Content uploaded by Mark D Griffiths
Author content
All content in this area was uploaded by Mark D Griffiths on Nov 10, 2023
Content may be subject to copyright.
Deep learning(s) in gaming disorder through
the user-avatar bond: A longitudinal study
using machine learning
VASILEIOS STAVROPOULOS
1,3
, DANIEL ZARATE
1
p,
MARIA PROKOFIEVA
2
, NOIRIN VAN DE BERG
4
,
LEILA KARIMI
1
, ANGELA GORMAN ALESI
5
,
MICHAELLA RICHARDS
6
, SOULA BENNET
7
and
MARK D. GRIFFITHS
8
1
Department of Psychology, Applied Health, School of Health and Biomedical Sciences,
RMIT University, Australia
2
Victoria University, Australia
3
National and Kapodistrian University of Athens, Greece
4
The Three Seas Psychology, Australia
5
Catholic Care Victoria, Australia
6
Mighty Serious, Australia
7
Quantum Victoria, Australia
8
International Gaming Research Unit, Psychology Department, Nottingham Trent University, UK
Received: July 17, 2023 •Revised manuscript received: September 4, 2023 •Accepted: October 10, 2023
ABSTRACT
Background and aims: Gaming disorder [GD] risk has been associated with the way gamers bond with
their visual representation (i.e., avatar) in the game-world. More specifically, a gamer’s relationship
with their avatar has been shown to provide reliable mental health information about the user in their
offline life, such as their current and prospective GD risk, if appropriately decoded. Methods:
To contribute to the paucity of knowledge in this area, 565 gamers (M
age
529.3 years; SD 510.6) were
assessed twice, six months apart, using the User-Avatar-Bond Scale (UABS) and the Gaming Disorder
Test. A series of tuned and untuned artificial intelligence [AI] classifiers analysed concurrently and
prospectively their responses. Results: Findings showed that AI models learned to accurately and
automatically identify GD risk cases, based on gamers’reported UABS score, age, and length of gaming
involvement, both concurrently and longitudinally (i.e., six months later). Random forests out-
performed all other AIs, while avatar immersion was shown to be the strongest training predictor.
Conclusion: Study outcomes demonstrated that the user-avatar bond can be translated into accurate,
concurrent and future GD risk predictions using trained AI classifiers. Assessment, prevention, and
practice implications are discussed in the light of these findings.
KEYWORDS
gaming disorder, avatar, user-avatar bond, machine learning, artificial intelligence, online gaming
INTRODUCTION
Since their commercial conception in the 1970s, videogames have become integrated into
modern popular culture (Will, 2019). Alongside a boom in technological advancements and
improved internet capabilities, the gaming industry has developed into a global community
allowing millions around the world, and in Australia (where the present study was carried
Journal of Behavioral
Addictions
DOI:
10.1556/2006.2023.00062
© 2023 The Author(s)
FULL-LENGTH REPORT
pCorresponding author.
E-mail: daniel.zarate.psychology@
gmail.com
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
out), to enjoy gaming as a shared activity (Statista, 2023;
Stavropoulos, Motti-Stefanidi, & Griffiths, 2022).
In the past two decades, gaming has greatly proliferated,
with recent nationwide data suggesting that approximately
70% of all Australians (i.e., 17 million) play videogames in
some form or frequency, while the vast majority of house-
holds (i.e., 8.6 million), including those with children, have
access to digital game devices (Brand, Todhunter, & Jervis,
2017). Alongside the growth of gaming, gaming pathologies
have begun to emerge (King et al., 2020). Literature high-
lights that while most gamers enjoy positive outcomes such
as psychomotor/dexterity, cognitive, health, and educational
benefits (Granic, Lobel, & Engels, 2014;Koulouris, Jeffery,
Best, O’Neill, & Lutteroth, 2020;Nuyens, Kuss, Lopez-Fer-
nandez, & Griffiths, 2017;Raith et al., 2021;Watson et al.,
2019), a minority of gamers may experience harmful effects
associated with excessive and/or disordered gaming (e.g.,
reduced educational/work performance, distress, loneliness;
Burleigh, Griffiths, Sumich, Stavropoulos, & Kuss, 2019;
Nuyens, Kuss, Lopez-Fernandez, & Griffiths, 2019;S
¸alvarlı
& Griffiths, 2022;Stavropoulos et al., 2019;Colder Carras,
Stavropoulos, Motti-Stefanidi, Labrique, & Griffiths, 2021;
Van Looy, 2015;
Spor
ci
c & Glavak-Tkali
c, 2018).
There is consensus that disordered gaming occurs as a
consequence of the interplay between factors related to the
individual players (e.g., personality, psychopathology), their
immediate and more distant environmental surroundings
(e.g., adverse family/peer interactions), as well as the game
applications themselves (e.g., reinforcement schedules; King
et al., 2019;Starcevic & Khazaal, 2020;Stavropoulos, Rennie,
Morcos, Gomez, & Griffiths, 2021). For instance, in relation
to individual factors, Király, Koncz, Griffiths, and Deme-
trovics (2023) highlighted disordered gaming risk factors
including gender (being male), age (being younger), per-
sonality traits (higher neuroticism, higher impulsivity, low
self-esteem), comorbidities (e.g., anxiety, autistic behav-
iours), motivation factors (e.g., escapism), and neurobio-
logical predispositions (e.g., reduced grey-matter volume in
the ventromedial and dorsolateral prefrontal brain areas). In
relation to environmental factors, disordered gaming risk
factors include poor quality of family relationships and
parental monitoring, childhood maltreatment and easy ac-
cess to gaming equipment, as well as pro-gaming peers and
broader cultural influences (Király et al., 2023). Finally, in
relation to specific structural characteristics of the game it-
self, disordered gaming risk factors include rewarding and
reinforcing gaming experiences through operant condition-
ing processes, online game delivery, monetization aspects
(e.g., buying/selling game winning equipment using offline
currencies), and distinct game genres (e.g., Massively
Multiplayer Online Role-Playing Games; MMORPGs;
involving character development, socialization, competition
and achievement elements; Király et al., 2023).
It should be noted that although higher gaming time has
been related to higher disordered gaming risk, scholars have
contended that it may not necessarily indicate disordered
gaming, unless it compromises functionality in the gamer’s
everyday life (e.g., employment, education, and family life;
Billieux, Flayelle, Rumpf, & Stein, 2019;Griffiths, 2010).
Consequently, it is emphasized that high gaming involve-
ment should be distinguished from disordered gaming
(Billieux et al., 2019;Griffiths, 2010). Such literature has led
to further calls for research examining the potentially
harmful consequences of excessive gaming, as well as better
identifying risk factors for developing problematic gaming
patterns (Király, Potenza, & Demetrovics, 2022).
Disordered gaming
The World Health Organization (WHO) officially included
gaming disorder (GD) in the 11th revision of the Interna-
tional Classification of Diseases (ICD-11; WHO, 2019). The
ICD-11 defines GD as a pattern of gaming behaviour
characterized by impaired control over gaming, increasing
priority given to gaming over other activities to the extent
that it takes precedence in daily life, and continuation/
escalation of gaming despite the occurrence of negative
consequences. The ICD-11 further states that a diagnosis of
GD must have a significant impairment to an individual’s
personal, family, social, educational, occupational and/or
other important areas of functioning (typically evident over
a period of at least 12 months). Given the increased recog-
nition of disordered gaming as a legitimate psychiatric
condition, research into more specific risk factors and po-
tential influencers of addictive gaming has greatly increased
(Bäcklund, Elbe, Gavelin, Sörman, & Ljungberg, 2022;Liao,
Chen, Huang, & Shen, 2022).
The WHO’s (2019) diagnostic classification of GD fol-
lowed the inclusion of the provisional diagnosis of internet
gaming disorder (IGD) in the fifth edition of the Diagnostic
and Statistical Manual for Mental Disorders (DSM-5;
American Psychiatric Association, 2013). According to the
DSM-5 (2013), and similar to WHO (2019) the criteria for
diagnosing IGD includes: preoccupation with gaming, with-
drawal symptoms when gaming is not possible, tolerance (i.e.,
needing to spend increasing amounts of time gaming), un-
successful attempts to control or reduce gaming, loss of in-
terest in other activities, continued excessive gaming despite
negative consequences, and significant impairment in per-
sonal, social, educational, or occupational areas of functioning
(with at least five of these criteria being met for more than a
year to be considered as having a gaming disorder).
In the present study, the ICD-11 criteria for GD (WHO,
2019) were employed for three compelling reasons: (i) it is
the only official (and not provisional) disordered gaming
diagnosis currently employed worldwide; (ii) it has been
supported that the ICD-11 diagnostic framework empha-
sizes more serious/pivotal (and a succinct number of)
GD symptoms, without compromising diagnostic validity
(Jo et al., 2019); and (iii) it provides consistency
and comparability in relation to empirical evidence
internationally (Pontes & Griffiths, 2019).
User-avatar bond
A number of scholars in the gaming studies field have reit-
erated that greater emphasis should be given to game-related
2Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
features. This includes the user-avatar bond (UAB), as a
potential GD risk factor in role-playing games (RPGs; Green,
Delfabbro, & King, 2021;Lemenager, Neissner, Sabo, Mann,
& Kiefer, 2020). RPGs have been consistently demonstrated
to be a genre of videogames that have a higher risk of GD
among individuals (Stavropoulos, Gomez, Mueller, Yucel, &
Griffiths, 2020;Stavropoulos, Pontes, Gomez, Schivinski, &
Griffiths, 2020;Szolin, Kuss, Nuyens, & Griffiths, 2022).
An avatar is a visual in-game representation of the player,
with the term originating from the Sanskrit word ‘avat
ara’,
referring to the embodiment of a deity in a human form
(Lochtefeld et al., 2002;Szolin et al., 2022).
Within the gaming context, the avatar facilitates a
process whereby the gamer may, to an extent, experience
embodiment with their gaming persona/figure, while they
are able to portray themselves in ways that align more with
their desired self-expressions (
Spor
ci
c&Glavak-Tkali
c,
2018;Stavropoulos, Gomez et al., 2020;Stavropoulos,
Pontes et al., 2020). Consequently, a complex psychological
attachment is facilitated between gamers and their avatars.
This increases game engagement and can also influence
some gamers’online and offline behaviours through sub-
conscious processes (e.g., altered perceptions, automatic
thoughts, and non-deliberate actions corresponding with
their avatar features; Burleigh, Stavropoulos, Liew, Adams,
&Griffiths, 2018;Liew, Stavropoulos, Adams, Burleigh, &
Griffiths, 2018;Ortiz de Gortari, Pontes, & Griffiths, 2015;
Ratan, Beyea, Li, & Graciano, 2020). Considering the
UAB’s particular strength/intensity, empirical research in-
dicates that factors such as age, and the duration of
engagement with the game world, may play a critical role
in the how an individual connects with their avatar
(e.g., younger gamers, with lengthier game involvement,
could be more UAB receptive/susceptible, due to more
dynamic/fluid personality features and time/emotional
game investment; Stavropoulos, Gomez et al., 2020;Stav-
ropoulos, Pontes et al., 2020;Stavropoulos, Ratan, & Lee,
2022;Rehbein, 2016).
Moreover, Blinka et al. (2008) noted that the UAB en-
compasses critical aspects and subdimensions. These entail
identification (e.g., the gamer becomes more like their
avatar, and they feel the same or alike), immersion (e.g., the
avatar’s needs in the world of the game [such as partici-
pating in a competition/task] are experienced as offline
needs by the gamer, and can even be prioritised to their
needs outside of the game [such as sleeping and/or eating]
in the case of disordered gaming), and compensation/
idealization (e.g., the avatar is who/how the gamer would
like to have been in their offline life, but they may not be
in a position to; the avatar may express an individual’s
ideal self).
Additionally, it has been argued that the need of some
gamers, who might be experiencing low-self-esteem and/or
may be dissatisfied by their offline self, could lead them to
escape their discomfort through their idealized avatars
within the game world (Stavropoulos, Gomez et al., 2020;
Stavropoulos, Pontes et al., 2020;Stavropoulos, Ratan et al.,
2022). Such avatar-mediated mood modification tendencies
may cause some gamers to immerse/over-engage with (and
emotionally depend on) their in-game character, fuelling
their GD risk (Stavropoulos, Gomez et al., 2020;Stavro-
poulos, Pontes et al., 2020;Stavropoulos, Ratan et al., 2022).
These findings are reinforced by other notable studies (e.g.,
those examining wishful avatar identification; Burleigh et al.,
2018;Green et al., 2021;Liew et al., 2018;Yee, Bailenson, &
Ducheneaut, 2009).
It has also been proposed that the UAB could operate as
a form of ‘digital phenotype’, meaning a digital/gamified
footprint of an individual’s mental health, that, if analysed,
can be translated into information not only concerning the
gamer’s risk of GD, but also for other psychopathological
conditions (e.g., depression, anxiety [Loi, 2019; Stavropoulos
et al., 2021; Zarate, Stavropoulos, Ball, de Sena Collier, &
Jacobson, 2022]). Despite the consistent associations be-
tween GD and the UAB in the extant literature, the trans-
lation of the UAB into GD risk has never to date, to the best
of the authors’knowledge, been investigated (Burleigh et al.,
2018;Liew et al., 2018;Ortiz de Gortari et al., 2015;Ratan
et al., 2020).
The present study
Analytical advancements in the field of machine learning
(ML) can support artificial intelligence (AI) applications,
which allow the automatic prediction/translation of one
form of information/data into another (e.g., a gamer’sUAB
into GD risk; Horton & Kleinman, 2015;Kuhn & Wick-
ham, 2020). To achieve such predictions, ML/AI pro-
cedures require training on related data, where predictors
(e.g., UAB sub-dimensions) and outcomes (e.g., GD risk)
are known, such that they can learn how to interpret/use
the first variable to identify the latter (in the form of su-
pervised algorithms; Horton & Kleinman, 2015;Kuhn &
Silge, 2022;Kuhn & Wickham, 2020).Afterthisstageis
completed, a new set of data is examined by the trained
AI/ML, where the accuracy of its predictions is validated
(i.e., while in the first stage of the process AI learns to
detect GD risk based on the UAB, in the second stage it
makes predictions to demonstrate their learning quality;
Kuhn & Silge, 2022).
Indeed, recent research examples have aimed to use
ML/AI to diagnose GD via Resting Brain State, MRI, PET
and EEG data with encouraging findings (Han et al., 2021;
Song et al., 2021). Taking these into consideration, the
present study innovatively examined a recently collected
longitudinal dataset using AI/ML classifiers, aiming to
translate gamers reported UAB identification, immersion,
and compensation/idealization into their present and pro-
spective (i.e., six months later) GD risk, while also taking
into consideration their age and years of videogame
engagement. In particular the choice of a longitudinal design
was chosen over cross-sectional data collection because it
allows the examination of the direction of causality between
the behaviours examined, while additionally enabling
the potential translation of the user-avatar bond into pro-
spective GD risk (Zarate, Dorman, Prokofieva, Morda, &
Journal of Behavioral Addictions 3
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Stavropoulos, 2023). Consequently, the following research
questions (RQs) were formulated:
RQ1: How can, if at all, ML/AI applications be trained to
identify whether a gamer presents with current GD risk,
based on their UAB reported identification, immersion,
compensation, age, and length of gaming involvement
(i.e., concurrent GD phenotype)?
RQ2: How can, if at all, ML/AI applications be trained to
identify whether a gamer presents with future GD risk
(i.e., six months later), based on their UAB reported
identification, immersion, compensation, age, and length
of gaming involvement (i.e., prospective GD phenotype)?
METHODS
Participants
A sample of 627 gamers were initially recruited. Of these,
seven were excluded as preview-only responses, 19 as spam,
one as a bot, 12 due to lack of consent, eight for failing
validity questions (e.g., claimed they played non-existing
games; e.g., Risk of Phantom), and 15 for insufficient re-
sponses. Therefore, the final sample comprised 565 role-
playing-gamers (M
age
529.3 years SD 510.6, Min
age
512,
Max
age
568; Males
cisgender
5283, 50.1%), who were
longitudinally assessed in the community, six months
apart (two time-points, T1 and T2). With regards to de-
mographics at T1, 271 (55.3%) reported being full-time
employed, 176 (36%) had an undergraduate degree, 359
(73.6%) stated heterosexual orientation, 410 (72.5%) iden-
tified as of Australian/English ancestry, 142 (25.1%) resided
with their family of origin, and 148 (30.2%) were single.
With regards to gaming patterns at T1, they reported
having been a gamer for on average for 5.62 years (Min5<1
year, Max 530 years; SD 54.49), for an average of 2.23 h
daily during weekdays (Min
5
<1 h, Max 515 h; SD 51.82)
and 3.39 h during the weekend (Min
5
<1 h, Max 518;
SD 52.40). Considering social media use patterns at T1,
they reported having been a social media user for an average
of 7.06 years (Min5<1 year, Max517; SD 57.06), spending
an average time of 2.55 h during weekdays (Min5<1 h,
Max 515 h; SD 52.16), and 3.01 h during the weekend
(Min5<1 h, Max 516 h; SD 52.48) with 145 (26%)
reporting Facebook as their preferred platform. The
maximum random sampling error for a sample of 565 at the
95% confidence interval (z51.96) equalled ±4.12% satis-
fying Hill’s (1998) recommendations. Missing values of
the analysed variables at T1 ranged between 3 (0.5% not
stating their age) to 16 (2.83% not answering Item 9 on the
User-Avatar Bond Scale), and were missing completely
at random in the broader dataset (MCAR
test
538.4,
p50.14
[9 missing patterns]
;Little (1988).
Attrition between waves was 276 participants (48.8%).
Therefore, retention/attrition were studied in relation to par-
ticipants’sociodemographic information considering statisti-
cal significance and effect size (Cohen’sd, very small∼0.01,
small∼0.20, medium∼0.50, large, 0.80, very large∼1.20;
Sawilowsky, 2009); Cramer’sV>0.255very strong,
>0.15 5strong, >0.10 5moderate, >0.05 5weak, >0 no or
very weak). Low to moderate effect-sizes were found
regarding the associations between attrition and gender
(χ
2
54.26, df 56, p50.642, Cramer’sV50.087),
sexual orientation (χ
2
57.75, df 54, p50.101, Cramer’s
V50.126), ancestry (χ
2
58.94, df 54, p50.063, Cramer’s
V50.126), romantic relationship engagement (χ
2
53.76,
df 54, p50.440, Cramer’sV50.088), educational status
(χ
2
511.2, df 57, p50.129, Cramer’sV50.152),
employment status (χ
2
57.58, df 56, p50.271, Cramer’s
V50.124), number of years spent gaming (t
Welch’s
53.509,
df 5526, p<0.001, Cohen’sd50.296), average daily gaming
time during the week (t
Student
50.873, df 5555, p50.383,
Cohen’sd50.0741), average daily gaming time during
the weekend (t
Student
50.159, df 5553, p50.874, Cohen’s
d50.0135), number of years spent using social media (t
Student
52.501, df 5556, p50.013, Cohen’sd50.2118), average
daily social media use time during the week (t
Student
5
2.313, df 5543, p50.021, Cohen’sd50.1983), average
daily social media use time during the weekend (t
Welch
5
2.447, df 5501, p50.015, Cohen’sd50.2111), and age
(t
Student
54.967, df 5560, p<0.001, Cohen’sd50.4192).
Tables 1 and 2provide detailed description of the sample
at T1.
Measures
In addition to data concerning demographics, gaming use,
and social media use, the following data were collected.
Gaming Disorder Test (GDT-4; Pontes et al., 2021). The
GDT-4 assesses the diagnostic features/severity of disordered
gaming with a design directly modelled on the WHO (2019)
conceptualisation. There are four items each addressing a
Table 1. Participant’s age, gaming/social media use years and daily week and weekend consumed time at T1
Age
Number of
years spent
gaming
Mean daily
gaming time
in the week
Mean daily
gaming time
at the weekend
Number of years
spent using
social media
Mean daily social
media use time
in the week
Mean daily social
media use time
at the weekend
N 562 556 557 555 558 545 543
Mean 29.3 5.62 2.23 3.39 7.06 2.55 3.01
SD 10.6 4.49 1.82 2.40 4.41 2.16 2.48
Min 12.0 0.00 0.00 0.00 0.00 0.00 0.00
Max 68.0 30.0 15.0 18.0 17.0 15.0 16.0
4Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Table 2. Participants’sociodemographic, gaming and social media use information at T1
N
Total
NProportion p
Gender Man (cisgender) 283 565 0.501 1.000
Woman (cisgender) 259 565 0.458 0.053
Man (transgender) 4 565 0.007 <0.001
Woman (transgender) 1 565 0.002 <0.001
Nonbinary 12 565 0.021 <0.001
Not Listed 3 565 0.005 <0.001
Prefer not to say 3 565 0.005 <0.001
Sexual Orientation Heterosexual-Straight 359 488 0.736 <0.001
Homosexual 36 488 0.074 <0.001
Bisexual 75 488 0.154 <0.001
Asexual 5 488 0.010 <0.001
Other 13 488 0.027 <0.001
Ancestry Aus./Engl. 412 565 0.552 0.015
Chinese 20 565 0.035 <0.001
German 7 565 0.012 <0.001
Indian 10 565 0.018 <0.001
Other 118 565 0.209 <0.001
Occupational Status Full-time employed 271 490 0.553 0.021
Part-time employed 77 490 0.157 <0.001
Student 64 490 0.131 <0.001
Trainee 2 490 0.004 <0.001
Not currently working 32 490 0.065 <0.001
On temporary leave (education leave, public service leave,
training, maternity leave)
5 490 0.010 <0.001
Other 39 490 0.080 <0.001
Educational Status Professional degree (i.e., MD, JD, etc. completed) 10 489 0.020 <0.001
PhD degree (completed) 17 489 0.035 <0.001
Postgraduate studies (MSc completed) 67 489 0.137 <0.001
Undergraduate university course (completed) 176 489 0.360 <0.001
Intermediate between secondary level and university
(e.g., technical training)
97 489 0.198 <0.001
Senior secondary school (Years 11–12) 101 489 0.207 <0.001
Secondary school (Years 7–10) 9 489 0.018 <0.001
Other 12 489 0.025 <0.001
Livingwith_w1 Family of origin (two parents/partners, only child) 34 564 0.060 <0.001
Family of origin (two parents/partners and siblings) 108 564 0.191 <0.001
Mother (only child, parent divorced-separated-widowed) 19 564 0.034 <0.001
Mother and sibling(s) (parent divorced-separated-widowed) 17 564 0.030 <0.001
Father (only child, parent divorced-separated-widowed) 6 564 0.011 <0.001
Father and sibling(s) (parent divorced-separated-widowed) 5 564 0.009 <0.001
With partner 149 564 0.264 <0.001
Alone 61 564 0.108 <0.001
With friend(s) 28 564 0.050 <0.001
Temporary accommodation 4 564 0.007 <0.001
Other 18 564 0.032 <0.001
With partner and children 115 564 0.204 <0.001
Relationship status Single 148 490 0.302 <0.001
In a romantic relationship (A romantic relationship is
defined as a romantic commitment of particular intensity
between two individuals of the same or the opposite sex
(When you like a guy [girl] and he [she] likes you back).
157 490 0.320 <0.001
Engaged 24 490 0.049 <0.001
Married 145 490 0.296 <0.001
De facto 16 490 0.033 <0.001
Partner games together Yes 99 344 0.288 <0.001
No 245 344 0.712 <0.001
Partner uses social media together Yes 227 340 0.677 <0.001
No 113 340 0.333 <0.001
(continued)
Journal of Behavioral Addictions 5
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
particular symptom (e.g., “I have had difficulties controlling
my gaming activity”)usingafive-point Likert-type scale from
1(Never)to5(Very often). Total scores range from 4 to 20
with higher scores indicating greater GD severity. Participants
at GD risk were classified those with more than 3/5 (Often)in
¾of the GDT-4 items (Pontes et al., 2021). The internal
consistency coefficients were sufficient across both study
waves (Cronbach’sα
GDT wave 1
50.808, McDonald’sω
GDT
wave 1
50.812, Cronbach’sα
GDT wave 2
50.854, McDonald’s
ω
GDT wave 2
50.862).
User-Avatar-Bond Questionnaire (UAB-Q; Blinka,
2008)
The UAB-Q was used to assess different gamer-avatar bond
dimensions. The 12 UAB-Q items are answered on a 5-point
Likert scale from 1 (strongly disagree)to5(strongly agree)
comprising three factors: identification (four items; “Both me
and my character are the same”), immersion (five items:
“Sometimes I think just about my character while not
gaming”), and compensation (three items: “I would rather be
like my character”). The scores range from 12 to 60 with
higher scores indicating stronger UAB experiences both
overall and on the respective subscales. The internal con-
sistency coefficients were sufficient across both study waves
(Cronbach’sα
UAB-Q wave 1
50.804; McDonald’sω
UAB-Q
wave 1
50.813, Cronbach’sα
UAB-Q wave 2
50.849; McDo-
nald’sω
UAB-Q wave 2
50.867, Cronbach’sα
Ident. wave 1
5
0.701; McDonald’sω
Ident. wave 1
50.729, Cronbach’sα
Ident.
wave 2
50.770; McDonald’sω
Ident. wave 2
50.789 Cronbach’s
α
Immers. wave 1
50.717; McDonald’sω
Immers. wave 1
50.727,
Cronbach’sα
Immers. wave 2
50.764; McDonald’sω
Immers.
wave 2
50.775, Cronbach’sα
Comp. wave 1
50.604; McDo-
nald’sω
Comp. wave 1
50.656, Cronbach’sα
Comp. wave 2
5
0.660; McDonald’sω
Comp. wave 2
50.709).
Procedure
Approvals were granted by the Victorian University Hu-
man Research Ethics Committee [HRE21-044], the
Department of Education and Training of The Victorian
State Government, Australia [2022_004542], and the Mel-
bourne Archdiocese of Catholic Schools [1179]. Partici-
pants were sampled from the community (e.g., RMIT,
Table 2. Continued
N
Total
NProportion p
Social media users Yes 550 565 0.973 <0.001
No 15 565 0.027 <0.001
Facebook users No 168 565 0.297 <0.001
Facebook 397 565 0.703 <0.001
Twitter users No 320 565 0.566 0.002
Twitter 245 565 0.434 0.002
Instagram users No 195 565 0.345 <0.001
Instagram 370 565 0.655 <0.001
Pinterest users No 469 565 0.830 <0.001
Pinterest 96 565 0.170 <0.001
TikTok users No 368 565 0.651 <0.001
Tik Tok 197 565 0.349 <0.001
Most preferred social media Facebook 145 557 0.260 <0.001
Twitter 66 557 0.118 <0.001
Instagram 135 557 0.242 <0.001
Pinterest 5 557 0.009 <0.001
Tik Tok 99 557 0.178 <0.001
Other, please define which 107 557 0.192 <0.001
Gaming with best friend No 336 565 0.595 <0.001
Yes 229 565 0.405 <0.001
Using social media with best friend No 189 565 0.335 <0.001
Yes 376 565 0.665 <0.001
Gaming with other friends No 312 565 0.552 0.015
Yes 253 565 0.448 0.015
Using social media with offline friends No 154 565 0.273 <0.001
Yes 411 565 0.727 <0.001
Gaming with family members No 406 565 0.719 <0.001
Yes 159 565 0.281 <0.001
Using social media with family
members
Yes 472 564 0.837 <0.001
No 92 564 0.163 <0.001
Note.Hₐis proportion ≠0.5.
6Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Victoria, Melbourne and Deakin Universities), Victorian
public and catholic schools, Australian gamers’groups
(e.g., Aus Gaymers Network), venues (e.g., Fortress Mel-
bourne), and online forums (e.g., AusGamers), as well as
advertising via YouTube videos. Gamers older than 12 years
were eligible to voluntarily/anonymously participate and
were provided with the plain language information state-
ment describing the study aims, risks and their participa-
tion rights (e.g., withdrawal without any penalties and/or
repercussions at any point) and provided their informed
consent. For adolescents (i.e., 12–18 years), these were
firstly addressed by their responsible parent/guardian and
secondly by the adolescents themselves. Data collection
involved three data-streams, paired via a non-identifiable
code, unique for each participant: (i) a battery of de-
mographic, internet/gaming/social media use questions,
and psychometric questionnaires/scales available via an
online Qualtrics link; (ii) wearing an actigraphy tracker
(Fitbit) for seven days to monitor physical activity/sleep
(e.g., daily steps and sleep duration), that was electronically
paired with the other data-streams via a unique code (i.e.,
records were automatically collected via the Fitbit portal
based on the participant’s code and those not owning a
Fitbit were provided with a device during a mutually
arranged/agreed meeting with the research team) and;
(iii) carrying a mobile monitoring application, called Aware
Light (Van Berkel, D’Alfonso, Susanto, Ferreira, & Kosta-
kos, 2023) recording screen on/off time, number and length
of calls (i.e., duration) and texts (i.e., length in characters)
for seven days (i.e., Light Aware data were also matched
with the other data-streams through the unique participant
code). The procedure was repeated four times, once every
six months, with the present study being based on the first
two completed collection waves (for detailed information
see Supplementary Materials 1.
Data analysis
To address RQ1 (i.e., concurrent GD digital phenotype;
identifying present GD risk based on an individual’sage,
number of years spent gaming, and reported avatar
identification, immersion and compensation/idealization)
machine learning (ML) procedures using the Tidymodels
package were conducted in R-Studio (Horton & Klein-
man, 2015;Kuhn & Wickham, 2020). Firstly, data were
balanced considering Yes/No GD risk cases to improve
learning/ML-prediction using the synthetic minority
oversampling technique (SMOTE; DMwR package;
Torgo & Torgo, 2013). This algorithm introduces addi-
tional cases of the minority group by taking into
consideration a potential number (k) of their nearest
neighbours based on Euclidean distance (Chawla, Bowyer,
Hall, & Kegelmeyer, 2002).
Practically, k-NN operates by identifying the distance
between a suggested case and all other data cases considered.
Firstly, it chooses a number (k) of cases nearest to the point
of interest. Then, it attaches the most frequent class to that
point (e.g., Yes/No GD risk; Chawla et al., 2002). Secondly,
data were split into 4/5 training and 1/5 testing, stratifying
Yes/No GD risk proportions to be equal across the splits,
while adopting a conservative bell-shaped Bayesian prior
distribution. It should be noted that when adopting a
Bayesian perspective, a potential distribution/variability is
required for every model parameter before proceeding to
data analysis. The range of these values was carefully/
modestly/conservatively suggested here to follow a Cauchy
shape (i.e., t-shape with seven degrees of freedom; Muth,
Oravecz, & Gabry, 2018).
Finalized training and testing datasets were similar
regarding Yes/No GD risk proportions (χ
2
50, df 51, p51).
For cross-validation and ML hyperparameters’tuning, training
data were additionally divided 10 times (i.e., folds) and training
data bootstrapped versions were also created. Thirdly, the ML
recipe (i.e., predictive equation) was introduced, such that: (i)
thebinaryYes/NoGDriskatT1wastheoutcomeandage,
number of years spent gaming, avatar-identification, avatar-
immersion and avatar-identification were the independent
predictors; (ii) a minimum ratio of 50% GD risk cases was
maintained across all samples tested, including the cross-vali-
dation and bootstrapped training data versions; and (iii) zero
variance, strongly sparse/skewed, and potentially highly inter-
correlated predictors were excluded, to solidify findings. It
should also be highlighted that the latter did not effectively
exclude any predictor in the current recipe.
Predictors were also scaled and centred prior to the
recipe to accommodate classification (i.e., 0 5mean and
15Standard Deviation [SD]; Kuhn & Wickham, 2020).
Fourthly, a series of supervised ML models (i.e., models
where the outcome is known in the training step/stage)
recommended for binary classification (see Table 3)
were introduced, alongside the null model (i.e., no ML
prediction) in their tuned and their untuned versions,
where hyperparameters were appropriately adjusted
(Kuhn & Wickham, 2020). A hyper-parameter constitutes
an ML parameter, the value of which needs to have been
specified prior to the learning ML being trained, in contrast
to simple parameters which are “learned”during the
training of the model. Therefore, hyperparameters pose
external model configurations (i.e., not based on the data)
employed for the estimation of model parameters. Fine-
tuned hyperparameters increase the capacity of a learning
model to perform with higher accuracy, and are achieved
through a “grid”process in tidymodels (Kuhn & Wick-
ham, 2020).
Fifthly, model and recipeswerecombinedtocreate
different workflows, which were: (i) trained in the default
versions on the training data; (ii) tuned considering their
hyperparameters via the bootstrapped versions the training
data, and; (iii) tested across their default/tuned versions on
the testing data. To address RQ2 (i.e., prospective GD
digital phenotype in six months), the same procedure
was repeated with GD T2 being the outcome/dependent
variable. Findings were compared based on their confusion
matrices, accuracy, precision, the area under the curve,
Journal of Behavioral Addictions 7
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Table 3. ML models trained, tuned and tested
Type Operation Hyperparameters tuned
R-package/engine
employed
Least Absolute Shrinkage
Selection Operator (LASSO)
LASSO constitutes a regression
analysis based, supervised ML
classifier, that applies variable
selection and regularization to
increase prediction accuracy. It
achieves that via reducing noises and
selecting certain features to
regularize the model. From a
calculation perspective lasso
considers the magnitude rate of the
coefficient, as a penalty to the loss
function. Therefore, the loss function
is amended to reduce model
complexity via restraining the sum of
predictors’coefficients [Loss
function 5OLS þA (penalty) X
summation (addition of s size[s] of
coefficients)].
penalty 5To perform regularization
(i.e., L1), LASSO considers/adds a
penalty to the size of regression
coefficients (i.e., predictor effects),
aiming to minimize them. The
optimum penalty value is obtained
via the tuning process.
glmnet
K Nearest Neighbours (k-NN) Th k-NN algorithm entails a
supervised, non-parametric
classification/prediction, that relies
on estimating proximity/relevance/
distance of one case with “k”others,
as per their Euclidean distance.
Alternatively, k-NN classifies/
categorizes a case taking into
consideration its neighbouring cases
(i.e., similarity of a case with
previously identified cases).
neighbors 5The number (k) of
neighbouring points to be
considered in order to optimize the
learning/prediction performance of
the algorithm, as defined via the
tuning process.
knn
Support Vector Machine Kernel
(SVM-K)
Kernel ML is based on pattern
examination/analysis and is mostly
known via its popular support-vector
machine (SVM) version. The kernel
function refers to a mathematic
procedure, which enables SVM to
pursue deep learning via conducting
bidimensional classifications of uni-
dimensional data through the
projection of a lower-dimension to a
higher one. Subsequently, a
kernelized SVM employs a linear
computation to address non-linear/
classification problems.
cost 5In SVM, cost resembles/
postulates the logistic function via a
piecewise linear. In practice, the cost
hyperparameter programs/guides the
algorithm’s optimization regarding
the rate/size of misclassification
allowed in the training sample.
Higher cost values indicate tighter
margins and the opposite.
degree 5The degree hyperparameter
dictates the flexibility/boundaries of
prediction(s), such that higher values
allow higher flexibility.
scale_factor 5The scaling hyper-
parameter of categorical/
classification kernel(s) reflects the
optimum normalization patterns/
process (i.e., kernel width) required
to avoid any data modification.
kernlab
X Gradient Boosting (XGB) XGBoost is recommended for
structured/tabular data. It
implements gradient boosted
decision trees to optimize prediction.
XGBoost does so via providing a
parallel tree boosting that integrates/
considers weak prediction/learner
models/decision trees. However, and
in contrast to random forest bagging
mtry 5The number of independent
variables to be randomly assessed at
each decision tree split.
min_n 5An integer/value/number
for the least data points in a node
(i.e., tree branch) that enables further
split.
tree_depth 5The value defining the
highest tree depth (i.e., subsequent
xgboost
(continued)
8Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Table 3. Continued
Type Operation Hyperparameters tuned
R-package/engine
employed
of generated trees, XG-Boosting
operates in a sequential manner,
with any subsequent tree being
influenced by the previous/last tree
outcome.
splits) suggested to optimize
prediction.
Learn rate (i.e., shrinkage) 5The
value/rate required for the boosting
adaptation to occur over successive
iterations. loss_reduction 5The
reduction rate of the loss function
suggested to progress with tree splits.
sample_size 5The amount/
proportion of data required to be
utilized in the algorithm’sfitting
process over each iteration.
Random Forests Random Forest is a flexible and
broadly employed supervised,
ensemble (i.e., composite) ML
model, that integrates/considers the
results of numerous decision trees
(i.e., bagging), while being trained/
learning to address a prediction/
classification task. Practically,
random forests conduct a meta-
estimation that averages/considers
the outcomes of multiple decision
tree classifiers, implemented on
different data sub-samples, to
improve accuracy and deter over-
fitting.
mtry 5The number of independent
variables to be randomly assessed at
each decision tree split.
min_n 5An integer/value/number
for the least data points in a node
(i.e., tree branch) that enables further
split.
ranger
Naïve Bayes Naïve Bayes operates as a
probabilistic, supervised, ML
classifier, which functions
generatively. This suggests that it
aims to model the data class
distribution, while assuming
conditional independence
probability (i.e., data characteristics/
measures are independent) to predict
the way a specific class would
generate input data.
smoothness 5This refers to the
Kernel component Smoothness,
which defines the density value
required for the algorithm to
converge quicker, to the real density
of random numeric predictors.
Laplace 5Laplace transformation/
smoothing refers to a technique/
strategy/method that addresses the
problem/risk of zero probability in
the algorithm.
naivebayes
Logistic Regression Logistic Regression is also
considered a supervised ML classifier
that employs a logistic function to
predict/model binary/dichotomous
dependent outcomes.
penalty 5In logistic regression, as
with LASSO, the regularization
penalty hyperparameter aims to
address generalization error and
therefore reduce overfitting risks. As
such, it enhances the probability of
simpler concluded models.
mixture 5A regularization
parameter value ranging between
0 and 1 to enhance model accuracy
[mixture 1 corresponds with LASSO;
0 with ridge regression and in the
interim with elastic modelling in
between LASSO and ridge].
glm
Note: Glmnet is derived from “Friedman et al. (2010). Package ‘glmnet’.CRAN R Repositary.”; Ranger is derived from “Wright and Ziegler
(2017). Package ‘ranger’.”Kernlab is derived from “Karatzoglou, Smola, and Hornik (2023). Package ‘kernlab’.CRAN R Project”. Xgboost is
derived from “Chen et al. (2023). Package ‘xgboost’.R version,90,1–66.”. All other engines ae derived from “Kuhn, M., & Silge, J. (2022).
Tidy Modeling with R."O’Reilly Media, Inc.".
Journal of Behavioral Addictions 9
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
recall, and f-measures (see yardstick r package; Kuhn,
Vaughan, & Vaughan, 2020).
1
Preceding the analysis, estimation for the sample size was
also considered from the overfitting perspective of the
developed models. In machine learning, overfitting refers to
the modelling error occurring, when a function used in a
model is too closely aligned to a limited set of data points.
This indicates insufficiency of the data and results in a
model generating accurate predictions for training data but
not for new/testing data (Chawla et al., 2002). The present
study addressed overfitting by considering the imbalance in
the dataset, and using the Synthetic Minority Over-Sampling
TEchnique (SMOTE; Chawla et al., 2002;Torgo & Torgo,
2013), applying early stopping in Random Forest applica-
tion, as well as use of regularization technique LASSO.
Further measures included cross-validation and hyper-
parameter tuning of the developed models (see Table 3).
Ethics
All procedures performed in the study involving human
participants were in accordance with the ethical standards of
the institutional and/or national research committee and
with the 1964 Helsinki declaration and its later amendments
or comparable ethical standards. The paper does not contain
any studies with animals performed by any of the authors.
Informed consent was obtained from all individual partici-
pants included in the study.
RESULTS
Before addressing RQ1 and RQ2, Yes/No GD risk
wave_1
participants were identified with N
no_GD_Risk
5430
(80.22%) and N
Yes_GD_Risk
5106 (19.78%). For RQ1, to
accommodate ML learning, oversampling of the minority
class was conducted using k-NN SMOTE (Chawla et al.,
2002;Torgo & Torgo, 2013) resulting in a balanced dataset
(i.e., N
Yes_GD_Risk
5530; 50%). Data were then split into
80% training and 20% testing and the proportions of Yes/No
GD risk were compared across the two parts showing
non-significant differences (χ
2
50, df 51, p51; Cramer’s
V50.00; 50% Yes GD risk across both training and testing).
The prediction recipe was introduced, scaling of predictors
was conducted, descriptives of the training, testing and
whole dataset were estimated (see bake recipe section;
Supplementary Material 2), while 10 sub-divisions and
bootstrapped versions of the training data were produced for
cross-validation and hyperparameter tuning (see folds
&
train_boot section, Supplementary Material 2). Models and
workflows of the Null, LASSO, SVM-Kernel, Random For-
ests, Naïve Bayes, and Logistic Regression (see Table 3)in
their default hyperparameter versions (i.e., untuned) were
then introduced, trained on the training data, and tested on
the testing data. Table 4 summarizes their performance
suggesting that, while all classifiers performed/learned
acceptably and better than the null model, except LASSO,
Random Forests learning outperformed other classifiers with
excellent indicators across all criteria (see Fig. 1). Immersion
was the most significant predictor for Random Forests (i.e.,
>25 points) with all other predictors exceeding 10 points
(see VIP section, Supplementary Material 2).
To optimize learning and modelling capacity, the versions
of LASSO, SVM-Kernel, Random Forests, Naïve Bayes and
Logistic Regression, as well as XGB and k-NN were later
tuned (see Table 3 regarding their respective hyper-
parameters’functions), trained on the training data and tested
on the testing data. Table 5 summarizes the tuned hyper-
parameters’values per classifier and Table 6 their perfor-
mance. Results suggest that, while all classifiers performed/
learned acceptably and better than the null model, including
LASSO, Random Forests learning outperformed other clas-
sifiers comparatively with excellent indicators across all
criteria, followed by XGB, SVM-Kernel, and k NN (see Fig. 2).
The same process was repeated for RQ2 with Random
Forests again outperforming other classifiers in both their
tuned and untuned versions. Tables 7–9summarize the per-
formance of the untuned versions, the tuned hyperparameters’
values, and the performance of the tuned classifiers respec-
tively. Figures 3 and 4visualize the performance of the tuned
and untuned models (see Supplementary Material 3 and 4).
DISCUSSION
The present longitudinal study employed a relatively large,
normative sample of gamers to train AI/ML automated
1
Accuracy reflects the ratio of correctly predicted cases, across the total
number of cases. It is produced through the accumulation of the true
positive and the true negative cases divided by the sum of all true positive,
true negative, false positive and false negative cases. Accuracy values closer
to 1 are considered desirable. Accuracy >0.90 5Excellent; 70%<Accu-
racy<90% 5Very good; 60%<Accuracy<70% 5Good; Accuracy<60% is
poor (Allwright, 2022).
Area under the curve (AUC) refers to the area under the receiver operating
characteristic (ROC) curve, as the latter is visualized in an orthogonal axis
system/graph, where the horizontal line captures the false positive rate
(FPR; 1 –specificity) and the vertical axis the sensitivity (True positive
rate [TPR]; values closer to 1 are considered better/improved). AUC <0.5
5No discrimination; 0.5<AUC<0.7 5Poor discrimination; 0.7<AUC<0.8
5Acceptable discrimination; 0.8<AUC<0.9 5Excellent discrimination;
AUC>0.9 5Outstanding discrimination (Statology, 2021).
Positive Predictive Value [PPV] or Precision is irrespective of the preva-
lence of a condition, and reflects the proportion/ratio of all the true positive
classified cases divided by the addition of the true positive and the false
positive cases (i.e., how many of those classified as positive were actually
positive? Values closer to 1 are considered better/improved).
Recall or sensitivity is associated to the prevalence of a condition and
reflects the proportion/ratio of all the true positive classified cases divided
by the sum of all the true positive and the false negative classified cases (i.e.,
how many of the true positive cases have been recalled? Values closer to 1
are considered better/improved).
Specificity reflects the proportion/ratio of all the true negative classified
cases divided by the sum of all the true negative and the false positive
classified cases (i.e., how many of the true negative cases have been
correctly classified? Values closer to 1 are considered better/improved).
F-Measure or F1-score/F-Score reflects the ratio of the multiplication of
recall and precision, multiplied by two and then divided by the accumu-
lation of recall and precision, such that the balance between precision and
recall achieved by the model is captured. Higher values are considered
better/improved (Jiao & Du, 2016).
10 Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
procedures to identify an individual’s concurrent and pro-
spective (i.e., six months later) GD risk, based on their age,
number of years spent gaming, and reported avatar identi-
fication, immersion, and compensation/idealization. Five
untuned (i.e., in their default versions) and seven tuned
(i.e., ML/AI hyper-parameters/calculation features specif-
ically adjusted to improve learning) recommended, and
widely employed ML classifiers, were comparatively exam-
ined twice (i.e., current and prospective GD risk; Blinka,
2008;Kuhn & Silge, 2022).
The data were split into training and testing parts for the
AIs to be trained and assessed respectively, while a predic-
tion recipe was introduced. The models were trained, tuned,
and tested, such that their capacity to learn whether an in-
dividual presents or not to be at GD risk at present and six
months later, could be confirmed. Findings demonstrated
that while all AI classifiers tested in the present study, were
able to learn and performed better than the null model
(i.e., random prediction), Random Forests had the strongest
learning potential. Of the UAB aspects identified, immersion
was the most important predictor of GD risk.
Gaming disorder and user-avatar bond
The present study’sfindings align with previous studies
suggesting that stronger/higher UAB experiences are more
likely to associate with excessive/disordered/problematic
gaming, when/if there is a tendency for the individual to
‘escape from reality’, as a result of identity-related issues
including poor self-concept, psychological vulnerability, and
‘wishful identification’(i.e., compensation for negative self-
perceptions; Green et al., 2021;Lemenager et al., 2020;
Spor
ci
c & Glavak-Tkali
c, 2018;Stavropoulos, Gomez et al.,
2020;Stavropoulos, Pontes et al., 2020;Van Looy, 2015).
Moreover, scholars have supported that one of the most
important indicators of GD is the process of transporting the
0.4
0.5
0.6
0.7
0.8
0.9
1
Null Model Random
Forests
Logisc
Regression
LASSO Naïve Bayes SVM Kernel
roc_auc ppv f_meas recall accuracy
Fig. 1. Untuned classifiers performance across the criteria (GD Wave 1)
Table 4. Null model and untuned algorithms performance on testing data (GD Wave 1)
Null model Random forests Logistic regression LASSO Naïve Bayes SVM Kernel
ROC_AUC 0.5 0.975 0.701 0.5 0.788 0.741
PPV 0.5 0.942 0.641 0.5 0.8 0.664
F_meas 0.667 0.933 0.673 0.667 0.712 0.685
Recall 1 0.925 0.708 1 0.642 0.708
Accuracy 0.5 0.934 0.656 0.5 0.741 0.675
Table 5. Hyperparameter tuning summary across classifiers
(GD Wave 1)
Type
Hyperparameters
tuned
Tuning
results
Least Absolute Shrinkage
Selection Operator (LASSO)
penalty 0.00139
K Nearest Neighbours (k-NN) neighbors 10
Support Vector Machine Kernel
(SVM-K)
cost 32
scale_factor 1
X Gradient Boosting (XGB) mtry 1
min_n 6
tree_depth 15
Learn rate
(i.e., shrinkage)
11
loss_reduction 0.0425
sample_size 0.171
Random Forests mtry 1
min_n 6
Naïve Bayes smoothness 0.5
Laplace 0
Logistic Regression penalty 0.00234
mixture 0.55
See Table 3 for detailed information regarding the classifiers
applied.
Journal of Behavioral Addictions 11
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
players’psyche into the gaming environment, that is, the
facilitation of a true “detachment from reality and the actual
self”(
Spor
ci
c & Glavak-Tkali
c, 2018, p. 8).
Relatedly, the player-avatar connection/interaction is
maintained by identification and idealisation, and subse-
quently strengthened through both the immersive qualities
of the game itself, and the ‘escape motives’of players (Green
et al., 2021;Lemenager et al., 2020;
Spor
ci
c & Glavak-Tkali
c,
2018;Stavropoulos, Gomez et al., 2020;Stavropoulos, Pontes
et al., 2020;Stavropoulos, Ratan et al., 2022). Therefore, the
immersion factor, expressing the experience of the avatar’s
needs as offline needs of the gamer, can be seen as advancing
UAB understanding, while sharpening the explanatory
framework for players vulnerable to GD (Stavropoulos,
Ratan et al., 2022). It is perhaps unsurprising that of all
the UAB aspects considered within the present study, im-
mersion was found to be the strongest predictor of GD risk.
In other words, whether a gamer resembles their avatar
(i.e., identification) or wishes to be like their avatar
(i.e., compensation/idealization) appears to induce lower GD
risk, compared to the extent that a gamer fuses with their
avatar’s needs, experiencing them as theirs (Ratan et al.,
2020). The latter increases more their GD likelihood and
Table 6. Tuned algorithms performance on testing data (GD Wave 1)
Null model Random forests Logistic regression LASSO Naïve Bayes SVM Kernel XGB k-NN
ROC_AUC 0.5 0.981 0.704 0.704 0.811 0.96 0.955 0.939
PPV 0.5 0.951 0.647 0.647 0.755 0.959 0.873 0.966
F_meas 0.667 0.938 0.676 0.676 0.725 0.916 0.889 0.876
Recall 1 0.925 0.708 0.708 0.698 0.877 0.906 0.802
Accuracy 0.5 0.939 0.66 0.66 0.736 0.92 0.887 0.887
0.4
0.5
0.6
0.7
0.8
0.9
1
Null
Model
Random
Forests
Logisc
Regression
LASSO Naïve
Bayes
SVM
Kernel
XGB k-NN
roc_auc ppv f_meas recall accuracy
Fig. 2. Tuned classifiers performance across the criteria (GD Wave 1)
Table 7. Null model and untuned algorithms performance on testing data (GD Wave 2)
Null model Random forests Logistic regression LASSO Naïve Bayes SVM Kernel
ROC_AUC 0.5 0.959 0.718 0.724 0.744 0.708
PPV 0.5 0.897 0.667 0.629 0.735 0.68
F_meas 0.667 0.897 0.643 0.65 0.673 0.63
Recall 1 0.897 0.621 0.672 0.621 0.586
Accuracy 0.5 0.897 0.655 0.638 0.698 0.655
Table 8. Hyperparameter tuning summary across classifiers
(GD Wave 2)
Type
Hyperparameters
tuned
Tuning
results
Least Absolute Shrinkage
Selection Operator (LASSO)
penalty 0.00569
K Nearest Neighbours (k-NN) neighbors 10
Support Vector Machine Kernel
(SVM-K)
cost 32
scale_factor 1
X Gradient Boosting (XGB) mtry 1
min_n 3
tree_depth 11
Learn rate
(i.e., shrinkage)
0.00268
loss_reduction 0.495
sample_size 0.336
Random Forests mtry 1
min_n 6
Naïve Bayes smoothness 0.5
Laplace 0
Logistic Regression penalty 0.0264
mixture 0.35
See Table 3 for detailed information regarding the classifiers
applied.
12 Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
presents an opportunity for AI to better learn to detect those
at risk of GD.
Furthermore, the methods employed in the present study
expand and advocate for the ML/AI translation of the UAB
into GD risk, while considering the age of the gamer and the
number of years they have spent gaming. Findings suggest
that the UAB could operate as a diagnostic indicator of GD
risk both at present and prospectively (six months later),
when addressed using trained ML/AI procedures. This
aligns with past literature recommending the careful
decoding/interpretation of the health/mental health infor-
mation likely embedded in the UAB (Stavropoulos et al.,
2021). Indeed, the avatar’s customization by the gamer,
allows conscious and less conscious projections of the
gamer’s wishes and characteristics into the avatar, such that
avatars and the way the gamers bond with them may prove
to be a valuable source of information (Stavropoulos, Ratan
et al., 2022).
These interpretations reinforce (and align with) the
proposed notion of ‘digital phenotype’, suggesting that an
individual’s cyber-behaviour and choices, such as their user-
avatar customization and bond, may operate as a unique
‘footprint’of what they are experiencing offline, if/when
appropriately translated (Loi, 2019; Stavropoulos et al., 2021;
Zarate et al., 2022). This possibility is additionally
strengthened by the work of Lemenager et al. (2020), who
reported: (i) a consistent association between disordered
gaming and bonding with the avatar, and; (ii) enhanced
Table 9. Tuned algorithms performance on testing data (GD Wave 2)
Null model Random forests Logistic regression LASSO Naïve Bayes SVM Kernel XGB k-NN
ROC_AUC 0.5 0.959 0.72 0.721 0.773 0.95 0.85 0.904
PPV 0.5 0.883 0.673 0.685 0.792 0.981 0.8 0.947
F_meas 0.667 0.898 0.655 0.661 0.717 0.946 0.741 0.75
Recall 1 0.914 0.638 0.638 0.655 0.914 0.69 0.621
Accuracy 0.5 0.897 0.664 0.672 0.741 0.948 0.759 0.793
0.4
0.5
0.6
0.7
0.8
0.9
1
Null
Model
Random
Forests
Logisc
Regression
LASSO Naïve
Bayes
SVM
Kernel
XGB k-NN
roc_auc ppv f_meas recall accuracy
Fig. 4. Tuned classifiers performance across the criteria (GD Wave 2)
0.4
0.5
0.6
0.7
0.8
0.9
1
Null Model Random
Forests
Logisc
Regression
LASSO Naïve Bayes SVM Kernel
roc_auc ppv f_meas recall accuracy
Fig. 3. Untuned classifiers performance across the criteria (GD Wave 2)
Journal of Behavioral Addictions 13
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
activation of brain regions during times an individual is
consumed by thoughts regarding their avatar. Interestingly,
the notion of game transfer phenomena, described as the
tendency of gamers to experience altered/involuntary cog-
nitions/thoughts, behaviours and perceptions outside of
their gaming sessions, has also been associated with suffering
from a medical condition and/or drug abuse, indirectly
advocating for the health phenotyping/footprint potential
of gaming behaviours including UAB (Ortiz de Gortari &
Griffiths, 2015).
IMPLICATIONS, LIMITATIONS, AND FURTHER
RESEARCH
The automation of the decoding of such information using
trained AI/ML procedures demonstrated here, likely revo-
lutionizes the potential use of the UAB as a cyber-pheno-
type, meaning a source of information about the health/
mental health of the user outside the game. More specif-
ically, findings of the present study may: (i) pave the way for
large-scale, avatar-mediated (and therefore, more gamer-
friendly), low-cost, ML/AI-facilitated GD risk diagnostic
procedures; (ii) help in the development of more effective
GD prevention strategies, through the targeting of AI-
detected GD risk gamer groups based on the way they bond
with their avatars and; (iii) encourage the implementation of
AIs for evaluation of user information potentially embedded
within the UAB. In particular, from a conceptual perspec-
tive, and in relation to the notion of digital phenotype, the
use of ML/AI to show the GD diagnostic potential of the
UAB, expands past studies in the field, suggesting the need
for exploration of further health and mental information
likely embedded within the UAB, independent of GD risk
(e.g. depression, anxiety; Lemenager et al., 2020;Loi, 2019;
Ortiz de Gortari & Griffiths, 2015).
Overall, the present study suggests that GD risk can be
predicted using ML/AI algorithms, that are capable of
combining different variables on a large scale with reduced
rates of misdiagnosis, providing more accurate diagnostic
and/or risk indicators. In turn, these techniques may provide
clinically relevant insights into assessment and save signifi-
cant time for clinicians. Furthermore, from a GD treatment
perspective, the present findings argue in favour of the uti-
lization of the user-avatar bond when addressing GD
symptoms. As Tisseron (2009) suggested, the UAB can
provide the map for more accurate case formulation that can
in turn drive more effective GD treatment plans, when and
where avatars are involved. For instance, by observing avatar
characteristics, possessions, and needs/commitments in the
virtual world (e.g. using the empty chair technique to invite
the ‘avatar’to talk in a disordered gamer’s session), clini-
cians may be able to work collaboratively with the treatment
seeker/receiver to understand what they could be missing in
their offline lives and plan how to pursue it to reduce their
game-dependency (Tisseron, 2009). However, the findings
of the present study should be interpreted taking into
account the limitations of the present study, which utilised a
rather small, community-sourced sample and relied exclu-
sively on self-reported data, that might invite potential biases
and confounding variables effects.
CONCLUSION
Despite such limitations, the present study innovatively
aimed to unlock the mental health diagnostic potential,
likely embedded within the UAB, through the pioneering
use of a sequence of different ML classifiers and emphasizing
an individual’s disordered gaming risk. It did so while
abiding with open science principles (i.e., accessible code
and findings), such that research teams in the field can
employ ML/AI to other already collected datasets related to
the UAB to corroborate or negate the present findings.
Furthermore, and in the context of the present study, ML/AI
is converted from a game mechanic employed by industry
to increase game engagement, and thus likely GD risk
(Millington, 2009) into a GD protective factor.
Funding sources: VS received funding by the RMIT Univer-
sity, Early Career Researcher Fund ECR 2020, number
68761601, and the Australian Research Council, Discovery
Early Career Researcher Award, 2021, number DE210101107.
Authors’contributions: VS contributed to the paper’s
conceptualization, data curation, formal analysis, method-
ology, project administration, and writing of the original
draft. DZ, MP, NVDB, LK, AGA, AG, SB and MDG
contributed to writing, reviewing, and editing the final draft.
Conflict of interest: The authors declare that they have no
known competing financial interests or personal relation-
ships that could have appeared to influence the work
reported in this paper. MDG has received research funding
from Norsk Tipping (the gambling operator owned by
the Norwegian government). MDG has received funding
for a number of research projects in the area of gambling
education for young people, social responsibility in
gambling and gambling treatment from Gamble Aware
(formerly the Responsibility in Gambling Trust), a chari-
table body which funds its research program based on
donations from the gambling industry. MDG undertakes
consultancy for various gambling companies in the area of
player protection and social responsibility in gambling. MR
currently works for Mighty Serious, a gaming company
focused on videogames aiming to drive positive behavioral
change.
SUPPLEMENTARY MATERIAL
Supplementary data to this article can be found online at
https://doi.org/10.1556/2006.2023.00062.
14 Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
REFERENCES
Allwright, S. (2022, May 14). What is a good accuracy score in
machine learning? Retrieved July 12, 2023, from: https://
stephenallwright.com/good-accuracy-score/.
American Psychiatric Association. (2013). Diagnostic and statistical
manual of mental disorders (5th ed.). American Psychiatric
Association.
Bäcklund, C., Elbe, P., Gavelin, H. M., Sörman, D. E., &
Ljungberg, J. K. (2022). Gaming motivations and gaming dis-
order symptoms: A systematic review and meta-analysis.
Journal of Behavioral Addictions,11(3), 667–688. https://doi.
org/10.1556/2006.2022.00053.
Billieux, J., Flayelle, M., Rumpf, H. J., & Stein, D. J. (2019). High
involvement versus pathological involvement in video games:
A crucial distinction for ensuring the validity and utility of
gaming disorder. Current Addiction Reports,6, 323–330.
https://doi.org/10.1007/s40429-019-00259-x.
Blinka, L. (2008). The relationship of players to their avatars in
MMORPGs: Differences between adolescents, emerging adults
and adults. Cyberpsychology: Journal of Psychosocial Research
on Cyberspace,2(1), 5. Full text: http://cyberpsychology.eu/
view.php?cisloclanku52008060901&article55.
Brand, J. E., Todhunter, S. & Jervis, J. (2017). Digital Australia
2018. Eveleigh, NSW: IGEA. Retrieved July 12, 2023, from:
https://igea.net/wp-content/uploads/2021/10/DA22-Report-
SUMMARY-FINAL-17-10-21.pdf.
Burleigh, T. L., Griffiths, M. D., Sumich, A., Stavropoulos, V., &
Kuss, D. J. (2019). A systematic review of the co-occurrence of
gaming disorder and other potentially addictive behaviors.
Current Addiction Reports,6, 383–401. https://doi.org/10.1007/
s40429-019-00279-7.
Burleigh, T. L., Stavropoulos, V., Liew, L. W., Adams, B. L., &
Griffiths, M. D. (2018). Depression, internet gaming disorder,
and the moderating effect of the gamer-avatar relationship:
An exploratory longitudinal study. International Journal of
Mental Health and Addiction,16, 102–124. https://doi.org/10.
1007/s11469-017-9806-3.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P.
(2002). Smote: Synthetic minority over-sampling technique.
Journal of Artificial Intelligence Research,16, 321–357. https://
doi.org/10.1613/jair.953.
Chen, T., He, T., Benesty, M., Khotilovic, V., Tang, Y., Cho, H., …
Yuan, J. (2023). XGBoost package. CRAN.
Colder Carras, M., Stavropoulos, V., Motti-Stefanidi, F., Labrique, A.,
and Griffiths, M. D. (2021). Draconian policy measures are
unlikely to prevent disordered gaming. Journal of Behavioral
Addictions,10, 849–853. https://doi.org/10.1556/2006.2021.
00075.
Friedman, J., Hastie, T., Tibshirani, R., Narasimhan, B., Tay, K.,
Simon, N., …Yang, J. (2010). Regularization paths for gener-
alized linear models via coordinate descent. Journal of Statis-
tical Software,33(1), 1–22. https://doi.org/10.18637/jss.v033.
i01.
Granic, I., Lobel, A., & Engels, R. C. M. E. (2014). The benefits of
playing video games. American Psychologist,69(1), 66–78.
https://doi.org/10.1037/a0034857.
Green, R., Delfabbro, P. H., & King, D. L. (2021). Avatar identifi-
cation and problematic gaming: The role of self-concept clarity.
Addictive Behaviors,113, 106694. https://doi.org/10.1016/j.
addbeh.2020.106694.
Griffiths, M. D. (2010). The role of context in online gaming excess
and addiction: Some case study evidence. International Journal
of Mental Health and Addiction,8, 119–125. https://doi.org/
10.1007/s11469-009-9229-x.
Han, X., Wei, L., Sun, Y., Hu, Y., Wang, Y., Ding, W., …Zhou, Y.
(2021). MRI-based radiomic machine-learning model may
accurately distinguish between subjects with internet gaming
disorder and healthy controls. Brain Sciences,12(1), 44. https://
doi.org/10.3390/brainsci12010044.
Hill, R. (1998). What sample size is “enough”in internet survey
research? Interpersonal Computing and Technology: An Elec-
tronic Journal for the 21st Century,6(3–4).
Horton,N.J.,&Kleinman,K.(2015).UsingRandRStudiofor
data management, statistical analysis, and graphics. CRC
Press.
Jiao, Y., & Du, P. (2016). Performance measures in evaluating
machine learning based bioinformatics predictors for classifi-
cations. Quantitative Biology,4,320–330. https://doi.org/
10.1007/s40484-016-0081-2.
Jo, Y. S., Bhang, S. Y., Choi, J. S., Lee, H. K., Lee, S. Y., &
Kweon, Y. S. (2019). Clinical characteristics of diagnosis for
internet gaming disorder: Comparison of DSM-5 IGD and
ICD-11 GD diagnosis. Journal of Clinical Medicine,8(7), 945.
https://doi.org/10.3390/jcm8070945.
Karatzoglou, A., Smola, A., & Hornik, K. (2023). Kernlab: Kernel-
based machine learning lab. R Package. https://cran.r-project.
org/web/packages/kernlab/index.html.
King, D. L., Chamberlain, S. R., Carragher, N., Billieux, J., Stein, D.,
Mueller, K., …Delfabbro, P. H. (2020). Screening and assess-
ment tools for gaming disorder: A comprehensive systematic
review. Clinical Psychology Review,77, 101831. https://doi.org/
10.1016/j.cpr.2020.101831.
King, D. L., Delfabbro, P. H., Perales, J. C., Deleuze, J., Király, O.,
Krossbakken, E., & Billieux, J. (2019). Maladaptive player-game
relationships in problematic gaming and gaming disorder:
A systematic review. Clinical Psychology Review,73, 101777.
https://doi.org/10.1016/j.cpr.2019.101777.
Király, O., Koncz, P., Griffiths, M. D., & Demetrovics, Z. (2023).
Gaming disorder: A summary of its characteristics and aeti-
ology. Comprehensive Psychiatry,122, 152376. https://doi.org/
10.1016/j.comppsych.2023.152376.
Király, O., Potenza, M. N., & Demetrovics, Z. (2022). Gaming
disorder: Current research directions. Current Opinion in
Behavioral Sciences,47, 101204. https://doi.org/10.1016/
j.cobeha.2022.101204.
Koulouris, J., Jeffery, Z., Best, J., O’Neill, E. & Lutteroth, C. 2020,
Me vs. Super(wo)man: Effects of customization and identifi-
cation in a VR exergame. In: Proceedings of the 2020 CHI
conference on human factors in computing systems. Vol. 2020-
April (pp. 1–17), CHI conference on human factors and
computing systems, Association for Computing Machinery,
New York, USA. https://doi.org/10.1145/3313831.3376661.
Kuhn, M., & Silge, J. (2022). Tidy modeling with R. O’Reilly Media,
Inc.
Journal of Behavioral Addictions 15
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Kuhn, M., Vaughan, D., & Vaughan, M. D. (2020). Package
‘yardstick’(version 1.2.0). Publisher of the software.
Lemenager, T., Neissner, M., Sabo, T., Mann, K., & Kiefer, F.
(2020). “Who am i”and “how should i be”: A systematic review
on self-concept and avatar identification in gaming disorder.
Current Addiction Reports,7, 166–193. https://doi.org/10.1007/
s40429-020-00307-x.
Liao, Z., Chen, X., Huang, Q., & Shen, H. (2022). Prevalence of
gaming disorder in East Asia: A comprehensive meta-analysis.
Journal of Behavioral Addictions,11(3), 727–738. https://doi.
org/10.1556/2006.2022.00050.
Liew, L. W., Stavropoulos, V., Adams, B. L., Burleigh, T. L., &
Griffiths, M. D. (2018). Internet gaming disorder: The interplay
between physical activity and user–avatar relationship. Behav-
iour
&
Information Technology,37(6), 558–574. https://doi.org/
10.1080/0144929X.2018.1464599.
Little, R. J. A. (1988). A test of missing completely at random for
multivariate data with missing values. Journal of the American
Statistical Association,83(404), 1198–1202. https://doi.org/10.
1080/01621459.1988.10478722.
Lochtefeld, J. (2002). Avatar. In The illustrated Encyclopedia of
Hinduism (Vol. 1, pp. 72–73). Rosen Publishing.
Loi, M. (2019). The digital phenotype: A philosophical and ethical
exploration. Philosophy
&
Technology,32(1), 155–171. https://
doi.org/10.1007/s13347-018-0319-1.
Millington, I., & Funge, J. (2009). Artificial intelligence for games.
CRC Press.
Muth, C., Oravecz, Z., & Gabry, J. (2018). User-friendly Bayesian
regression modeling: A tutorial with rstanarm and shinystan.
The Quantitative Methods for Psychology,14(2), 99–119.
https://doi.org/10.20982/tqmp.14.2.p099.
Nuyens, F., Kuss, D. J., Lopez-Fernandez, O., & Griffiths, M. D.
(2017). The experimental analysis of problematic video gaming
and cognitive skills: A systematic review. Journal de Thérapie
Comportementale et Cognitive,27(3), 110–117. https://doi.org/
10.1016/j.jtcc.2017.05.001.
Nuyens, F. M., Kuss, D. J., Lopez-Fernandez, O. & Griffiths, M. D.
(2019). The empirical analysis of non-problematic video
gaming and cognitive skills: A systematic review. International
Journal of Mental Health and Addiction,17, 389–414. https://
doi.org/10.1007/s11469-018-9946-0.
Ortiz de Gortari, A. B., & Griffiths, M. D. (2015). Game transfer
phenomena and its associated factors: An exploratory empirical
online survey study. Computers in Human Behavior,51,
195–202.https://doi.org/10.1016/j.chb.2015.04.060.
OrtizdeGortari,A.B.,Pontes,H.M.,&Griffiths, M. D.
(2015). The Game Transfer Phenomena Scale: An instru-
ment for investigating the nonvolitional effects of video
game playing. Cyberpsychology, Behavior, and Social
Networking,18(10), 588–594. https://doi.org/10.1089/cyber.
2015.0221.
Pontes, H., Schivinski, B., Sindermann, C., Li, M., Becker, B.,
Zhou, M., & Montag, C. (2021). Measurement and conceptu-
alization of gaming disorder according to the World Health
Organization Framework: The development of the Gaming
Disorder Test. International Journal of Mental Health and
Addiction,19, 508–528. https://doi.org/10.1007/s11469-019-
00088-z.
Pontes, H. M., & Griffiths, M. D. (2019). A new era for gaming
disorder research: Time to shift from consensus to consistency.
Addictive Behaviors,11, 100242. https://doi.org/10.1016/j.
addbeh.2019.106059.
Raith, L., Bignill, J., Stavropoulos, V., Millear, P., Allen, A.,
Stallman, H. M., …Kannis-Dymand, L. (2021). Massively
multiplayer online games and well-being: A systematic litera-
ture review. Frontiers in Psychology,12, 698799. https://doi.org/
10.3389/fpsyg.2021.698799.
Ratan, R., Beyea, D., Li, B. J., & Graciano, L. (2020). Avatar char-
acteristics induce users’behavioral conformity with small-to-
medium effect sizes: A meta-analysis of the Proteus effect.
Media Psychology,23(5), 651–675. https://doi.org/10.1080/
15213269.2019.1623698.
Rehbein, F., Staudt, A., Hanslmaier, M., & Kliem, S. (2016). Video
game playing in the general adult population of Germany: Can
higher gaming time of males be explained by gender specific
genre preferences? Computers in Human Behavior,55, 729–735.
https://doi.org/10.1016/j.chb.2015.10.016.
S
¸alvarlı,S
¸.
_
I., & Griffiths, M. D. (2022). The association between
internet gaming disorder and impulsivity: A systematic review
of literature. International Journal of Mental Health and
Addiction,20, 92–118. https://doi.org/10.1007/s11469-019-
00126-w.
Sawilowsky, S. (2009). New effect size rules of thumb. Journal of
Modern Applied Statistical Medhods,8(2), 597–599. https://doi.
org/10.22237/jmasm/1257035100.
Song, K. R., Potenza, M. N., Fang, X. Y., Gong, G. L., Yao, Y. W.,
Wang, Z. L., …Zhang, J. T. (2021). Resting‐state connectome‐
based support‐vector‐machine predictive modeling of internet
gaming disorder. Addiction Biology,26(4), e12969. https://doi.
org/10.1111/adb.12969.
Spor
ci
c, B., & Glavak-Tkali
c, R. (2018). The relationship between
online gaming motivation, self-concept clarity and tendency
toward problematic gaming. Cyberpsychology: Journal of Psy-
chosocial Research on Cyberspace,12(1), 4. https://doi.org/
10.5817/CP2018-1-4.
Starcevic, V., & Khazaal, Y. (2020). Problematic gaming, person-
ality, and psychiatric disorders. Frontiers in Psychiatry,10,
1004. https://doi.org/10.3389/fpsyt.2019.01004.
Statista (2023). Online games –worldwide. Retrieved July 12, 2023,
from: https://www.statista.com/outlook/dmo/digital-media/
video-games/online-games/worldwide.
Statology. (2021, September 9), What is considered a good AUC
score? Retrieved July 12, 2023, from: https://www.statology.org/
what-is-a-good-auc-score/.
Stavropoulos, V., Adams, B. L., Beard, C. L., Dumble, E., Trawley, S.,
Gomez, R., & Pontes, H. M. (2019). Associations between
attention deficit hyperactivity and internet gaming disorder
symptoms: Is there consistency across types of symptoms, gender
and countries? Addictive Behaviors Reports,9, 100158. https://
doi.org/10.1016/j.abrep.2018.100158.
Stavropoulos, V., Gomez, R., Mueller, A., Yucel, M., & Griffiths, M.
(2020). User-avatar bond profiles: How do they associate with
disordered gaming? Addictive Behaviors,103, 106245. https://
doi.org/10.1016/j.addbeh.2019.106245.
Stavropoulos, V., Motti-Stefanidi F. & Griffiths M. D. (2022).
Risks and opportunities for youth in the digital era. European
16 Journal of Behavioral Addictions
Unauthenticated | Downloaded 11/10/23 07:08 AM UTC
Psychologist,27(2), 86–101. https://doi.org/10.1027/1016-9040/
a000451.
Stavropoulos, V., Pontes, H. M., Gomez, R., Schivinski, B., &
Griffiths, M. (2020). Proteus effect profiles: How do they relate
with disordered gaming behaviours? Psychiatric Quarterly,91,
615–628. https://doi.org/10.1007/s11126-020-09727-4.
Stavropoulos, V., Ratan, R., & Lee, K. M. (2022). User-avatar bond:
Risk and opportunities in gaming and beyond. Frontiers in
Psychology,13, 923146. https://doi.org/10.3389/fpsyg.2022.
923146.
Stavropoulos, V., Rennie, J., Morcos, M., Gomez, R., &
Griffiths, M. D. (2021). Understanding the relationship between
the Proteus effect, immersion, and gender among World of
Warcraft players: An empirical survey study. Behaviour and
Information Technology,40(8), 821–836. https://doi.org/10.080/
0144929X.2020.1729240.
Szolin, K., Kuss, D., Nuyens, F., & Griffiths, M. (2022). Gaming
disorder: A systematic review exploring the user-avatar rela-
tionship in videogames. Computers in Human Behavior,128,
107124. https://doi.org/10.1016/j.chb.2021.107124.
Tisseron, S. (2009). L’ado et ses avatars. Adolescence,27(3), 591–600.
Torgo, L., & Torgo, M. L. (2013). DMwR: Functions and data for
data mining with R. R package version 0.4, 1. CRC Press.
Van Berkel, N., D’Alfonso, S., Susanto, R., Ferreira, D., &
Kostakos, V. (2023). AWARE-Light: A smartphone tool for
experience sampling and digital phenotyping. Personal and
Ubiquitous Computing,27,435–445. https://doi.org/10.1007/
s00779-022-01697-7.
Van Looy, J. (2015). Online games characters, avatars, and identity.
In: Mansell R. & Hwa Ang, P. (Eds.), The international
encyclopedia of digital communication and society (pp. 748–
758). Hoboken, NJ: Wiley-Blackwell. https://doi.org/10.1002/
9781118767771.wbiedcs106.
Watson, D. S., Krutzinna, J., Bruce, I. N., Griffiths, C. E.,
McInnes, I. B., Barnes, M. R., & Floridi, L. (2019). Clinical
applications of machine learning algorithms: Beyond the black
box. BMJ,364, l886. https://doi.org/10.1136/bmj.l886.
Wills, J. (2019). Gamer nation: Video games and American culture.
Johns Hopkins University Press.
World Health Organization (2019). International classification of
Diseases, Eleventh Revision. Retrieved September 1, 2023, from:
https://icd.who.int/browse11.
Wright, M., & Ziegler, A. (2017). Ranger: A fast implementation of
Random Forests for high dimensional data in Cþþ and R.
Journal of Statistical Software,77(1), 1–17. https://doi.org/10.
18637/jss.v077.i01.
Yee, N., Bailenson, J. N., & Ducheneaut, N. (2009). The Proteus
effect: Implications of transformed digital self-representation
on online and offline behavior. Communication Research,36(2),
285–312. https://doi.org/10.1177/0093650208330254.
Zarate, D., Dorman, G., Prokofieva, M., Morda, R., &
Stavropoulos, V. (2023). Online behavioral addictions: Longi-
tudinal network analysis and invariance across men and
women. Technology Mind and Behavor,4(2), 1–12. https://doi.
org/10.1037/tmb0000105.supp.
Zarate, D., Stavropoulos, V., Ball, M., de Sena Collier, G., &
Jacobson, N. C. (2022). Exploring the digital footprint of
depression: A PRISMA systematic literature review of the
empirical evidence.