings to analyze public perceptions of
adolescence on Twitter. The first one was a theory-driven approach in which I tested how the
word embeddings of adolescence were associated with two sets of pre-defined descriptors (i.e.,
stereotypical vs., neutral descriptors, positive vs. negative descriptors) in comparison with other
terms (i.e., “adult” and “people”). Results suggested that although the overall construct of
adolescence as reflected by the composite measure (i.e., the centroid of the word embeddings of
“adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and “tween”) did not show
higher tendency towards stereotypical/negative descriptors vs. neutral/positive descriptors,
compared with the word embeddings of “adult” and “people”, different terms varied in their
associations with stereotypical/negative descriptors vs. neutral/positive descriptors. Specifically,
the word embeddings of “juvenile,” “adolescent,” and “puberty” showed higher levels of
tendency towards stereotypical descriptors (vs. neutral descriptors) compared with “teen,”
“youth,” and “youngster”, indicating relatively stereotypical perceptions associated with these
terms. The word embedding of “juvenile” showed higher levels of tendency towards negative
descriptors (vs. positive descriptors) compared with other terms, indicating negative perceptions
associated with the term of “juvenile”. The word embedding of “youth” and “youngster” showed
higher levels of tendency towards positive descriptors (vs. negative descriptors) compared with
other terms, indicating positive perceptions.
Based on my exploration of the data and the qualitative content analysis in the next
section, the finding that the term of “juvenile” was associated with more stereotypical and
negative descriptors may not indicate Twitter users’ perceptions, but rather reflect the language

64



habit that “juvenile” is usually used in delinquency or criminal events, which are also associated
with some stereotypical descriptors (e.g., “rebellious,” “risk-taking”) and negative descriptors
(e.g., “aggressive,” “foolish”).The findings that the terms of adolescent and puberty also showed
more stereotypical perceptions may also be impacted by language habit as the two terms were
usually used in academic language and the tweets were mostly about research and popular
science. In this sense, stereotypical descriptors that were associated with the terms of adolescent
and puberty may not reflect Twitter users’ stereotypical perceptions, but rather indicate that
tweets talking about research and knowledge on adolescence are more focused on negative
aspects, such as rebellion and risky behavior. Indeed, Rich (2003) suggested that contemporary
research on adolescence still focuses more on negative aspects. Except for juvenile, all the other
terms of adolescence did not show significant tendency towards negative (vs. positive)
descriptors, suggesting that the overall public perceptions of adolescence is not negative but
neutral on Twitter. Although I identified several negative themes in the qualitative analysis, such
negative themes and tweets were found to be not prevalent in the dataset, and there were other
positive and neutral themes. Furthermore, the terms of youth and youngster showed positive
perceptions, with the qualitative analysis identified primarily positive themes associated with the
two terms, such as youth with positive characteristics and youngster sports players getting
attention.
These findings suggested that adolescence is a comprehensive construct and may not be
fully captured by a single term. Gross and Hardin (2007) used the term “teenager” in comparison
with “adult” and conducted an IAT experiment, and results suggested that participants held an
implicit stereotype that they tended to associate “teenager” with stereotypical descriptors (vs.
neutral descriptors; the stereotypical descriptors they used were the same as I used in the current

65



study) compared to “adult”. The results of my dissertation also suggested that the word
embedding of “teen” showed higher levels of tendency towards stereotypical descriptors (vs.
neutral descriptors; mean difference = -.09) compared to the word embedding of “adult” (mean
difference = -.19). However, Gross and Hardin’s (2007) finding may be limited as they only
tested the term of “teenager”. Switching to other terms such as “youth” may result in less
stereotypical results, as indicated by my dissertation. Therefore, future work analyzing
perceptions of adolescence may consider the use of different terms rather than using a single
term to indicate adolescence.
Furthermore, existing research analyzing individual perceptions of adolescence generally
indicated that people hold overall negative and stereotypical perceptions of adolescence (e.g.,
Buchanan & Holmbeck, 1998) or focused on the negative and stereotypical aspects (e.g.,
Buchanan & Hughes, 2009; Qu et al., 2016). However, my dissertation suggested that the overall
construct of adolescence (i.e., the centroid of the word embeddings of “adolescent,” “teen,”
“youth,” “juvenile,” “puberty,” “youngster,” and “tween”) was not perceived as negative or
stereotypical. This finding may inform future research to focus more on positive or neutral
perceptions of adolescence.
Besides, although theoretical frameworks such as positive youth development have
advocated positive perceptions of adolescence (e.g., Lerner et al., 2015) and non-profit
organizations such as the FrameWorks Institute has been calling for shifting the frame of how the
public perceive adolescence (Kendall-Taylor, 2020), existing research has rarely suggested ways
to shift public perceptions of adolescence into more positive forms. My finding that the terms of
“youth” and “youngster” tended to be more positive and unstereotypical may inform education
and communication practices to shift the framing of adolescence by using these terms more when

66



referencing adolescence. Since the terms of “youth” and “youngster” are already perceived as
positive and unstereotypical by the public, using these terms to reference adolescence may
influence the readers to associate positive perceptions with adolescence, therefore re-framing
public perceptions of adolescence. Communication research has already demonstrated that
shifting the terms can help re-frame public perceptions of social events such as climate change
movements (e.g., Newell et al., 2015).
A Mixed Methods Analysis of Themes of Public Perceptions
Before discussing the results, I note that the descriptors and themes associated with terms
of adolescence should be understood as reflecting unique associations rather than common
associations. For instance, it is possible that positive descriptors and themes also exist in tweets
containing the term of “teen,” but if they are more prevalent in tweets containing the term of
“youth,” they will be more likely to be identified with the term of “youth” rather than the term of
“teen”. Therefore, although I discussed each term of adolescence and its unique association with
certain descriptors and themes, the results should be understood conjointly across different terms
of adolescence, which were captured by the composite measure of adolescence. I also note that
the quantitative analysis of valence of public perceptions may not be consistent with the mixed
methods analysis of themes of public perceptions, as the prior analysis was theory-driven based
on pre-defined descriptors from existing literature but the latter analysis was data-driven based
on the outputs of word embeddings. Most of the pre-defined descriptors in the quantitative
analysis did not appear to be closely associated with terms of adolescence in the mixed methods
analysis.
Quantitative analysis of textual data may be restricted in its ability to extract nuances
from the texts and therefore ignores the meaningful contexts of texts. To better dig information

67



from the data, I developed a mixed-methods approach in which I first extracted descriptors that
were closely associated with the word embeddings of adolescence using the word embedding
model, and then conducted qualitative content analysis by coding the descriptors and referencing
example tweets containing the descriptors. The qualitative content analysis identified distinct
patterns of themes for different word embeddings of adolescence, echoing the findings that
different word embeddings of adolescence showed different levels of stereotypes in quantitative
analysis.
Specifically, the word embedding of “adolescent” was found to be associated with themes
of research on adolescence to be advocated and concern for adolescent health, reflecting the
academia’s interests in adolescence. Indeed, adolescent health has been acknowledged as a
central issue in contemporary study of adolescence (e.g., Patton et al., 2016) and an emerging
line of work has begun to investigate the role of social media on adolescent health (e.g., Shaw et
al., 2015). My dissertation suggested that discussion of the word “adolescent” on Twitter was
centered around the themes of adolescent research and health, which may add to this line of work
and inform future study to analyze specific contents of discussion about adolescent health on
Twitter. While researchers studying health communication have investigated general health
communication on Twitter, such as the topics of health communication and the role of health
organizations (e.g., Park et al., 2016), such investigation is rare on adolescent health
communication (e.g., Barry et al., 2016). My dissertation provides initial clue that there was
discussion about adolescent health that served as an eminent theme in Twitter users’ discussion
of adolescence, and future work may analyze such discussion to promote the understanding of
adolescent health communication on Twitter as well as to facilitate positive health
communication.

68



The word embedding of “teen” was found to be associated with the theme of hate and
bias for teenagers, suggesting that although contemporary research on adolescence has advocated
a positive view of adolescence (e.g., positive youth development; Benson, 1997; Lerner et al.,
2005) and emphasized multiple themes of adolescent development such as ecological systems
(Bronfenbrenner, 1979, 1995) and identity development (Erikson, 1968), the storm-and-stress
view of adolescence (Hall, 1904) may still be influential and the negative/stereotypical
perceptions of adolescence as identified in previous research (e.g., Buchanan & Holmbeck, 1998;
Qu et al., 2016) may still be prevalent in contemporary social media, with social media users
explicitly expressing hate and bias towards adolescence or implicitly associating adolescence
with undesirable characteristics. This type of posts may further strengthen the storm-and-stress
view of adolescence by influencing other social media users who join the discussion or read the
messages. This type of posts may also impact adolescent well-being by hurting adolescent social
media users who encounter these messages. Indeed, adolescents’ use of social media has been
found impact their well-being and one of the underlying mechanisms is through their browsing
behavior, in which negative information may increase negative emotions such as depressive
mood (Weinstein, 2018). Education and communication practitioners may take steps to reduce or
counter such posts in order to shift the negative framing of adolescence and protect adolescents
from reading negative information.
Themes associated with “youth” were found to be quite positive, including youth as an
important element of society, youth with positive characteristics, and youth to be empowered and
supported. These themes are in line with the framework of positive youth development (PYD).
Specifically, the themes I identified met Benson’s (1997) advocate of empowering youth from
multiple social contexts (e.g., community, school, and family). The descriptors and example

69



tweets I coded acknowledged the value of youth in society and proposed empowerment of youth.
The relatively positive themes of youth on Twitter may partly reflect PYD’s efforts of proposing
positive framing of youth. They can also be used by education and communication practitioners
to strengthen positive perceptions of adolescence and counter negative perceptions in the future.
Themes associated with “juvenile,” “puberty,” and “youngster” were found to be closely
associated with specific fields (i.e., justice, physiology, and sports), including juvenile justice and
delinquency, puberty as a period of physical development and sexual maturation, and youngster
sports players getting attention, due to the relatively narrower usage in the English language. The
term of “tween” was found to be associated with the theme of tweens’ everyday life, suggesting
that “tween” may be a neutral term that is broadly linked with various aspects of adolescence.
These themes highlighted a diverse range of themes that Twitter users associated with
adolescence, rather than just the negative or stereotypical perceptions that were focused on in
previous studies (e.g., Buchanan & Holmbeck, 1998; Qu et al., 2016)
The composite measure of adolescents was found to capture multiple themes associated
with different terms of adolescence, with positive and neutral themes being more evident
compared to negative themes. This finding suggested that the overall public perceptions of
adolescence may be positive on Twitter, echoing my quantitative analysis in the prior section.
While existing research investigating perceptions of adolescence mainly focused on the negative
aspects (e.g., Buchanan, 2003; Qu et al., 2016), my analysis suggested that positive and neutral
perceptions may be more evident on social media. Although negative perceptions of adolescence
may impact adolescent development such that parents who held negative perceptions of
adolescents tended to have their own children showing parallel behavior (e.g., Jacobs et al., 2005)
and adolescents who held negative perceptions of their age group tended to have less family

70



responsibility and school engagement (Qu et al., 2016), positive and neutral perceptions of
adolescence may also influence adolescent development, such as the PYD advocates positive
perceptions of adolescence to facilitate positive youth development (e.g., Benson, 1997).
Therefore, future work may focus more on the positive and neutral aspects of public perceptions
of adolescence to fill in the gap of literature and better promote positive youth development.
Overall, combining word embeddings with qualitative content analysis, I was able to
generate multiple meaningful themes associated with different terms of adolescence. While some
of the themes can be indicated from existing literature, such as hate and bias for teenagers as
indicated by the storm-and-stress view of adolescence (Hall, 1904), I also identified other themes
that has rarely been discussed in previous research, such as youth as an important element of
society, and youth to be empowered and supported. These themes suggested that more aspects of
public perceptions of adolescence are present in social media, rather than just the storm-andstress view. Some of the themes are worth attention from future researchers and practitioners to
create a more positive and healthy environment for positive youth development. For instance, the
positive themes associated with “youth” (i.e., youth as an important element of society, youth
with positive characteristics, and youth to be empowered and supported) may be utilized by
education and communication practitioners to re-frame public perceptions of adolescence into
more positive forms by referencing adolescence as youth. While existing researchers have
emphasized the negative impact of negative perceptions of adolescence (e.g., Buchanan, 2003),
my dissertation also contributes to this line of work by identifying the theme of hate and bias for
teenagers, suggesting that such negative and stereotypical perception still exist on social media
and practitioners may take actions to reduce such perception to facilitate positive youth
development.

71



Methodologically, while there is an emerging line of work that advocates the use of
computational methods to combine computational methods with qualitative analysis (e.g.,
Nelson, 2020), to my knowledge, my dissertation is among the first set of studies that integrated
machine learning with qualitative content analysis. I leveraged the word embedding model’s
ability to synthesize information from big data and capture latent meanings of terms. To embed
abstract results generated by the word embedding model into more detailed contexts, I developed
a coding approach through which coders read the descriptors and example tweets to identify
qualitative themes from the results. This approach was able to identify meaningful themes that
captured public perceptions of adolescence on Twitter and demonstrated advantages over pure
quantitative or qualitative analysis. Compared to pure quantitative analysis, my approach
combined quantitative results (i.e., the top 60 descriptors that were most closely associated with
the term of adolescence) with qualitative analysis of the descriptors and example tweets, which
helped me to put the descriptors into specific contexts of tweets and better understand how and
why the descriptors were associated with adolescence. Without the qualitative analysis,
quantitative results of how a certain term of adolescence was associated with a certain descriptor
would be difficult to interpret. Compared to pure qualitative analysis, my approach was able to
synthesis information from millions of tweets and generate the most relevant descriptors
associated with adolescence. Without the quantitative word embedding model, it would be
impossible to process millions of tweets and identify the most relevant themes associated with
adolescence. Future work can leverage or further develop this analytic approach to better analyze
big textual data and promote the understanding of social/psychological constructs in big data.

72



Public Perceptions of Subgroups of Adolescents
Quantitative Analysis of Valence of Public Perceptions
Applying the same analytic approach to test the quantitative association between different
word embedings of subgroups of adolescence and pre-defined sets of descriptors, I also found
that different word embedings of subgroups of adolescence varied in their associations with
stereotypical/negative descriptors vs. neutral/positive descriptors. Specifically, focusing on
race/ethnicity (i.e., Black, Latinx, Asian, and White), Latinx adolescents showed higher levels of
tendency towards stereotypical descriptors (vs. neutral descriptors), yet they also showed slightly
higher levels of tendency towards positive descriptors (vs. negative descriptors). The other
race/ethnic groups (i.e., Black, Asian, and White) did not show significant tendency towards
stereotypical descriptors (vs. neutral descriptors) or positive descriptors (vs. negative descriptors).
I discussed potential explanation in the next section of themes of public perceptions. Male
adolescents and female adolescents did not show much difference in their associations with
stereotypical/negative descriptors vs. neutral/positive descriptors. The two subgroups both
appeared to be relatively neutral compared with other subgroups of adolescence. The results
suggested that gender-related bias and stereotypes may not be evident in tweets related to
adolescence on Twitter. Focusing on sexual orientation (i.e., gay, lesbian, and LGBT), LGBT
adolescents were found to show higher levels of tendency towards positive descriptors (vs.
negative descriptors). I also discussed potential explanation in the next section of themes of
public perceptions.
Although there is a handful research analyzing public perceptions of adolescence for the
general adolescent population (e.g., Buchanan, 2003), few existing studies have explicitly
investigated public perceptions of subgroups of adolescence, probably due to the complexity that

73



such investigation requires examining both group identity and the developmental period of
adolescence. Therefore, my analytic approach may inform future work to analyze complex
constructs by combining different word embeddings to generate new constructs that captured the
latent meanings from different concepts. For instance, my dissertation combined the composite
measure of adolescence with the word embeddings of group membership indicator (i.e., “boy”
and “girl”). Future work may explore other combination of word embeddings (e.g., combining
the word embedding of “adolescence,” “boy,” and “Black”) to analyze public perceptions of
more specific subgroups of adolescents.
A Mixed Methods Analysis of Themes of Public Perceptions
Subgroups by race/ethnicity. I observed both similarities and notable differences in the
themes associated with adolescents of different racial/ethnic subgroups. For Black, Latinx, and
White adolescents, I observed themes about social justice movements, race/ethnicity and racism,
with Black and Latinx adolescents’ participation in anti-racism movements been discussed or
promoted, and White adolescents’ passively involved into such discussion. Racism is an
important context for adolescent development in the United States, with White individuals being
prioritized over racial/ethnic minorities in the society (García Coll et al., 1996; Roberts, 2012),
which can operate through media (Jones, 1997). Racial/ethnic minority adolescents who are
more stressful experiences of racism tend to have more engagement in anti-racism actions (e.g.,
Hope et al., 2022). Racial/ethnic minority adolescents’ participation in discussions related to
social movements may also serve as a socialization process, especially for the development of
their racial/ethnic identity when engaging in anti-racism movements (Umaña‐Taylor et al., 2014).
Therefore, experiences of racism may engage Black and Latinx adolescents into social justice
and anti-racism movements while also involving While adolescents into discussion, leading to

74



the themes I observed. These themes also reflected racial/ethnic adolescents’ active civic
engagement. Indeed, adolescent civic engagement has become an important theme in adolescent
development (Chan et al., 2014), with adolescents highly engaged in social movements such as
the Black Lives Matter movement (Baskin-Sommers et al., 2021). Future research on
racial/ethnic minority adolescents’ engagement in social justice movements may pay attention to
the context of social media as currently there are few studies focusing this context.
Although experiences of racism and youth participation in anti-racism movements are
also salient for Asian adolescents (e.g., Ballard, 2016), these themes were not identified in my
analysis. This suggests that Asian adolescents’ racial/ethnic experiences may be overlooked on
Twitter. Indeed, recent research has already begun to discuss how racism against Asian youth
may be overlooked compared to anti-Blackness (Lee et al., 2022). Future work may further
explore the lack of representations and voices for Asian youth on social media and develop
intervention programs to address this issue, such as identifying the reasons why voices for Asian
youth are missing and developing programs to encourage Asian youth to express their voices.
Education and communication practitioners may pay more attention to racial/ethnic minority
groups other than the Black and Latinx to facilitate the discussion and movements of anti-racism
for these groups.
For White adolescents, the theme of hate and bias for teenagers was specifically
associated with this subgroup rather than others, suggesting that the storm-and-stress view of
adolescence may be primarily associated with White adolescents. This finding resonates with a
long-lasting discussion regarding the nature of storm and stress in adolescence. Whereas Stanley
Hall (1904) suggested that storm and stress is universal yet acknowledging cultural variations,
Margret Mead (1928) argued that storm and stress is not prevalent in non-western cultures. More

75



recent theoretical work generally acknowledged cultural variations and individual differences of
adolescent development. For instance, Bronfenbrenner (1979, 1995) highlighted the impact of
ecological systems such as culture and community on adolescent development. The positive
youth development framework also emphasizes the role of culture on positive development,
suggesting that the meanings of positive development may differ across cultural contexts (Shek
et al., 2019). While most of existing research has not empirically compared the storm-and-stress
view of adolescence for different subgroups of adolescence, Qu and colleagues (2016) found that
U.S. youth (73% White) were more likely to see adolescence as a time of less family
responsibility and school engagement compared to Chinese adolescents. My dissertation did not
evaluate the actual levels of storm and stress in various subgroups, but suggested that the
perceptions of adolescence as storm and stress may be primarily linked with White adolescents
on Twitter, echoing Qu and colleagues’ (2016) work.
Subgroups by gender. For both gender groups, I identified the theme of adolescents’
everyday life, yet this theme was more evident for male adolescents compared to female
adolescents, whereas descriptors that indicate online sexual abuse and sexual objectification
were more prevalent for female adolescents compared to male adolescents, indicating potential
gender inequity. I discussed potential issues in the section of Sexual Content and Potential Issues
of Online Sexual Abuse and Sexual Objectification.
Subgroups of sexual minority adolescents. For LGBT adolescents, descriptors and
example tweets expressed support for this subgroup, which may promote positive development
for LGBT adolescents. Research has shown that LGBTQ+ adolescents use social media for selfexpression and seeking social support (Selkie, et al., 2020) and social media use has been found
to impact their sexual identity and well-being such as reducing their mental health concerns and

76



promoting positive emotions (e.g., Berger et al., 2022). Therefore, future work and practices may
examine how discussion about LGBTQ+ adolescents may promote the well-being of them and
facilitate positive discussion on Twitter.
Sexual Content and Potential Issues of Online Sexual Abuse and Sexual Objectification
Unexpectedly, this dissertation study identified descriptors and tweets that may indicate
online sexual abuse and sexual objectification of adolescents in online spaces. Although I did not
proceed into coding these descriptors as they are outside of the scope of the current study and do
not necessarily reflect public perceptions of the nature of adolescence or adolescent development,
I discussed potential implications of the findings here.
For the word embedding of “teen”, I found that many descriptors were related to sexual
contents, with the example tweets displaying teenagers as an attractive point to distribute sexual
contents, even though the actual contents may be performed by adults. This may indicate sexual
objectification (e.g., Fredrickson & Roberts, 1997) of teenagers. The original objectification
theory suggested that women’s body tends to be treated as an object for the pleasure and use of
others in society, which can be reflected by the representation of women in the media
(Fredrickson & Roberts, 1997). Recent work on sexual objectification has paid attention to the
objectification of adolescents (e.g., Slater & Tiggemann, 2010). My finding may contribute to
this line of work by showing that adolescents are objectified in sexual contents on the Internet.
On the one hand, using teenagers as attractive points in sexual contents may indicate loopholes in
the current law and justice system on child sexual abuse material. Whereas visual depiction of
sexually explicit conduct involving someone under 18 years of age has been clearly identified as
violating federal law on child sexual abuse material (Quayle & Taylor, 2004), textual depiction
of sexual content involving someone under 18 years of age may receive less supervision. The

77



finding of my dissertation highlighted the need to supervise textual contents that may relate to
child sexual abuse material on the Internet. On the other hand, the association between teen and
sexual contents may also reflect sexual development and an emerging interest in sexuality among
adolescent users themselves (Ponton & Judice, 2004). Sexual contents that used teenagers as
attractive points may appeal to adolescent Twitter users and impact their well-being. Adolescents’
use of sexual material and exposure to sexual contents on the Internet has also been found to
impact adolescent development. For example, research shows that increasing use of Internet
pornography is associated with decreased male adolescents’ academic performance six months
later (Beyens et al., 2014), and more frequent use of sexually explicit Internet material led to
more permissive attitudes and more frequent experience with sexual behavior (Doornwarrd et al.,
2015). Therefore, the spread of sexual messages about teenagers on Twitter may also impact
adolescent well-being, and future work and practices may target this phenomenon to promote
positive sexual development in adolescence.
For the word embedding of adolescent subgroups, I found that almost all the top 60
descriptors associated with Asian adolescents, gay adolescents, and lesbian adolescents were
related to sexual contents, with the example tweets displaying these groups as an attractive point
to distribute sexual contents. Although I also identified similar descriptors for other subgroups
(e.g., Black and White adolescents), they were less prevalent. Similarly, I found that descriptors
related to sexual contents were more prevalent for the word embedding of female adolescents
compared to male adolescents. Moreover, when the descriptors and example tweets for male
adolescents contained sexual contents, they were targeting more specifically for gay adolescents.
These findings highlight certain adolescent subgroups (e.g., female, Asian, gay, and lesbian
adolescents) as vulnerable groups that may be particularly targeted by sexual objectification and

78



child sexual abuse materials, who may need particular attention in research and practices that
seek to address inequality in online spaces.
Overall, these findings suggest that potential issues of sexual objectification of
adolescents and the spread of child sexual abuse materials may exist on Twitter. They may not
necessarily reflect Twitter users’ perceptions of adolescence, as many of the tweets may be
posted by robot or commercial accounts that tried to propagandize sexual materials, yet they also
reflected how adolescents are depicted in public spaces, as the tweets were all publicly available
in 2019. For education and communication practitioners as well as law makers, my findings may
inform practices to address the issues of sexual objectification of adolescents and the spread of
sexual abuse materials on Twitter, with the goal of reducing relevant messages on social media.
Moreover, Asian adolescents may require more attention as the issues were more evident among
them compared to other racial/ethnic groups. Female adolescents may require more attention as
they were more frequently targeted for sexual contents compared to male adolescents.
Practitioners may also examine homosexual contents targeting male adolescents, gay and lesbian
adolescents.
Implications and Limitations
Theoretical and Methodological Implications
While prior research on perceptions of adolescence mainly explored individuals’ negative
perceptions of adolescence in the family and school contexts, the current study shifted the
literature to the broader context of social media and explored multiple themes of perceptions,
which may contribute to a more comprehensive understanding of public perceptions of
adolescence. Future work may build upon the current study to explore more themes and nuances
in public perceptions of adolescence as well as to explore this issue in other contexts such as

79



TikTok, a more popular social media platform among adolescents. Whereas Twitter data are textbased and public perceptions of adolescence can be extracted using textual analysis, data from
TikTok are video-based and may require a different methodology for data analysis.
Although developmental theories such as the ecological systems theory (Bronfenbrenner,
1979) have long emphasized the importance of macrosystem such as sociocultural values on
adolescent development, existing research on adolescence has focused more on individuals and
their proximal environments (e.g., family and school; Steinberg & Morris, 2001). While scholars
have increasingly realized the importance of values and ideologies in the macrosystem and their
impact on micro-level processes (e.g., structural racism; Rogers et al., 2021), few studies have
directly examined factors in the macrosystem. Leveraging big data on social media and machine
learning, the current study focused on the broader context of social media and explored how the
developmental period of adolescence is perceived in this broader context. This approach may
implicate future research to pay more attention to macro-level factors that impact development
and directly examine such factors using big data and machine learning.
Conventional social science research tends to employ a deductive or theory-driven
approach, in which researchers fit data to the theory (Grimmer et al., 2021). This approach has
also been endorsed by existing research on perceptions of adolescence, with most studies
adopted the storm-and-stress view of adolescence and focused more on the negative perceptions
of adolescence. However, the current study employed an inductive or data-driven approach that
first utilized an unsupervised, data-driven algorithm to extract information from the data, which
may enable us to transcend the storm-and-stress view of adolescence and discover more themes
around perceptions of adolescence from the data. The descriptors as identified in my dissertation
were purely data-driven, and the selection process was not impacted by researchers’ prior

80



knowledge or bias (e.g., researchers believing the storm-and-stress view may select negative
descriptors over positive ones). This approach may inform future research to leverage datadriven approaches for research questions that have not been directly addressed by theories.
Practical Implications
While the PYD framework propose to focus on the positive aspects of adolescent
development, which inherently advocates positive perceptions of adolescence, American adults
and adolescents have still been found to endorse negative and stereotyped perceptions of
adolescence (Buchanan & Holmbeck, 1998; Qu et al., 2016), which leads to undesirable
developmental outcomes (Buchanan & Hughes, 2009; Qu et al., 2018). Furthermore, American
media tends to present a negative image of adolescents, which may in turn strengthen the
negative perceptions of adolescence (Nichols & Good, 2004). Kendall-Taylor (2020) has called
for shifting the negative framing of adolescence to promote positive development, yet such
efforts are still scarce. Qu and colleagues’ (2020) intervention program shifted adolescents’
stereotyped perceptions of their age group and promoted their positive development in the school
context. Further work is needed to facilitate positive perceptions of adolescence in other contexts
such as social media, which has had an increasing impact on adolescents (Best et al., 2014). The
current study may lay the foundation for this line of work by facilitating the understanding of
public perceptions of adolescence on social media, therefore informing media practitioners and
providing insights on how to tackle the negative perceptions of adolescence in the future. For
instance, the theme of hate and bias for teenagers that I identified suggests that the negative and
stereotypical view of adolescence exists on social media and may impact adolescent social media
users, especially for White adolescents. It is necessary to take actions to reduce such contents on
social media, such as changing the recommendation algorithms of social media to make such

81



contents less viable to other users. While Twitter already has policy for hateful conduct that seeks
to make hateful content less visible, hateful or biased content on adolescence existed in my
dataset, suggesting that their algorithms can be improved to create a better environment for
adolescents. The positive themes of youth (i.e., youth as an important element of society, youth
with positive characteristics, and youth to be empowered and supported) suggested that Twitter
users have an overall positive view of youth. Education and communication practitioners may
use the term of “youth” more to reference adolescence to facilitate the positive framing of
adolescence.
In addition, while existing research generally focuses on perceptions of the general
adolescent population, the current study paid special attention to adolescent subgroups by gender,
race/ethnicity, and sexual orientation, which may contribute to the understanding of the
developmental contexts of these adolescent subgroups. By highlighting the unique perceptions
associated with different adolescent subgroups, the current research may also inform intervention
programs to facilitate a better environment for adolescent subgroups on social media. For
instance, I identified themes related to anti-racism for Black and Latinx adolescents but not for
Asian adolescents. Education and communication practitioners may continue to facilitate
positive discussion around anti-racism for Black and Latinx adolescents while addressing the
lack of voices of Asian adolescents and encourage them to speak on social media. For sexual
minority adolescents, I found the theme of increasing awareness and supporting LGBT
adolescents, such theme can be further promoted to improve the well-being of LGBTQ+
adolescents.

82



Limitations
The current study leveraged big data and machine learning to analyze public perceptions
of adolescence on Twitter. Although the use of big data and unsupervised machine learning
algorithms may allow the extraction of meaningful information from the data, such an analytic
approach also has major methodological limitations.
First, unlike conventional research that directly surveys participants about their
perceptions of adolescence, the current study used word embeddings to analyze the latent
meanings of words in tweets and interpreted the results as reflecting Twitter users’ perceptions of
adolescence, which is an indirect way of investigating perceptions of adolescence. Therefore, the
results may not always reflect public perceptions of adolescence. For instance, while I found that
the term of “adolescent” was more closely associated with stereotypical descriptors (e.g.,
rebellious, risk-taking) compared to neutral descriptors, it is difficult to distinguish whether such
association reflects stereotypical perceptions (e.g., Twitter users think that adolescents are
rebellious) or actual disparities (e.g., Twitter users discuss the finding that adolescents tend to
have higher levels of rebellious behavior). I tried to address this issue by referencing example
tweets for the mixed-methods analysis to identify themes of public perceptions, during which the
coders were instructed to read the example tweets and analyze if the descriptors in the tweets
reflected certain perceptions of adolescence. However, this issue can be further complicated by
the fact that there are many robotic users and commercial accounts on Twitter (Alothali et al.,
2018). Tweets posted by these accounts may not reflect public perceptions of adolescence, as
their depiction of adolescence may be of propaganda purposes. I did not address this issue in the
current study but future work may remove posts from robotic or commercial accounts in data
analysis.

83



Second, it is difficult to ensure the quality of Twitter data and exclude undesirable data
(Salvatore et al., 2021), such as tweets irrelevant to adolescence, which may contribute
meaningless information to the word embedding model. In my previous work analyzing public
perceptions of adolescence using more than 20,000 tweets through qualitative content analysis, I
found that almost 50% of the tweets were irrelevant to adolescence. The use of word embedding
model may partly resolve this issue, as the model seeks to extract repeated information from the
data as capturing the latent meanings of words, whereas irrelevant tweets tend to be random and
may not contribute repeated information to the model.
Third, although the word embedding model is able to extract meaningful information
from the data, compared to qualitative analysis, such extraction can be superficial and missing
important, nuanced information by aggregating all the data to generate a universal representation
of the words. The current study sought to address for this limitation by integrating word
embeddings with qualitative content analysis. I also referenced example tweets when coding the
descriptors, thus providing the contexts of the descriptors. However, the qualitative content
analysis was mainly based on the output of the word embedding model (i.e., the descriptors were
selected based on the model), therefore I may still fail to extract nuanced information or identify
important themes if the word embedding model ignored such information during model training.
Fourth, although Twitter is a platform for opinion expression, it can only be used to
understand clearly conveyed perceptions but not the aspects that users find difficult or not
willing to express. Therefore, the perceptions of adolescence as identified in my dissertation
mainly reflect Twitter users’ perceptions that they expressed via tweets only.
Fifth, although the current study explicitly focuses on the context of Twitter, the
demographics of Twitter users are not representative of the U.S. population such as more male

84



users (Mislove et al., 2011). Future work may explore public perceptions of adolescence on other
platforms, such as Facebook and Tik Tok, to facilitate the understanding of the topic in more
contexts.
Sixth, my analysis only utilized the textual part of the data whereas the meta information
such as user characteristics was left. Future work may incorporate meta information into analysis,
such as comparing perceptions of adolescence between individual and organizational users.
Finally, the data I extracted are publicly available tweets posted in 2019. Public
perceptions of adolescence and the digital environment of social media may be constantly
shifting, therefore the results of the current study can only reflect public perceptions of
adolescence on Twitter in 2019. What’s more, the data were collected before the COVID-19
pandemic, public perceptions of adolescence may shift due to the pandemic and require future
research to examine potential change.
Conclusion
The current study applied word embeddings, a machine learning model, to analyze
millions of tweets about adolescence, in which different words were represented as highdimensional vectors by the model as reflecting their latent meanings. The model was able to
identify the latent construct of adolescence by associating different terms of adolescence close to
each other. The association between a certain term of adolescence and another descriptor about
adolescence (e.g., “rebellious”) was interpreted as reflecting public perceptions of adolescence. I
first employed a theory-driven and quantitative approach to examine if terms of adolescence
were perceived as stereotypical (vs. neutral) or positive (vs. negative) compared to the term of
“adult” and “people,” finding that although the overall construct of adolescence was not
perceived as stereotypical or negative, perceptions varied by different terms of adolescence,

85



suggesting that adolescence is a complex and comprehensive construct that cannot be captured
by a single item. Existing research may be limited by their use of single term to reference
adolescence. Certain terms of adolescence such as youth may be used more when referencing
adolescence as it was perceived as less stereotypical and more positive, therefore promoting
positive framing of adolescence.
I then developed an innovative method that employed a data-driven and mixed-methods
approach to analyze word embeddings, in which I first generated lists of descriptors that were
most closely associated with terms of adolescence and then conducted qualitative content
analysis on the descriptors while referencing example tweets containing the descriptors to better
understand how and why the descriptors were associated with adolescence. I was able to identify
various meaningful themes from the descriptors, reflecting not only negative perceptions of
adolescence, but also neutral and positive ones, with the themes suggesting an overall positive
and neutral perception of adolescence on Twitter. The themes revealed several perceptions of
adolescence that have rarely been addressed in literature, such as youth as an important element
of society. Some themes can also inform future research and practices, such as addressing bias
and offensive messages targeting adolescence on Twitter, leveraging the positive perceptions of
youth to promote positive framing of adolescence.
I also generated word embeddings for subgroups of adolescents based on race/ethnicity,
gender, and sexual orientation, and applied the same set of analysis on subgroups of adolescents.
The results suggested both similarities and differences existed in the perceptions of different
subgroups of adolescents. For instance, While Black, Latinx adolescents were involved in
discussion about social justice and anti-racism movements, reflecting experiences of racism and
participation in anti-racism movements among them, such discussion was missing for Asian

86



adolescents, suggesting lack of voices for them. White adolescents were found to be uniquely
associated with the theme of hate and bias, suggesting that the storm-and-stress view of
adolescence may mainly target White adolescents. Twitter users’ discussed increasing awareness
and supporting LBGT adolescents for LGBT adolescents, suggesting positive discussion for
sexual minority adolescents. These results can also inform future research and practices, such as
increasing the voices of Asian adolescents, promoting the well-being of racial/ethnic minority
adolescents by engaging them in anti-racism movements, and promoting positive development
for sexual minority adolescents.
Overall, my dissertation leveraged big data and machine learning to shift the literature
from focusing on individual perceptions of adolescence in the family and school contexts to the
broader context of social media. The analytic approach I used or developed was more inductive
or data-driven, allowing me to capture several perceptions of adolescence that has rarely been
addressed in literature, where as existing research mainly focused on the negative perceptions.
My dissertation was also among the first set of studies that combined machine learning outputs
with qualitative analysis, making the abstract and simple outputs more interpretable and putting
them into specific contexts of the data. This approach may inform future studies to leverage big
textual data to analyze public perceptions or meanings of other psychological/sociological
constructs. My dissertation can also inform future practices to promote positive framing of
adolescence and create a better digital environment for adolescent development, such as reducing
harmful and bias messages about adolescence on Twitter and promoting positive framing of
adolescence by using “youth” to reference adolescences. My dissertation also have several
limitations regarding data interpretation, representativeness, and quality, the use of the word
embedding model, and the timing of data collection.

87



REFERENCES
Abbott-Chapman, J., Denholm, C., & Wyld, C. (2008). Gender differences in adolescent risk
taking: are they diminishing? An Australian intergenerational study. Youth & Society,
40(1), 131-154.
Alothali, E., Zaki, N., Mohamed, E. A., & Alashwal, H. (2018, November). Detecting social bots
on Twitter: A literature review. In 2018 International Conference on Innovations in
Information Technology (IIT) (pp. 175-180). IEEE.
Altikulaç, S., Lee, N. C., van der Veen, C., Benneker, I., Krabbendam, L., & van Atteveldt, N.
(2019). The teenage brain: Public perceptions of neurocognitive development during
adolescence. Journal of Cognitive Neuroscience, 31(3), 339-359.
Anderson, M., & Jiang, J. (2018). Teens, social media & technology 2018. Pew Research
Center, 31, 2018.
Arnett, J. J. (1999). Adolescent storm and stress, reconsidered. American Psychologist, 54(5),
317.
Arnett, J. J. (2000). Emerging adulthood: A theory of development from the late teens through
the twenties. American Psychologist, 55(5), 469.
Aubrun, A., & Grady, J. (2000). Aliens in the living room: how TV shapes our understanding of
‘Teens’. Washington, DC: Frameworks Institute.
Austin, J., & Willard, M. (Eds.). (1998). Generations of youth: Youth cultures and history in
twentieth-century America. New York, NY: New York University Press.
Ballard, P. J. (2016). Longitudinal links between discrimination and civic development among
Latino and Asian adolescents. Journal of Research on Adolescence, 26(4), 723-737.
Barry, A. E., Bates, A. M., Olusanya, O., Vinal, C. E., Martin, E., Peoples, J. E., ... & Montano, J.
R. (2016). Alcohol marketing on Twitter and Instagram: Evidence of directly advertising
to youth/adolescents. Alcohol and Alcoholism, 51(4), 487-492.
Baskin-Sommers, A., Simmons, C., Conley, M., Chang, S. A., Estrada, S., Collins, M., ... &
Casey, B. J. (2021). Adolescent civic engagement: lessons from Black lives matter.
Proceedings of the National Academy of Sciences, 118(41), e2109860118.
Beede, D. N., Julian, T. A., Langdon, D., McKittrick, G., Khan, B., & Doms, M. E. (2011).
Women in STEM: A gender gap to innovation (No. 04-11, p. 11). Rochester, NY: US
Department of Commerce, Education and Statistics Administration.
Benner, A. D., Bakhtiari, F., Wang, Y., & Schulenberg, J. (2022). Party, academic, or prepped for
college? School norm profiles and adolescent well‐being using national data. Journal of
Research on Adolescence, 32(4), 1388-1403.

88



Benson, P. (2006). All kids are our kids: What communities must do to raise caring and
responsible children and adolescents. (2nd ed). San Francisco, CA: Jossey-Bass.
Berger, M. N., Taba, M., Marino, J. L., Lim, M. S., & Skinner, S. R. (2022). Social media use
and health and well-being of lesbian, gay, bisexual, transgender, and queer youth:
systematic review. Journal of Medical Internet Research, 24(9), e38449.
Best, P., Manktelow, R., & Taylor, B. (2014). Online communication, social media and
adolescent wellbeing: A systematic narrative review. Children and Youth Services Review,
41, 27-36.
Beyens, I., Vandenbosch, L., & Eggermont, S. (2014). Early adolescent boys’ exposure to
Internet pornography: Relationships to pubertal timing, sensation seeking, and academic
performance. The Journal of Early Adolescence, 35(8), 1045-1068.
Bleakley, A., Hennessy, M., Fishbein, M., & Jordan, A. (2009). How sources of sexual
information relate to adolescents' beliefs about sex. American Journal of Health Behavior,
33(1), 37-48.
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. Journal of Machine
Learning Research, 3(Jan), 993-1022.
Bostrom, M. (2004). A Developmental Perspective: An Analysis of Qualitative Research
Exploring Views of Youth Programs. Washington, DC: FrameWorks Institute.
Bronfenbrenner, U. (1979). The ecology of human development: Experiments by nature and
design. Cambridge, MA: Harvard University Press.
Bronfenbrenner, U. (1995). Developmental ecology through space and time: A future perspective.
In Moen, P., Elder, G., Luscher, K. (Eds.), Examining lives in context: Perspectives on the
ecology of human development. Washington, DC: American Psychological Association.
Bronfenbrenner, U. (Ed.). (2005). Making human beings human: Bioecological perspectives on
human development. Thousand Oaks, CA: Sage.
Bronfenbrenner, U., & Morris, P. A. (1998). The ecology of developmental processes. In W.
Damon & R. M. Lerner (Eds.), Handbook of child psychology: Theoretical models of
human development (pp. 993–1028). Hoboken, NJ: John Wiley & Sons Inc.
Brown, B. B., & Larson, J. (2009). Peer relationships in adolescence. In R. M. Lerner & L.
Steinberg (Eds.), Handbook of adolescent psychology: Contextual influences on
adolescent development (pp. 74–103). Hoboken, NJ: John Wiley & Sons, Inc..
Buchanan, C. M. (2003). Mothers’ generalized beliefs about adolescents: Links to expectations
for a specific child. Journal of Early Adolescence, 23(1), 29-50.

89



Buchanan, C. M., Eccles, J. S., Flanagan, C., Midgley, C., Feldlaufer, H., & Harold, R. D. (1990).
Parents' and teachers' beliefs about adolescents: Effects of sex and experience. Journal of
Youth and Adolescence, 19, 363-394.
Buchanan, C. M., & Holmbeck, G. N. (1998). Measuring beliefs about adolescent personality
and behavior. Journal of Youth and Adolescence, 27(5), 607-627.
Buchanan, C. M., & Hughes, J. L. (2009). Construction of social reality during early adolescence:
Can expecting storm and stress increase real or perceived storm and stress?. Journal of
Research on Adolescence, 19(2), 261-285.
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from
language corpora contain human-like biases. Science, 356(6334), 183-186.
Chan, W. Y., Ou, S. R., & Reynolds, A. J. (2014). Adolescent civic engagement and adult
outcomes: An examination among urban racial minorities. Journal of Youth and
Adolescence, 43, 1829-1843.
Christensen, C. (2011). Twitter revolutions? Addressing social media and dissent.
Communication Review, 14(3), 155-157.
Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological
Measurement, 20(1), 37-46.
Cormode, G., & Krishnamurthy, B. (2008). Key differences between Web 1.0 and Web 2.0. First
Monday, 13(6).
Crosnoe, R., & Johnson, M. K. (2011). Research on adolescence in the twenty-first century.
Annual Review of Sociology, 37, 439-460.
Damon, W. (2004). What is positive youth development?. The Annals of the American Academy
of Political and Social Science, 591(1), 13-24.
Doornwaard, S. M., Bickham, D. S., Rich, M., ter Bogt, T. F. M., & van den Eijnden, R. J. J. M.
(2015). Adolescents’ use of sexually explicit internet material and their sexual attitudes
and behavior: Parallel development and directional effects. Developmental Psychology,
51(10), 1476–1488.
Dornbusch, S. M. (1989). The sociology of adolescence. Annual Review of Sociology, 15(1),
233-259.
Dowler, E., Green, J., Bauer, M., & Gasperoni, G. (2006). Assessing public perception: Issues
and methods. In Health hazard and public debate: Lessons for risk communication from
BSE/CJD saga. Geneva, Switzerland: World Health Organization.
Eccles, J. S., & Roeser, R. W. (2011). Schools as developmental contexts during adolescence.
Journal of Research on Adolescence, 21(1), 225-241.

90



Eichstaedt, J. C., Schwartz, H. A., Kern, M. L., Park, G., Labarthe, D. R., Merchant, R. M., ... &
Seligman, M. E. (2015). Psychological language on Twitter predicts county-level heart
disease mortality. Psychological Science, 26(2), 159-169.
Elder, G. H. (1998). The life course as developmental theory. Child Development, 69(1), 1-12.
Ellemers, N. (2018). Gender stereotypes. Annual Review of Psychology, 69, 275-298.
Entman, R. M. (1993). Framing: Towards clarification of a fractured paradigm. Journal of
Communication, 43(4), 51-58.
Erikson, E. H. (1968). Identity, youth, and crisis. New York, NY: Norton.
Fredrickson, B. L., & Roberts, T. A. (1997). Objectification theory: Toward understanding
women’s lived experiences and mental health risks. Psychology of Women Quarterly, 21,
173–206.
Freud, A. (1946). The Ego and the Mechanisms of Defense. New York, NY: International
Universities Press.
Freud, A. (1958). Adolescence. Psychoanalytic Study of the Child, 15, 255-278.
Friedman, R. (1975). The vicissitudes of adolescent development and what it activates in adults.
Adolescence, 10, 520-526.
García Coll, C., Crnic, K., Lamberty, G., Wasik, B. H., Jenkins, R., Garcia, H. V., & McAdoo, H.
P. (1996). An integrative model for the study of developmental competencies in minority
children. Child development, 67(5), 1891-1914.
Gilliam, F. D, & Bales, S. N. (2001). Strategic Frame Analysis: Reframing America's Youth.
Social Policy Report, XV(3), 3-15.
Garg, N., Schiebinger, L., Jurafsky, D., & Zou, J. (2018). Word embeddings quantify 100 years
of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences,
115(16), E3635-E3644.
Gilbert, K. E. (2012). The neglected role of positive emotion in adolescent psychopathology.
Clinical Psychology Review, 32(6), 467-481.
Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. (1998). Measuring individual differences in
implicit cognition: the implicit association test. Journal of Personality and Social
Psychology, 74(6), 1464.
Grimmer, J., Roberts, M. E., & Stewart, B. M. (2021). Machine learning for social science: An
agnostic approach. Annual Review of Political Science, 24, 395-419.
Gross, E. F., & Hardin, C. D. (2007). Implicit and explicit stereotyping of adolescents. Social
Justice Research, 20(2), 140-160.
91



Hall, G. S. (1904). Adolescence: Its psychology and its relation to physiology, anthropology,
sociology, sex, crime, religion, and education (Vols. I & II). Englewood Cliffs, NJ:
Prentice-Hall.
Hall, G. S. (1916). Adolescence. New York, NY: Appleton.
Harter, S. (1998). The development of self-representations. In W. Damon & N. Eisenberg (Ed.),
Handbook of child psychology: Social, emotional, and personality development (pp. 553–
617). Hoboken, NJ: John Wiley & Sons, Inc..
Hays, R., & Daker-White, G. (2015). The care. data consensus? A qualitative analysis of
opinions expressed on Twitter. BMC Public Health, 15(1), 838.
Heintz-Knowles, K. E. (2000). Images of youth: A content analysis of adolescents in prime-time
entertainment programming. Washington, DC: Frameworks Institute.
Heller, J. R., & Johnson, H. L. (2013). Parental opinion concerning school sexuality education in
a culturally diverse population in the USA. Sex Education, 13(5), 548-559.
Henig, R. M. (2010). What is it about 20-somethings. The New York Times, 8(18), 2010.
Hines, A. R., & Paulson, S. E. (2006). Parents' and teachers' perceptions of adolescent storm and
stress: Relations with parenting and teaching styles. Adolescence, 41(164).
Holmbeck, G. N. (1996). A model of family relational transformations during the transition to
adolescence: Parent-adolescent conflict and adaptation. In J. A. Graber, J. Brooks-Gunn,
& A. C. Petersen (Eds.), Transitions through adolescence: Interpersonal domains and
context (pp. 167-199). Mahwah, NJ: Erlbaum.
Holmbeck, G. N., & Hill, J. P. (1988). Storm and stress beliefs about adolescence: Prevalence,
self-reported antecedents, and effects of an undergraduate course. Journal of Youth and
Adolescence, 17(4), 285-306.
Hope, E. C., Volpe, V. V., Briggs, A. S., & Benson, G. P. (2022). Anti‐racism activism among
Black adolescents and emerging adults: Understanding the roles of racism and
anticipatory racism‐related stress. Child Development, 93(3), 717-731.
Hutto, C., & Gilbert, E. (2014, May). Vader: A parsimonious rule-based model for sentiment
analysis of social media text. In Proceedings of the International AAAI Conference on
Web and Social Media (Vol. 8, No. 1, pp. 216-225).
Jacobs, J. E., Chhin, C. S., & Shaver, K. (2005). Longitudinal links between perceptions of
adolescence and the social beliefs of adolescents: Are parents’ stereotypes related to
beliefs held about and by their children?. Journal of Youth and Adolescence, 34(2), 61-72.
Jacobucci, R., & Grimm, K. J. (2020). Machine learning and psychological research: The
unexplored effect of measurement. Perspectives on Psychological Science, 15(3), 809816.
92



Java, A., Song, X., Finin, T., & Tseng, B. (2007). Why we twitter: understanding microblogging
usage and communities. In Proceedings of the 9th WebKDD and 1st SNA-KDD 2007
workshop on Web mining and social network analysis (pp. 56-65). ACM.
Jeffrey, C., & Dyson, J. (2009). Telling young lives: Portraits of global youth. Philadelphia, PA:
Temple University Press.
Jimenez‐Sotomayor, M. R., Gomez‐Moreno, C., & Soto‐Perez‐de‐Celis, E. (2020). Coronavirus,
ageism, and Twitter: An evaluation of tweets about older adults and COVID‐19. Journal
of the American Geriatrics Society, 68(8), 1661-1665.
Jones, J. M. (1997). Prejudice and racism (2nd ed.). McGraw Hill Companies.
Jussim, L. (1986). Self-fulfilling prophecies: A theoretical and integrative review. Psychological
Review, 93(4), 429.
Jussim, L. (1991). Social perception and social reality: A reflection-construction model.
Psychological Review, 98(1), 54.
Karami, A., Bennett, L. S., & He, X. (2018). Mining public opinion about economic issues:
Twitter and the us presidential election. International Journal of Strategic Decision
Sciences, 9(1), 18-28.
Kendall-Taylor, N. (2020). Shifting the frame to change how we see young people. Journal of
Adolescent Health, 66(2), 137-139.
Kern, M. L., Eichstaedt, J. C., Schwartz, H. A., Park, G., Ungar, L. H., Stillwell, D. J., ... &
Seligman, M. E. (2014). From “Sooo excited!!!” to “So proud”: Using language to study
development. Developmental Psychology, 50(1), 178.
Klaassen, M. J., & Peter, J. (2015). Gender (in) equality in internet pornography: A content
analysis of popular pornographic internet videos. The Journal of Sex Research, 52(7),
721-735.
Kozlowski, A. C., Taddy, M., & Evans, J. A. (2019). The geometry of culture: Analyzing the
meanings of class through word embeddings. American Sociological Review, 84(5), 905949.
Kuruvilla, S., Bustreo, F., Kuo, T., Mishra, C. K., Taylor, K., Fogstad, H., ... & Costello, A.
(2016). The Global strategy for women’s, children’s and adolescents’ health (2016–2030):
a roadmap based on evidence and country experience. Bulletin of the World Health
Organization, 94(5), 398.
Larson, R. W. (2000). Toward a psychology of positive youth development. American
Psychologist, 55(1), 170.

93



Lee, H., Lei, Q., Su, G., & Zhang, S. (2022). Acknowledging Anti‐Blackness, Overlooking
Anti‐Asian Racism: Missed Developmental Opportunities for Chinese American Youth.
Journal of Research on Adolescence, 32(3), 1064-1082.
Lerner, R. M., Lerner, J. V., Almerigi, J. B., Theokas, C., Phelps, E., Gestsdottir, S., ... & von Eye,
A. (2005). Positive youth development, participation in community youth development
programs, and community contributions of fifth-grade adolescents: Findings from the
first wave of the 4-H study of positive youth development. Journal of Early Adolescence,
25(1), 17-71.
Lerner, R. M., Steinberg, L. (2009). Handbook of adolescent psychology, volume 1: Individual
bases of adolescent development. Somerset County, NJ: John Wiley & Sons.
Levy, O., & Goldberg, Y. (2014). Neural word embedding as implicit matrix factorization.
Advances in Neural Information Processing Systems, 27.
MacKay, M., Parlee, B., & Karsgaard, C. (2020). Youth engagement in climate change action:
Case study on indigenous youth at COP24. Sustainability, 12(16), 6299.
Marino, C., Gini, G., Angelini, F., Vieno, A., & Spada, M. M. (2020). Social norms and emotions in problematic social media use among adolescents. Addictive Behaviors Reports,
11, 100250.
McFarland, D. A., Lewis, K., & Goldberg, A. (2016). Sociology in the era of big data: The ascent
of forensic social science. American Sociologist, 47(1), 12-35.
McLoyd V., & Steinberg L. (Ed.). (1998). Studying minority adolescents: Conceptual,
methodological, and theoretical issues. Mahwah, NJ: Erlbaum.
Mead, M. (1928). Coming of age in Samoa: A psychological study of primitive youth for western
civilization. New York, NY: Dell.
Meeus, W. (2011). The study of adolescent identity formation 2000–2010: A review of
longitudinal research. Journal of Research on Adolescence, 21(1), 75-94.
Memmott-Elison, M. K., Holmgren, H. G., Padilla-Walker, L. M., & Hawkins, A. J. (2020).
Associations between prosocial behavior, externalizing behaviors, and internalizing
symptoms during adolescence: A meta-analysis. Journal of Adolescence, 80, 98-114.
Mendle, J., Beltz, A. M., Carter, R., & Dorn, L. D. (2019). Understanding puberty and its
measurement: ideas for research in a new generation. Journal of Research on
Adolescence, 29(1), 82-95.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed
representations of words and phrases and their compositionality. Advances in Neural
Information Processing Systems, 26.

94



Mislove, A., Lehmann, S., Ahn, Y. Y., Onnela, J. P., & Rosenquist, J. (2011). Understanding the
demographics of Twitter users. In Proceedings of the International AAAI Conference on
Web and Social Media (Vol. 5, No. 1, pp. 554-557).
National Academies of Sciences, Engineering, and Medicine. (2019). The promise of
adolescence: Realizing opportunity for all youth. Washington, DC: The National
Academies Press.
Nelson, L. K. (2020). Computational grounded theory: A methodological framework.
Sociological Methods & Research, 49(1), 3-42.
Newell, P., Bulkeley, H., Turner, K., Shaw, C., Caney, S., Shove, E., & Pidgeon, N. (2015).
Governance traps in climate change politics: re‐framing the debate in terms of
responsibilities and rights. Wiley Interdisciplinary Reviews: Climate Change, 6(6), 535540.
Nichols, S. L., & Good, T. L. (2004). American’s teenagers—Myths and realities: Media images,
schooling, and the social costs of careless indifference. Mahwah, NJ: Lawrence Erlbaum
Associates.
Park, H., Reber, B. H., & Chon, M. G. (2016). Tweeting as health communication: health
organizations’ use of Twitter for health promotion and public engagement. Journal of
Health Communication, 21(2), 188-198.
Patton, G. C., Sawyer, S. M., Santelli, J. S., Ross, D. A., Afifi, R., Allen, N. B., ... & Viner, R. M.
(2016). Our future: a Lancet commission on adolescent health and wellbeing. The Lancet,
387(10036), 2423-2478.
Pavlova, A., & Berkers, P. (2022). “Mental Health” as defined by Twitter: Frames, emotions,
stigma. Health Communication, 37(5), 637-647.
Pennebaker, J. W., Francis, M. E., & Booth, R. J. (2001). Linguistic inquiry and word count:
LIWC 2001. Mahway: Lawrence Erlbaum Associates, 71(2001), 2001.
Perrin, A., & Anderson, M. (2019). Share of US adults using social media, including Facebook,
is mostly unchanged since 2018. Pew Research Center.
Phinney, J. S., & Kohatsu, E. L. (1997). Ethnic and racial identity development and mental
health. In J. Schulenberg, J. L. Maggs, & K. Hurrelmann (Eds.), Health risks and
developmental transitions during adolescence (pp. 420–443). Cambridge, United
Kingdom: Cambridge University Press.
PettyJohn, M. E., Muzzey, F. K., Maas, M. K., & McCauley, H. L. (2019). # HowIWillChange:
Engaging men and boys in the# MeToo movement. Psychology of Men &
Masculinities, 20(4), 612.
Ponton, L. E., & Judice, S. (2004). Typical adolescent sexual development. Child and Adolescent
Psychiatric Clinics, 13(3), 497-511.
95



Qu, Y., Pomerantz, E. M., Wang, M., Cheung, C., & Cimpian, A. (2016). Conceptions of
adolescence: Implications for differences in engagement in school over early adolescence
in the United States and China. Journal of Youth and Adolescence, 45(7), 1512-1526.
Qu, Y., Pomerantz, E. M., McCormick, E., & Telzer, E. H. (2018). Youth's conceptions of
adolescence predict longitudinal changes in prefrontal cortex activation and risk taking
during adolescence. Child Development, 89(3), 773-783.
Qu, Y., Pomerantz, E. M., & Wu, G. (2020). Countering youth's negative stereotypes of teens
fosters constructive behavior. Child Development, 91(1), 197-213.
Quayle, E., & Taylor, M. (2004). Child pornography: An internet crime. Routledge.
Rich, G. J. (2003). The positive psychology of youth and adolescence. Journal of Youth and
Adolescence, 32(1), 1.
Roberts, D. (2012). Fatal invention: How science, politics, and big business re-create race in the
twenty-first century. The New Press.
Rogers, L. O., Niwa, E. Y., Chung, K., Yip, T., & Chae, D. (2021). M (ai) cro: Centering the
macrosystem in human development. Human Development, 65(5-6), 270-292.
Rudkowsky, E., Haselmayer, M., Wastian, M., Jenny, M., Emrich, Š., & Sedlmair, M. (2018).
More than bags of words: Sentiment analysis with word embeddings. Communication
Methods and Measures, 12(2-3), 140-157.
Salvatore, C., Biffignandi, S., & Bianchi, A. (2021). Social media and twitter data quality for
new social indicators. Social Indicators Research, 156, 601-630.
Savin-Williams, R. C. (2011). Identity development among sexual minority youth. In S. J.
Schwartz, K. Luyckx & V. L. Vignoles (Eds.), Handbook of identity theory and research
(pp.671-689). New York, NY: Springer.
Sawyer, S. M., Azzopardi, P. S., Wickremarathne, D., & Patton, G. C. (2018). The age of
adolescence. The Lancet Child & Adolescent Health, 2(3), 223-228.
Selkie, E., Adkins, V., Masters, E., Bajpai, A., & Shumer, D. (2020). Transgender adolescents'
uses of social media for social support. Journal of Adolescent Health, 66(3), 275-280.
Shaw, J. M., Mitchell, C. A., Welch, A. J., & Williamson, M. J. (2015). Social media used as a
health intervention in adolescent health: A systematic review of the literature. Digital
Health, 1, 2055207615588395.
Shek, D. & Chan, C. (2011). Perceptions of adolescents among teachers and social workers in
the context of training programs in Hong Kong. International Journal of Adolescent
Medicine and Health, 23(4), 373-378.

96



Shek, D., Dou, D., Zhu, X., & Chai, W. (2019). Positive youth development: Current
perspectives. Adolescent Health, Medicine and Therapeutics, 131-141.
Sigel, I. E. (Ed.). (1985). Parental belief systems: The psychological consequences for children.
Hillsdale, NJ: Erlbaum.
Slater, A., & Tiggemann, M. (2010). Body image and disordered eating in adolescent girls and
boys: A test of objectification theory. Sex Roles, 63, 42-49.
Smetana, J. G., & Rote, W. M. (2019). Adolescent–parent relationships: Progress, processes, and
prospects. Annual Review of Developmental Psychology, 1, 41-68.
Smith, E. J. (1991). Ethnic identity development: Toward the development of a theory within the
context of majority/minority status. Journal of Counseling & Development, 70(1), 181188.
Spencer, S. J., Logel, C., & Davies, P. G. (2016). Stereotype threat. Annual Review of Psychology,
67, 415-437.
Staff, P. T. (1995). The invention of adolescence. Psychology Today. Retrieved from
https://www.psychologytoday.com/intl/articles/199501/the-invention-adolescence
Starr, C. R., & Simpkins, S. D. (2021). High school students’ math and science gender
stereotypes: relations with their STEM outcomes and socializers’ stereotypes. Social
Psychology of Education, 24, 273-298.
Steele, C. M., & Aronson, J. (1995). Stereotype threat and the intellectual test performance of
African Americans. Journal of Personality and Social Psychology, 69(5), 797.
Steinberg, L., & Morris, A. S. (2001). Adolescent development. Annual Review of Psychology,
52(1), 83-110.
Twitter. (2022a). The conversation: Twitter trends 2022. Retrieved from
https://marketing.twitter.com/en/insights/the-conversation-twitter-trends-2022
Twitter. (2022b). One-time Historical PowerTrack jobs. Retrieved from
https://developer.twitter.com/en/docs/tutorials/one-time-historical-powertrack-jobs
Uhls, Y., Ellison, N., & Subrahmanyam, K. (2017). Benefits and costs of social media in
adolescence. Pediatrics, 140, S67-S70.
Umaña‐Taylor, A. J., Quintana, S. M., Lee, R. M., Cross Jr, W. E., Rivas‐Drake, D., Schwartz, S.
J., ... & Ethnic and Racial Identity in the 21st Century Study Group. (2014). Ethnic and
racial identity during adolescence and into young adulthood: An integrated
conceptualization. Child Development, 85(1), 21-39.
United Nations Educational, Scientific and Cultural Organization. (1985). World Congress of
youth—final report. 1985.

97



Vaccari, C., Valeriani, A., Barberá, P., Bonneau, R., Jost, J. T., Nagler, J., & Tucker, J. A. (2015).
Political expression and action on social media: Exploring the relationship between
lower-and higher-threshold political activities among Twitter users in Italy. Journal of
Computer-Mediated Communication, 20(2), 221-239.
Vaismoradi, M., Jones, J., Turunen, H., & Snelgrove, S. (2016). Theme development in
qualitative content analysis and thematic analysis. Journal of Nursing Education and
Practice, 6(5), 100-110.
Vliegenthart, R., & van Zoonen, L. (2011). Power to the frame: Bringing sociology back to
frame analysis. European Journal of Communication, 26(2), 101-115.
Vogels, E., Gelles-Watnick, R., & Massarat, N. (2022). Teens, social media and technology 2022.
Retrieved from https://www.pewresearch.org/internet/2022/08/10/teens-social-mediaand-technology-2022/
Wallace, S. A., Townsend, T. G., Glasgow, Y. M., & Ojie, M. J. (2011). Gold diggers, video
vixens, and jezebels: Stereotype images and substance use among urban African
American girls. Journal of Women's Health, 20(9), 1315-1324.
Wang, Y., Chen, M., & Lee, J. H. (2019). Adolescents’ social norms across family, peer, and
school settings: Linking social norm profiles to adolescent risky health behaviors.
Journal of Youth and Adolescence, 48, 935-948.
Waters, S. K., Cross, D. S., & Runions, K. (2009). Social and ecological structures supporting
adolescent connectedness to school: A theoretical model. Journal of School Health,
79(11), 516-524.
Way, N., Hernández, M. G., Rogers, L. O., & Hughes, D. L. (2013). “I’m not going to become no
rapper” stereotypes as a context of ethnic and racial identity development. Journal of
Adolescent Research, 28(4), 407-430.
Weinberg, T. (2009). The new community rules: Marketing on the social web. Sebastopol, CA:
O'Reilly Media, Inc.
Weinstein, E. (2018). The social media see-saw: Positive and negative influences on adolescents’
affective well-being. New Media & Society, 20(10), 3597-3623.
White, R. M., Witherspoon, D. P., Wei, W., Zhao, C., Pasco, M. C., Maereg, T. M., ... & Winkler,
E. (2021). Adolescent development in context: A decade review of neighborhood and
activity space research. Journal of Research on Adolescence, 31(4), 944-965.
Youniss, J., Bales, S., Christmas‐Best, V., Diversi, M., McLaughlin, M., & Silbereisen, R. (2002).
Youth civic engagement in the twenty‐first century. Journal of Research on Adolescence,
12(1), 121-148.

98



Zheng, D., Ni, X. L., & Luo, Y. J. (2019). Selfie posting on social networking sites and female
adolescents’ self-objectification: The moderating role of imaginary audience ideation. Sex
Roles, 80, 325-331.
Zhou, Y., & Paul, B. (2016). Lotus blossom or dragon lady: A content analysis of “Asian women”
online pornography. Sexuality & Culture, 20, 1083-1100.

99



APPENDIX 1: TABLES
Table 1
Number of Tweets Containing Different Terms of Adolescence
Terms

Number of Tweets

Adolescent

85,022

Teen

1,694,602

Youth

1,500,895

Juvenile

104,857

Puberty

62,309

Youngster

133,874

Tween

57,928

Example Tweet
teenage #pregnancy nutrition-nutritional needs for the
adolescent mum-to-be
group conducts another search for missing teens
there is a thin line between enjoying your youth and
destroying your future
earlier this month, congress passed the juvenile justice
and delinquency prevention act that now awaits
presidential signature
hrt gives you the hormone levels of a typical male going
through puberty
he feels the youngster is in a 'very positive moment'
mentally
where are the just-right books for tweens?

100



Table 2
Sentiment Scores of Tweets Containing Different Terms of Adolescence
Positive
Negative
Neutral
Overall
Terms
Sentiment
Sentiment
Sentiment
Sentiment
Mean (SD)
Mean (SD)
Mean (SD)
Mean (SD)
Adolescent
.14
(.13)
.09
(.12)
.76
(.16)
.13
(.60)
Teen
.15
(.15)
.13
(.14)
.72
(.17)
.07
(.62)
Youth
.17
(.15)
.08
(.11)
.75
(.16)
.28
(.59)
Juvenile
.13
(.13)
.13
(.14)
.74
(.16)
-.02
(.63)
Puberty
.15
(.15)
.10
(.12)
.75
(.17)
.10
(.54)
Youngster
.20
(.16)
.08
(.11)
.72
(.16)
.33
(.57)
Tween
.15
(.13)
.08
(.11)
.77
(.15)
.19
(.60)
Note. For positive, negative, and neutral sentiment, the scores range from 0 to 1, with higher
scores indicating higher levels of sentiment. For overall sentiment, the scores range from -1 to 1,
with scores higher than .05 indicating positive sentiment, scores lower than -.05 indicating
negative sentiment, and scores between -.05 and .05 indicating neutral sentiment.

101



Table 3
Average Cosine Similarity Scores for Different Word Embeddings of Adolescence in their
Associations with Stereotypical Descriptors and Neutral Descriptors
Stereotypical
Neutral
Paired Sample
Mean
Descriptors
Descriptors
t-tests
Difference
Mean (SD) Mean (SD)
t
p
Cohen’s d
Adolescent
.18
(.07)
.28
(.05)
-.09
-2.94
.012
-1.57
Teen
.19
(.09)
.46
(.04)
-.27
-7.27
.001
-3.89
Youth
.12
(.03)
.36
(.10)
-.25
-6.04
.001
-3.22
Juvenile
.17
(.07)
.25
(.03)
-.08
-3.01
.013
-1.61
Puberty
.17
(.04)
.27
(.04)
-.11
-4.58
.001
-2.45
Youngster
.15
(.09)
.33
(.06)
-.18
-4.35
.001
-2.33
Tween
.15
(.05)
.31
(.07)
-.16
-4.84
.001
-2.59
Composite
.26
(.06)
.51
(.06)
-.25
-7.67
.001
-4.10
Adult
.20
(.03)
.39
(.04)
-.19
-9.25
.001
-4.95
People
.20
(.07)
.53
(.08)
-.33
-7.98
.001
-4.26
Note. Stereotypical descriptors included “rebellious,” “reckless,” “risk-taking,” “stubborn,”
Word
Embedding

“rude,” “impulsive,” “selfish,” “crazy”. Neutral descriptors included “get,” “say,” “still,” “also,”
“thing,” “would,” “as,” “go”. Composite: composite measure of adolescence (i.e., the centroid of
the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and
“tween”).

102



Table 4
Average Cosine Similarity Scores for Different Word Embeddings of Adolescence in their
Associations with Positive Descriptors and Negative Descriptors
Positive
Negative
Paired Sample
Mean
Descriptors
Descriptors
t-tests
Difference
Mean (SD) Mean (SD)
t
p
Cohen’s d
Adolescent
.15
(.05)
.13
(.05)
.02
0.77
.455
0.41
Teen
.17
(.07)
.14
(.05)
.03
0.90
.384
0.48
Youth
.18
(.07)
.10
(.04)
.08
2.58
.025
1.38
Juvenile
.12
(.07)
.20
(.07)
-.07
-2.13
.051
-1.14
Puberty
.13
(.07)
.13
(.03)
.00
0.11
.915
0.06
Youngster
.20
(.07)
.14
(.06)
.06
1.96
.071
1.05
Tween
.16
(.06)
.14
(.06)
.02
0.76
.459
0.41
Composite
.25
(.07)
.23
(.06)
.03
0.79
.445
0.42
Adult
.17
(.08)
.16
(.04)
.02
0.50
.626
0.27
People
.18
(.09)
.22
(.05)
-.03
-0.81
.434
-0.43
Note. Positive descriptors included “agreeable,” “bright,” “diligent,” “gentle,” “helpful,” “honest,”
Word
Embedding

“kind,” “nice”. Negative descriptors included “aggressive,” “careless,” “dishonest,” “foolish,”
“rude,” “lazy,” “selfish,” “silly”. Composite: composite measure of adolescence (i.e., the
centroid of the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,”
“youngster,” and “tween”).

103



Table 5
Summary of Themes for Different Word Embeddings of Adolescence
Word
Embedding

Theme

Top Five Descriptors

Example Tweets

Research on adolescence to be advocated

“meta-analysis,” “systematic review,”
“springerlink,” “comorbidity,”
“psychosocial”

Concern for adolescent health

“mum-to-be,” “nutrition,” “comorbidity,”
“health,” “over-nutrition”

“Important paper that robustly challenges claims
for high incidence of bipolar disorder in children &
adolescents made by some US studies.”
“Sometimes children and adolescents suffer from
depressive symptoms or are even have a
#depression.”

Hate and bias for teenagers

“fuck,” “fucking,” “horny,” “dirty,”
“worthless”

Youth as an important element of society

“support,” “country,” “nation,” “future,”
“community”

Youth with positive characteristics

“leader,” “great,” “young,” “good,”
“well”

Youth to be empowered and supported

“support,” “hope,” “help,” “opportunity,”
“work”

Juvenile

Juvenile justice and delinquency

“delinquent,” “delinquency,” “justice
reform,” “justice,” “peer justice”

Puberty

Puberty as a period of physical
development and sexual maturation

“blocker,” “hormone,” “hrt,” “estrogen,”
“testosterone”

“Leaders like Hong Hoang, who mobilized a youthled movement to create a greener world after
becoming the first Vietnamese person to visit
Antarctica.”
“Youths contribute to economy and the growth of
Nigeria.”
“End 2018 being the change you wish to see. help
us spread kindness & empower youth. together, we
can prevent #bullying.”
“Earlier this month, congress passed the juvenile
justice and delinquency prevention act that now
awaits presidential signature.”
“Sophia began identifying as a boy at 11 years
old...to halt the effects of female puberty.”

Youngster

Youngster sports players getting attention

“Arsche,” “player,” “Sanju Samson,”
“Minamino,” “best”

“A mixture of first team players and youngsters will
all get their chance to impress ahead of 2019!”

Tween

Tweens’ everyday life

“kids book,” “childrens book,” “gift
finder,” “birthday gifts,” “school library”

Adolescent

Teen

Youth

104

“Fuck teens in your area.”

“What do you buy? gift ideas for teens who love
dogs #giftfinder #teensandtweens”



Table 6
Average Cosine Similarity Scores for Different Word Embeddings of Subgroups of Adolescence in
their Associations with Stereotypical Descriptors and Neutral Descriptors
Stereotypical
Neutral
Paired Sample
Mean
Descriptors
Descriptors
t-tests
Difference
Mean (SD) Mean (SD)
t
p
Cohen’s d
Boy
.21
(.07)
.52
(.05)
-.30
-9.04
.001
-4.83
Girl
.21
(.07)
.53
(.04)
-.32
-9.69
.001
-5.18
Black
.18
(.09)
.46
(.04)
-.28
-7.45
.001
-3.98
Latinx
.17
(.05)
.29
(.04)
-.13
-5.03
.001
-2.69
Asian
.17
(.07)
.37
(.03)
-.21
-7.16
.001
-3.83
White
.23
(.09)
.49
(.03)
-.26
-7.05
.001
-3.77
Gay
.21
(.06)
.39
(.04)
-.18
-7.14
.001
-3.82
Lesbian
.19
(.04)
.34
(.06)
-.14
-5.53
.001
-2.96
LGBT
.18
(.06)
.32
(.03)
-.14
-5.71
.001
-3.05
Composite
.26
(.06)
.51
(.06)
-.25
-7.67
.001
-4.10
Adult
.20
(.03)
.39
(.04)
-.19
-9.26
.001
-4.95
People
.21
(.07)
.53
(.08)
-.33
-7.98
.001
-4.26
Note. Stereotypical descriptors included “rebellious,” “reckless,” “risk-taking,” “stubborn,”
Word
Embedding

“rude,” “impulsive,” “selfish,” “crazy”. Neutral descriptors included “get,” “say,” “still,” “also,”
“thing,” “would,” “as,” “go”. Composite: composite measure of adolescence (i.e., the centroid of
the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,” “youngster,” and
“tween”).

105



Table 7
Average Cosine Similarity Scores for Different Word Embeddings of Subgroups of Adolescents in
their Associations with Positive Descriptors and Negative Descriptors
Positive
Negative
Paired Sample
Mean
Descriptors
Descriptors
t-tests
Difference
Mean (SD) Mean (SD)
t
p
Cohen’s d
Boy
.21
(.09)
.20
(.05)
.01
0.15
.884
0.08
Girl
.21
(.09)
.19
(.05)
.02
0.57
.578
0.31
Black
.18
(.08)
.15
(.04)
.02
0.72
.486
0.39
Latinx
.16
(.04)
.13
(.05)
.04
1.46
.165
0.78
Asian
.16
(.07)
.13
(.05)
.01
0.46
.655
0.24
White
.23
(.10)
.23
(.05)
.00
0.02
.981
0.01
Gay
.18
(.08)
.20
(.05)
-.01
-0.37
.716
-0.20
Lesbian
.18
(.08)
.17
(.04)
.01
0.37
.718
0.20
LGBT
.21
(.08)
.15
(.05)
.06
1.75
.107
0.94
Composite
.25
(.07)
.23
(.06)
.03
0.79
.445
0.42
Adult
.17
(.08)
.16
(.04)
.02
0.50
.626
0.27
People
.18
(.09)
.22
(.05)
-.03
-0.81
.434
-0.43
Note. Positive descriptors included “agreeable,” “bright,” “diligent,” “gentle,” “helpful,” “honest,”
Word
Embedding

“kind,” “nice”. Negative descriptors included “aggressive,” “careless,” “dishonest,” “foolish,”
“rude,” “lazy,” “selfish,” “silly”. Composite: composite measure of adolescence (i.e., the
centroid of the word embeddings of “adolescent,” “teen,” “youth,” “juvenile,” “puberty,”
“youngster,” and “tween”).

106



Table 8
Summary of Themes for Different Word Embeddings of Subgroups of Adolescence
Word Embedding
Black adolescents

Latinx adolescents

White adolescents

Male adolescents
Female adolescents
LGBT adolescents

Theme
Black adolescents engaging in
social justice movements
Concern for Black
adolescents’ life
Increasing racial equity for
Latinx adolescents
Hate and Bias for White
adolescents
White adolescents involved in
discussion about race and
racism
Male adolescents’ everyday
life
Female adolescents are
attractive in sexual contents
Increasing awareness and
supporting LGBT adolescents

Top Five Descriptors
“bbconly,” “say no to white boys,” “history,”
“black twitter movement,” “black lives
matter,” “black is better”
“poncho_free,” “melaninpoppin,” “school,”
“family,” “group”
“representationmatters,” “racialjustice,”
“hispanicheritagemonth,” “educolor,”
“minoritymentalhealthmonth”

Example Tweets
“Stop telling the Black youth half a bread is better
than nothing, we want the whole bakery.”
“Almost half of Black youth in the US grow up in
the bottom 20% of the income distribution.”
“New reports confirm #soda companies
disproportionately target #latino & #black youth
with their #caffeine laced.”

“fuck,” “fucking,” “wrong,” “bad,” “weird”

“Stupid white girls.”

“racist,” “supremacist,” “supremacy,” “white
nationalism,” “white supremacist terrorism”

“If you are a white female teen and have a black
boyfriend, don't have fear, at least one of your
friends will follow your example soon!”

“school,” “family,” “son,” “group,” “home”

“A teenage boy on holiday”

“love,” “nudeselfie,” “nicetits,” “nudegirls,”
“buyynudes”
“pridemonth,” “homophobia,” “loveislove,”
“lgbtqoftwitter,” “nooutsiders”

“Do you like adorable cute teengirls? of course
you do! enjoy these leaked collection.”
“#hope #loveislove one man's advice for friends
and relatives of lgbtq youth.”

107



APPENDIX 2: FIGURES
Figure 1
Conceptual Illustration of Representing Words as Word Embeddings

108



Figure 2
Conceptual Illustration of Identifying Themes of Perceptions Using Word Embeddings

109



Figure 3
Relative Location of Different Terms of Adolescence in the Word Embedding Model

Note. The word embedding model represent different terms with 300 dimensions, which cannot
be directly visualized. I used t-distributed Stochastic Neighbor Embedding to reduce the
dimensions to two, therefore plotting the relative location of the terms of adolescence in a twodimensional space with some other words that were closely associated with them.

110



Figure 4
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of
“Adolescent”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “adolescent”.

111



Figure 5
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of “Teen”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “teen”. I identified descriptors that indicated online sexual abuse of adolescents.
These descriptors were removed and not coded for public perceptions of adolescence. The
removed descriptors are available upon request from the author.

112



Figure 6
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of “Youth”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “youth”.

113



Figure 7
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of “Juvenile”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “juvenile”.

114



Figure 8
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of “Puberty”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “puberty”.

115



Figure 9
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of
“Youngster”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “youngster”.

116



Figure 10
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of “Tween”

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of “tween”.

117



Figure 11
Word Cloud Figure of the Top 60 Descriptors Associated with the Composite Measure of
Adolescent

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the
composite measure of adolescent. The composite measure of adolescence was constructed as the
centroid of the word embeddings of adolescent,” “teen,” “youth,” “juvenile,” “puberty,”
“youngster,” and “tween”. I identified descriptors that indicated online sexual abuse of
adolescents. These descriptors were removed from the figure and not coded for public
perceptions of adolescence. The removed descriptors are available upon request from the author.

118



Figure 12
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of Black
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of Black adolescents. I identified descriptors that indicated online sexual abuse of
adolescents. These descriptors were removed from the figure and not coded for public
perceptions of adolescence. The removed descriptors are available upon request from the author.

119



Figure 13
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of Latinx
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of Latinx adolescents.

120



Figure 14
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of White
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of White adolescents. I identified descriptors that indicated online sexual abuse of
adolescents. These descriptors were removed from the figure and not coded for public
perceptions of adolescence. The removed descriptors are available upon request from the author.

121



Figure 15
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of Male
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of male adolescents. I identified descriptors that indicated online sexual abuse of
adolescents. These descriptors were removed from the figure and not coded for public
perceptions of adolescence. The removed descriptors are available upon request from the author.

122



Figure 16
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of Female
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of female adolescents. I identified descriptors that indicated online sexual abuse of
adolescents. These descriptors were removed from the figure and not coded for public
perceptions of adolescence. The removed descriptors are available upon request from the author.

123



Figure 17
Word Cloud Figure of the Top 60 Descriptors Associated with the Word Embedding of LGBT
Adolescents

Note. Larger font sizes indicate higher levels of closeness between the descriptor and the word
embedding of LGBT adolescents

124