Psychology of intimate relationships week 7

The Link Between Narcissism and Aggression: A Meta-Analytic Review

Sophie L. Kjærvik and Brad J. Bushman
School of Communication, The Ohio State University

This meta-analytic review examines the link between narcissism and aggression, and whether the link is stronger
under provocation conditions. A total of 437 independent studies were located, which included 123,043 partici-
pants. Narcissism was related to both aggression (r = .26, [.24, .28]) and violence (r = .23, [.18, .27]). As
expected, the narcissism-aggression link was stronger under provocation conditions (r = .29, [.23, .36]) than
under no provocation conditions (r = .12, [.05, .18]), but was even significant in the absence of provocation.
Both “normal” and “pathological” narcissism were related to aggression. All three dimensions of narcissism (i.e.,
entitlement, grandiose narcissism, vulnerable narcissism) were related to aggression. Narcissism was related to all
forms of aggression (i.e., indirect, direct, displaced, physical, verbal, bullying), and to both functions of aggres-
sion (i.e., reactive, proactive). The relation between narcissism and aggression was significant for males and
females, for people of all ages, for students and nonstudents, and for people from individualistic and collectivistic
countries. Significant results were obtained in experimental, cross-sectional, and longitudinal studies, in published
and unpublished studies, and in studies that assessed aggression using different types of measures (i.e., self-report,
other-report, observation). Overall results were robust to publication bias and the presence of outliers.
Theoretically, these results indicate that provocation is a key moderator of the link between narcissism and
aggression. Individuals high in narcissism have “thin skins” and are prone to aggression when they are provoked.
Practically, these results suggest that narcissism is an important risk factor for aggression and violence.

Public Significance Statement
This meta-analytic review found that individuals with high levels of narcissism are prone to aggres-
sive and violent behavior, especially when they are provoked. There was a triangulation of results
across different research methodologies. The relationship was robust for different forms and func-
tions of aggression, different types and dimensions of narcissism, and for males and females of dif-
ferent ages from different countries.

Keywords: narcissism, aggression, violence, ego-threat, meta-analysis

Supplemental materials: https://doi.org/10.1037/bul0000323.supp

When science discovers the center of the universe, a lot of people will
be disappointed to find they are not it.—Bernard Baily, American
comic book artist

This quote seems to accurately describe people with high levels
of narcissism. The personality trait narcissism was named after the
mythological Greek figure Narcissus, who was a handsome, self-
absorbed, vain young man who fell in love with his own image
reflected in still water. “I burn with love for—me!” Narcissus
exclaimed.

In this meta-analytic review, narcissism is defined as “entitled
self-importance” (Krizan & Herlache, 2018, p. 6). People with
high levels of narcissism think they are special people who
deserve special treatment. They have an exaggerated and inflated
sense of their own importance. Indeed, narcissism was initially la-
beled the “God-complex” (Jones, 1913).

Narcissism Is Not a Dichotomy

Overwhelming empirical evidence suggests that narcissism is
continuous variable rather than a dichotomy (for a review see Kri-
zan & Herlache, 2018). Thus, we will not refer to individuals with
high levels of narcissism as “narcissists.” Calling someone a nar-
cissist implies that narcissism is a dichotomous variable, which it

This article was published Online First May 24, 2021.
Brad J. Bushman https://orcid.org/0000-0002-1266-5101
We thank Alexandra Blouin, Cecily Exline, David Fuller, Jessica

Kaplan, Claire Teets, and Tess Wells for their help coding studies. We
thank Eddie Brummelman, Josh Miller, Aaron Pincus, Sander Thomaes,
and Aiden Wright for their helpful feedback on a draft of this article.
We thank Sven Kepes for his help conducting the sensitivity analyses,
Gertrudes Velasquez and Betsy Becker for their help conducting the
analyses for correlated effects, and Matthew Sweitzer for his help with the
violin plot. Finally, we would like to thank Golnoosh Behrouzian for her
help with Table SM1, the references, and computing intercoder reliability.
The data for this meta-analytic review are available on Open Science
Framework (OSF) via this link: https://osf.io/m6xek/.
Correspondence concerning this article should be addressed to Brad J.

Bushman, School of Communication, The Ohio State University, 3016
Derby Hall, 154 North Oval Mall, Columbus, OH 43210, United States.
Email: [email protected]

477

Psychological Bulletin

© 2021 American Psychological Association 2021, Vol. 147, No. 5, 477–503
ISSN: 0033-2909 https://doi.org/10.1037/bul0000323

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

is not. Even the latent structure of narcissistic personality disorder
(NPD) is continuous (Aslinger et al., 2018).
Because narcissism is a continuous variable, it has a distribu-

tion. However, the distribution of narcissism differs for children
and adults. In children ages 7–17, narcissism tends to have a nor-
mal, bell-shaped distribution (Thomaes & Brummelman, 2016). In
adults 18 and older, narcissism tends to have a positively skewed
distribution with a long right tail (Krizan & Herlache, 2018). In
other words, most adults have relatively low levels of narcissism,
some adults have moderate levels, and a few adults have extremely
high levels.

Narcissism Is Not a Unitary Construct

There is growing consensus that narcissism is not a unitary con-
struct. Rather, narcissism has important dimensions, although not
all researchers agree on what those dimensions are. Two independ-
ent groups of scholars recently suggested that the narcissism spec-
trum can best be represented using three correlated factors (Wright
& Edershile, 2018). One group proposed the Narcissism Spectrum
Model (Krizan & Herlache, 2018) to describe these three compo-
nents, whereas the other group proposed the Trifurcated model of
Narcissism (Weiss et al., 2019). Both groups identified an essential
core of narcissism, as well as two peripheral components. One
group called the core component entitlement (Krizan & Herlache,
2018), whereas the other group called it antagonistic (Miller et al.,
2017). In terms of the two peripheral components, one group
called them grandiosity and vulnerability (Krizan & Herlache,
2018), whereas the other group called them extraverted and neu-
rotic (Miller et al., 2017).
Data show that a three-factor model fits the data better than a

single-factor or two-factor model (Krizan & Herlache, 2018). In
this meta-analytic review, we focus on these three components,
and we use the terms entitlement, grandiose narcissism, and vul-
nerable narcissism. Individuals high in grandiose narcissism tend
to have high levels of self-esteem, whereas those high in vulnera-
ble narcissism tend to have low levels of self-esteem (Wink,
1991). The grandiose dimension of narcissism is characterized by
self-assuredness, imposingness, attention seeking, entitlement,
exhibitionism, self-indulgence, and disrespect for the needs of
others (Bushman et al., 2009; Wink, 1991). An individual high in
grandiose narcissism is often referred to as “hard-headed, aggres-
sive, and a show-off” (Wink, 1991, p. 569). The vulnerable dimen-
sion is characterized by hypersensitivity to evaluations from
others, defensiveness, bitterness, anxiousness, self-indulgence,
conceitedness, arrogance, and an insistence on having one’s own
way (Wink, 1991). An individual high in vulnerable narcissism is
often referred to as “unpredictable and prone to act out unexpect-
edly” (Greene, 1991, p. 283). In the past, more emphasis has been
given to the grandiose component of narcissism than to the vulner-
able component (Cain et al., 2008; Miller et al., 2011; Wright &
Edershile, 2018), both in the clinical (i.e., NPD) and subclinical
domains.
Clinically, the American Psychiatric Association’s (APA) Diag-

nostic and Statistical Manual of Mental Disorders-Fifth Edition
(DSM–5; APA, 2013) has consistently emphasized grandiose fea-
tures of narcissism. Likewise, the most popular self-report meas-
ures of subclinical narcissism, such as the Narcissistic Personality
Inventory (NPI; Raskin & Terry, 1988), have emphasized the

grandiose features of narcissism. Unfortunately, this has led to an
unbalanced and incomplete picture of narcissism (Wright & Eder-
shile, 2018). More recently, researchers have begun developing
and measuring the vulnerable features of narcissism. This meta-
analytic review will examine the relation between these three fac-
tors of narcissism (i.e., entitlement, grandiose narcissism, vulnera-
ble narcissism) and aggression.

Pathological Levels of Narcissism

Some scholars have suggested that narcissism has both nor-
mal and pathological expressions (e.g., Pincus & Lukowitsky,
2010). Normal narcissism is sometimes called “subclinical nar-
cissism” (e.g., Wallace & Baumeister, 2002). Pathological lev-
els of narcissism are associated with clinically relevant mental
health problems such as depression, suicidal ideation, self-
harm, pathological gambling, homicidal ideation, stalking
behavior, aggression, and violence (Schoenleber et al., 2015).
Those with pathological levels of narcissism have significant
regulatory deficits and maladaptive strategies to cope with self-
image threats. Thus, individuals with pathological narcissism
might be especially prone to lash out at others in an aggressive
manner when they are threatened.

In clinical and psychiatric research, pathological expressions of
narcissism are typically operationalized using the APA’s DSM–5
(APA, 2013) as NPD, which is defined as a “pervasive pattern of
grandiosity (in fantasy or behavior), a constant need for admira-
tion, and lack of empathy” (p. 669). Although the latent structure
of NPD is continuous (Aslinger et al., 2018), NPD is a dichoto-
mous categorization made by mental health professionals typically
following clinical interviews. Researchers have also developed a
self-report measure of pathological narcissism called the Patholog-
ical Narcissism Inventory (PNI; Pincus, 2013; Pincus et al. 2009).
The PNI measures two dimensions of narcissism—grandiosity and
vulnerability. The grandiosity and vulnerability subscales of the
PNI are highly correlated, and individuals high in “pathological
narcissism” are thought to have problems with both grandiosity
and vulnerability (e.g., Pincus, 2013), although this is an area in
need of further research (e.g., Gore & Widiger, 2016; Hyatt et al.,
2018). Other self-report measures of narcissism capture so-called
“normal narcissism.”

Other scholars disagree that the distinction between normal and
pathological narcissism is a meaningful one (e.g., Miller et al.
2017). They argue that whether narcissism is pathological depends
on how extreme, inflexible, and pervasive it is, and on whether it
causes distress and impairment to the self and to others. These
scholars also point out that so-called “normal” measures of narcis-
sism, such as the NPI (Raskin & Terry, 1988), are highly corre-
lated with clinical symptoms of NPD and ratings of psychopathy
(Krusemark et al., 2018; Miller et al., 2009) and generate empiri-
cal profiles consistent with expert ratings (Miller et al., 2014,
2016). In addition, so-called normal narcissism is often correlated
with maladaptive behaviors, such as aggression and violence (e.g.,
Bushman & Baumeister, 2002).

This meta-analytic review examines the link between both nor-
mal and pathological levels of narcissism and aggression. We will
explore whether the link between pathological narcissism (i.e.,
with high PNI scores, NPD) and aggression is stronger than the
link between normal narcissism and aggression.

478 KJÆRVIK AND BUSHMAN

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

Aggression and Violence

The main outcome variable in this meta-analytic review is aggres-
sion, which is defined as “any form of behavior directed toward the
goal of harming or injuring another living being who is motivated to
avoid such treatment” (Baron & Richardson, 1994, p. 7). Aggression
can take different forms (Bushman & Huesmann, 2010). For example,
physical aggression involves harming others physically (e.g., hitting,
kicking, stabbing, or shooting), whereas verbal aggression involves
harming others with spoken or written words (e.g., yelling, screaming,
swearing, or name calling). Aggression can be expressed directly
when the victim is present, or indirectly when the victim is absent (e.
g., spreading gossip or rumors behind someone’s back). Aggression
can also be displaced onto an innocent target. Bullying involves
repeated acts of aggression, which can occur offline or online (called
cyberbullying). Offline, the bully is generally large, strong, and intimi-
dating so they can “push other people around.” Online, the bully does
not need to possess these characteristics. Because there has been some
debate about whether there is a meaningful distinction between bully-
ing and cyberbullying (e.g., Olweus & Limber, 2018), we coded
whether the bullying occurred offline or online. We also coded trolling,
which involves intentionally provoking others online. In this meta-ana-
lytic review, we coded the form of aggression for exploratory
purposes.
Aggression can also serve different functions. Researchers have

distinguished between reactive and proactive aggression (Bush-
man & Anderson, 2001). Reactive aggression (also called hostile,
emotional, affective, and impulsive aggression) is “hot tempered”
annoyance-based aggression that is an “end in itself.” Proactive
aggression (also called instrumental and premeditated aggression)
is “cold-blooded” incentive-based aggression that is a “means to
some other end” (e.g., money, status, power, reputation, revenge,
prestige). In this meta-analytic review, we also coded the function
of aggression. Because individuals high in narcissism are hyper-
sensitive to provocation, we expect narcissism to be positively
relate to reactive aggression. But narcissism might also be posi-
tively related to proactive aggression, especially when it comes to
obtaining status (Grapsas et al., 2020). Some scholars have also
suggested that the distinction between reactive and proactive
aggression is not particularly useful because aggression can be
both annoyance-based and incentive-based, and because the two
types of aggression tend to be very highly correlated (see Bush-
man & Anderson, 2001). Thus, narcissism should be related to
both reactive and proactive aggression.
We also examined the link between narcissism and violence,

which is defined as any aggressive behavior intended to cause
extreme physical harm, such as injury or death (Bushman & Hues-
mann, 2010). All violent behaviors are aggressive, but only behav-
iors intended to cause extreme physical harm are violent. For
example, hitting someone is violent, but calling someone an insult-
ing name is not violent, although it is aggressive. We expected
narcissism to be more strongly related to aggression than to vio-
lence, because rare and extreme behaviors have a restricted range
and are more difficult to predict.

Narcissism and Aggression

There is a large body of research examining the link between
narcissism and aggression. Individuals with high levels of

narcissism have several characteristics that predispose them to
behave in an aggressive manner. Some aspects of narcissism are
directly related to aggression, such as being antagonistic and dis-
agreeable (Krizan & Herlache, 2018; Vize et al., 2019), and lack-
ing empathy for others (Czarna et al., 2015; Hepper et al., 2014;
Jones, 1913).

How people view themselves can also influence how they treat
others, including whether they treat them in an aggressive manner.
People with high levels of narcissism view themselves as superior
beings. Viewing oneself as superior implies viewing others as in-
ferior (Brummelman et al., 2018; Grapsas et al., 2020). Of course,
it is easier to aggress against others if one considers them to be in-
ferior beings.

This meta-analytic review integrates this body of research and
examines possible moderator variables. Theoretically, the most
important moderator variable is provocation.

Provocation

Some theoretical models predict high levels of aggression
among individuals high in narcissism in certain contexts and situa-
tions. According to the threatened egotism model (Baumeister et
al., 1996), individuals high in narcissism levels are especially
prone to aggress against others who evaluate them negatively,
such as by shaming, humiliating, embarrassing, or criticizing
them. Although negative evaluations can be unpleasant for every-
one, they are especially unpleasant and threatening for individuals
high in narcissism due to their inflated egos and their “thin skins.”
The model also predicts that individuals high in narcissism are
likely to reject negative evaluations from others, which allows
them to maintain their inflated self-views. Finally, the model pre-
dicts that individuals high in narcissism will be more likely to
retaliate against those who evaluate them negatively by aggressing
against them.

According to the Status Pursuit in Narcissism (SPIN) model
(Grapsas et al., 2020), individuals high in narcissism are also
prone to aggress against others in their attempts to obtain status.
For example, narcissistic youth, especially boys, are prone to bully
others to obtain dominance or status in their peer group (Reijntjes
et al., 2016). Most models of narcissism, including the threatened
ego model, propose that the main motive of people high in narcis-
sism is the desire to create and maintain an overly positive, grandi-
ose self-image, which is an intrapersonal motive. In contrast, the
SPIN model proposes that this desire to have a grandiose self-
image serves a higher-order social motive—to obtain social status,
which is an interpersonal motive. From an evolutionary perspec-
tive, this makes good sense. A grandiose self-image helps people
accrue higher status and social benefits by convincing others of
their superior skills and abilities.

The recent literature seems to suggest that narcissistic aggres-
sion stems from provocation more generally (e.g., Rasmussen,
2016), although some studies find that narcissistic aggression can
also occur in the absence of provocation (e.g., Reidy et al., 2010).
In this meta-analytic review, we code provocation as a key moder-
ator variable. We predict the link between narcissism and aggres-
sion to be stronger when participants are provoked than when they
are not provoked. If narcissism is related to aggression in nonpro-
voking contexts, it would have interesting theoretical implications.
Perhaps seeing others as inferior beings is sufficient to ignite

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 479

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

aggression in individuals high in narcissism, even in the absence
of external provocation.
We also code whether the provocation is related to status (e.g.,

respect, prominence, or reputation) or affiliation (e.g., belonging,
liking, or closeness). The SPIN model (Grapsas et al., 2020) would
predict higher levels of narcissistic aggression for status-related
provocations than for affiliation-related provocations.

Exploratory Moderator Variables

In addition to theoretical moderators described above, this meta-
analytic review also considered several additional characteristics
that might moderate the relation between narcissism and aggres-
sion, including source, participant, and design characteristics.

Source Characteristics

As one measure of publication bias, we coded whether each
study was published in a peer reviewed journal or not. In addition,
we used more formal methods for assessing the impact of publica-
tion bias by conducting a comprehensive battery of sensitivity
analyses.
We also coded the year the data were collected. We explore

whether the relation between narcissism and aggression is stable
over time.

Participant Characteristics

Previous research has reported that males have higher levels of
narcissism than females (Grijalva et al., 2015). When it comes to
gender differences in aggression, males exhibit higher levels of
physical aggression than females do, whereas females exhibit
higher levels of indirect aggression than males do (Björkqvist,
2018). This meta-analytic review explores whether there are gen-
der differences in the relation between narcissism and aggression.
Consistent with evolutionary theory, several scholars have sug-

gested that narcissism levels reach a developmental high in adoles-
cence (e.g., Bleiberg, 1994), and decrease across the life span
(Foster, Campbell, & Twenge, 2003, Trzesniewski & Donnellan,
2010), until they reach a developmental low in older age (Carter &
Douglass, 2018; Robins et al., 2002). This meta-analytic review
explores whether age moderates the relation between narcissism
and aggression.
Researchers in psychology and related fields have questioned

whether the results obtained from studies that use samples of stu-
dents from universities and colleges generalize to other popula-
tions (e.g., Ashraf & Merunka, 2017; Briones & Benham, 2017;
Espinosa & Ortinau, 2016; Flere & Lavri�c, 2008; Hanel & Vione,
2016; James & Sonner, 2001; Landers & Behrend, 2015; Sears,
1986). Some studies find differences between student and nonstu-
dent samples, whereas other studies find no differences. This
meta-analytic review explores whether the relation between narcis-
sism and aggression differs for students and nonstudents.
Higher narcissism levels have been found in people from indi-

vidualistic countries than in people from collectivistic countries
(Foster, Campbell, & Twenge, 2003; Foster, Shrira, & Campbell,
2003; Twenge & Campbell, 2009). This meta-analytic review
explores whether the relation between narcissism and aggression
differs for people from individualistic and collectivistic countries.

Design Characteristics

This meta-analytic review explores the relation between narcis-
sism and aggression in experimental, cross-sectional, and longitu-
dinal studies. For experimental studies, we also coded whether the
experiment was conducted in a laboratory, online, or field setting.
There has been some debate about whether online and offline
experiments produce similar results (e.g., Horton et al., 2011;
Reips, 2000). Finally, we explore whether similar results are
obtained when aggression is assessed via self-report, other-report
(e.g., parent, teacher, or peer), and behavioral observation.

Previous Meta-Analytic Reviews

Although other meta-analytic reviews have examined the link
between narcissism and aggression (Bettencourt et al., 2006; Geel
et al., 2017; Hyatt et al., 2019; Lambe et al., 2018; Moor & Ander-
son, 2019; Muris et al., 2017; Rasmussen, 2016), this meta-ana-
lytic review is far more comprehensive than past meta-analytic
reviews. For example, it includes many more studies, it includes
unpublished studies, it examines normal and pathological narcis-
sism, it examines different dimensions of narcissism, it examines
different forms and functions of aggression, it includes provoca-
tion as a key moderator variable, and it includes several other ex-
ploratory moderator variables. It controls for correlated effects. It
also includes a comprehensive battery of sensitivity analyses to
examine the effects of publication bias and outliers on the obtained
results.

Method

Literature Search

Five computer databases were used to locate relevant studies
published through November 15, 2020: Medline, PsycINFO, Edu-
cational Resources Information Center (ERIC), PubMed, and Pro-
Quest. We searched for unpublished studies in ProQuest to
address potential publication bias (i.e., the file drawer problem;
Rosenthal, 1979). The terms used for the literature search were
(“dark core” OR “dark personality traits” OR “dark traits” OR
“dark tetrad” OR “dark triad” OR ego* OR entitlement OR gran-
dios* OR narcissis* OR NPD) AND (aggress* OR agonistic OR
“antisocial behavior” OR “antisocial behaviour” OR “anti-social
behavior” OR “anti-social behaviour” OR assault* OR attack*
OR bullie* OR bully* OR conflict* OR coerc* OR cruel* OR der-
ogate* OR dominant* OR “externalising behaviour”* OR “exter-
nalizing behavior”* OR “externalising problems”* OR
“externalizing problems”* OR fight* OR harass* OR hostil* OR
“negative reciprocity” OR rape* OR reveng* OR retaliat* OR
sadis* OR troll* OR venge* OR victim* OR violen*). The asterisk
allows terms to have all possible endings (e.g., the term aggress*
will retrieve studies that used the terms aggress, aggressed,
aggressor, aggressive, and aggression).

Most of the aggression terms are the ones used in the journal
Aggressive Behavior, which is the flagship journal of the Interna-
tional Society for Research on Aggression. We also used search
terms from previous meta-analytic reviews of narcissism and
aggression. In addition, we added search terms based on a

480 KJÆRVIK AND BUSHMAN

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

thesaurus search and our reading of the literature. The search terms
were limited to be in the Title and/or Abstract.
We also were going to include the terms “conduct problems,”

“oppositional behavior,” and “oppositional behaviour” as aggres-
sion terms, but such measures often lump aggressive behavior in
with other antisocial behaviors (e.g., stealing, lying, cheating). We
excluded these terms because this meta-analytic review focuses
specifically on aggression.
To obtain studies that we might have been missed, two addi-

tional steps were taken. First, we searched the reference sections
of previous meta-analytic reviews (Bettencourt et al., 2006; Geel
et al., 2017; Hyatt et al., 2019; Lambe et al., 2018; Moor & Ander-
son, 2019; Muris et al., 2017; Rasmussen, 2016). Second, we
searched the reference sections of all retrieved studies.

Inclusion Criteria

Four inclusion criteria were used. First, the study had to include
a measure of narcissism. Second, the study needed to examine nar-
cissism at the individual level rather than at the group or collective
level (e.g., de Zavala, 2011; de Zavala et al., 2019). Third, the
study had to include a measure of aggression and/or violence.
Fourth, participants had to be at least 7 years old. One cannot
accurately measure narcissism in children under about age 7.
Young children typically hold inflated views of themselves and
their competencies. One reason is that they are not yet well able to
compare their own competencies with those of others, or to incor-
porate such self-other comparisons into their self-views (Ruble &
Frey, 1991). Moreover, young children find it difficult to distin-
guish their ideal self (the person they want to be) from their actual

self (the person they really are; Harter, 2012). From about age 7,
however, children develop more realistic self-views, and it is from
this age that individual differences in narcissism can be meaning-
fully assessed (Thomaes & Brummelman, 2016).

Preferred Reporting Items for Systematic Reviews and
Meta-Analyses (PRISMA) Flowchart

The literature search produced 11,862 research reports, but not
all of them were relevant to this meta-analytic review. To deter-
mine whether research reports were relevant, we read their titles
and abstracts. As shown in the PRISMA flowchart in Figure 1, the
final sample included 320 research reports (each marked with an
asterisk in the References section), which included 437 independ-
ent studies involving 123,043 participants.

Theoretical Moderator Variables

Normal Versus Pathological Narcissism

Pathological narcissism can be determined using the classifica-
tion of NPD or the PNI (Pincus, 2013; Pincus et al. 2009; includ-
ing the brief PNI; Schoenleber et al. 2015).

Narcissism Dimensions

We examined three dimensions of narcissism: (1) entitlement, (2)
grandiose narcissism, and (3) vulnerable narcissism. We also examined
pathological narcissism, which includes elevated levels of both grandi-
osity and vulnerability. To code these different aspects of narcissism,
we used classifications provided by other researchers based on several
different narcissism scales (e.g., Crowe et al., in press; Krizan & Herl-

Figure 1
PRISMA Flowchart of Literature Search

Research reports
identified through

PsycINFO, Medline,
ERIC, PubMed,

ProQuest
(n = 11,862)

Additional research
reports identified
through reference
sections of other
reviews (n = 129)

Research reports
excluded after reading
title/abstract or more

(n = 9,965)

Research reports
excluded due to

duplication
(n = 1,672)

Total number of
research reports

identified
(n = 11,991)

Research reports were
excluded due to
insufficient data

(n = 34)

Research reports
included in the final

review
(n = 320)

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 481

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

ache, 2018; Weiss et al. 2019; Wright & Edershile, 2018; see Table
SM1 in online supplementary materials).
Because we were interested in dimensions of narcissism that are

often formed using different subscales, we asked 83 authors to
send us correlations for different narcissism subscales and aggres-
sion if they were not included in the original research report. Of
these 83 authors, 43 (52%) responded to our requests and provided
the missing subscale data, 28 (34%) did not respond to our initial
or follow-up e-mail, seven (8%) said the data were inaccessible
(e.g., due to coronavirus disease 2019 [COVID-19], data stored on
an old computer), and five (6%) could not be contacted (e.g., were
deceased, no forwarding e-mail). Thus, the analyses based on sub-
scales should be interpreted with some caution.

Aggression and Violence

This meta-analytic review examined the relation between nar-
cissism and both aggression and violence. In laboratory experi-
ments, the most widely used behavioral measure of aggression is
the competitive reaction time (RT) task (Warburton & Bushman,
2019), where participants deliver unpleasant noise blasts through
headphones to an ostensible opponent each time they beat them on
a RT task. Displaced aggression can also be measured using this
task, but the target is an innocent person rather than the provoca-
teur (e.g., Bushman & Baumeister, 1998; Twenge & Campbell,
2003). The most common self-report measure of aggression is the
Aggression Questionnaire (AQ; Buss & Perry, 1992), which meas-
ures both physical aggression (e.g., “If somebody hits me, I hit
back”) and verbal aggression (e.g., “I can’t help getting into argu-
ments when people disagree with me”). We coded the form of
aggression (i.e., indirect, direct, displaced, physical, verbal, trol-
ling, bullying). For bullying, we coded whether it occurred offline
or online (i.e., cyberbullying).
We also coded the function of aggression (i.e., reactive, proac-

tive). One common self-report measure of reactive, and proactive
aggression is the Reactive-Proactive Aggression Questionnaire
(Raine et al., 2006), which contains 11 items to measure reactive
aggression (e.g., “Gotten angry or mad or hit others when teased”)
and 12 items to measure proactive aggression (e.g., “Had fights
with others to show who was on top”).
Violence is often measured using convictions for violent crimes

(e.g., Fisher & Hall, 2011; Westhead & Egan, 2015). One common
self-report measure of intimate partner violence is the Conflict Tac-
tics Scale (Straus, 1979; e.g., “Kicked, bit, or punched partner”).

Provocation

We coded whether participants were provoked or not. If participants
were provoked, we coded whether the provocation was
status-related (e.g., respect, prominence, or reputation), or affiliation-
related (e.g., belonging, liking, or closeness). A common status-related
provocation consists of giving participants bogus negative feedback
about an essay they wrote (e.g., Bushman & Baumeister, 1998); or
bogus negative results on an intelligence test (e.g., Smalley & Stake,
1996). A common affiliation-related provocation consists of telling
participants that nobody has chosen them as a partner for an upcoming
group task (e.g., Twenge & Campbell, 2003), or excluding them from
a ball-toss game called cyberball (e.g., Chester & DeWall, 2016).
It is important to note that provocation is confounded with the

function of aggression. Specifically, studies that involved provoca-
tion measured reactive aggression.

Exploratory Moderator Variables

Source Characteristics

Publication Status. Publication status (i.e., peer reviewed or
not peer reviewed) was analyzed as a one test of publication bias,
although more formal procedures were also used (see Sensitivity
Analyses portion of Results section).

Year. We coded the year the data were collected. If the
authors did not report the year the data were collected, we coded
the year the article was published.

Participant Characteristics

Gender. If studies reported separate results for males and
females, we coded them separately. If studies did not report sepa-
rate results for males and females, we contacted authors and asked
for the separate results.

Age. We coded the mean age of participants. Some studies
that used students as participants did not report their mean age.
Thus, the mean age of the available student samples in this meta-
analytic review was computed (M = 20.69) and imputed for miss-
ing values.

Student Sample. We coded whether participants were stu-
dents from a university or college or not.

Country. We coded whether participants were from an indi-
vidualistic or collectivistic country.

Design Characteristics

Type of Design. We coded the type of design researchers
used (i.e., experimental, cross-sectional, or longitudinal). Wave 1
of a longitudinal study was coded as a cross-sectional study. If a
longitudinal study included more than two waves of data, waves
after wave 1 were averaged.

Type of Setting. For experiments, we coded whether it was
conducted online, in a laboratory setting, or in a naturalistic field
setting.

Type of Report. We coded whether aggression was assessed
via self-report, other-report (e.g., parent, teacher, peer, or signifi-
cant other), or behavioral observation.

Intercoder Reliability

All studies were coded by two independent raters (the lead
author and trained undergraduate research assistants), called “dou-
ble coding” (Cooper, 2016). To assess intercoder reliability, the
intraclass coefficient was used for continuous characteristics and
the kappa coefficient was used for categorical characteristics
(Orwin & Vevea, 2009). The median reliability coefficient was
.69. The median percent agreement on the coded study characteris-
tics was 81%. Any coding disagreements were resolved via discus-
sion between the two authors.

Analysis Strategy

In this meta-analytic review, both narcissism and aggression
were conceptualized as continuous variables. Thus, the correlation
coefficient was used as the effect-size index. Because the distribu-
tion of the correlation coefficient is not normally distributed unless
the population correlation coefficient equals zero, Fisher’s z-

482 KJÆRVIK AND BUSHMAN

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

transformation was applied to each correlation coefficient before
pooling them. Each z-transformed value was weighted by the
inverse of its variance (i.e., N-3). Thus, larger studies received
more weight when effect-size estimates were combined.
Data were analyzed using Comprehensive Meta-Analysis

(CMA) software (Version 3; Borenstein et al., 2011) and R (R
Core Team, 2020). Random-effects models were used, which
assume that effect sizes differ from population means by both sub-
ject-level sampling error and study-level variability (Borenstein et
al., 2011). Although random-effects models are more conservative
than fixed-effects models, they require fewer statistical assump-
tions and allow for generalizations to a broader set of studies than
only the ones included in the meta-analytic review (Borenstein et
al., 2011).
Conventional meta-analytic methods rely on the assumption

that effect size point estimates are independent (Hedges et al.,
2010). In this meta-analytic review, the independence assumption
was often violated because many studies provided more than one
correlation, resulting in clusters of correlations from these respec-
tive studies. For example, we examined the relation between three
factors of narcissism (i.e., entitlement, grandiose narcissism, vul-
nerable narcissism) and aggression. Some studies measured more
than one factor of narcissism using the same sample of partici-
pants, which violates the independence assumption. In addition.
some studies measured both aggression and violence, both func-
tions of aggression (i.e., reactive and proactive aggression), or
multiple forms of aggression (e.g., physical and verbal aggression)
using the same sample of participants. Other examples include
experiments used within-subjects designs, longitudinal studies that
reported both cross-sectional and longitudinal correlations, and
studies that used both self-report and other-report measures of
aggression. There were 1,048 correlations from the 437 studies.
Thus, we calculated robust variance estimates to adjust standard
errors of dependent effect size point estimates and confidence
intervals (CIs; Hedges et al., 2010; Tipton 2013, 2015) with the
metafor v. 2.4-0 (Viechtbauer, 2010) and robumeta v. 2.0 (Fisher
et al., 2017) packages in R (R Core Team, 2020).
For distributions with fewer than 10 effect sizes, we only pro-

vide descriptive statistics. Most meta-analysts urge caution when
interpreting results from distributions with fewer than 10 effect
sizes (e.g., Kepes et al., 2012; Sterne et al., 2011).
We also conducted comprehensive battery of sensitivity analy-

ses to determine whether the obtained results were robust to publi-
cation bias and outliers. The sensitivity analyses are summarized
after the main results.

Results

Study Characteristics

This meta-analytic review included 437 independent studies
involving 123,043 participants. Of these 437 studies, 360 were
published in peer-reviewed journals and 77 were unpublished. The
studies were conducted between 1985 and 2020. The average age
of participants ranged from 8.3 to 61.7 years. Of the 437 studies,
183 used student samples from colleges or universities, and 254
used nonstudent samples. About 88% of studies were conducted in
individualistic countries.

Overall Results

The average correlation for the 437 studies was .25 with a 95%
CI ranging from .24 to .27, which excludes the value zero. The
25th, 50th, and 75th percentiles were .14, .24, and .33, respec-
tively. As can be seen in the violin plot (see Figure 2), the distribu-
tion is symmetric (i.e., the median and mean are almost identical,
the median cuts the box in half, the whiskers are about the same
length, and the mild outliers are symmetrically distributed in the
two tails of the distribution). The correlations varied widely, rang-
ing from �.39 to .76, suggesting that the study results were hetero-
geneous. Nearly all correlations (94%) were positive.

Although random effects models were used for theoretical rea-
sons, there were also good empirical reasons for using them. The
observed effects were heterogeneous, Q(436) = 3484.31, p ,
.0001. The I2 statistic indicated that 87% of the heterogeneity in
effects was due to between-study differences. According to the
Cochrane Collaboration (Higgins & Green, 2011), an I2 greater
than 75% suggests considerable heterogeneity. The coded charac-
teristics for the individual studies are included in Table SM2 in
online supplemental materials.

The Link Between Narcissism and Aggression and
Violence

The correlation between narcissism and violence, defined as
aggression intended to cause extreme physical harm such as injury
or death (Bushman & Anderson, 2001).was .23 [.18 to .27], k =
65. The correlation between narcissism and less extreme acts of
aggression was .26 [.24, .27], k = 376. The two correlations did
not significantly differ in size, Q(1)= 1.46, p = .23.

Normal Versus Pathological Narcissism

The average correlation between normal narcissism and aggres-
sion was .25 [.23, .27], k = 386. The average correlation between
pathological narcissism and aggression was .29 [.20, .37] for the
21 studies that used NPD and was .28 [.22, .33] for the 30 studies
that used the PNI. The correlations based on the NPD and PNI
were combined (r = .28, [.23, .33], k = 51) because they did not
differ, Q(1)= 0.053, p = . 82. Likewise, normal and pathological
narcissism (i.e., NPD studies plus PNI studies) did not differ in
their relation to aggression, Q(1)= 1.63, p = .20.

Narcissism Dimensions

We examined the relation between three dimensions of narcis-
sism and aggression. The correlation between entitlement, which
is the core of narcissism, and aggression was .26 [.23, .29], k =
100. The correlation between grandiose narcissism and aggression
was .24 [.22, .26], k = 358. The correlation between vulnerable
narcissism and aggression was .26 [.22, .30], k = 89. The magni-
tude of the correlations did not significantly differ, Q(2) = 1.60, p
= .45.

Forms of Aggression

Narcissism was related to indirect aggression (r = .24, [.18,
.29], k = 59), direct aggression (r = .22, [.20, .25], k = 205), and

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 483

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

displaced aggression (r = .18, [.05, .31], k = 12), and the correla-
tions did not significantly differ in size, Q(2) = 2.52, p = .28.
Narcissism was related to both physical aggression (r = .24,

[.21, .26], k = 175) and verbal aggression (r = .21, [.18, .25], k =
69), and the correlations did not significantly differ in size, Q(1) =
1.03, p = .31.
Narcissism was more strongly related to offline bullying (r =

.29, [.21, .37], k = 21) than to online bullying (i.e., cyberbullying;
r = .20, [.17, .24], k = 18), Q(1) = 4.37, p = .04. However, both
CIs excluded zero. Thus, there was a significant positive relation
between narcissism and bullying both offline and online.
Only nine studies measured the link between narcissism and

trolling. Descriptively, these nine studies found a significant posi-
tive correlation between narcissism and trolling, r = .19 [.14, .24].

Functions of Aggression

The average correlation between narcissism and proactive
aggression (r = .34, [.30, .38], k = 59) was stronger than the aver-
age correlation between narcissism and reactive aggression (r =
.26, [.24, .29]), k = 179, Q(1) = 10.99, p = .001. However, both
CIs excluded zero. Thus, there was a significant positive relation
between narcissism and both cold-blooded and hot-headed
aggression.

To more directly examine this moderator, we also analyzed the
32 studies that measured both reactive and proactive aggression
using the same sample of participants via the Reactive-Proactive
Aggression Questionnaire (Raine et al., 2006). This analysis found
that the relationship between narcissism and proactive aggression
(r = .30, [.26, .34]) did not differ from the relationship between

Figure 2
Violin Plot

Note. The density of the distribution at a given correlation is denoted by its width. The
mean correlation is denoted by a solid line, and the 95% confidence interval (CI) bounds are
denoted by dashed lines. The center of the violin plot contains a boxplot. Mild outliers
are denoted by solid black circles. There were no extreme outliers. See the online article for
the color version of this figure.

484 KJÆRVIK AND BUSHMAN

T
hi
s
do

cu
m
en
t
is
co
py

ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og

ic
al
A
ss
oc
ia
ti
on

or
on

e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd

ed
so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

narcissism and reactive aggression (r = .29, [.24, .34]), Q(1) =
0.13, p = .72.

Provocation

A total of 38 studies included both a provocation condition
and a no provocation condition. As expected, narcissism was
more strongly related to aggression under provocation condi-
tions (r = .29, [.23, .36]) than under no provocation conditions
(r = .12, [.05, .18]), Q(1) = 9.84, p = .002. However, both CIs
excluded the value zero. Thus, there was a significant positive
relation between narcissism and aggression even in the ab-
sence of provocation.
An additional 20 studies included only a provocation condition.

Of the 58 studies that included some form of provocation, 55
could be coded status-related (e.g., respect, prominence, or reputa-
tion) or affiliation-related (e.g., belonging, liking, or closeness).
Contrary to the SPIN model (Grapsas et al., 2020), narcissism was
more strongly related to aggression when the provocation was
affiliation-related (r = .39, [.28, .48], k = 20) than when the provo-
cation was status-related (r = .22, [.17, .27], k = 35), Q(1) = 7.56,
p = .006. However, both CIs excluded zero. Thus, there was a sig-
nificant positive relation between narcissism and aggression
regardless of whether the provocation was affiliation-related or
status-related.

Source Characteristics

Year

The year the data were collected did not significantly influ-
ence the magnitude of the relation between narcissism and
aggression, b = �.0005, [�.0033, .0022], k = 437. Likewise,
the year the data were collected from student samples did not
significantly influence the magnitude of the relation between
narcissism and aggression, b = �.0003, [�.0034, .0028], k =
183. Thus, the association between narcissism and aggression
was stable over the years examined (1985 to 2020) for all sam-
ples and for student samples.

Publication Status

The average correlation did not differ for published (r = .26,
[.24, .27], k = 360) and unpublished (r = .24, [.20, .27], k = 77)
studies, Q(1) = 0.95, p = .33. Thus, publication bias did not
contaminate the meta-analytic results using this measure (for
other measures see Sensitivity Analyses portion of Results
section).

Participants Characteristics

Gender

The relation between narcissism and aggression did not signifi-
cantly differ for males (r = .24, [.20, .27], k = 108) and females
(r = .21, [.16, .26], k = 51), Q(1) = 0.58, p = .45.

Age

Participant age did not significantly influence the magnitude of
the relationship between narcissism and aggression, b �.0004
[.0010, �.0024], k = 427. Thus, the link between narcissism and
aggression was stable across the ages studied (mean ages ranged
from 8.3 to 61.7 years). The remaining 10 studies did not report
mean age of participants.

University Students

The link between narcissism and aggression was slightly but
significantly larger for nonstudent samples (r = .27, [.25, .29], k =
254) than for student samples (r = .23, [.21, .25], k = 183), Q(1) =
5.15, p = .02. However, both CIs excluded zero. Thus, there was a
significant positive relation between narcissism and aggression for
both students and nonstudents.

Country

The average correlation between narcissism and aggression did
not significantly differ for participants from individualistic coun-
tries (r = .25, [.23, .27], k = 384) and those from collectivistic
countries (r = .26, [.21, .31], k = 50), Q(1) = 0.14, p = .71.

Design Characteristics

Type of Design

Narcissism was positively related to aggression in cross-sec-
tional studies (r = .26, [.24, .28], k = 358), experimental studies
(r = .22, [.18, .26], k = 76), and longitudinal studies (r = .19, [.11,
.27], k = 13). The between-groups test was significant, Q(2) =
7.21, p = .03. However, all three CIs excluded zero. Thus, there
was a significant positive relation between narcissism and aggres-
sion regardless of the type of design researchers used.

Type of Setting

For experimental studies, we coded whether the setting was
online, in a laboratory, or in the field. However, there were no field
experiments, and only seven online experiments, which had an av-
erage correlation of .29 [.23, .34]. The 69 lab experiments had an
average correlation of .20 [.16, .24].

Type of Report

We coded whether aggression was assessed via self-report, other-
report (e.g., parent, teacher, peer, or significant other), or behavioral ob-
servation. The correlation was largest for other-report measures (r = .30,
[.22, .35], k = 48), followed by self-report measures (r = .26, [.24, .27], k
= 312), followed by behavioral measures (r = .21, [.17, .25], k = 93).
The between-groups test was significant, Q(2) = 6.28, p = .04. However,
all three CIs excluded the value zero. Thus, there was a significant posi-
tive relation between narcissism and aggression regardless of the type of
report used to measure aggression.

Sensitivity Analyses

We conducted sensitivity analyses to assess whether publication
bias, outliers, or both affected our results. Both publication bias
and outliers can adversely affect meta-analytic results and associ-
ated conclusions (Kepes et al., 2013). In fact, publication bias has

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 485

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

been referred to as the potentially greatest threat to the validity of
meta-analytic results (Rothstein et al., 2005). We performed a
comprehensive battery of sensitivity analyses to determine the
degree to which our obtained results were affected by publication
bias, outliers, or both (e.g., Kepes et al., 2017).
All sensitivity analyses were conducted using Comprehensive

Meta-Analysis (Borenstein et al., 2011) and R syntax (Kepes &
McDaniel, 2015; Viechtbauer, 2017). First, we calculated the na-
ïve observed meta-analysis mean estimate (�roÞ for each distribu-
tion as well as the associated statistics (e.g., 95% CI, 90%
prediction interval [PI], weighted sum of squared deviations from
the mean, ratio of true heterogeneity to total variation). By “na-
ïve,” we mean the meta-analytic results without any “adjustment”
for potential biases (Copas & Shi, 2000). Next, we conducted a
one sample removed analyses to examine the influence of each
individual effect size on the obtained meta-analytic results. Then,
we assessed the potential for publication bias in each distribution.
We used several methods, including trim-and-fill (fixed effects
and random effect), cumulative meta-analysis (Kepes et al., 2017),
selection models (Vevea & Woods, 2005), PET-PEESE, and influ-
ence diagnostics to identify potential outliers. All identified
outlier(s) were excluded, and we reran all publication bias meth-
ods. Thus, when outliers were identified we performed all publica-
tion bias methods twice: (1) once on the original distribution, and
(2) once on the distribution without identified outlier(s). Then, the
results were averaged before making any conclusions. Specifi-
cally, when determining whether a particular naïve mean estimate
was robust to the presence of publication bias and/or outliers, we
calculated the average and the median estimates of all five publica-
tion bias methods with and without outliers. We used the median
as well as the average because the median minimizes the potential
undue influence of an estimate from any individual method on the
overall results and conclusions. To estimate the severity of poten-
tial publication bias, we used guidelines suggested by other schol-
ars (Kepes et al., 2012; Kepes & McDaniel, 2015). Specifically,
we calculated the difference between the naïve estimate and the
average estimates of all five publication bias methods with and
without outliers, as well as the corresponding percent change dif-
ference. Any change of more than 20% was considered “moder-
ate” and any change of more than 40% was considered “severe.”
In addition to the cumulative meta-analysis by precision (Kepes

et al., 2012), we reported the cumulative meta-analytic mean of
the most precise effect sizes. For the selection models, a prior
models with p value cut-points were used to model moderate and
severe instances of publication bias (Vevea & Woods, 2005). Pub-
lication bias assessment methods were not used on distributions
containing 10 or fewer effect sizes because all methods become
less stable with small samples (Kepes et al., 2012).
The results from the sensitivity analyses are displayed in Table

1. Columns 1–3 display the distribution analyzed, the number of
samples (k), and the number of participants (N). Columns 4–10
report the naïve meta-analytic results, including the random effects
(RE) meta-analytic mean (�roÞ, the 95% CI, the 90% PI, Cochran’s
Q statistics, I2, tau (s), and the one-sample removed statistics (me-
dian mean, minimum, or maximum). Column 11–18 display the
results from the trim-and-fill analyses; for the fixed effects (FE) as
well as the RE model. For each model, the table includes the side
of the funnel plot on which the imputed samples are located
(FPS), the number of imputed samples (ik), the trim-and-fill

adjusted mean effect size (t&fFE �ro, t&fRE�roÞ, and the respective
95% CI. Column 19 contains the cumulative mean for the five
most precise samples (pr5�roÞ. Columns 20 and 21 contain the
results from the moderate (smm�roÞ and severe (sms�roÞ selection
models. Column 22 contains the result of the PET-PEESE (pp�ro)
analysis. Figures SM1 to SM36 in online supplemental materials
contain forest plots that display the cumulative meta-analyses and
funnel plots that display the magnitude of the effect size by preci-
sion, with trim-and-fill imputations (for interpretation guidelines,
see Kepes et al., 2012). CIs and Q-tests for correlated effects are
based on robust variance estimates.

Neither publication bias nor outliers appear to influence the overall
effect size. It could be overestimated by an average of .02 (7%). As
can be seen in the violin plot (see Figure 2), there were no extreme
outliers in the data set. The mild outliers were evenly distributed in the
positive and negative distribution tails. Removing these 12 mild out-
liers did not change the naïve mean correlation or the CI limits.

Many other distributions were not affected by publication bias or
outliers. Most of the aggression distributions were unaffected by publi-
cation bias or outliers, including the distributions for aggression, nearly
all forms of aggression (i.e., indirect aggression, displaced aggression,
physical aggression, verbal aggression, bullying, cyberbullying), both
functions of aggression (i.e., reactive aggression, proactive aggression),
and provoked aggression regardless of whether the provocation was
affiliation-related or status-related. None of the narcissism distributions
were affected by publication bias or outliers (i.e., normal narcissism,
pathological narcissism, entitlement, grandiose narcissism, vulnerable
narcissism). The distributions for both published and unpublished stud-
ies were unaffected by publication bias or outliers. Many distributions
for participant characteristics were unaffected by publication bias,
including the distributions for males, students, nonstudents, and partici-
pants from both individualistic and collectivistic countries. The distri-
butions for all three types of research designs (i.e., experimental, cross-
sectional, longitudinal) were unaffected by publication bias or outliers.
Finally, the distribution for self-report measures of aggression was
unaffected by publication bias or outliers.

However, some distributions were affected by publication bias
and/or outliers. After deletion of identified outliers and removal of
publication bias, the naïve mean correlation for violence (r = .23)
might be overestimated by .04 to .16 (20–70%)

1

and the naïve
mean correlation for unprovoked aggression (r = .12) might be
overestimated by .05 to .07 (50–62%). Other distributions were
affected mainly by publication bias. After removal of publication
bias, the naïve mean correlation for direct aggression (r = .22)
might have been underestimated by .02 (8%) or overestimated by
.08 (38%), the naïve mean correlation for females (r = .21) might

1
For example, the mean naïve correlation for violence was r = .23 with

outliers and r = .21 without outliers (see Table 1). The text states these
estimates “might be overestimated by .04 to .16 (20– 70%).” We calculated
these values as follows. For these calculations we use three digits to avoid
rounding error, although we only report 2 digits in the text. First, we
calculated the average estimates of all five publication bias methods with
outliers [(.16 þ .23 þ .19 -.42 þ .18)/5 = .068] and without outliers [(.15 þ
.21 þ .18 þ .12 þ .18)/5 = .168]. Next, we subtracted these average
estimates from the naïve mean with outliers (.23 – .068 = .162) and without
outliers (.21 – .168 = .042) to obtain the average differences. To obtain the
percentages, we divided the average differences by the mean naïve
correlation with outliers (.162/.23 = .70 = 70%) and without outliers (.042/
.21 = .20 = 20%).

486 KJÆRVIK AND BUSHMAN

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

T
ab

le
1

S
en
si
ti
vi
ty

A
n
a
ly
si
s
R
es
u
lt
s
fo
r
A
ll
O
u
tc
o
m
es

a
n
d
M
o
d
er
a
to
rs

M
et
a-
an
al
ys
is

P
ub
li
ca
ti
on

bi
as

an
al
ys
es

F
E
tr
im

an
d
fi
ll

R
E
tr
im

an
d
fi
ll

C
M
A

S
el
ec
ti
on

m
od
el
s

P
E
T
–

P
E
E
S
E

D
is
tr
ib
ut
io
n

k
N

� r
o

95
%

C
I

90
%

P
I

Q
I2

s
os
r
� r
o

F
P
S

ik
t&

f F
E

� r
o

t&
f F

E
95
%

C
I

F
R
S

ik
t&

f R
E

� r
o

t&
f R

E
95
%

C
I

p
r 5 � r
o

sm
m

� r
o

sm
s

� r
o

pp
� r
o

A
ll
ou
tc
om

es
43
7

12
3,
04
3

.2
5

[.
24
,.
27
]

.0
0,
.4
8

3,
48
3.
31

87
.4
8

.1
6

.2
5

.2
5,
.2
5

.2
5

[.
24
,.
27
]

R
88

.3
0

[.
29
,.
32
]

.2
2

.2
3

.1
9

.3
1

A
ll
ou
tc
om

es
(w

/o
)

42
5

11
8,
87
1

.2
4

[.
23
,.
26
]

.0
3,
.4
4

2,
50
1.
08

83
.0
5

.1
3

.2
4

.2
4,
.2
4

.2
4

[.
23
,.
26
]

R
66

.2
8

[.
26
,.
29
]

.2
2

.2
2

.2
0

.2
7

A
gg
re
ss
io
n
vs
.

vi
ol
en
ce

A
gg
re
ss
io
n

37
6

10
7,
80
0

.2
6

[.
24
,.
27
]

.0
1,
.4
8

3,
58
5.
35

87
.8
6

.1
4

.2
6

.2
6,
.2
6

R
39

.2
8

[.
27
,.
30
]

R
74

.3
0

[.
29
,.
32
]

.2
2

.2
4

.2
0

.3
1

A
gg
re
ss
io
n
(w

/o
)

36
9

10
6,
14
7

.2
5

[.
24
,.
27
]

.0
2,
.4
5

2,
42
5.
87

84
.8
3

.1
4

.2
5

.2
5,
.2
5

R
36

.2
7

[.
26
,.
29
]

R
64

.2
9

[.
27
,.
31
]

.2
2

.2
3

.2
1

.3
0

V
io
le
nc
e

65
16
,1
99

.2
3

[.
18
,.
27
]

�.
06
,.
48

72
7.
80

89
.4
2

.1
7

.2
3

.2
2,
.2
3

L
12

.1
6

[.
11
,.
21
]

.2
3

[.
18
,.
27
]

.1
8

.1
9

�.
42

.2
6

V
io
le
nc
e
(w

/o
)

63
15
,7
63

.2
1

[.
17
,.
25
]

�.
04
,.
43

39
4.
39

84
.2
8

.1
5

.2
0

.2
0,
.2
1

L
12

.1
5

[.
10
,.
20
]

.2
1

[.
17
,.
25
]

.1
8

.1
8

.1
2

.2
1

D
im

en
si
on

of
na
rc
is
si
sm

N
or
m
al

38
6

10
8,
60
7

.2
5

[.
23
,.
27
]

.0
0,
.4
7

3,
68
9.
78

87
.8
2

.1
5

.2
5

.2
5,
.2
5

.2
5

[.
24
,.
27
]

R
75

.3
0

[.
28
,.
32
]

.2
2

.2
3

.1
9

.3
0

N
or
m
al
(w

/o
)

37
2

10
3,
23
4

.2
4

[.
22
,.
25
]

.0
4,
.4
2

1,
88
2.
91

80
.3
0

.1
2

.2
4

.2
4,
.2
4

.2
4

[.
22
,.
25
]

.2
4

[.
22
,.
25
]

.2
2

.2
2

.2
0

.2
6

P
at
ho
lo
gi
ca
l

51
14
,3
32

.2
8

[.
23
,.
33
]

.0
1,
.5
2

55
0.
63

87
.9
5

.1
5

.2
8

.2
7,
.2
9

.2
8

[.
24
,.
33
]

.2
8

[.
24
,.
33
]

.3
1

.2
7

.2
5

.3
4

P
at
ho
lo
gi
ca
l
(w

/o
)

50
14
,0
71

.2
7

[.
23
,.
32
]

.0
2,
.4
9

37
6.
81

87
.0
0

.1
6

.2
7

.2
6,
.2
8

L
1

.2
6

[.
22
,.
31
]

.2
7

[.
23
,.
32
]

.3
1

.2
6

.2
4

.3
1

D
im

en
si
on

of
na
rc
is
si
sm

E
nt
it
le
m
en
t

10
0

24
,2
21

.2
6

[.
23
,.
29
]

.0
5,
.4
4

56
2.
00

82
.3
4

.1
3

.2
6

.2
5,
.2
6

.2
6

[.
23
,.
28
]

.2
6

[.
23
,.
28
]

.2
2

.2
4

.2
2

.2
8

E
nt
it
le
m
en
t
(w

/o
)

98
23
,9
43

.2
5

[.
22
,.
27
]

.0
6,
.4
1

39
1.
49

75
.2
2

.1
1

.2
5

.2
4,
.2
5

.2
5

[.
22
,.
27
]

.2
5

[.
22
,.
27
]

.2
2

.2
3

.2
2

.2
7

G
ra
nd
io
se

35
8

10
3,
65
1

.2
4

[.
22
,.
26
]

.0
0,
.4
6

3,
45
5.
26

88
.3
9

.1
5

.2
4

.2
4,
.2
5

.2
4

[.
23
,.
26
]

R
63

.2
9

[.
27
,.
30
]

.2
2

.2
2

.1
8

.2
7

G
ra
nd
io
se

(w
/o
)

34
8

10
0,
80
6

.2
4

[.
22
,.
25
]

.0
3,
.4
2

1,
94
9.
72

82
.1
8

.1
3

.2
4

.2
3,
.2
4

L
2

.2
3

[.
22
,.
24
]

.2
4

[.
22
,.
25
]

.2
2

.2
2

.2
0

.2
5

V
ul
ne
ra
bl
e

89
22
,8
00

.2
6

[.
22
,.
30
]

�.
01
,.
50

88
6.
98

87
.7
1

.1
5

.2
6

.2
6,
.2
7

R
23

.3
3

[.
29
,.
37
]

R
22

.3
3

[.
29
,.
36
]

.2
4

.2
4

.2
1

.3
7

V
ul
ne
ra
bl
e
(w

/o
)

88
22
,5
39

.2
6

[.
22
,.
29
]

.0
0,
.4
9

65
5.
28

86
.7
2

.1
6

.2
6

.2
5,
.2
6

R
15

.3
0

[.
27
,.
33
]

R
19

.3
1

[.
28
,.
35
]

.2
4

.2
4

.2
1

.3
5

F
or
m

of
ag
gr
es
si
on

D
ir
ec
t

20
5

41
,0
05

.2
2

[.
20
,.
25
]

�.
04
,.
46

1,
48
1.
31

84
.8
6

.1
5

.2
2

.2
2,
.2
3

R
53

.3
0

[.
27
,.
32
]

R
45

.2
8

[.
26
,.
31
]

.3
0

.1
9

�.
39

.3
0

D
ir
ec
t
(w

/o
)

20
1

40
,2
13

.2
2

[.
20
,.
24
]

�.
01
,.
42

94
9.
26

78
.9
3

.1
4

.2
2

.2
2,
.2
2

R
52

.2
9

[.
26
,.
31
]

R
39

.2
7

[.
25
,.
29
]

.3
0

.1
9

.1
4

.2
9

In
di
re
ct

59
14
,0
92

.2
4

[.
18
,.
29
]

�.
01
,.
45

45
4.
61

86
.9
3

.1
5

.2
4

.2
2,
.2
4

.2
4

[.
19
,.
28
]

.2
4

[.
19
,.
28
]

.2
4

.2
1

.1
7

.3
1

In
di
re
ct
(w

/o
)

58
13
,6
62

.2
2

[.
19
,.
26
]

.0
2,
.4
1

26
1.
98

78
.2
4

.1
2

.2
3

.2
2,
.2
3

L
1

.2
2

[.
18
,.
26
]

.2
2

[.
19
,.
26
]

.2
4

.2
0

.1
7

.2
5

D
is
pl
ac
ed

12
2,
32
1

.1
8

[.
05
,.
31
]

.0
3,
.3
1

25
.8
1

59
.8
9

.0
1

.1
6

.1
9,
.1
7

L
1

.1
6

[.
09
,.
23
]

L
2

.1
5

[.
08
,.
22
]

.1
8

.1
6

.1
5

.1
8

P
hy
si
ca
l

17
5

36
,8
59

.2
4

[.
21
,.
26
]

�.
03
,.
47

1,
36
2.
27

85
.6
7

.1
5

.2
4

.2
3,
.2
4

R
27

.2
8

[.
25
,.
30
]

R
32

.2
9

[.
26
,.
31
]

.2
7

.2
1

�.
40

.3
0

P
hy
si
ca
l
(w

/o
)

16
9

34
,7
83

.2
2

[.
19
,.
24
]

.0
1,
.4
0

69
0.
47

75
.6
7

.1
2

.2
2

.2
1,
.2
2

R
2

.2
2

[.
19
,.
24
]

.2
2

[.
19
,.
24
]

.1
9

.1
9

.1
5

.2
5

V
er
ba
l

69
28
,9
74

.2
1

[.
18
,.
25
]

.0
1,
.4
0

77
4.
03

86
.3
2

.1
2

.2
0

.2
2,
.2
1

L
24

.1
3

[.
09
,.
17
]

R
21

.2
7

[.
23
,.
30
]

.2
7

.2
0

.1
7

.2
3

V
er
ba
l
(w

/o
)

67
22
,1
41

.2
1

[.
18
,.
23
]

.0
6,
.3
5

23
9.
29

72
.4
2

.0
9

.2
1

.2
0,
.2
1

R
23

.2
7

[.
24
,.
29
]

R
8

.2
3

[.
20
,.
25
]

.3
2

.2
0

.1
9

.2
7

O
ff
li
ne

bu
ll
yi
ng

21
11
,8
16

.2
9

[.
21
,.
37
]

.0
4,
.5
1

31
0.
96

94
.9
2

.1
7

.3
0

.2
7,
.3
1

L
2

.2
6

[.
19
,.
33
]

.2
9

[.
23
,.
35
]

.2
9

.2
8

.2
6

.3
0

O
nl
in
e
bu
ll
yi
ng

18
13
,8
81

.2
0

[.
17
,.
24
]

.0
4,
.3
6

12
3.
39

76
.3
8

.0
7

.2
0

.2
0,
.2
1

L
9

.1
1

[.
06
,.
16
]

L
3

.1
8

[.
14
,.
23
]

.1
8

.2
0

.1
9

.0
4

O
nl
in
e
bu
lly
in
g
(w
/

o)
17

7,
15
7

.2
1

[.
18
,.
25
]

.1
2,
.3
0

35
.4
9

54
.9
2

.0
5

.2
1

.2
1,
.2
2

L
3

.2
0

[.
16
,.
23
]

.2
1

[.
18
,.
25
]

.2
2

.2
1

.2
1

.2
2

F
un
ct
io
n
of

ag
gr
es
si
on

R
ea
ct
iv
e

17
9

37
,0
18

.2
6

[.
24
,.
29
]

.0
1,
.4
8

1,
35
2.
07

83
.1
6

.1
4

.2
6

.2
6,
.2
7

R
25

.3
0

[.
27
,.
32
]

R
35

.3
1

[.
28
,.
33
]

.2
4

.2
4

.2
1

.3
4

R
ea
ct
iv
e
(w

/o
)

17
6

36
,3
71

.2
5

[.
23
,.
28
]

.0
3,
.4
5

86
5.
74

79
.7
9

.1
4

.2
5

.2
5,
.2
6

R
19

.2
8

[.
26
,.
30
]

R
28

.2
9

[.
27
,.
31
]

.2
4

.2
4

.2
1

.3
1

P
ro
ac
ti
ve

59
20
,2
51

.3
4

[.
30
,.
38
]

.1
1,
.5
4

71
2.
45

90
.7
3

.1
5

.3
4

.3
3,
.3
4

.3
4

[.
30
,.
38
]

.3
4

[.
30
,.
38
]

.3
4

.3
3

.3
2

.4
2

(t
a
b
le
co
n
ti
n
u
es
)

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 487

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

T
ab
le

1
(c
o
n
ti
n
u
ed
)

M
et
a-
an
al
ys
is

P
ub
li
ca
ti
on

bi
as

an
al
ys
es

F
E
tr
im

an
d
fi
ll

R
E
tr
im

an
d
fi
ll

C
M
A

S
el
ec
ti
on

m
od
el
s

P
E
T
–

P
E
E
S
E

D
is
tr
ib
ut
io
n

k
N

� r
o

95
%

C
I

90
%

P
I

Q
I2

s
os
r
� r
o

F
P
S

ik
t&

f F
E

� r
o

t&
f F

E
95
%

C
I

F
R
S

ik
t&

f R
E

� r
o

t&
f R

E
95
%

C
I

p
r 5 � r
o

sm
m

� r
o

sm
s

� r
o

pp
� r
o

P
ro
ac
ti
ve

(w
/o
)

57
19
,6
53

.3
2

[.
29
,.
35
]

.1
3,
.4
9

34
0.
95

83
.5
8

.1
2

.3
2

.3
2,
.3
3

R
9

.3
5

[.
32
,.
38
]

.3
2

[.
29
,.
35
]

.3
4

.3
2

.3
1

.3
7

R
P
Q
R
ea
ct
iv
e

32
14
,2
80

.2
9

[.
23
,.
34
]

.0
5,
.4
9

31
6.
79

90
.2
1

.1
5

.2
9

.2
8,
.2
9

L
15

.1
8

[.
12
,.
24
]

L
7

.2
4

[.
19
,.
30
]

.2
4

.2
8

.2
8

.2
8

R
P
Q
P
ro
ac
ti
ve

32
14
,2
80

.3
0

[.
26
,.
34
]

.1
1,
.4
7

21
3.
09

85
.4
5

.1
2

.3
0

.2
9,
.3
0

R
10

.3
5

[.
31
,.
39
]

R
9

.3
4

[.
30
,.
38
]

.3
4]

.3
0

.2
9

.3
7

R
P
Q
P
ro
ac
ti
ve

(w
/

o)
30

12
,8
15

.2
8

[.
25
,.
31
]

.1
6,
.3
8

86
.4
7

66
.4
6

.0
7

.2
8

.2
7,
.2
8

R
9

.3
2

[.
29
,.
32
]

R
1

.2
8

[.
25
,.
31
]

.2
7

.2
8

.2
7

.3
1

P
ro
vo
ca
ti
on

P
ro
vo
ke
d

38
3,
93
0

.2
9

[.
23
,.
36
]

�.
04
,.
57

31
9.
75

81
.8
4

.1
8

.2
9

.2
6,
.3
0

R
7

.2
9

[.
22
,.
37
]

.2
9

[.
22
,.
36
]

.3
8

.2
6

.2
0

.5
0

P
ro
vo
ke
d
(w

/o
)

37
3,
77
0

.2
6

[.
21
,.
32
]

.0
4,
.4
6

10
1.
28

64
.4
6

.1
4

.2
6

.2
5,
.2
7

.2
6

[.
21
,.
31
]

.2
6

[.
21
,.
32
]

.3
8

.2
4

.2
1

.3
3

U
np
ro
vo
ke
d

38
3,
97
1

.1
2

[.
05
,.
18
]

�.
26
,.
47

54
3.
43

84
.1
2

.1
9

.1
0

.1
0,
.1
3

R
15

.2
3

[.
16
,.
30
]

R
14

.2
2

[.
15
,.
29
]

.1
8

.0
5

�.
45

.4
7

U
np
ro
vo
ke
d
(w

/o
)

37
3,
81
1

.1
0

[.
05
,.
15
]

�.
09
,.
29

82
.6
3

56
.4
3

.1
1

.1
0

.0
9,
.1
1

R
11

.1
7

[.
11
,.
22
]

R
10

.1
6

[.
11
,.
22
]

.1
8

.0
6

�.
32

.2
1

S
ta
tu
s

35
4,
00
1

.2
2

[.
17
,.
27
]

.0
2,
.4
1

89
.7
6

62
.1
2

.1
2

.2
2

.2
1,
.2
3

R
1

.2
3

[.
18
,.
28
]

.2
2

[.
17
,.
27
]

.3
0

.2
0

.1
7

.3
3

A
ff
il
ia
ti
on

20
2,
90
5

.3
9

[.
28
,.
48
]

.0
2,
.6
6

11
4.
04

83
.3
4

.2
3

.3
9

.3
3,
.4
0

L
4

.2
9

[.
16
,.
42
]

.3
9

[.
16
,.
42
]

.2
8

.3
6

.3
3

.5
3

A
ff
il
ia
ti
on

(w
/o
)

19
1,
93
5

.3
3

[.
26
,.
39
]

.1
5,
.4
8

36
.4
6

50
.6
4

.1
0

.3
3

.3
0,
.3
4

L
4

.2
8

[.
20
,.
36
]

L
4

.2
8

[.
20
,.
36
]

.2
8

.3
1

.3
0

.3
1

P
ub
li
ca
ti
on

st
at
us

P
ub
li
sh
ed

36
0

10
9,
54
1

.2
6

[.
24
,.
27
]

.0
0,
.4
8

3,
15
3.
88

88
.6
2

.1
6

.2
6

.2
5,
.2
6

.2
6

[.
24
,.
27
]

R
71

.3
1

[.
29
,.
32
]

.2
2

.2
3

.2
0

.3
1

P
ub
li
sh
ed

(w
/o
)

35
0

10
5,
54
8

.2
4

[.
23
,.
26
]

.0
3,
.4
4

2,
23
5.
00

84
.3
8

.1
4

.2
4

.2
4,
.2
5

.2
4

[.
23
,.
26
]

R
50

.2
8

[.
26
,.
30
]

.2
2

.2
3

.2
1

.2
7

U
np
ub
li
sh
ed

77
13
,5
02

.2
4

[.
20
,.
27
]

.0
1,
.4
4

32
7.
44

76
.7
9

.1
4

.2
4

.2
3,
.2
4

R
17

.2
9

[.
25
,.
32
]

R
17

.2
9

[.
25
,.
32
]

.2
1

.2
1

.1
7

.3
0

U
np
ub
li
sh
ed

(w
/o
)

75
13
,2
32

.2
2

[.
19
,.
26
]

.0
2,
.4
1

26
5.
62

72
.1
4

.1
2

.2
2

.2
2,
.2
3

R
18

.2
7

[.
24
,.
31
]

R
14

.2
6

[.
23
,.
30
]

.2
1

.2
1

.1
8

.2
8

G
en
de
r

M
al
e

10
8

18
,8
25

.2
4

[.
20
,.
27
]

�.
07
,.
50

74
6.
84

85
.6
7

.1
9

.2
4

.2
3,
.2
4

R
12

.2
7

[.
23
,.
30
]

R
19

.2
9

[.
25
,.
32
]

.2
5

.2
0

�.
40

.3
2

M
al
e
(w

/o
)

10
6

18
,2
12

.2
2

[.
19
,.
26
]

�.
04
,.
46

57
8.
49

81
.8
5

.1
6

.2
2

.2
2,
.2
3

R
3

.2
3

[.
20
,.
27
]

R
12

.2
6

[.
22
,.
29
]

.2
5

.2
0

.1
3

.2
8

F
em

al
e

51
10
,6
79

.2
1

[.
16
,.
26
]

�.
05
,.
45

31
1.
69

83
.9
6

.1
6

.2
1

.2
0,
.2
1

.2
1

[.
20
,.
21
]

R
10

.2
6

[.
21
,.
30
]

.1
2

.1
8

�.
39

.2
5

F
em

al
e
(w

/o
)

50
10
,4
18

.2
0

[.
16
,.
24
]

�.
01
,.
39

21
2.
23

76
.9
1

.1
3

.2
0

.1
9,
.2
0

.2
0

[.
16
,.
24
]

.2
0

[.
16
,.
24
]

.1
2

.1
7

�.
36

.2
0

U
ni
ve
rs
it
y
st
ud
en
ts

S
tu
de
nt
s

18
3

42
,1
70

.2
3

[.
21
,.
25
]

.0
4,
.4
0

73
1.
81

75
.1
3

.1
2

.2
3

.2
3,
.2
3

R
13

.2
5

[.
23
,.
27
]

R
26

.2
6

[.
24
,.
27
]

.1
6

.2
2

.1
9

.2
7

S
tu
de
nt
s
(w

/o
)

17
9

40
,6
19

.2
3

[.
21
,.
24
]

.0
8,
.3
6

50
1.
80

64
.5
3

.0
9

.2
3

.2
2,
.2
3

R
7

.2
3

[.
22
,.
25
]

R
11

.2
4

[.
22
,.
26
]

.1
6

.2
1

.2
0

.2
5

N
on
st
ud
en
ts

25
4

80
,8
73

.2
7

[.
25
,.
29
]

�.
02
,.
51

2,
73
0.
36

90
.7
3

.1
8

.2
7

.2
6,
.2
7

.2
7

[.
25
,.
29
]

R
50

.3
2

[.
30
,.
34
]

.3
2

.2
4

.2
0

.3
2

N
on
st
ud
en
ts
(w

/o
)

25
0

80
,0
30

.2
6

[.
24
,.
28
]

.0
0,
.5
0

2,
40
4.
86

89
.6
5

.1
7

.2
6

.2
6,
.2
6

.2
6

[.
24
,.
28
]

R
46

.3
1

[.
29
,.
33
]

.3
2

.2
4

.2
1

.3
1

C
ou
nt
ry

In
di
vi
du
al
is
ti
c

38
4

95
,2
71

.2
5

[.
23
,.
27
]

.0
0,
.4
7

2,
64
3.
17

85
.5
1

.1
6

.2
5

.2
5,
.2
5

R
81

.3
0

[.
29
,.
32
]

R
80

.3
0

[.
29
,.
32
]

.3
5

.2
3

.1
9

.3
2

In
di
vi
du
al
is
ti
c
(w

/o
)

37
1

90
,5
11

.2
4

[.
22
,.
25
]

.0
4,
.4
2

1,
68
4.
13

78
.0
3

.1
2

.2
4

.2
4,
.2
4

R
44

.2
7

[.
25
,.
28
]

R
50

.2
7

[.
26
,.
28
]

.3
0

.2
2

.2
0

.2
8

C
ol
le
ct
iv
is
ti
c

50
27
,4
66

.2
6

[.
21
,.
31
]

�.
02
,.
50

79
6.
10

93
.8
4

.1
7

.2
6

.2
5,
.2
7

L
21

.1
2

[.
06
,.
18
]

.2
6

[.
21
,.
31
]

.2
9

.2
4

.2
1

.2
5

C
ol
le
ct
iv
is
ti
c
(w

/o
)

48
26
,8
53

.2
4

[.
20
,.
28
]

�.
01
,.
46

61
5.
19

92
.3
6

.1
5

.2
4

.2
3,
.2
5

L
20

.1
2

[.
07
,.
18
]

.2
4

[.
20
,.
28
]

.2
9

.2
3

.2
1

.2
3

T
yp
e
of

de
si
gn

C
ro
ss
-s
ec
ti
on
al

35
8

11
4,
65
6

.2
6

[.
24
,.
28
]

.0
0,
.4
9

4,
17
7.
49

90
.1
2

.1
5

.2
6

.2
6,
.2
6

.2
6

[.
24
,.
28
]

.2
6

[.
24
,.
28
]

.2
2

.2
4

.2
0

.3
1

C
ro
ss
-s
ec
tio
na
l
(w
/

o)
34
7

11
1,
83
7

.2
5

[.
23
,.
26
]

.0
3,
.4
5

2,
39
9.
93

85
.5
8

.1
4

.2
5

.2
5,
.2
5

.2
5

[.
23
,.
26
]

.2
5

[.
23
,.
26
]

.2
2

.2
3

.2
1

.2
7

E
xp
er
im

en
ta
l

76
7,
15
4

.2
2

[.
18
,.
26
]

.0
2,
.3
9

18
0.
82

58
.3
0

.1
1

.2
2

.2
2,
.2
2

R
18

.2
8

[.
24
,.
31
]

R
13

.2
6

[.
22
,.
39
]

.2
6

.1
9

�.
37

.3
2

E
xp
er
im

en
ta
l
(w

/o
)

75
7,
05
4

.2
2

[.
19
,.
26
]

.4
2,
.3
9

15
3.
79

51
.8
8

.1
1

.2
2

.2
2,
.2
3

R
18

.2
8

[.
24
,.
31
]

R
12

.2
6

[.
22
,.
29
]

.2
6

.2
0

.1
5

.3
2

L
on
gi
tu
di
na
l

13
5,
87
0

.1
9

[.
11
,.
27
]

�.
05
,.
40

11
6.
45

88
.9
4

.1
2

.1
9

.1
6,
.2
0

R
6

.2
8

[.
20
,.
35
]

R
3

.2
1

[.
14
,.
28
]

.2
7

.1
7

.1
4

.3
2

D
es
ig
n
ty
pe

of
se
tt
in
g

L
ab

69
5,
56
3

.2
0

[.
16
,.
24
]

�.
01
,.
40

15
4.
87

56
.0
9

.1
3

.2
0

.2
0,
.2
1

R
13

.2
5

[.
21
,.
29
]

R
10

.2
4

[.
20
,.
28
]

.3
3

.1
7

�.
38

.3
2

(t
a
b
le
co
n
ti
n
u
es
)

488 KJÆRVIK AND BUSHMAN

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.

have been overestimated by .13 to .14 (64–68%), the naïve
mean correlation for laboratory studies of aggression (r =
.20) might have been overestimated by .08 (39–40%), the na-
ïve mean correlation for other reports of aggression (r = .30)
might have been overestimated by .08 to .15 (38–52%) and
the naïve mean correlation for behavioral measures of aggres-
sion (r = .21) might have been overestimated by .08 to .11
(37–57%).

Discussion

This meta-analytic review found a significant positive rela-
tionship between narcissism and aggression. The average cor-
relation was, r = .25, which is close to Cohen’s (1988)
conventional value for a “medium” effect (r = .30) and to
Lipsey and Wilson’s (1993) empirical value for the 75th per-
centile in the social sciences (r = .30). It exceeds the r = .12
benchmark value, which is reserved for effects labeled signif-
icant, important, notable, or consequential (Promising Prac-
tices Network, 2014). Thus, the link between narcissism and
aggression is not trivial in size.
Both normal and pathological narcissism were related to

aggression, and the correlations did not significantly differ in
size. These findings suggest that narcissism levels need not
be pathological to be a risk factor for aggression.
Narcissism is not a unitary construct, and scholars have

identified a core component of narcissism (i.e., entitlement),
as well as the two peripheral components (i.e., grandiose nar-
cissism, vulnerable narcissism). In this meta-analytic review,
all three components of narcissism were related to aggres-
sion, and the correlations were of similar size. These findings
suggest that entitlement, grandiose narcissism, and vulnerable
narcissism are all risk factors for aggression.
Narcissism was related to all forms of aggression measured

in the studies included in this meta-analytic review (i.e.,
direct, indirect, displaced, physical, verbal, bullying offline
and online, trolling). Thus, individuals high in narcissism are
not particularly picky when it comes to how they aggress
against others. However, more research is needed to ascertain
the relation between narcissism and trolling, which was
measured in only nine studies.
Narcissism was also related to both reactive and proactive

aggression. Thus, individuals with high levels of narcissism
are inclined to engage in both annoyance-based and incen-
tive-based aggression. As noted previously, the two types of
aggression are highly correlated, and aggression often has
mixed motives (see Bushman & Anderson, 2001).
The correlation between narcissism and violence was only

slightly smaller (r = .23) than the correlation between narcis-
sism and less serious forms of aggression (r = .26). Thus, nar-
cissism is also a risk factor for violent behavior. This finding
is consistent with other research suggesting that narcissism
might be a risk factor for extremely violent acts such as mass
shootings (Bushman, 2018).
As expected, provocation was a key moderator of the link

between narcissism and aggression. Individuals high in nar-
cissism are especially likely to aggress against others when
they are provoked, insulted, humiliated, criticized, orT

ab
le

1
(c
o
n
ti
n
u
ed
)

M
et
a-
an
al
ys
is

P
ub
li
ca
ti
on

bi
as

an
al
ys
es

F
E
tr
im

an
d
fi
ll

R
E
tr
im

an
d
fi
ll

C
M
A

S
el
ec
ti
on

m
od
el
s

P
E
T
–

P
E
E
S
E

D
is
tr
ib
ut
io
n

k
N

� r
o

95
%

C
I

90
%

P
I

Q
I2

s
os
r
� r
o

F
P
S

ik
t&

f F
E

� r
o

t&
f F

E
95
%

C
I

F
R
S

ik
t&

f R
E

� r
o

t&
f R

E
95
%

C
I

p
r 5 � r
o

sm
m

� r
o

sm
s

� r
o

pp
� r
o

L
ab

(w
/o
)

68
5,
46
3

.2
1

[.
17
,.
25
]

.0
2,
.3
9

13
9.
48

51
.9
6

.1
2

.2
1

.2
1,
.2
2

R
13

.2
6

[.
21
,.
30
]

R
9

.2
4

[.
20
,.
28
]

.3
3

.1
8

�.
38

.3
2

T
yp
e
of

re
po
rt

S
el
f-
re
po
rt

31
2

10
5,
72
5

.2
6

[.
24
,.
27
]

.0
2,
.4
6

2,
93
0.
01

87
.3
7

.1
3

.2
6

.2
6,
.2
6

.2
6

[.
24
,.
27
]

.2
6

[.
24
,.
27
]

.2
2

.2
4

.2
2

.2
9

S
el
f-
re
po
rt
(w

/o
)

30
6

10
2,
75
3

.2
5

[.
24
,.
27
]

.0
5,
.4
3

1,
86
3.
32

83
.6
3

.1
2

.2
5

.2
5,
.2
5

L
4

.2
5

[.
23
,.
26
]

.2
5

[.
24
,.
27
]

.2
2

.2
4

.2
3

.2
7

O
th
er
-r
ep
or
t

48
8,
88
4

.3
0

[.
22
,.
35
]

�.
10
,.
61

70
1.
99

92
.1
7

.2
1

.3
0

.2
8,
.3
0

R
4

.3
3

[.
26
,.
39
]

.3
0

[.
23
,.
36
]

.3
2

.2
6

�.
51

.4
8

O
th
er
-r
ep
or
t
(w

/o
)

47
8,
72
4

.2
8

[.
22
,.
34
]

�.
08
,.
58

45
9.
28

89
.9
8

.2
2

.2
8

.2
7,
.2
9

R
4

.3
1

[.
25
,.
37
]

.2
8

[.
22
,.
34
]

.3
2

.2
4

�.
48

.4
4

B
eh
av
io
ra
l

93
10
,7
62

.2
1

[.
17
,.
25
]

�.
12
,.
49

60
5.
73

78
.7
9

.1
6

.2
0

.1
9,
.2
0

R
13

.2
4

[.
20
,.
29
]

R
15

.2
5

[.
20
,.
29
]

.4
2

.1
6

�.
41

.3
3

B
eh
av
io
ra
l
(w

/o
)

91
10
,1
13

.1
9

[.
15
,.
23
]

�.
06
,.
42

30
7.
94

70
.7
7

.1
5

.1
9

.1
8,
.1
9

L
1

.1
9

[.
15
,.
22
]

R
10

.2
2

[.
18
,.
26
]

.2
3

.1
6

�.
39

.2
2

N
o
te
.

w
/o

=
w
it
ho
ut

ou
tl
ie
rs
;
k
=
nu
m
be
r
of

co
rr
el
at
io
n
co
ef
fi
ci
en
ts
in

th
e
an
al
yz
ed

di
st
ri
bu
ti
on
.
W
he
n
k
,

10
,
th
e
di
st
ri
bu
ti
on

w
as

no
t
an
al
yz
ed

be
ca
us
e
th
e
nu
m
be
r
of

st
ud
ie
s
w
as

to
o
sm

al
l.
N
=

m
et
a-
an
al
yt
ic
sa
m
pl
e
si
ze
;
r� o

=
ra
nd
om

-e
ff
ec
ts
w
ei
gh
te
d
m
ea
n
ob
se
rv
ed

co
rr
el
at
io
n;

90
%

P
I
=
90
%

pr
ed
ic
ti
on

in
te
rv
al
;
Q
=
w
ei
gh
te
d
su
m

of
sq
ua
re
d
de
vi
at
io
ns

fr
om

th
e
m
ea
n;

I2
=
ra
ti
o
of

tr
ue

he
te
r-

og
en
ei
ty

to
to
ta
l
va
ri
at
io
n;

be
tw
ee
n-
sa
m
pl
e
st
an
da
rd

de
vi
at
io
n;

os
r
=
on
e
sa
m
pl
e
re
m
ov
ed
,
in
cl
ud
in
g
th
e
m
in
im

um
an
d
m
ax
im

um
ef
fe
ct
si
ze

an
d
th
e
m
ed
ia
n
w
ei
gh
te
d
m
ea
n
ob
se
rv
ed

co
rr
el
at
io
n;

tr
im

an
d
fi
ll
=
tr
im

-a
nd
-f
il
l
an
al
ys
is
;
F
P
S
=
fu
nn
el
pl
ot

si
de

(i
.e
.,
si
de

of
th
e
fu
nn
el
pl
ot

in
w
hi
ch

sa
m
pl
es

w
er
e
im

pu
te
d;

L
=
le
ft
;
R
=
ri
gh
t)
;
ik
=
nu
m
be
r
of

tr
im

-a
nd
-f
il
l
sa
m
pl
es

im
pu
te
d;

t&
fF
E
r� o

=
fi
xe
d-

ef
fe
ct
s
tr
im

-a
nd
-f
il
l–
ad
ju
st
ed

ob
se
rv
ed

m
ea
n;

t&
fF
E
95
%

C
I
=
fi
xe
d-
ef
fe
ct
s
tr
im

-a
nd
-f
il
l–
ad
ju
st
ed

95
%

co
nf
id
en
ce

in
te
rv
al
;
t&

fR
E
r� o

=
ra
nd
om

-e
ff
ec
ts
tr
im

-a
nd
-f
il
l–
ad
ju
st
ed

ob
se
rv
ed

m
ea
n;

t&
fR
E

95
%

C
I
=
ra
nd
om

-e
ff
ec
ts
tr
im

-a
nd
-f
il
l–
ad
ju
st
ed

95
%

co
nf
id
en
ce

in
te
rv
al
;
C
M
A

=
cu
m
ul
at
iv
e
m
et
a-
an
al
ys
is
;
pr
5
r� o

=
m
et
a-
an
al
yt
ic

m
ea
n
es
ti
m
at
e
of

th
e
fi
ve

m
os
t
pr
ec
is
e
ef
fe
ct
s;
sm

m
r� o

=
on
e-
ta
il
ed

m
od
er
at
e
se
le
ct
io
n
m
od
el
’s

ad
ju
st
ed

ob
se
rv
ed

m
ea
n;

sm
s
r� o

=
on
e-
ta
il
ed

se
ve
re

se
le
ct
io
n
m
od
el
’s

ad
ju
st
ed

ob
se
rv
ed

m
ea
n;

P
E
T
-P
E
E
S
E
=
pr
ec
is
io
n-
ef
fe
ct

te
st
–
pr
ec
is
io
n
ef
fe
ct

es
ti
m
at
e
w
it
h
st
an
da
rd

er
ro
r;
P
E
T
-P
E
E
S
E
r� o

=
P
E
T
-P
E
E
S
E
ad
ju
st
ed

ob
se
rv
ed

m
ea
n;

n/
a
=
no
t
ap
pl
ic
ab
le

(b
ec
au
se

sm
s
r� o

pr
es
en
te
d
no
ns
en
si
ca
l
re
su
lt
s
be
ca
us
e
of

hi
gh

va
ri
an
ce

es
ti
m
at
es
).
C
Is
,
P
Is
,
an
d
Q
s
fo
r
co
rr
el
at
ed

ef
fe
ct
s
ar
e
ba
se
d
on

ro
bu
st
va
ri
an
ce

es
ti
m
at
es
.

NARCISSISM AND AGGRESSION: A META-ANALYTIC REVIEW 489

T
hi
s
do
cu
m
en
t
is
co
py
ri
gh
te
d
by

th
e
A
m
er
ic
an

P
sy
ch
ol
og
ic
al
A
ss
oc
ia
ti
on

or
on
e
of

it
s
al
li
ed

pu
bl
is
he
rs
.

T
hi
s
ar
ti
cl
e
is
in
te
nd
ed

so
le
ly

fo
r
th
e
pe
rs
on
al
us
e
of

th
e
in
di
vi
du
al
us
er

an
d
is
no
t
to

be
di
ss
em

in
at
ed

br
oa
dl
y.