Validation of the German Five-Factor Narcissism Inventory and Construction of a Brief Form Using Ant Colony Optimization

When thinking of narcissism, most people think of exaggerated feelings of self-worth alongside dominant and cold interpersonal behavior (Buss & Chiodo, 1991; Miller, Lynam, Siedor, et al., 2018), which is in line with the concept of grandiose narcissism as studied in social/personality and clinical psychology. Grandiose narcissism as a trait was devised from the diagnostic criteria for Narcissistic Personality Disorder (NPD) in the third edition of the Diagnostic and Statistical Manual of Mental Disorders (3rd ed.; DSM-III; American Psychiatric Association [APA], 1980), which Raskin and Hall (1979, 1981) translated into self-report items in the Narcissistic Personality Inventory (NPI). It encompasses statements such as “I think I am a special person” or “I am more capable than other people” (Raskin & Terry, 1988). NPI grandiose narcissism can be characterized mainly by high extraversion and low agreeableness in the FFM (e.g., Ackerman et al., 2011), paralleling the notion of highly narcissistic individuals as “disagreeable extraverts” (Paulhus, 2001, p. 228). While the NPI has long been the standard measure of trait grandiose narcissism and is still being extensively used, one of its problems is that the total score intermingles agentic and antagonistic aspects of narcissism (Back et al., 2013), which have substantially different nomological networks (see below).

Newer models conceptualize grandiose narcissism along intrapersonal and interpersonal dynamics and ground them in broad personality models: Morf and Rhodewalt (2001) describe Narcissism in terms of the dynamic processes involved in the construction and maintenance of a grandiose self—the central self-regulatory goal in narcissism that is played out in the social arena (see also Morf et al., 2011). Building on this line of thought, in the narcissistic admiration and rivalry concept (NARC), Back and colleagues (2013) describe characteristic strategies of either self-enhancement or self-protection which serve to maintain this grandiose self. Viewed from an individual differences perspective, these strategies reflect in either more agentic or more antagonistic patterns of experience and behavior (Back, 2018; see also Grapsas et al., 2020). These agentic and antagonistic aspects, respectively, can be thought of as narcissistic variants of broad FFM dimensions, which served as the theoretical rationale for the construction of the FFNI (Glover et al., 2012), as discussed in more detail shortly.

Complemental to the study of grandiose narcissism, vulnerable narcissism was studied in terms of a largely independent trait in the past decades (Fossati et al., 2009). While the concept of narcissistic vulnerability had long been part of clinical theories (Pincus & Lukowitsky, 2010), Wink (1991) noted in nonclinical factor-analytic studies of several narcissism scales that they sort into two largely orthogonal dimensions which he called “grandiosity-exhibitionism” and “vulnerability-sensitivity” (p. 590). Whereby the first of these aligns with the construct of grandiose narcissism, Wink’s “second face” of narcissism constitutes the independent yet related construct of vulnerable narcissism¹ in terms of “introversion, defensiveness, anxiety, and vulnerability to life’s traumas” (p. 590). Despite their opposing experiential and behavioral tendencies, Wink noted both share “common narcissistic characteristics of conceit, self-indulgence, and disregard for the needs of others” (p. 596), thereby already anticipating contemporary personality models (see below). Building on this distinction, Hendin and Cheek (1997) constructed the Hypersensitive Narcissism Scale (HSNS) for the assessment of vulnerable narcissism. It encompasses items such as “I can become entirely absorbed in thinking about my personal affairs, my health, my cares or my relations to others” (p. 592). Individuals scoring high on the HSNS are characterized by neuroticism, disagreeableness, and introversion in the FFM (Jauk et al., 2017; Miller et al., 2011). Although the HSNS is a widely used measure of vulnerable narcissism and is valid with respect to expert ratings (Miller et al., 2014), a shortcoming is that—as to the NPI—it intermingles discernible aspects of vulnerable narcissism which have substantially different nomological networks, in this case, antagonistic and neurotic aspects (Miller et al., 2011).

The FFNI has been found to be a reliable and valid measure of different aspects of narcissism. Its grandiose aspects correlate highly (r = .74) with other trait measures of grandiose narcissism such as the NPI (Raskin & Hall, 1979) but also with clinically oriented self-report measures of NPD (.55 ≤ r ≤ .60) such as the Minnesota Multiphasic Personality Inventory (MMPI; Morey et al., 1985) or Personality Diagnostic Questionnaire (PDQ; Bagby & Farvolden, 2004; see initial publication by Glover et al., 2012). Similarly, its vulnerable aspects correlate substantially with the HSNS (Hendin & Cheek, 1997) as a trait measure of vulnerable narcissism (r = .65) and the vulnerable subscale of the Pathological Narcissism Inventory (r = .74; PNI; Pincus et al., 2009; see Glover et al., 2012). Grandiose as well as vulnerable aspects are related to DSM-based ratings of NPD (grandiose: .45 ≤ r ≤ .64, vulnerable: .24 ≤ r ≤ .36; Miller, Few, Wilson, et al., 2013). For the FFNI-SF, similar validity coefficients were observed across several samples, for instance with the NPI (.62 ≤ r ≤ .73) and the HSNS (.59 ≤ r ≤.71; Sherman et al., 2015). The FFNI-SF factors further show correlations with the grandiose (.27 ≤ r ≤ .55) and vulnerable (.70 ≤ r ≤ .82) subscales of the PNI (Fossati et al., 2018; Sherman et al., 2015). As predicted by the Narcissism Spectrum Model (Krizan & Herlache, 2018), both grandiose and vulnerable narcissism show substantial correlations with psychological entitlement (grandiose: r = .54, vulnerable: r = .33; Hart et al., 2020). The FFNI also shows discriminant validity across studies and samples, as grandiose and vulnerable aspects are unrelated or only moderately related to measures such as the HSNS or the NPI in crossed-over analyses (grandiose: .16 ≤ r ≤ .36, vulnerable: −.07 ≤ r ≤ .12; Glover et al., 2012; Sherman et al., 2015).

The FFNI grandiose and vulnerable factors are both associated with other socially aversive personality traits, among them the “Dark Triad” traits Machiavellianism and psychopathy. Correlations with these traits are substantial in full-scale measures (grandiose: .42 ≤ r ≤ .64, vulnerable: .41 ≤ r ≤ .57; Miller, Gentile, & Campbell, 2013) and short measures (Short Dark Triad, SDT; Jones & Paulhus, 2014; .76 ≤ r ≤ .84; Wehner et al., 2021), and substantial even when variance of the other two traits is partialed out (.31 ≤ r ≤ .41; Somma et al., 2020). FFNI vulnerability displays lower correlations to SDT narcissism (.11 ≤ r ≤ .40; Wehner et al., 2021).

As outlined above, the FFNI scales can not only be described at the two-factor level of grandiose and vulnerable narcissism, but also at the three-factor level of agentic, antagonistic, and neurotic narcissism from an FFM perspective. Within this framework, the FFNI factors display the expected pattern of correlations with FFM measures (absolute rs from .51 to .61; Miller et al., 2016; Rogoza et al., 2021), and they align conceptually with the distinction between admiration and rivalry (Back, 2018). Evidence also points to the expected relations with admiration and rivalry (from the Narcissistic Admiration and Rivalry Questionnaire, NARQ; Back et al., 2013) on an empirical basis (admiration: .53 ≤ r ≤ .70, rivalry: .63 ≤ r ≤ .71; Miller et al., 2016; Rogoza et al., 2021). The FFNI further displays expected convergent validity with related FFM traits at facet level, with relations being stronger for clinical measures (selected facets .63 ≤ r ≤ .83 with the Personality Inventory for DSM-5, PID-5; Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5; APA, 2013; Krueger et al., 2012)) than for nonclinical measures (selected facets’ absolute correlations ranging between .35 ≤ r ≤ .75 with the NEO-Personality Inventory—Revised, NEO-PI-R; Costa & McCrae, 1992; see Helle & Mullins-Sweatt, 2019).

Regarding the nomological network of the FFNI in other aspects of personality, interpersonal experience and behavior, and psychological functioning, one of the most general notions is that grandiose narcissism—particularly its agentic aspects—is related to approach motivation (Spencer et al., 2018) and egotistical, but—from the perspective of the individual—largely adaptive functioning (cf. Kaufman et al., 2020), whereas vulnerable narcissism—particularly its neurotic aspects—is related to avoidance motivation (Spencer et al., 2018) and a wide array of problematic patterns of experience and behavior, psychopathological symptoms, and generally reduced personality functioning (reduced self- and interpersonal functioning; APA, 2013; for data across different inventories, see Fossati et al., 2018; Miller, Lynam, Vize, et al., 2018). This is evident, for instance, in positive associations of FFNI grandiose narcissism with self-reported mental toughness (Papageorgiou, Gianniou, et al., 2019), status pursuit (Zeigler-Hill et al., 2021), self-reported social success (Oltmanns et al., 2018), hubristic but also authentic pride (Kaufman & Jauk, 2020), negative associations with experiencing feelings of guilt (Kaufman et al., 2020), fear (Jauk & Kaufman, 2018), or depression (Papageorgiou, Denovan, & Dagnall, 2019), and positive associations with adaptive coping, self-esteem (due to its agentic aspects; Jauk & Kaufman, 2018; Miller et al., 2016), and life satisfaction (again due to agentic aspects; Kaufman et al., 2020). FFNI grandiose narcissism is unrelated to explicit negative views of the self or others (Kaufman et al., 2020), unrelated (Kaufman et al., 2020), or weakly to moderately related to attachment anxiety and avoidance (Miller, Gentile, & Campbell, 2013), but its antagonistic aspects do relate to negative views of the self and others, a cynical-distrustful social attitude, and attachment anxiety and avoidance (Kaufman et al., 2020; Szymczak et al., 2020). Consequently, the FFNI grandiose dimension is associated with a cold-dominant interpersonal style (peaking in the arrogant-calculating octant of the interpersonal circumplex; Miller, Gentile, & Campbell, 2013), which can be hypothesized to draw more upon its antagonistic aspects, and game-playing (ludus) and obsessive (mania) love styles (Miller, Gentile, & Campbell, 2013). Also, FFNI grandiose narcissism—particularly its antagonistic aspects—is related to self-reports of anger, aggression, and externalizing symptoms (Sauls & Zeigler-Hill, 2020; Vize et al., 2019, 2020), laboratory-based measures of aggression (Chester & Lasko, 2019; Hyatt et al., 2019) or measures of racial prejudice (Infante et al., 2019). This makes the FFNI—particularly FFNI antagonistic narcissism—a potentially relevant predictor in forensic psychology and criminology (cf. DeLisi, 2019), which is supported by recent evidence showing that antagonism in general (Niemeyer et al., 2022) and antagonistic narcissism specifically (Niemeyer et al., 2021) are highly relevant predictors of criminal behavior.

FFNI vulnerable narcissism, on the contrary, shows associations with experiencing less positive and more negative affect (Jauk & Kaufman, 2018), experiencing shame (Di Sarno et al., 2020), lower authentic and higher hubristic pride (Kaufman & Jauk, 2020), higher attachment anxiety and avoidance (particularly its neurotic aspects; Kaufman et al., 2020; Miller, Gentile, & Campbell, 2013), feelings of social exclusion (Mazinani et al., 2021), a cold-submissive interpersonal style (peaking in the aloof-introverted octant of the interpersonal circumplex; Miller, Gentile, & Campbell, 2013), which can be hypothesized to relate particularly to its neurotic aspects. FFNI vulnerable narcissism further relates to game-playing (ludus) as well as obsessive (mania) love styles (like the grandiose dimension; Miller, Gentile, & Campbell, 2013) and pathologically altruistic behavior (Kaufman & Jauk, 2020). It is associated with interpersonal problems and social dysfunction (Fossati et al., 2018; Oltmanns et al., 2018) and a wide array of psychopathology (particularly internalizing pathology such as anxiety or depression; Jauk & Kaufman, 2018; Kaufman et al., 2020; Miller, Gentile, & Campbell, 2013; Papageorgiou, Denovan, & Dagnall, 2019). Regarding psychological adjustment, FFNI vulnerable narcissism is associated with lower self-esteem (due to its neurotic aspects; Jauk & Kaufman, 2018; Miller et al., 2016) and lowered well-being (as evident in lower life satisfaction, feelings of purpose, or self-acceptance, for instance, which are mainly due to neurotic, but also antagonistic aspects; Kaufman et al., 2020). Like grandiose narcissism, FFNI vulnerable narcissism is related to aggression, though associations are mainly driven by its antagonistic aspects (Vize et al., 2019). Ultimately, FFNI vulnerable narcissism is related to suicidal ideation (Brioschi et al., 2020).

While FFNI grandiose narcissism is broadly associated with egotistical-adaptive functioning from the perspective of the individual, it is also associated with feelings of alienation, a weak sense of self, maladaptive defense, and dissociation (Kaufman et al., 2020; Şar & Türk-Kurtça, 2021), reports of childhood trauma (Şar & Türk-Kurtça, 2021), and substance use (Miller, Gentile, & Campbell, 2013; Welker et al., 2019). Also, when contrasting self- and informant-reports, it becomes evident that others do not see grandiose narcissism as favorably as highly grandiose individuals do themselves (despite generally high agreement on grandiosity between self and informants; Oltmanns et al., 2018).

We gathered data from eight samples of German-speaking participants, four of which completed the long, and four of which the short version of the FFNI alongside different validity measures. The rationale was to pool several samples to obtain the largest possible sample size for analyses of the factorial structure and validity of the FFNI (see Analysis Plan). Altogether, the samples comprised N = 2,921 participants (2,140 women, 778 men, 3 other) with a mean age of 25.59 years (SD = 7.67). Among those, N = 1,823 completed the long, and N = 1,098 the short FFNI. All samples were convenience samples assessed online, with the exception of sample 8, which includes young criminal offenders incarcerated in a correctional institution in the German state of Mecklenburg-Vorpommern. All data were collected in accordance with the relevant local guidelines and regulations and with the Declaration of Helsinki. There were no missing data (incomplete data sets were excluded a-priori). Beyond the scales listed in Table 1, further personality scales which are not analyzed here (due to redundancy or nonrelevance to the research questions) were administered in some samples.

Among the pooled sample, N = 1,745 (59.74%) of participants were students, 475 (16.26%) were employed, 50 (1.71%) were apprentices or secondary / high school pupils, 90 (3.08%) retired or unemployed, and the 68 (2.33%) participants of sample 8 were incarcerated (information on vocational status was unavailable for 493 participants, or 16.88% of the pooled sample). Table 1 provides detailed characteristics of the individual samples and the measures included therein. Self-reports on previously professionally diagnosed mental disorders were available in samples 6 and 7. Of the 466 participants in those samples, 363 (77.90%) reported no history of mental disorder, 55 (11.80%) reported a past disorder, and 48 (10.30%) a present disorder.

The FFNI is a multidimensional measure of narcissism grounded in the FFM. It comprises 148 items (FFNI-LF)/60 items (FFNI-SF) belonging to 15 facets which were constructed as narcissistic variants of FFM facets (for examples, see introduction; Glover et al., 2012). Within the two-factor model, these facets can be summarized to reflect factors of grandiose narcissism (indifference, exhibitionism, authoritativeness, grandiose fantasies, manipulativeness, exploitativeness, entitlement, lack of empathy, arrogance, acclaim seeking, thrill seeking) and vulnerable narcissism (reactive anger, shame, need for admiration, distrust; Glover et al., 2012). Within the three-factor model, the FFNI provides factor scores of agentic narcissism (acclaim seeking, authoritativeness, grandiose fantasies, exhibitionism), antagonistic narcissism (manipulativeness, exploitativeness, entitlement, lack of empathy, arrogance, reactive anger, distrust, thrill seeking), and neurotic narcissism (shame, indifference [reversed], need for admiration; Miller et al., 2016) (see Note 2).

The FFNI was initially translated into German independently by several members of the research groups from Bern and Münster. Next, a final consensus version was created by three psychologists all experts in narcissism and one of whom is a bi-lingual native speaker. This process involved an iterative approach in which each translated item was evaluated and in case of deviant translations the best wording was determined through discussion until consensus was reached. Following this, a back-translation was performed by a different bi-lingual native speaker, upon which any discrepancies again were resolved through discussion. Items of the German FFNI are available in Supplemental Table S1 and on the Open Science Framework (OSF): https://osf.io/ae7vr/

This section provides information on the validity measures included in this study. Table 1 shows the samples in which the measures were used, Table 4 displays internal consistencies.

Our rationale was to conduct analyses based on the largest set of available data across the samples. We pooled samples for analyses of reliability and factor structure for the different FFNI versions, leading to pooled data of N = 1,823 participants who filled in the FFNI-LF and N = 1,098 who filled in the FFNI-SF. Sample sizes for validity analyses differ depending on which inventories were included in the samples (see Tables 1 and 4). For analyses of internal consistency and factor structure of the short versions (novel brief form described in Study 2 and recently published FFNI-SSF), we used all available data of participants who filled in either the FFNI-LF or the FFNI-SF (N = 2,921).

The analysis proceeded in three steps—descriptives and internal consistencies, confirmatory examination of higher- and lower-order factor-structure, followed by convergent and discriminant validity. We first analyzed internal consistencies and intercorrelations of the German FFNI in its different versions (including the brief form presented in Study 2) at the two-factor, three-factor, and facet level. Next, we set up confirmatory factor analysis (CFA) models to analyze the factor structure at the different levels (correlated facets models, two-factor, and three-factor models). As the FFNI items display substantial cross-loadings, we complemented traditional CFA models with exploratory structural equation models (ESEMs; Asparouhov & Muthén, 2009). ESEMs combine the confirmatory with an exploratory factor analysis framework and allow for cross-loadings and rotation of solutions. We used target rotation to maximize the main loadings and minimize cross-loadings. As the ESEM technique does not directly allow the implementation of hierarchical models, we used the ESEM-within-CFA approach (Morin & Asparouhov, 2018) for models with more than two hierarchical levels (for a similar analysis strategy, see Morf et al., 2017).⁵ Here, the first-order factor loadings are first estimated in ESEM. These loadings (including cross-loadings) are then used as constraints for an estimation of the higher-order model in CFA. As not all item-level data were normally distributed, we attempted to estimate models using weighted least square mean and variance adjusted (WLSMV) estimators. However, many of the models could not be estimated using WLSMV (nonconvergence); therefore, we used maximum likelihood (ML) estimation for CFA and ESEM models instead.

Finally, we analyzed correlations with validity measures assessing narcissism and related constructs, nonclinical and clinical personality constructs, interpersonal experience and behavior, as well as psychological adjustment and maladjustment.

Table 2 shows the scale averages, internal consistencies, and intercorrelations of the FFNI (both two- and three-factor models constructed based on Glover et al., 2012 and Miller et al., 2016; see Method section) for the different versions. The scores of both models displayed high internal consistencies in the LF and SF (.95 ≤ α ≤ .89), the FFNI-SSF displayed somewhat lower but still good internal consistencies given its length (.60 ≤ α ≤ .73). At facet level, internal consistencies for the LF were between .70 ≤ α ≤ .89, those for the SF were between .69 ≤ α ≤ .90 (SSF has only one item per facet).

The intercorrelations among the scores according to the two- and three-factor models were generally as expected and similar for the different scale versions: grandiose and vulnerable narcissism displayed no or slight positive associations (for recent meta-analysis, see Jauk et al., 2021), grandiose narcissism was equally highly correlated with its constituent factors agentic narcissism and antagonistic narcissism and moderately negatively associated with neurotic narcissism. Vulnerable narcissism was unrelated or moderately related to agentic narcissism, moderately related to its constituent factor antagonistic narcissism (though lower in the SSF than the other versions), and strongly related to its constituent factor neurotic narcissism. Facet intercorrelations were, with few exceptions, highly similar for the different versions and can be seen in Table 2.

To evaluate the factor structure of the German FFNI in a confirmatory framework, we tested CFA, ESEM, as well as ESEM-within-CFA models within the two- and three-factor structure. In the following, we discuss results for the FFNI-LF, -SF, and the -SSF here; results for the novel FFNI-BF are discussed in Study 2. Table 3 displays the complete fit statistics for the different models. Within the CFA models, fit was generally better for models taking the full hierarchical structure into account (factor -> facet -> item; models 4 & 7), compared to those which solely account for either the facet structure (factor -> facet; models 2 & 5) or item uniqueness (factor -> item; models 3 & 6). Within these hierarchical models, fit indices only partially indicated acceptable fit for the FFNI-LF in the two-factor (model 4: RMSEA =.047, CFI =.676, SRMR =.108) and three-factor solution (model 7: RMSEA = .046, CFI =.690, SRMR =.097). For the FFNI-SF, we observed less then optimal yet largely acceptable fit in the two-factor (model 4: RMSEA =.056, CFI =.821, RMSEA =.110) and in the three-factor solution (model 7: RMSEA =.052, CFI =.843, RMSEA =.092). For the SSF, hierarchical models cannot be estimated because facets are represented by single indicators. CFA models with items loading on factors did not generally display acceptable fit for this scale version in the two-factor (model 3: RMSEA =.101, CFI =.675, SRMR =.088) or three-factor solution (model 6: RMSEA =.077, CFI =.816, SRMR =.069). Fit was acceptable for this scale version, however, when using ESEM models as discussed next.

We proceeded the analysis with ESEM models, which allow for cross-loadings. These models generally outperformed traditional CFA models regarding model fit. Within the ESEM models, again, those taking the full hierarchical structure into account (ESEM-within-CFA; models 11 & 14) displayed the best data fit. Here, we observed good fit for the FFNI-LF in both the two-factor (model 11: RMSEA =.028, CFI =.906, SRMR =.028) and three-factor solution (model 14: RMSEA =.028, CFI =.907, SRMR = .022). Also, the FFNI-SF displayed good fit in the two-factor (model 11: RMSEA =.037, CFI =.951, SRMR =.051) and three-factor solution (model 14: RMSEA =.033, CFI =.960, SRMR =.035). Regarding the FFNI-SSF, ESEM models with cross-loadings of items on factors for the FFNI-SSF displayed largely acceptable fit in the two-factor solution (model 10: RMSEA =.079, CFI =.830, SRMR =.050) and good fit in the three-factor solution (model 13: RMSEA = .048, CFI =.948, SRMR =.023).

Taken together, though many of the models did not show acceptable fit by conventional standards, hierarchical CFA and ESEM models did so for different versions of the FFNI. Fit of ESEM-within-CFA models was generally superior to those of traditional higher-order CFA models, indicating that considering cross-loadings at item-level improves model fit. However, this also shows that the item sets of existing versions might not be optimally suited to capture the intended structure in an unidimensional manner. While ESEM models can effectively capture variance related to different aspects at item-level, we also note that they may run at the risk of over-fitting the data (all theoretically possible cross-loadings are being freely estimated). In Study 2, we thus developed a novel short form by selecting a set of items which best represented the intended structure of the FFNI in terms of model fit and factor loadings.

Turning to factor loadings (range displayed in Table 3, coefficients available in the OSF), within the evaluated models, all theoretically assumed (i.e., original subfactor assignment) loadings were significant, with the exceptions of the ESEM-within-CFA model (model 11) for the two-factor solution in the FFNI-LF, in which the facets indifference and arrogance did not load significantly on the grandiose narcissism factor, and the ESEM-within-CFA model for the three-factor solution (Model 14) in the FFNI-LF, in which the facets arrogance, reactive anger, and distrust did not load significantly on antagonistic narcissism. This indicates that considering cross-loadings at item-level, though improving model fit, unveils problems in the factorial structure of the FFNI-LF, in the way that these facets do no longer represent the designated higher-level factors (i.e., item-level variance can now be allocated to different facets). Such problems were not evident in the corresponding models of the FFNI-SF, where all facet loadings were significant, showing that shortened versions of the FFNI are better suited to capture the intended factor structure.

To investigate the criterion validity of the German FFNI with other measures of narcissism and related constructs, general as well as clinical personality measures, interpersonal experiences and behavior, as well as measures of psychological adjustment and maladjustment, we inspected correlations with the respective scales (see Table 4).

We used the samples and measures described in Study 1 for the ACO item selection, cross-validation of the solution, and further factorial structure and criterion validity tests of the brief form. Specifically, for the item selection process (see below), we used data from those samples in which the FFNI-LF was administered (samples 1-4, N = 1,823, see Table 1). We then cross-validated the solution using data from the samples in which the FFNI-SF was administered (samples 5-8, N = 1,098, see Table 1). Finally, we pooled all samples (N = 2,921) for further tests of the factorial structure in which we also evaluated the models described in Study 1, and examined criterion validity.

We constructed a 30-item brief form of the FFNI using ACO (Olaru et al., 2019; Schroeders et al., 2016). ACO is a metaheuristic algorithm that selects and evaluates item sets based on (combinatorial) scale-level properties such as model fit or reliability. Traditional short scale development procedures generally rely on item-level information derived in the full scale (e.g., factor loadings of the 60-item FFNI-SF), even though the criteria of interest can only be derived for combinations of items (e.g., reliability). In addition, selecting items based on properties in the initial full scale relies on the assumption that these do not change as soon as items are removed (e.g., initial factor loadings corresponding to the short scale loadings). ACO selects and evaluates combinations of items and learns over the course of several iterations which items are particularly well suited to optimize the given criteria (e.g., model fit). Because it selects items based on scale-level properties of the final instead of the initial scale, it is superior to traditional item selection techniques (for more details, see Olaru et al., 2015, 2019).

Here, we used ACO to select a 30-item brief form with two items per facet. As the measurement model for which ACO selected the items, in keeping with the FFNI idea of narcissistic variants of FFM traits (Glover et al., 2012), we specified a three-factor model with agentic, antagonistic, and neurotic narcissism as factors encompassing 16, 8, and 6 items (i.e., two items per facet; see method section). To account for the facet structure of the scale (i.e., items from a common facet correlating higher with each other), we added residual correlations between item-pairs from a common facet. This is equivalent to specifying nested orthogonal facet factors in a bifactor model. We chose this approach over a higher-order model for the ACO item selection because higher-order models with only two indicators per first-order factor caused convergence problems and yielded a large number of negative residual variances in initial ACO runs. The used model is more robust and considers variance at the factor and facet level, thus capturing the intended factor- and facet structure of the FFNI.

Based on this model, we optimized (1) the model fit, more specifically RMSEA, CFI, and SRMR, (2) the reliability (factor saturation McDonald’s ω) of the three factors, and (3) their correlations with external criteria. Regarding the latter, we maximized the correlations between the FFNI factors of agentic narcissism with NPI leadership/authority as well as NARQ admiration, antagonistic narcissism with NPI entitlement/exhibitionism, NARQ rivalry, and the PES, and neurotic narcissism with the HSNS and PNI vulnerable dimension.⁶ Each criterion was logit transformed to a scale of 0 to 1 (for an illustration, see Olaru et al., 2019), and the average across the three criteria was used as the overall optimization criterion. The exact optimization function can be found in the R script in the OSF (https://osf.io/ae7vr/).

As item selection pool, we used the FFNI-SF (Sherman et al., 2015). We chose the short form as a starting point to ensure that the current scale was nested in the FFNI-SF, and also because this would give us the largest sample size to work with (see above). From the initial 60-item pool, we removed six highly content redundant items⁷ to ensure that pairs of semantically similar items were not selected to increase the internal consistency of the scale at the cost of construct coverage. We further excluded two reverse-coded items⁸ from the initial pool, as those would otherwise have been the only reverse-coded items in the pool, leading to a 52-item selection pool.

We used ACO with a comparison of 50 models per iteration (i.e., “ants”) and a stopping criterion of 30 iterations without improvement to the previously best solution. Because ACO is a probabilistic algorithm, a single run may not be sufficient to find an optimal solution. We thus ran it 10 times with different random generator seeds and used the best solution of the ten runs as final solution. To ensure that the solution was robust and generalizable across samples, we used cross-validation. Specifically, we selected the brief form solely in the combined sample in which participants filled in the long form (selection sample; N = 1,823). To evaluate the robustness of the model fit, reliability and correlations of the scale, we then estimated these values for the final brief form on the combined sample in which participants filled in the short form (validation sample; N = 1,098). Because these samples had been collected independently from each other in the context of different studies, this approach should yield a good indication of the generalizability of the selected scale properties. Finally, for consistency with the analyses presented in Study 1, we re-evaluated the factorial structure in the overall pooled sample within the models presented there.

The 30 items selected by ACO for the Brief Form can be found in Supplemental Table S1 and in the OSF (https://osf.io/ae7vr/). The BF showed largely satisfactory CFA model fit on the selection sample (RMSEA = .069; CFI = .900; SRMR = .072), as well as in the validation sample (RMSEA = .073; CFI = .904; SRMR = .075), and also in the pooled sample (see next section). Factor saturation omega was on average ω = .76 in the selection sample (agentic narcissism ω = .75; antagonistic narcissism ω = .75; neurotic narcissism ω = .79), and ω = .77 for all three factors in the validation sample. The average correlation between the three factors and the related scales used in the optimization was comparable across the selection and validation scale (selection: r = .58; validation: r = .61; for single correlations see Table 4). The three brief form scale scores correlated on average r = .92 with the scores of the original 60-item short form (agentic narcissism r = .94; antagonistic narcissism r = .92; neurotic narcissism r = .92). In summary, the 30-item brief form yielded satisfactory psychometric properties, which were robust across different samples, and was highly correlated with the 60-item version, thus maintaining construct coverage.

The 30-item FFNI-BF displayed good internal consistencies ranging from .72 ≤ α ≤ .84 at factor-level (two- and three-factor models) and from .60 ≤ α ≤ .88 at facet level. Internal consistencies were generally in between those of the FFNI-SF and FFNI-SSF, as it can be expected given the different scale lengths. The pattern of intercorrelations among the factors of the two- and three-factor models was generally highly similar to those of other scale versions (see Table 2).

To re-evaluate the factor structure of the FFNI-BF within the hierarchical CFA, ESEM, and ESEM-within-CFA framework, we tested the same models⁹ as described in Study 1 in the overall pooled sample. As Table 3 shows, similar to the long form, we observed generally better fit for models which consider the nested structure, and which consider cross-loadings of items at the facet level. The FFNI-BF showed largely acceptable and—according to most indices—superior fit than other scale versions in higher-order CFA models, in particular for the three-factor model for which the items were selected (model 4, two-factor: RMSEA = .070, CFI = .814, SRMR = .102; model 7, three-factor: RMSEA = .054, CFI = .892, SRMR = .072). For consistency reasons we also estimated the FFNI-BF in ESEM-within-CFA models (model 11, two-factor: RMSEA = .051, CFI = .955, SRMR = .037; model 14, three-factor: RMSEA = .040, CFI = .973, SRMR = .026). We note, however, that ESEM-within-CFA models, despite their good fit, also yielded parameter estimates which were out of bounds. This likely results from over-fitting a model whose loadings have already been optimized (leading to collinearity when further residual variance is accounted for by the model), in combination with the small number of indicators per facet. The use of ESEM thus seems neither appropriate nor necessary for the FFNI-BF, given the model used for the construction thereof has already been optimized. We recommend the use of more parsimonious traditional hierarchical CFA models or correlated factor model with residual correlations between items of the same facet (selection model used) in future studies. We further note that although fit was better for the optimized three-factor solution than for two-factor models, fit of the two-factor models was similar to the FFNI-SF, and superior to the FFNI-LF. Supplemental Table S1 displays loadings of the FFNI-BF (hierarchical CFA three-factor model).

The FFNI-BF displayed generally highly similar validity as did longer scale versions regarding measures of narcissism and related constructs, general and clinical personality measures, measures of interpersonal functioning, as well as adaptive and maladaptive psychological adjustment (see Table 4). We observed strong convergence between the FFNI-BF grandiose narcissism factor and other measures of grandiose narcissism (NPI, NARQ) and, to a smaller extent, the PNI grandiosity scales (as in the longer versions). Grandiose narcissism was also correlated with psychological entitlement and composite measures of narcissism (DAPP, SDT narcissism) to a similar extent as in other scale versions. FFNI-BF vulnerable narcissism was correlated strongly to other vulnerable narcissism scales (HSNS and PNI vulnerability) and was also correlated to psychological entitlement and composite measures of narcissism. Of note, the correlation between vulnerable narcissism and psychological entitlement (PES) was within the expected range in the FFNI-BF (r = .26) while being weak in the FFNI-SSF (r = .11), and similar differences were evident for the entitlement/exploitativeness factor of the NPI, the entitlement rage facet of the PNI, and more generally the disagreeableness/antagonism measures (NEO-PI-R and PID-5). This is likely a result of including the correlation with entitlement—a core characteristic of both forms of narcissism (Krizan & Herlache, 2018)—as an optimization criterion. Within the three-factor model, we also observed strong convergent validity of the FFNI-BF agentic, antagonistic, and neurotic narcissism factors with other measures of agentic, antagonistic, and neurotic narcissism. Specifically, agentic narcissism correlated most strongly with NPI leadership/authority and NARQ admiration, whereas antagonistic narcissism correlated most strongly with NPI entitlement/exploitativeness and NARQ rivalry. Both displayed substantial associations with composite measures of narcissism and psychological entitlement (PES), the latter being more pronounced for antagonistic than for agentic narcissism, as theoretically intended (Weiss et al., 2019). FFNI-BF neurotic narcissism, as in the other scale versions, displayed the highest correlations with PNI vulnerability (though also substantially related to PNI grandiosity), was unrelated to psychological entitlement (PES), and slightly negatively related to agentic measures of narcissism, which has implications for the interpretation of this factor (see general discussion).

Correlations of the FFNI-BF with other personality measures and further external validity criteria were generally highly similar to those of the other scale versions reported in Study 1, with some notable exceptions: while agentic narcissism correlates most strongly with aspects of extraversion related to social dominance (NEO-PI-R assertiveness), the longer versions also display a substantial correlations with aspects related to positive emotionality (NEO-PI-R positive emotions, r = .30), which is only weakly the case in the novel BF (r = .12) or the recently developed SSF (r = .14). Similarly, agentic narcissism is less strongly related to authentic pride in the shorter versions (r = .43 vs. r = .22 / r = .23), and less strongly related to self-esteem (r = .29 vs. r = .12 / r = .14) as well as life satisfaction (r = .28 vs. r = .08 / r = .08). Taken together, this shows that shorter scales draw less strongly upon adaptive aspects of extraversion, which potentially make them more suitable measures of agentic-extraverted narcissism, rather than extraversion per se (see general discussion).

Finally, the novel BF (and also the SSF) displayed largely the same pattern of mean differences between groups with and without mental disorders diagnoses or incarceration status (see Table 5). Diagnoses were related to higher scores on vulnerable narcissism and its constituent factor neurotic narcissism (but not the other factors) and all of its facets, as well as the antagonistic narcissism facet distrust, and lower scores on lack of empathy. Offenders had higher scores on grandiose narcissism and its constituent factor antagonistic narcissism (but not agentic narcissism) and lower scores on vulnerable narcissism (note that this was not the case in longer versions) and its constituent factor neurotic narcissism (in all versions). At facet level, incarceration status was related to higher scores in the facets reactive anger (though only by trend), indifference, grandiose fantasies, manipulativeness (by trend), exploitativeness, entitlement, lack of empathy, arrogance, thrill seeking, and distrust, and lower scores in shame and need for admiration. The trend-level difference observed for authoritativeness in Study 1 was not evident in the BF. Taken together, the FFNI-BF is well suited to capture different characteristic FFNI profiles (see Study 1) at factor and facet level.