The Development of Machiavellianism, Psychopathy, and Narcissism in Young Adulthood

We focus on Machiavellianism, psychopathy, and narcissism because each of these traits has an extensive history in social scientific research and because there is great interest in studying these three traits together (Dinić & Jevremov, 2021). This so-called “Dark Triad” (Paulhus & Williams, 2002) has been criticized for redundancy among Machiavellianism and psychopathy (Persson et al., 2019) or with honesty-humility (Hodson et al., 2018). However, the meta-analytic and attenuation-corrected overlaps between Machiavellianism and psychopathy (r = .66; Vize et al., 2018) and with honesty-humility (−.69 for Machiavellianism, −.63 for psychopathy, and −.63 for narcissism; Howard & Van Zandt, 2020) are high, but far from indicating complete redundancy. For each of the three traits, there are still various aspects left that are theoretically and empirically unique to it (Horsten et al., 2021; Jones & Mueller, 2022; Jones & Paulhus, 2017; Marcus et al., 2018; Miao et al., 2019; Volmer et al., 2019).

Recent research emphasized a general dispositional tendency (named “Dark Core” or “Dark Factor”) underlying Machiavellianism, psychopathy, and narcissism alongside various other traits (Bader et al., 2021; Moshagen et al., 2018). In the present study, we deemed it worthwhile to consider each trait as a distinct entity possessing both—unique features as well as shared features.

The construct of Machiavellianism emerged from the study of individual differences in manipulative tendencies and is inspired by Niccolò Machiavelli’s 16th century book “Il Principe” (Christie & Geis, 1970). It captures the degree to which individuals use deceitful means in order to obtain personal benefits and rationalize this behavior by having a negative view of human nature. Individuals high in Machiavellianism can be described as immoral, cynical, and manipulative (Collison et al., 2018; Dahling et al., 2009; Grosz, Harms, et al., 2020; Monaghan et al., 2020; Rauthmann, 2013).

The conceptualization of subclinical psychopathy could merely be based on personality features such as guiltlessness, fearlessness, lack of forethought, and dishonesty and does not a require a history of antisocial and criminal behaviors (Lilienfeld & Andrews, 1996). However, the commonly used self-report measure SRP-III (based on Hare’s Psychopathy Checklist; Hare, 1985) does not only consist of personality features like remorselessness (“callous affect”) or impulsiveness (“erratic lifestyle”), but also captures antisocial tendencies (“interpersonal manipulation”) and illegal behavior (“criminal tendencies”; Mahmut et al., 2011). Aspects such as impulsivity and thrill-seeking are more central to psychopathy than to Machiavellianism (Grosz, Harms, et al., 2020).

The construct of narcissism is inspired by the 2,000-year-old myth of Narcissus. According to a recent conceptualization (Back et al., 2013), subclinical narcissists strive for a grandiose self through an assertive/self-enhancing strategy (narcissistic admiration) and/or by antagonistic/self-protective means (narcissistic rivalry). Individuals high in narcissistic admiration perceive themselves as grandiose, promote themselves with self-confident behavior, feel entitled to others’ admiration, and leave a mostly positive first impression. Individuals high in narcissistic rivalry think competitively, distrust others, see them as worthless, feel entitled to others’ undivided attention, and behave coldly and revengeful (Back et al., 2013).

Traits by definition assume a substantial degree of stability. First and foremost, this concerns the order of individuals on a particular trait. Rank-order stability (or “differential” stability) may not be confused with the absence of mean-level change (or “absolute” stability; Morey & Hopwood, 2013). For instance, mean levels of intelligence may increase strongly over time in an entire cohort of children while the relative positions of each child in repeatedly assessed distributions of intelligence scores may remain quite stable.

We expected that the rank order of young adults on Machiavellianism, psychopathy, narcissistic rivalry, and narcissistic admiration would be highly stable over two years. All of these traits are characterized by specific patterns of social interaction as a manifestation of negative world views, affective predispositions, or status striving. We believe that this behavior is unlikely to fluctuate on a short-term basis because important relationships with friends or family may be built around said views and habits. Others may even stabilize high (vs. low) levels on these traits by reinforcing and reciprocating (vs. avoiding and questioning) typical behavior.

Large datasets on the Big Five suggest that latent test-retest correlations over a few years range in the .70s (Ferguson, 2010; Seifert et al., 2021). Similar findings were reported using short measures for Machiavellianism, psychopathy, and narcissism (Zettler et al., 2021). We posited:

Theory and empirical findings both suggest that mean levels of Machiavellianism, psychopathy, and narcissistic rivalry are likely to decrease over time in young adults. In the present study, we draw on the neo-socioanalytic model (Roberts & Nickel, 2021) which comprises the maturity principle and the social investment principle. The maturity principle is based on the empirical observation that agreeableness, conscientiousness, and emotional stability increase in young adulthood and interprets these changes as reflecting maturation (Bleidorn et al., 2022; Hoff et al., 2020; Lüdtke et al., 2011; Roberts et al., 2006). The social investment principle attributes this assumed maturation to young adults’ mastery of universal life tasks such as establishing one’s education, employment, or family (Roberts et al., 2005). It posits that these life tasks are associated with specific social roles such as being a student, an employee, or a romantic partner. It assumes that individuals invest effort to fulfill others’ expectations associated with a role and may get rewarded in return. For instance, being a student may require agreeable behavior which might lead to good relationships with fellow students or advisers. As those cultural socialization experiences are assumed to promote maturation (Bleidorn et al., 2013), maturity is defined in social (rather than biological) terms (Otto & Kaiser, 2014). While genetic factors contribute substantially to phenotypic personality stability in young adulthood, environmental factors contribute to stability, too (Briley & Tucker-Drob, 2014), and, more importantly, to developmental changes in phenotypic personality (McGue et al., 1993; Roberts & Nickel, 2021). Please note that these theoretical assumptions serve as a potential explanation for the patterns of mean-level development examined here but are not being investigated themselves in the present study.

Here, we propose that the maturity principle applies to young adults’ Machiavellianism, psychopathy, and narcissistic rivalry as well. As described above, these traits are all characterized by patterns of antisocial behavior which is unlikely to be tolerated in students, colleagues, or romantic partners. For Machiavellianism, positive experiences at university may disprove negative views of human nature. For psychopathy, breaking rules may result in punishment. For narcissistic rivalry, cold and revengeful behavior may lead to escalating interpersonal conflicts with aversive consequences.

Some longitudinal studies have found that Machiavellianism (Grosz, Göllner, et al., 2019; Rogoza et al., 2021; Zettler et al., 2021), psychopathy (Zettler et al., 2021), or narcissism (Chopik & Grimm, 2019; Rogoza et al., 2021; Wetzel et al., 2020; Zettler et al., 2021) decreased with age whereas others have not reported reliable decreases in Machiavellianism (Davis et al., 2022; Klimstra et al., 2020; Sijtsema et al., 2019), psychopathy (Davis et al., 2022; Klimstra et al., 2020; Rogoza et al., 2021; Sijtsema et al., 2019), or narcissism (Carson et al., 2019; Davis et al., 2022; Fang et al., 2021; Farrell & Vaillancourt, 2021; Grosz, Göllner, et al., 2019; Klimstra et al., 2020; Orth & Luciano, 2015; Sijtsema et al., 2019). Large-scale cross-sectional studies mainly suggest that Machiavellianism, psychopathy, and narcissism all decrease with age (Foster et al., 2003; Götz et al., 2020; Hartung et al., 2022; Kawamoto et al., 2020; Klimstra et al., 2020; Weidmann et al., 2023). However, several studies found that Machiavellianism, psychopathy, or narcissism first increased at some point during adolescence before they started to decrease in adulthood (K. S. Carlson & Gjerde, 2009; Götz et al., 2020; Klimstra et al., 2020; Sijtsema et al., 2019; Tuvblad et al., 2016). According to the disruption hypothesis, it is not uncommon that a trait develops in different directions in adolescence and young adulthood (Brandes et al., 2021; Soto, 2016). Other studies reported significant decreases even in adolescence (Rogoza et al., 2021). Taken together, the cited evidence suggests that Machiavellianism, psychopathy, and narcissism indeed decrease throughout adulthood. However, it is not clear for each of these traits whether or when there is some kind of a turning point at a young age at which their decrease begins.

Many issues remain unresolved. First, the most consistent evidence came from cross-sectional studies which arguably provide less information about developmental trajectories than longitudinal studies. Second, virtually all of the cited studies used short or very short measures of Machiavellianism, psychopathy, and/or narcissism which may have limited the conceptual representation of these traits. Third, narcissism was most often considered as a unidimensional trait without distinguishing between dimensions of narcissism with antagonistic (narcissistic rivalry) versus agentic (narcissistic admiration) behavioral dynamics, which may have concealed potential uniqueness in their development. Fourth, very few studies focused specifically on young adulthood. Fifth, some longitudinal studies collected data in only two waves of measurement and were thus unable to examine nonlinear development. Sixth, even the studies that did focus on young adulthood and included three or more waves of measurement did not report analyses that accounted for the possibility of nonlinear development. This may have resulted in overlooking the turning point in these traits at which their decline begins—something that has been found in cross-sectional studies.

The present study aimed to address these issues to the best of its capacity limited by design aspects (sample size, attrition, and study length) and sample composition (with regard to gender, age, student status, and geographic region). To account for the possibility of nonlinear development in young adulthood, the analyses involved identifying a best-fitting growth function to approximate the development of each trait in this particular sample of young adults. To this end, we collected data in four waves of measurement and selected final models for each trait according to preregistered criteria. During this model-building process, we considered whether form, direction, and steepness of growth curves differ as a function of participants’ age when they entered the study. As participants were between 18 and 30 years old at Time 1, this study’s analytical approach is theoretically able to detect whether a trait increases in 18-year-olds but decreases in 30-years-olds. Aside from that, this study employed full-length measurement instruments and distinguished narcissistic rivalry from narcissistic admiration.

Overall, we expected that mean levels of young adults’ personality change in a direction that can be interpreted as greater maturity. We posited:

Furthermore, we argue that the maturity principle is unlikely to involve a change in young adults’ narcissistic admiration. As described above, narcissistic admiration reflects a self-enhancing strategy intended to raise social potency. Individuals high in narcissistic admiration show charismatic behavior (Back et al., 2013; Rogoza & Fatfouta, 2020). They are likely to gain popularity (Leckelt et al., 2020) and status (Grosz, Leckelt, et al., 2020; Leckelt et al., 2019), especially in relationships with minimal acquaintance (Back et al., 2010; Dufner et al., 2019; Grijalva et al., 2015). In this regard, narcissistic admiration seems compatible with some requirements of young adults’ social roles as friends, partners, students, or colleagues. However, short-term social benefits may wane over time (Dufner et al., 2019) due to narcissistic admiration’s less desirable aspects such as entitlement (Back et al., 2013) or seeking validation. Results in university exams may disconfirm inflated self-views of one’s abilities (Zajenkowski et al., 2020). Taken together, narcissistic admiration may neither be particularly beneficial or detrimental to the fulfillment of age-graded social roles.

Longitudinal research documented substantial increases in self-esteem (Orth et al., 2018) as well as in the social dominance facet of extraversion (Roberts et al., 2006) during young adulthood, which both overlap with narcissistic admiration. Longitudinal research on narcissistic admiration itself is scarce. The few available studies did not detect any changes in young adulthood (Carson et al., 2019; Grosz, Göllner, et al., 2019) or late adolescence (Fang et al., 2021). As described in the preregistration and methods, we conducted an equivalence test (Lakens et al., 2018) and posited:

It is important to distinguish conceptually between mean-level change and individual differences in developmental paths. Mean-level change reflects the aggregate of change across all individuals. Note that even with a constant amount (or lack) of mean-level change, individuals can still differ systematically from each other with regard to the direction or steepness of their own developmental paths (and, hence, the amount of their own personal change). A contrasting view would be that all individuals follow the exact same developmental trajectory and differ merely in the starting levels of their paths. Unsystematic deviations from a person’s developmental path may always be present and by definition do not follow a discernible pattern.

We expected people to differ in their pace of becoming more mature (steepness of developmental trajectories) or whether they become more mature at all (direction of developmental trajectories). Maturation itself might depend on the amount (or lack) of experiences associated with an individual’s social roles that are assumed to stimulate maturation (Bleidorn et al., 2013) and/or on their perception of these experiences. Machiavellianism might decrease more quickly if individuals experience trustworthiness in fellow students, colleagues, or supervisors. Psychopathy might decline more quickly if individuals get caught during cheating (Ternes et al., 2019) or underperform in exams due to a lack of preparation. Narcissistic rivalry might decrease more quickly if a person gets involved in a grueling conflict with other students or supervisors. Narcissistic admiration may change systematically in either direction depending on the development of one’s social status (Mahadevan et al., 2019). The occurrence of critical experiences could partly depend on a person’s study subject (Krishnan, 2008), their employment status (Grosz, Göllner, et al., 2019), and on having a leadership role (Wetzel et al., 2020).

Previous research on systematic individual differences in the developmental paths of Machiavellianism, psychopathy, or narcissism is scarce. We identified four studies that reported significant Wald tests for the variance of the linear time slope for all three traits (Rogoza et al., 2021), for Machiavellianism (Klimstra et al., 2020), for narcissistic rivalry and narcissistic admiration (Carson et al., 2019), and for (subclinical grandiose) narcissism in general (Farrell & Vaillancourt, 2021). However, as variances have a boundary at 0 (i.e., they cannot be negative), conventional Wald tests are inappropriate and model comparisons need to be conducted instead (Hoffman, 2015, p. 210). Based on theoretical considerations rather than previous findings, we posited:

If individuals differ systematically from each other in the direction or steepness of their developmental paths on two (or more) particular traits, then it is possible for the slopes of both developmental trajectories to be related with each other. For example, if Person A exhibits a steeper decline of Machiavellianism than Person B, this may systematically cooccur with a steeper decline of psychopathy in Person A as well or with a steeper increase of life satisfaction (as compared to Person B).

The presence of correlated change suggests that the factors behind individual differences in change in one trait are likely to be (partially) identical to the factors behind individual differences in change in another trait. Put differently, correlated change between two traits such as Machiavellianism and psychopathy is likely to reflect the effect of at least one common cause that must have produced systematic variation in change in both traits simultaneously (with or without a direct causal relationship between the two traits). Whenever theoretical explanations of personality change draw on global principles such as young adults’ investment in normative social roles, this may imply some degree of relatedness of change in all traits that are expected to be affected by the same theoretical mechanism (e.g., Klimstra et al., 2013).

Various analytical techniques exist for modeling correlated change (Allemand & Martin, 2016). Note that within a latent variable framework, latent growth curve modeling (LGCM) allows a clearer interpretation of the estimated coefficients than latent change models. This is because LGCM is able to separate variation due to systematic development from two other sources of variation—shared influences of the measurement occasion and unique state fluctuations (for the latter two sources of variation, see Schermelleh-Engel et al., 2004). Latent change models, in contrast, conflate all of these types of variation, introducing ambiguity to the interpretation of the estimated coefficients.

Previous research investigating correlated change in personality traits using LGCM is relatively scarce (for an exception, see for instance Hoff et al., 2020), especially research that employed so called “second-order” LGCM which includes measurement models for each trait (for an exception, see Grosz, Nagengast, et al., 2019). This may be due to the requirement of LGCM of at least three waves of measurement or due to the potential decrease in the probability of a statistically significant result (as compared to latent change models) as a consequence of disentangling systematic trait development from shared influence of the measurement occasion and, in turn, disregarding the latter source of covariation.

Beyond their unique features, Machiavellianism, psychopathy, and narcissistic rivalry all share a high degree of commonalities. Strong commonalities are best documented for Machiavellianism and psychopathy (e.g., Egan et al., 2014; Miller et al., 2017; Muris et al., 2017; Persson, 2019; Vize et al., 2018), but are likely to be present in pairings with narcissistic rivalry as well (e.g., Back et al., 2013; Dinić et al., 2021). Previous research has documented a common factor underlying Machiavellianism, psychopathy, and an undifferentiated conceptualization of narcissism (Jones & Figueredo, 2013; Kowalski et al., 2021; McLarnon & Tarraf, 2017; Schreiber & Marcus, 2020; Volmer et al., 2019). Such a general factor underlies many other traits as well (Bader et al., 2021; Moshagen et al., 2020) and retains its meaning even if any one particular trait is omitted (Moshagen et al., 2018).

We expected the slopes of each person’s developmental paths on Machiavellianism, psychopathy, and narcissistic rivalry to be positively related with each other for two reasons. Most importantly, sharing substantial commonalities means that these traits partly consist of identical (or perfectly related) aspects. Any systematic change in one of these aspects should concern all three traits in a virtually identical way. For instance, as these traits are all characterized by patterns of antisocial behavior such as acting selfishly, an environment that disincentivizes selfishness may lead to a similar decline in all three traits.

Less importantly, even the unique (unrelated) aspects of each trait can be linked in their development to some degree given that we assume that the same theoretical mechanism is driving their change. Whenever one person has more of the critical experiences expected to drive change in multiple traits than another (e.g., due to occupying different social roles), systematic individual differences in the slopes of the developmental paths may arise that could even be consistent across distinct aspects of multiple traits (Klimstra et al., 2013).

A recent study found that change in the global factor underlying Machiavellianism, psychopathy, narcissism, and other traits was highly related with change in most of these traits while excluding each focal trait from the respective measurement model for the global factor (Zettler et al., 2021). That study used short measures, covered a time span of four years, and employed latent difference models. Based on the strong commonalities between Machiavellianism, psychopathy, and narcissistic rivalry, the shared theoretical mechanism assumed to cause change in their unique aspects, and the findings by Zettler et al. (2021), we posited:

Narcissistic admiration and self-esteem share essential features. On each of these constructs, high levels reflect a positive view of the self (Brummelman et al., 2016). One factor shaping this regard for oneself is an individual’s perceived social status. More specifically, hierometer theory posits that narcissism and self-esteem both track perceived social status in order to navigate status hierarchies and to regulate assertive behavior (Mahadevan et al., 2016, 2019). In young adulthood, we expected perceived social status to change systematically (with mean levels increasing) as many young adults attempt to master normative social roles. We theorized that systematic change in perceived social status acts as a common factor underlying individual differences in change in both narcissistic admiration and self-esteem, causing both constructs to be positively associated in their development.

Systematic change in perceived social status itself can be due to various factors. First, it should reflect individual differences in the achievements that young adults accumulate, for instance, throughout their studies. Second, it may reflect different speeds of maturation in terms of increases in the social dominance facet of extraversion (Roberts et al., 2006), which facilitates social interaction with others. Third, an individual’s mere subjective perception of their social status may change (while their actual social status remains constant). Individuals with high levels of narcissistic admiration tend to have an overly positive self-perception (Grijalva & Zhang, 2016; Zajenkowski et al., 2020). After receiving negative feedback from others (or a lack of positive feedback; E. N. Carlson & Lawless DesJardins, 2015; Dufner et al., 2013; Morf & Rhodewalt, 2001; Paulhus, 1998), and/or after getting worse than expected grades (Zajenkowski et al., 2020), an individual’s exaggerated status perception may become more realistic. Based on hierometer theory, all of these factors could contribute towards correlated change between narcissistic admiration and self-esteem by systematically altering perceived social status.

Overall, we expected narcissistic admiration and self-esteem to be positively associated in their development due to their common function as trackers of social status. This expectation is in line with the principle of parsimony, given that both constructs share essential features. However, narcissistic admiration and self-esteem are not identical and should partly diverge in their development as well. While narcissistic admiration is characterized by striving for grandiosity and superiority, self-esteem reflects the more general feeling of worthiness which can be achieved in many different ways (Back et al., 2013; Baumeister et al., 2003; Brummelman et al., 2016; Campbell et al., 2002; Sedikides, 2021). Cross-sectionally, narcissism and self-esteem are correlated around .30 to .40 (Back et al., 2013; Hyatt et al., 2018; Orth et al., 2016) and they are associated with different developmental experiences (Hyatt et al., 2018). In childhood, longitudinal research found different parenting styles to be related to their development (Brummelman et al., 2015; Wetzel & Robins, 2016). In adulthood, while both constructs are assumed to track perceived social status according to hierometer theory, only self-esteem is assumed to track social inclusion according to sociometer theory (Denissen et al., 2008; Gebauer et al., 2015; Leary & Baumeister, 2000; Mahadevan et al., 2016, 2019). Further factors complicate the interplay of narcissistic admiration and self-esteem. While narcissism tends to be associated with experiencing more stressful life events (Orth & Luciano, 2015) and getting divorced (Wetzel et al., 2020), stressful life events and relationship problems are inversely related with self-esteem (Harris & Orth, 2020; Orth & Luciano, 2015). Finally, the lack of mean-level change that we expect for narcissistic admiration as posited in Hypothesis 2d substantially differs from the strong increase documented for self-esteem throughout young adulthood (Orth et al., 2018).

Previous research on the longitudinal associations between narcissism and self-esteem reported positive cross-lagged paths between both constructs (Cichocka et al., 2019), was inconclusive (Cichocka et al., 2019; Jung et al., 2022; Leckelt et al., 2019; Orth & Luciano, 2015), or even found negatively correlated slopes between desire for fame and self-esteem (Grosz, Nagengast, et al., 2019). In light of the complexity of their relationship and the mixed evidence so far, we deemed it informative to test the following hypothesis:

Finally, we examined whether individual differences in young adults’ developmental paths on Machiavellianism, psychopathy, narcissistic rivalry, or narcissistic admiration systematically cooccur with individual differences in young adults’ developmental paths on life satisfaction. Again, we draw on assumed, yet unmeasured processes described by the social investment principle. More specifically, if it is indeed the case that some young adults invest the effort required for changing (part of) their personality in order to keep up with the demands imposed by age-graded social roles, then this should enable them to fulfill their social roles more successfully and, in turn, experience a sense of mastery or control over their life. In contrast, other young adults who do not become more mature are likely to experience setbacks associated with social roles that are too challenging for them and obstacles that they cannot overcome. In this way, developing more mature personality traits should be associated with a more positive trajectory on life satisfaction and could be considered “adaptive” or “healthy” (Bleidorn et al., 2020; Schwaba et al., 2022).

We are not aware of any systematic review of research on the factors underlying the development of life satisfaction in young adulthood. Some studies suggest that mean levels of life satisfaction tend to decline throughout young adulthood (e.g., Blanchflower & Graham, 2021; Cheng et al., 2017; Orben et al., 2020; Otterbach et al., 2018; Willroth et al., 2021). Rank-order stability of life satisfaction is generally high, but far from perfect (Fujita & Diener, 2005; Lucas & Donnellan, 2007), especially in young adulthood (Mann et al., 2021).

While the reasons underlying change in life satisfaction appear to be manifold, many studies reported findings in line with the assumption that mastery of age-graded social roles is vital for life satisfaction in young adulthood. For example, longitudinal change in young adults’ life satisfaction was positively predicted by partnership formation (Switek & Easterlin, 2018), marriage (Piper, 2015), or childbirth (Switek & Easterlin, 2018) and negatively predicted by unemployment (Piper, 2015). More generally, life satisfaction is positively associated with social goals (Rohrer et al., 2018), relationships with family and friends (Diener & Oishi, 2005; Helliwell & Putnam, 2004; Ponti & Smorti, 2019), romantic relationships (Gómez-López et al., 2019; Guarnieri et al., 2015), social interaction (Sun et al., 2020), or social support (Azpiazu Izaguirre et al., 2021; Calmeiro et al., 2018). Life satisfaction is also associated with self-control (Hofmann et al., 2014), self-efficacy (Burger & Samuel, 2017), proactive coping (Dwivedi & Rastogi, 2017), and academic achievement (Bücker et al., 2018; Joshanloo & Jovanović, 2020) as well as with changes in the Big Five towards greater maturity (Deventer et al., 2019; Magee et al., 2013; O’Connor, 2020; Specht et al., 2013; Tauber et al., 2016; van Aken et al., 2006).

Here we propose that the slope of young adults’ developmental path on Machiavellianism, psychopathy, or narcissistic rivalry is inversely related to the slope of their developmental trajectory on life satisfaction. Machiavellianism, psychopathy, and narcissistic rivalry are all characterized by patterns of antisocial behavior. We theorize that systematic decreases in these traits might facilitate the fulfillment of age-graded social roles which could prevent decreases in life satisfaction (Crocker et al., 2017; Eriksson et al., 2020). In contrast, failure to adapt to the requirements of important social roles could be associated with declining life satisfaction.

In meta-analyses of cross-sectional studies, Machiavellianism and psychopathy were associated with lower levels of emotional intelligence (Miao et al., 2019; Vize et al., 2018) and empathy (Vize et al., 2018), with more selfishness (Thielmann et al., 2020), antisocial tactics (Muris et al., 2017), and aggression (Vize et al., 2018), as well as with lower levels of general well-being (Muris et al., 2017). Narcissistic rivalry has been found to be related to lower levels of empathy (Back et al., 2013), trust in others (Fang et al., 2021), communal behavior (Back et al., 2013; Mielke et al., 2021; Rogoza & Fatfouta, 2020), and being perceived as trustworthy or likable (Back et al., 2013). Furthermore, it was related to more aggression and vindictiveness (Back et al., 2013; Kjærvik & Bushman, 2021) as well as to lower levels of life satisfaction (Leckelt et al., 2019; Zajenkowski et al., 2020). In summary, decreasing levels of Machiavellianism, psychopathy, and narcissistic rivalry might enable young adults to better fulfill their social roles and prevent dissatisfaction. We posited:

We had very different expectations regarding correlated change between narcissistic admiration and life satisfaction. Compared to Machiavellianism and psychopathy, narcissism shows significantly more desirable associations with well-being (Muris et al., 2017), internalizing symptoms (Vize et al., 2018), and emotional intelligence (Miao et al., 2019; Vize et al., 2018) in meta-analyses of cross-sectional studies. Narcissism was positively associated with prosocial behavior (Moshagen et al., 2018) and life satisfaction (Volmer et al., 2019) after controlling for its commonalities with Machiavellianism and psychopathy. Distinguishing narcissistic admiration from narcissistic rivalry appears to accentuate narcissistic admiration’s advantage over narcissistic rivalry (Back et al., 2013; Leckelt et al., 2019) and over traits like Machiavellianism and psychopathy in terms of psychological adjustment.

Narcissistic admiration and life satisfaction should be positively related in their development because both constructs are likely to be affected by individual differences in young adults’ development on perceived social status. Narcissistic admiration has been found to reflect perceived social status (Mahadevan et al., 2016, 2019; Mota et al., 2022). More generally, narcissistic admiration is associated with positive self-views (Back et al., 2013; Grijalva & Zhang, 2016; Humberg et al., 2019; Mielke et al., 2021) which are linked with pleasant feelings (Baumeister et al., 2003; Mota et al., 2022), psychological health (Dufner et al., 2019; Sedikides et al., 2004), and life satisfaction (Zajenkowski et al., 2020). Life satisfaction on the other hand is positively related with social status as well, both with objective indicators of academic achievement (Bücker et al., 2018; Joshanloo & Jovanović, 2020) and with subjective perceptions of status-related competencies (Dufner et al., 2019). We posited:

Note that we did not base our expectation of positively correlated change between narcissistic admiration and life satisfaction on assumed maturation in narcissistic admiration. As mentioned above, some aspects of narcissistic admiration such as charismatic behavior (Rogoza & Fatfouta, 2020) may indeed facilitate the fulfillment of age-graded social roles. For instance, narcissistic admiration has been linked with having many friends, earning more money, and obtaining a leadership position (Leckelt et al., 2019). Becoming more mature in these particular aspects of narcissistic admiration could indeed contribute towards positively correlated change with life satisfaction. However, other aspects of narcissistic admiration such as entitlement (Back et al., 2013) and seeking validation are likely to be inversely related with life satisfaction (Crocker & Park, 2004; Morf & Rhodewalt, 2001; Sedikides, 2021). For instance, self-enhancement values negatively predicted change in affective well-being (Grosz et al., 2021) and the slope of young adults’ developmental path on desire for fame was positively related with the slope of their developmental path on negative affect (Grosz, Nagengast, et al., 2019). Becoming more mature in those less desirable aspects of narcissistic admiration should contribute towards negatively correlated change with life satisfaction (as decreases in undesirable characteristics should facilitate role fulfillment), thus counterbalancing any potential positively correlated change between narcissistic admiration and life satisfaction due to assumed maturation in desirable aspects of narcissistic admiration.

The present study examined the development of Machiavellianism, psychopathy, and narcissism with preregistered hypotheses. It focused on young adulthood, which is dense with challenging social roles, and on students in particular. We believe that the student role can be a fit with some of the assumptions put forward by the social investment principle. It provides a challenging environment in many regards. Being a student involves meeting new people regularly, getting along with everyone, allocating enough time for studying, organizing oneself to fulfill all study requirements, as well as overcoming the challenges of living away from one’s parents.

Change towards greater maturity in traits like Machiavellianism or psychopathy can manifest itself in linear decreases or in an inverted U-shaped curve. We collected data in four waves of measurement to be able to explore the possibility of nonlinear change following a preregistered model building procedure. The analysis plan also took into account that participants were between 18 and 30 years old when they entered the study, exposing any age-related differences in their developmental paths throughout the two-year duration of the study.

Narcissism is a complicated construct. This study differentiated an antagonistic dimension of narcissism (narcissistic rivalry) from a self-enhancing one (narcissistic admiration), revealing potential uniqueness in their development.

The present research included repeated assessments of life satisfaction. Maintaining life satisfaction while navigating the assumed challenges of young adulthood may be associated with personality change interpreted as reflecting greater maturity. This would corroborate a view of maturation as desirable. Self-esteem was also measured repeatedly. Its overlap with narcissistic admiration makes it an important consideration in the entire constellation of traits.

Current knowledge about the development of Machiavellianism, psychopathy, and narcissism in young adulthood is very fragmented and provides few reliable answers. Previous longitudinal research often studied only one or two of these traits, used short or very short measures, did not model (potentially curvilinear) mean-level change, and/or did not specify measurement models. Correlated development with life satisfaction had not been examined. In an attempt to overcome these shortcomings, the current study may contribute incremental information about the development of Machiavellianism, psychopathy, narcissistic rivalry, and narcissistic admiration in young adulthood.

An extensive preregistration with detailed information about hypotheses, samples, measures, and analyses is available for this study at https://osf.io/98wqf/. The preregistration was created after data collection was completed and before any substantive analyses were performed. Any deviations from the preregistration are stated in the text.

This study collected data in four waves of measurement covering a time period of two years (at Months 0, 6, 12, and 24). Participants were recruited from two German universities via flyers and mailing lists and through advertisements on bulletin boards and in courses. The final sample consisted of 926 individuals (74.0% female, 25.8% male, 0.2% transgender) with a mean age of 21.89 years (SD = 2.82) at T1. Of those 926 individuals, 39% participated in all four waves, 20% took part in three waves, 17% participated in two waves, and 25% participated only once. To obtain this final sample, we applied various preregistered exclusion criteria which are described below. In addition to the preregistered exclusions, we removed two participants who reported an implausible pattern of age values and three more individuals who turned out to be outside the targeted age range of 18-30 years at T1 but had not been excluded before because they had reported a different (and conflicting) age at T2. Table S1 presents an overview of all exclusions as well as demographic information for all waves of data collection.

At T1, data collection took place in the laboratory. At T2 to T4, we invited participants using personalized invitation links sent via e-mail. At T1, we compensated participants with either research participation credit or a payment of 8 to 15 Euros depending on the duration of the assessment. Furthermore, we offered feedback on an intelligence test score (which is not part of the present study) to those who took that test. Compensation for participation at T2 to T4 was offered in form of Amazon vouchers over 8 Euros. In addition, participants could take part in a lottery of Amazon vouchers with values from 25 to 100 Euros.

At the first wave of data collection, a total of 989 individuals within our targeted age range of 18-30 years participated. We removed six participants who exhibited language difficulties and one individual who did not read the instructions. Of the remaining 982 participants, we removed one person who filled out the survey more than two SD quicker than the average participation time (M = 47:12 min, SD = 14:39 min), 110 individuals who responded incorrectly to one or both instructed response items, and two participants who reported implausible age values over time. The final sample consisted of 869 participants (M_age = 21.89 SD = 2.82, 75.1% female, 24.6% male, 0.2% transgender). Data from this sample have been used in Frick (2022), Grosz et al. (2020), Wetzel and Frick (2020), and Wetzel, Frick, and Brown (2021).

Approximately six months after T1, 614 individuals participated in the second wave, resulting in a 62% response rate. We excluded one individual whose ID had not been recorded. Furthermore, we removed 16 participants who filled out the survey more than two SD faster than the average participation time (M = 25:02 min, SD = 7:26 min), 44 individuals who responded incorrectly to the instructed response item, one person who reported implausible age values over time, and three participants who reported their age incorrectly at T2 (who were outside the targeted age range from 18-30 years at T1). The final sample consisted of 549 participants (M_age = 22.34, SD = 2.75, 77.0% female, 22.8% male, 0.2% transgender, 94% students). Data from this sample have been used in Wetzel and Frick (2020) and Grosz et al. (2020).

Approximately one year after T1, 619 individuals took part in the third wave, amounting to a 63% response rate. We excluded one person who exhibited language difficulties at T1. Of the remaining 618 participants, we removed six individuals who filled out the survey more than two SD quicker than the average participation time (M = 16:21 min, SD = 4:57 min), 45 participants who responded incorrectly to the instructed response item, and two participants who reported implausible age values over time. We obtained a final sample of 565 individuals (M_age = 22.74, SD = 2.78, 75.8% female, 24.1% male, 0.2% transgender, 92% students).

Approximately two years after T1, 584 individuals participated in the fourth wave for a response rate of 59%. We excluded 13 participants who filled out the survey more than two SD faster than the average participation time (M = 22:25 min, SD = 6:58 min), removed another 32 participants who responded incorrectly to the instructed response item, and excluded two participants who reported implausible age values over time. The final sample thus contained 537 participants (M_age = 23.73, SD = 2.88, 76.0% female, 23.8% male, 0.2% transgender, 83% students).

As attrition between T1 and T4 was not negligible, we tested whether attrition could be predicted by characteristics at T1 (which has not been planned in our preregistration). We modeled attrition at T4 as a dichotomous outcome. Multiple regression analysis revealed that age_T1 was not a significant predictor of attrition, β_std. = −.017, p = .80, whereas gender was, β_std. = −.134, p = .024, indicating that women were less likely than men to drop out of the study.

In four separate simple regression analyses, we predicted attrition at T4 by levels of each trait at T1. While Machiavellianism (β = 0.075, p = .11), narcissistic rivalry (β = 0.067, p = .16), and narcissistic admiration (β = 0.013, p = .48) did not significantly predict attrition at T4, psychopathy did (β = 0.043, p = .018), indicating that individuals with higher levels of psychopathy at T1 were more likely to drop out of the study.

To test our hypotheses, we used established scales as our primary measures of Machiavellianism, psychopathy, and narcissism. Additionally, we repeated all analyses using shorter measures of Machiavellianism, psychopathy, and narcissism and report the results for those measures in online supplementary files. All measures were assessed at T1 to T4 except for self-esteem, which was assessed at T2 to T4.

Machiavellianism. Respondents completed the MACH-IV scale (Christie & Geis, 1970) as our primary measure of Machiavellianism. The MACH-IV consists of 20 items reflecting interpersonal manipulation, cynicism, or immorality. Participants indicated their agreement with statements such as “It is wise to flatter important people” on a 6-point rating scale ranging from −3 = do not agree at all to 3 = strongly agree.

Psychopathy. Participants completed a modified version of Hare’s (1985) Self-Report Psychopathy Scale (SRP-III; Williams et al., 2003) as our primary measure of subclinical psychopathy. The SRP-III consists of 31 items reflecting antisocial behavior, impulsive thrill-seeking, interpersonal manipulation, and cold affect. Participants indicated their agreement with statements such as “I get in trouble for the same things” on a 5-point rating scale ranging from do not agree at all to strongly agree.

Narcissism. Grandiose narcissism was measured with the 18-item Narcissistic Admiration and Rivalry Questionnaire (NARQ; Back et al., 2013). The NARQ distinguishes two subdimensions of narcissism named admiration and rivalry, which are each measured with nine items. Sample items are “I show others how special I am” (admiration) or “I react annoyed if another person steals the show from me” (rivalry). Participants indicated the extent of their agreement with these statements on a 6-point rating scale ranging from does not apply at all to fully applies.

Short Dark Triad. We also included the 27-item Short Dark Triad scale (SD3; Jones & Paulhus, 2014) as a secondary measure of Machiavellianism, psychopathy, and narcissism. The SD3 assesses each trait with nine items such as “Whatever it takes, you must get the important people on your side” (Machiavellianism), “Payback needs to be quick and nasty” (psychopathy), and “People see me as a natural leader” (narcissism). Respondents indicated how much they agreed with each statement on a 5-point rating scale ranging from disagree strongly to agree strongly.

M7. As another secondary measure of Machiavellianism, we scored the M7 from the administered Machiavellianism instruments. The M7 is a seven-item measure which was developed to reduce the overlap between Machiavellianism and psychopathy (Grosz, Harms, et al., 2020). It uses three items from the MACH-IV and four items from the SD3. An example item reads “Whatever it takes, you must get the important people on your side.”

P7. We also scored the P7 (Grosz, Harms, et al., 2020) from the data. The P7 assesses subclinical psychopathy with five items from the SRP-III and two items from the SD3. A sample item is “It’s fun to see how far you can push people before they get upset.”

NARQ-S. As another secondary measure of narcissism, we also scored the NARQ-S (Leckelt et al., 2018) from the responses to the full NARQ. It measures admiration and rivalry with three items each.

Self-Esteem. Respondents completed the 10-item Rosenberg Self-Esteem Scale (Rosenberg, 1965) on a 4-point rating scale ranging from does not apply at all to fully applies. A sample item is “On the whole, I am satisfied with myself.”

Life Satisfaction. Participants completed the five-item Satisfaction With Life Scale (Diener et al., 1985) on a 7-point rating scale ranging from do not agree at all to completely agree. An example item reads “I am satisfied with my life.”

Further variables were assessed of which many were used and described in a study focusing on the forced-choice response format (see Wetzel & Frick, 2020). The experimental manipulations for two studies on the forced-choice response format are described in Wetzel and Frick (2020) and Wetzel et al. (2021).

We applied item factor analysis (Glockner-Rist & Hoijtink, 2003; Wirth & Edwards, 2007) to account for the categorical response format of all indicators of Machiavellianism, psychopathy, and narcissism. All models were fit to the data in Mplus version 8.4 (L. K. Muthén & Muthén, 1998–2017). Parameters were estimated using unweighted least squares estimation with a mean- and variance-adjusted chi-square test statistic (denoted ULSMV in Mplus). With categorical data, ULSMV tends to be fast and accurate (Forero et al., 2009; B. O. Muthén et al., 2015). We used the R (R Core Team, 2013) package MplusAutomation (Hallquist & Wiley, 2018) to determine which parameters violate measurement invariance and to analyze the Mplus output files.

Measurement Models. We specified measurement models for each trait in order to be able to account for measurement error in the indicators. To scale the continuous common factor underlying all indicators, we fixed its mean and variance at 0 and 1, respectively. To scale the continuous latent response variables underlying each categorical indicator, we used the conditional parameterization and fixed the residual variance of each indicator at 1 (Kamata & Bauer, 2008) which is referred to as theta parameterization in Mplus.

With respect to the factor structure of each trait, we followed the respective literatures for each of the questionnaires we applied. We preferred models with fewer dimensions to keep our analyses as simple as possible and to preserve statistical power for examining mean-level change. More specifically, we specified unidimensional measurement models for Machiavellianism and psychopathy because of a lack of stable multidimensional factor structures for our measures. For narcissism, we distinguished between rivalry and admiration as this distinction is well supported for the measure we used (Back et al., 2013). We excluded Items 4 and 15 from the SRP because they had nonsignificant factor loadings at T1. Based on a priori theoretical considerations and modification indices from CFA analyses with the T1 data, we allowed the following covariances between item residuals: For Machiavellianism, we correlated Item 4 (“Most people are basically good and kind”) with Item 14 (“Most people are brave”) and Item 6 (“Honesty is the best policy in all cases”) with Item 9 (“All in all, it is better to be humble and honest than important and dishonest”). For psychopathy, we correlated Item 6 (“I get a kick out of ‘conning’ someone”) with Item 18 (“Conning people gives me the ‘shakes’”) and Item 24 (“I cheated on school tests”) with Item 26 (“I plagiarized a school essay”). For narcissistic rivalry, we correlated Item 13 (“Most people won’t achieve anything”) with Item 17 (“Most people are somehow losers”), Item 4 (“I react annoyed if another person steals the show from me”) with Item 12 (“I can barely stand it if another person is at the center of events”), and Item 14 (“Other people are worth nothing”) with Item 17. Finally, for narcissistic admiration, we correlated Item 7 (“Most of the time I am able to draw people’s attention to myself in conversations”) with Item 16 (“I manage to be the center of attention with my outstanding contributions”) and Item 7 with Item 18 (“Mostly, I am very adept at dealing with other people”).

Measurement Invariance. When examining development over time, it is a necessary prerequisite that the measurement properties of the indicators remain constant over time. Otherwise, it would be ambiguous whether observed change over time can be attributed to changes in the construct of interest or merely to changes in the measurement properties of its indicators. We specified latent-state models (Steyer et al., 1999, 2015) and started with imposing strict measurement invariance, i.e., we constrained factor loadings, thresholds, and residual variances to equality over time for each indicator. Since residual variances have to be fixed to 1 at one measurement occasion, all residual variances were fixed to 1. Holding residual variances equal prevents a potential loss of statistical power to detect noninvariance in loadings and thresholds (Lubke & Dolan, 2003). We then allowed partial noninvariance by iteratively freeing all factor loadings and thresholds where noninvariance exceeded 0.25 for factor loadings and 0.375 for thresholds, corresponding to moderate to strong noninvariance in terms of effect size (Wetzel et al., 2017; Zieky, 1993), in the order of their modification indices. Those parameters that remained invariant established a common metric across measurement occasions (Reise et al., 1993). It was therefore only necessary to fix the mean and variance of the factor at T1 (or T2 for self-esteem) at 0 and 1, respectively, whereas means and variances were freely estimated for the factors at T2 to T4.

We used a modified version of the final partial invariance model for the measurement part of all further models. More specifically, we only fixed the residual variances of each indicator to 1 at T1 (or T2 for self-esteem) for identification and allowed the others to be estimated freely. In addition, we estimated correlations between all residuals of the same indicator over time to prevent inflated estimates of stability (Marsh & Hau, 1996).

In order to examine Hypothesis 1, we used the modified version of the final partial measurement invariance model. Additionally, we controlled for age differences at T1 (age_T1) in all latent state variables. Without controlling for age_T1, any correlation between latent state variables might not only reflect longitudinal rank-order stability but could also be inflated due to cross-sectional age differences creating a confounding source of common variance in all latent state variables. As displayed in Figure 1, we considered the correlation between the disturbances of the factors for T1 and T4 as our parameter of interest for testing Hypothesis 1. We used a Wald test with one-tailed significance testing against a lower bound of r = .30 (Gignac & Szodorai, 2016; see preregistration for a table with an overview of all hypotheses and the corresponding tests).

Mean-Level Change. We modeled longitudinal change using latent growth curve models (LGCM; McArdle, 2009). LGCM enabled us to account for measurement error in the indicators. Compared to growth curve modeling without measurement models, LGCM avoids conflating measurement error with occasion-specific state disturbances. LGCM also allowed us to base all analyses on the modified version of the partial measurement invariance model described above.

To be able to examine Hypothesis 2, we needed to identify a final model for each trait that best approximates the pattern of change observed in the data. The final model may include effects of a higher order than a linear time slope. To arrive at that final model, we drew on principles that are established in longitudinal multilevel modeling (also known as hierarchical linear models) which is conceptually similar to growth curve modeling (Geiser et al., 2013; MacCallum et al., 1997). As described in the preregistration, we built the final model in the following way (cf. Hoffman, 2015, p. 258). First, we started with the highest-order fixed time slope that was estimable and may be interpreted, which is a fixed cubic time slope in our data in a model with a random effect (i.e., a variance of a latent variable) only for the intercept. Whenever a higher-order term was in the model, all lower-order terms had to be in the model as well to allow a clear interpretation of the higher-order term (see Figure 2). We would then, step-by-step, remove all nonsignificant fixed effects according to Wald tests with two-tailed significance testing starting with the highest-order terms. We did not remove the fixed linear time slope or any nonsignificant effects that had significant higher-order terms above them.

So far, all models only had a random effect for the latent intercept variable in them. From there, we added a random effect to the linear time slope (by freely estimating its variance instead of fixing it at 0). We kept it if a chi-square difference test indicated that removing it significantly decreased model fit (this was the test of Hypothesis 3). If we kept it, then we would add a random variance to the quadratic time slope (if it was still in the model) and kept it if removing it significantly decreased model fit. With four measurement occasions, a model with a random variance for the cubic time slope would not be identified. Note that most of the above steps did not serve the purpose of hypothesis testing itself but rather aimed to identify the final model which we would then use for hypothesis testing. In case that the disturbance at T4 was estimated at a negative value, we fixed it at 0 because variances cannot be negative. In the other model involved in a model comparison according to Hypothesis 3, we fixed the disturbance at T4 to the value it was estimated at. This had not been planned in the preregistration.

When analyzing mean-level development, we took into account participants’ age differences at T1. Age_T1 varied between 18 to 30 years (M = 21.89, SD = 2.83). In longitudinal studies, age_T1 is not a regular covariate. Instead, it is the cross-sectional equivalent of what is investigated longitudinally. While we identified the final model for each trait (as described above), we controlled for age_T1 at each step. Alongside each longitudinal time slope, we included an effect of age_T1 of the same order on the intercept as well as an effect of age_T1 on all lower-order time slopes (which can be understood as an interaction of age_T1 with time). Effects of age_T1 were kept in the model as long as the respective time slope remained in the model. Controlling for age_T1 helped us to distinguish cross-sectional effects of being older from longitudinal effects of getting older (Hoffman, 2015, p. 448) and helped avoid imposing unintended contextual cohort effects on the model. Accounting for cross-sectional age differences may also enhance the detection of longitudinal change. For instance, when time slopes from different age cohorts go in different directions (as it is commonly observed with quadratic change), accounting for age_T1 could prevent them from canceling each other out.

We centered age_T1 at 22 years (close to its mean). At the centering point, each time slope is unconditional to any potential effects that age_T1 may have on it. Due to potential effects of age_T1 on a time slope, testing a time slope for significance only applies to individuals at the centering age_T1 of 22 years. To estimate the predicted mean for a time slope across the entire range of age_T1, we used the “model constraint” command available in Mplus which provides any requested parameters and their standard errors that derive from other estimated parameters in the model. For identification, we fixed intercept and disturbance (variance not accounted for by any effect of age_T1) of the latent intercept variable at 0 and 1, respectively.

To report an effect size of longitudinal mean-level development over the period of 2 years, we estimated Cohen’s d, $Cohen'sd=y^T4−y^T11+var(T1Disturbance)$⁠.

In the numerator, we subtracted the mean at T1 (fixed at 0) from the model-implied mean at T4 (calculated as $2⋅αLin$ plus any higher-order effects). In the denominator, we used the square root of the sum of the disturbance of the latent intercept variable (fixed at 1) and the estimated disturbance of the factor for T1 (occasion-specific variance which is unexplained by the intercept variable). In this way, the denominator should be similar to the estimate of the standard deviation in the population which, according to Feingold (2009), is the consensus to be used for estimating effect sizes in growth modeling.

We tested Hypotheses 2a-2c using Wald tests with one-tailed null-hypothesis significance testing of Cohen’s d against 0. To test Hypothesis 2d, we used an equivalence test (as described in Lakens et al., 2018) and used the two one-sided tests (TOST) procedure against symmetric equivalence bounds of ±0.075 for Cohen’s d. This means that we reject the null hypothesis that narcissistic admiration changes with a d of at least ±0.075 and accept the alternative hypothesis that narcissistic admiration does not change or changes significantly less that an effect of d of at least ±0.075 if both one-sided Wald tests are significant. We report the larger of the two p values.

The design of this study allowed testing for cohort and/or selection effects given that there is both longitudinal variability in time (i.e., four measurement occasions over two years) and cross-sectional variability in age_T1 (i.e., participants were aged 18-30 years at T1). More specifically, we tested whether characteristics of growth curves (position, direction, and acceleration) converged across longitudinal and cross-sectional data. This was done by requesting contextual effects using the “model constraint” command available in Mplus. Contextual effects are a combination of other estimated parameters in the model (cf. Hoffman, 2015, p. 449). They indicate nonconvergence between longitudinal and cross-sectional growth curve parameters and can be interpreted as cohort and/or selection effects. Whenever these analyses did not reveal significant contextual effects, we constrained any contextual effects to 0 (as in accelerated longitudinal designs; Bell, 1953; Duncan et al., 1996). This meant that the predicted mean for 18-year-olds at T4 equals the predicted mean for 20-year-olds at T1 and so forth. In that case, we reported additional estimates of Cohen’s d for the predicted mean-level development over the entire age span covered in this study (i.e., from age 18 to 32).

Correlated Change. In order to examine Hypotheses 4-6, we specified LGCMs based on the final model for each trait. To examine Hypothesis 4, we used bivariate LGCMs for each potential combination of Machiavellianism, psychopathy, narcissistic rivalry, and narcissistic admiration. Using bivariate LGCMs instead of multivariate LGCMs constitutes a deviation from the preregistration. A multivariate LGCM including all four traits simultaneously would have required even more computing power than the already complex bivariate models. As preregistered, we examined Hypotheses 5 and 6 with bivariate LGCMs with Machiavellianism, psychopathy, narcissistic rivalry, or narcissistic admiration on the one hand and self-esteem or life satisfaction on the other hand. In all models, we freely estimated all covariances between the latent intercept variables and all random time slopes that remained in the final model for each trait. We also estimated all covariances between the occasion-specific state disturbances of different traits at each measurement occasion (Schermelleh-Engel et al., 2004) to allow for linked fluctuations around the growth curve. The parameter of interest for testing Hypotheses 4-6 was the correlation between the random linear time slopes for each pair of traits under investigation. We used classic one-tailed significance tests of r against 0. In case that the final model for a particular trait did not include a random linear time slope, the respective hypothesis was automatically rejected. We also reported correlations between random quadratic time slopes in case that the final models for both traits involved in a pairwise comparison included such slopes.

To account for multiple testing, we controlled the false discovery rate (FDR) using the Benjamini-Hochberg procedure (Benjamini & Hochberg, 1995). Controlling the FDR is one way to deal with the problem of multiplicity which occurs, amongst others, when multiple outcomes (e.g., Machiavellianism, psychopathy, and narcissistic rivalry) are assessed to test the same hypothesis (e.g., that maturation occurs in young adults over 2 years). Controlling the FDR means that the expected proportion of false positives among all rejected hypotheses in an experiment or in a family of tests is smaller than or equal to a desired level which we choose to be .05. We controlled the FDR separately for each family of tests. We define a ‘family’ of tests as the repeated test of the same hypothesis involving different traits. We considered H1a-d, H2a-c, H3a-d, H4a-c, and H6a-c to be separate families of tests. For testing H2d, H5, and H6d, we used the conventional alpha-level of .05.

We estimated the statistical power of the present study to detect true mean-level change of different sizes. Monte Carlo simulations (L. K. Muthén & Muthén, 2002) with 1000 replications and realistic patterns of missing values were performed. The estimated parameters from the final model for each trait served as population values. In the final model, we had to collapse response categories so that each response category contained at least 2% of the responses. This prevented the generation of simulated datasets with varying amounts of empty response categories. We specified true effects of different sizes by changing the population value for the linear slope. These analyses were conducted post-hoc because we perceived a lack of reliable information about important population parameters. Those parameters include the variance of the linear time slope as well as the variance of the occasion-specific state disturbances for latent growth curve models with measurement models and categorical indicators. An informative overview of growth curve parameter estimates from previous longitudinal studies by Rast and Hofer (2014) only applies to growth curve studies using multilevel designs without specifying measurement models. Those models do not distinguish measurement error from occasion-specific state disturbances.

We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study. Sample size was determined by practical constraints. No analyses pertaining to the research questions in this paper were performed before ending data collection and before publishing the preregistration (https://osf.io/98wqf/). We report all data exclusions in the preregistration and in the methods section. The preregistration describes a detailed analysis plan, all measures that were used for this study, all hypotheses, and all corresponding tests. The analysis code (input and output) is available at https://osf.io/p8f67/. A reduced data set with age groups and item-level data for all measures of Machiavellianism, psychopathy, and narcissism is available at https://osf.io/j5ub7/. We can only provide a reduced data set because of confidentiality concerns (i.e., especially older participants in our sample might be re-identifiable in the full data set). Post-hoc power analyses were conducted using Monte Carlo simulations.

We will first report the results of our preliminary analyses on measurement invariance and then the results of our hypothesis tests on rank-order stability, mean-level development, and correlated change. All coefficients are unstandardized unless otherwise specified. For all hypothesis tests with FDR control, Table S2 comprises the required information in aggregate. In summary, all p-values < .05 were significant after controlling the false discovery rate.

For each trait at each wave, Table S3 reports indices of model fit for unidimensional measurement models with correlated residuals. When testing strict measurement invariance over time for each trait, the expected parameter changes for several thresholds exceeded our criterion for moderate to strong noninvariance (0.375). These thresholds are listed in Table S4. None of the expected parameter changes for factor loadings exceeded our noninvariance criterion (0.25). In the final partial measurement invariance models, one threshold was freed for Machiavellianism and narcissistic rivalry, two thresholds were freed for psychopathy, three thresholds were freed for narcissistic admiration, and four thresholds were freed for life satisfaction. No thresholds had to be freed for self-esteem.

In Hypotheses 1a to 1d, we expected that all traits would show high rank-order stability over two years in young adulthood. All of these hypotheses were confirmed. Figure 3 displays the correlations between measurements of each trait at T1 and T4 (after controlling for age_T1). The correlations ranged from .74 to .81. The lower bounds of the one-sided 95% CIs ranged from .69 to .76 and thus lay substantially above the preregistered threshold of .30. All estimated coefficients (including stabilities of self-esteem and life satisfaction) are reported in Table S5.

Means with two-sided 95% CIs (from the previous model that had been used to examine rank-order stability) are displayed in Figure 4 and reported in Table S6 for descriptive purposes. In order to test Hypotheses 2 and 3, we identified a final model that best approximated the pattern of change observed in the data for each trait separately. All estimated coefficients relevant to the development of Machiavellianism, psychopathy, narcissistic rivalry, and narcissistic admiration are reported in Table 1.

Model Building. Change in Machiavellianism was best described by a model with a random linear time slope which is highlighted in Figure 5. Table 1 reports all relevant coefficients estimated during the model building process as described in methods and preregistration. Figure 5 displays the respective model-implied growth curves. In the model with a random linear time slope, the disturbance at T4 was estimated at −0.074. As variances cannot have negative values, we fixed it at 0. This had not been planned in the preregistration. Fit indices for the final model were χ² (df = 3322, n = 926) = 5443.35, p < .001, RMSEA = .026, CFI = .90, SRMR = .071.

Hypothesis 2a. According to Hypothesis 2a, we expected mean levels to decrease over two years in young adults. Machiavellianism indeed decreased linearly with an unstandardized slope of α_LinMac = −0.095 (SE = 0.026) per year (see Table 1). Change over two years amounted to an estimated d_T1T4 = −0.179, one-sided 95% CI (−∞, −0.098], p < .001 (see Figure 9). Hypothesis 2a was confirmed.

Hypothesis 3a. According to Hypothesis 3a, we expected that young adults would differ in their development over two years. The variance around the linear time slope was estimated at ψ_{LinMac, LinMac} = 0.155 (SE = 0.022). Model comparisons indicated that a model without a random effect around the linear time slope fit significantly worse than a model with it, χ² (df = 1) = 27.10, p < .001. Hypothesis 3a was confirmed.

Cross-Sectional Differences in Age. In the final model, individual differences in age_T1 had a positive (unstandardized) relationship of β = 0.054 (SE = 0.015) with the random intercept variable. This means that model-implied starting points for the (declining) growth curves were 0.65 SD_IntMac higher for 30-year-olds compared to 18-year-olds. It is evident from the highlighted model in Figure 5 that cross-sectional effects of being older did not converge with longitudinal effects of getting older. The respective contextual effect was 0.148 (SE = 0.029), p < .001.

Model Building. Change in psychopathy was best described by a model with a fixed linear time slope which is highlighted in Figure 6. Estimated coefficients are reported in Table 1 and model-implied growth curves and displayed in Figure 6. Fit indices for the final model were χ² (df = 6822, n = 926) = 8525.27, p < .001, RMSEA = .016, CFI = .89, SRMR = .086.

Hypothesis 2b. According to Hypothesis 2b, we expected mean levels to decrease over two years in young adults. Psychopathy indeed decreased linearly with an unstandardized slope of α_LinPsy = −0.066 (SE = 0.028) per year. Change over two years amounted to an estimated d_T1T4 = −0.116, one-sided 95% CI (−∞, −0.037], p = .008. Hypothesis 2b was confirmed.

Hypothesis 3b. According to Hypothesis 3b, we expected that young adults would differ in their development over two years. The variance around the linear time slope was estimated at ψ_{LinPsy,LinPsy} = 0.046 (SE = 0.037). Model comparisons indicated that a model without a random effect around the linear time slope did not fit significantly worse than a model with it, χ² (df = 1) = 0.73, p = .39. Hypothesis 3b was not confirmed.

Cross-Sectional Differences in Age. In the final model, cross-sectional differences in age_T1 had a positive relationship of β = 0.067 (SE = 0.015) with the random intercept variable. Figure 6 illustrates that cross-sectional effects of being older did not converge with longitudinal effects of getting older. The respective contextual effect was 0.133 (SE = 0.032), p < .001.

Model Building. Change in narcissistic rivalry was best described by a model with a fixed quadratic time slope which is highlighted in Figure 7. Estimated coefficients are reported in Table 1 and model-implied growth curves are displayed in Figure 7. The final model fit the data relatively well, χ² (df = 722, n = 926) = 1201.37, p < .001, RMSEA = .027, CFI = .97, SRMR = .071.

Hypothesis 2c. According to Hypothesis 2c, we expected mean levels to decrease over two years in young adults. The model-implied growth curves had an inverted u-shaped trajectory without an overall decrease in the investigated time frame. More specifically, growth was best approximated as a combination of a positive fixed linear time slope of α_LinRiv@age22 = 0.327 (SE = 0.099) per year and a negative fixed quadratic time slope of α_QuadRiv = −0.138 (SE = 0.046). Taken together, change over two years amounted to an estimated d_T1T4@age22 = 0.090, one-sided 95% CI (−∞, 0.191], p = 1. Hypothesis 2c was not confirmed.

Hypothesis 3c. We expected that young adults would differ in their development over two years. The disturbance of the linear time slope was estimated at ψ_{LinRiv, LinRiv} = 0.059 (SE = 0.031). Model comparisons indicated that a model without a random effect on the linear time slope did not fit significantly worse than a model with it, χ² (df = 1) = 0.99, p = .32. Hypothesis 3c was not confirmed.

Cross-Sectional Differences in Age. The final model contained various relationships with age_T1 which are reported in Table 1. The age-dependent model-implied range was α_LinRiv@age18 = 0.290 to α_LinRiv@age30 = 0.400 for the linear time slope and d_T1T4@age18 = 0.025 to d_T1T4@age30 = 0.221 for change over two years. Figure 7 illustrates that cross-sectional effects of being older did not converge with longitudinal effects of getting older. The respective contextual effects were −0.290 (linear cohort → intercept, SE = 0.100), −0.150 (quadratic cohort → intercept, SE = 0.047), and 0.285 (linear cohort → linear time slope, SE = 0.091), all ps ≤ .004.

Model Building. Change in narcissistic admiration was best described by a model with a fixed linear time slope which is highlighted in Figure 8. Estimated coefficients are reported in Table 1 and model-implied growth curves are displayed in Figure 8. In the model with a random linear time slope, the disturbance at T4 was estimated at −0.014. As variances cannot have negative values, we fixed it at 0. This had not been planned in the preregistration. The final model fit the data well, χ² (df = 694, n = 926) = 1161.39, p < .001, RMSEA = .027, CFI = .97, SRMR = .050.

Hypothesis 2d. According to Hypothesis 2d, we expected that mean levels would not change over two years in young adults. The linear time slope was estimated at α_LinAdm = −0.022 (SE = 0.024) per year. Change over two years amounted to an estimated d_T1T4 = −0.038. As stated in methods and preregistration, we chose symmetric equivalence bounds of ±0.075 for the two one-sided tests (TOST) procedure. In line with Hypothesis 2d, the test against the upper bound (0.075) was significant, d_{T1T4vsUpperBound} = −0.113, p = .004. However, the test against the lower bound (−0.075) was not significant, d_{T1T4vsLowerBound} = 0.037, p = .20. The latter test indicates that we cannot rule out that narcissistic admiration decreased even more than our lower bound. For this reason, Hypothesis 2d was not confirmed.

Hypothesis 3d. We expected that young adults would differ in their development over two years. The variance around the linear time slope was estimated at ψ_{LinAdm, LinAdm} = 0.079 (SE = 0.015). Model comparisons indicated that a model without a random effect around the linear time slope did not fit significantly worse than a model with it, χ² (df = 1) = 0.87, p = .35. Hypothesis 3d was not confirmed.

Cross-Sectional Differences in Age. In the final model, cross-sectional differences in age_T1 had a nonsignificant negative relationship of β = −0.006 (SE = 0.015) with the random intercept variable. Figure 8 suggests that cross-sectional effects of being older may converge with longitudinal effects of getting older. In fact, the contextual effect of 0.016 (SE = 0.028) was not significant, p = .57. As described in methods and preregistration, we then constrained the contextual effect to be 0, effectively conducting an accelerated longitudinal analysis. Cohen’s d for the model-implied mean-level development over the entire age span of 14 years (i.e., from 18 to 32) was d_18-32 = −0.207 (SE = 0.216), p = .34.

We conducted the same analyses for the outcomes self-esteem and life satisfaction on an exploratory basis as described in the preregistration. Table S7 reports all estimated coefficients relevant to their development.

Self-Esteem. Change in self-esteem was best approximated by a model with a fixed linear time slope of α_LinSE = 0.056 (SE = 0.032) per year. Change over 1.5 years amounted to an estimated d_T2T4 = 0.079 (SE = 0.044). The variance around the linear time slope was estimated at ψ_{LinSE, LinSE} = 0.101 (SE = 0.073). Model comparisons indicated that a model without a random effect around the linear time slope did not fit significantly worse than a model with it, χ² (df = 1) = 0.31, p = .58. The final model fit the data relatively well, χ² (df = 460, n = 719) = 842.69, p < .001, RMSEA = .034, CFI = .98, SRMR = .051. The contextual effect of −0.057 (SE = 0.035) was not significant, p = .11. In the accelerated longitudinal analysis, Cohen’s d for the model-implied mean-level development over the covered age span of 13.5 years was d_18.5-32 = 0.322 (SE = 0.208), p = .12.

Life Satisfaction. Development in life satisfaction was best described by a model with a random linear time slope, α_LinLS = 0.038 (SE = 0.027) per year. Change over two years was estimated at d_T1T4 = 0.073 (SE = 0.051). The variance around the linear time slope was estimated at ψ_{LinLS, LinLS} = 0.158 (SE = 0.037). Model comparisons indicated that a model without a random effect around the linear time slope fit significantly worse than a model with it, χ² (df = 1) = 15.81, p < .001. The final model fit the data well, χ² (df = 237, n = 926) = 313.74, p < .001, RMSEA = .019, CFI = 1.00, SRMR = .037. The contextual effect of −0.114 (SE = 0.030) was significant, p < .001.

Testing hypotheses concerning correlated change requires the presence of individual differences in the linear time slopes (random effects, see Hypothesis 3) of both traits involved in a test because constant values (fixed effects) cannot covary with each other. As individual differences in change could not be confirmed for psychopathy, narcissistic rivalry, narcissistic admiration, and self-esteem, none of the hypotheses involving at least one of these traits could be tested and all of them were rejected automatically as described in methods and preregistration. For descriptive purposes, Table S8 reports covariances between linear time slopes for each combination of two traits from models with random linear time slopes (regardless of rejection of Hypothesis 3).

According to Hypothesis 4, we expected that changes in Machiavellianism, psychopathy, and narcissistic rivalry would be positively associated with each other over two years in young adults. As individual differences in linear time slopes could not be confirmed for psychopathy and narcissistic rivalry, none of these hypotheses could be tested and all of them were rejected automatically.

In Hypothesis 5, we expected changes in narcissistic admiration and self-esteem to be positively associated with each other. Again, this hypothesis had to be rejected automatically as the existence of a random effect around the linear time slope could be confirmed neither for narcissistic admiration nor for self-esteem.

Hypothesis 6a. Hypothesis 6a posits that changes in Machiavellianism would be negatively associated with changes in life satisfaction. The random linear time slopes of Machiavellianism and life satisfaction were indeed negatively related with each other, ψ_{LinMac, LinLS} = −0.059 (SE = 0.016), r = −.39 (SE = .11), 95% CI (−∞, −.21], p < .001 (see Figure 10). Hypothesis 6a was confirmed. Fit indices for the bivariate LGCM were χ² (df = 5152, n = 926) = 7421.35, p < .001, RMSEA = .022, CFI = .91, SRMR = .066.

Hypotheses 6b and 6c. According to Hypotheses 6b/6c, we expected changes in psychopathy/narcissistic rivalry to be negatively associated with changes in life satisfaction. These hypotheses could not be tested (see above) and were automatically rejected.

Hypothesis 6d. In Hypothesis 6d, we expected that changes in narcissistic admiration would be positively related to changes in life satisfaction. This hypothesis could also not be tested (see above) and was automatically rejected as well.

For Machiavellianism, the statistical power to detect mean-level change of different sizes according to H2a was 0.64 (for a true effect of d = −0.10), 0.92 (for d = −0.15), and 0.98 (for d = −0.20). Thus, the power for detecting mean-level change of at least d = −0.15 was very good.

For narcissistic admiration, the statistical power of detecting the absence of mean-level change of |d| ≥ 0.075 according to H2d was 0.12 (for a true effect of d = 0), 0.10 (for d = 0.025), and 0.06 (for d = 0.05). Figure 9 illustrates that the equivalence bounds fall very close together in comparison to the wide range of the confidence interval for the two one-sided tests. Setting close equivalence bounds of d = ±0.075 yielded very low power to detect the absence of meaningful mean-level change as posited in H2d, even in the best-case scenario of total absence of any mean-level change.

Simulations for psychopathy and narcissistic rivalry could not be conducted due to estimation problems. Considering the similarity with Machiavellianism, it is likely that the statistical power was approximately in the same range for psychopathy and narcissistic rivalry.

In addition to the reported results, we replicated all analyses with data from secondary measures of Machiavellianism, psychopathy, and narcissism. Table 2 presents the main findings concerning rank-order stability, mean-level development, and individual differences in change.

Correlated change between Machiavellianism, psychopathy, and narcissism as well as correlated change with self-esteem could not be examined as individual differences in change were not present for any of these traits except Machiavellianism. Correlated change between Machiavellianism and life satisfaction was not significant when Machiavellianism was measured with the Short Dark Triad (r = −.35, SE = .26) or the M7 (r = −.33, SE = .20).

More detailed information is available in the supplementary information. We report fit indices for measurement models (Table S9), an examination of measurement invariance (Table S10), rank-order stabilities (Table S11), and saturated means (Table S12). Table S13 provides all coefficients estimated during the model building stage.

The present study contributes towards a better understanding of young adults’ development on Machiavellianism, psychopathy, and narcissism. It documents mean-level decreases in Machiavellianism and psychopathy that are in line with the maturity principle. Inversely correlated change between Machiavellianism and life satisfaction suggests that decreases in Machiavellianism can be considered adaptive. Results involving narcissism were not conclusive.

It is not clear yet at what age Machiavellianism starts to decrease. This study found evidence for a linear decrease over two years that occurred already in 18-year-olds. If one were to project the effect size of d = −0.18 linearly to a longer period of time of, for example, ten years, it would quickly become quite substantial (95% CI for d_10years [−1.38, −0.41]). Model comparisons revealed systematic individual differences in the steepness or direction of participants’ developmental trajectories. This suggests that maturation in Machiavellianism could be an individual process that occurs only in some individuals and not in others (in the investigated time span).

Inversely correlated slopes between Machiavellianism and life satisfaction suggest that maturation in terms of decreasing Machiavellianism is compatible with a more beneficial development of life satisfaction. Young adults who did not report decreasing levels of Machiavellianism over time had a lower likelihood of stable or increasing life satisfaction. The social investment principle could be perceived as a plausible, yet untested explanation for this phenomenon (Roberts et al., 2005). As young adults are assumed to be confronted with life tasks associated with education, employment, and social relationships, they might be incentivized to cultivate desirable aspects of their personality. A decrease in Machiavellianism—reflected in higher levels of honesty and increased trust in others—appears to be compatible with age-graded tasks and with leading a satisfying life. Correlated change between Machiavellianism and life satisfaction had not been documented before.

Until now, no longitudinal research that we are aware of has focused on the development of psychopathy in young adulthood. The current study provides tentative evidence for a linear decrease in psychopathy over two years starting as early as in 18-year-olds. We found no evidence for systematic individual differences in participants’ developmental path. We conclude that it is more likely than not that the maturity principle applies to psychopathy in young adulthood.

Currently available findings concerning the development of narcissism vary substantially. It is not clear yet until what age around adolescence narcissism increases (Fang et al., 2021; Klimstra et al., 2020), when it remains mostly unchanged (e.g., Grosz, Göllner, et al., 2019), and at what age around young adulthood it starts to decrease.

In our sample, mean-level development of narcissistic rivalry was best approximated by an inverted U-shaped function. Such a trajectory would imply that narcissistic rivalry peaks in young adulthood. A purely speculative post-hoc explanation could be that establishing oneself in a new environment at university or work confronts individuals with new people that might be perceived as competitors. This may pose a temporary ego threat to some individuals and may lead to a self-defensive attitude at first which fades away over time. Please note that this finding might not be reliable or robust and our post-hoc explanation would not fit for those older participants in our sample who were already at university for some time.

With regard to narcissistic admiration, we did not expect any substantial mean-level change in young adulthood. Even if social investment affects the development of narcissistic admiration, some aspects of narcissistic admiration might be beneficial for the fulfillment of age-graded social roles whereas other aspects might be detrimental. The adaptiveness of narcissistic admiration may also depend on random variation in the circumstances of a given situation or on completely different factors. The present study found no conclusive evidence whether the maturity principle does or does not apply to narcissistic admiration in young adulthood.

Beyond the focus on particular traits, this study contributes to a broader understanding of maturation in young adulthood. Our results emphasize that maturation applies to antagonistic traits as well. Our final models for each trait differ with regard to the approximate timing of mean-level decreases. Whereas our final models for some traits (Machiavellianism and psychopathy) implied a linear decrease that started already at the age of 18 or earlier, our final model for another trait (narcissistic rivalry) implied an inverted U-shaped pattern. This means that an initial increase in mean levels turned into a decrease only during the time period covered by the current study. In general, curvilinear trajectories are not uncommon. According to the disruption hypothesis, some traits develop in different directions in adolescence versus young adulthood (Brandes et al., 2021; Soto, 2016). Our analytical approach allowed such differences between traits with regard to the timing of mean-level development to emerge.

The results for Machiavellianism raise the possibility that the extent of maturation might not only vary between traits, but also between individuals. While some participants in our sample became more mature (in terms of decreasing Machiavellianism), others developed in the opposite direction. The effect size of these individual differences in the linear time slope was substantial: A variance of 0.155 (SD = 0.39) around the linear time slope implies a range for individual change in Machiavellianism over two years from d = −0.92 to d = 0.56 for one standard deviation around the average change.

These individual differences in maturation with regard to Machiavellianism likely have important consequences. The present study found negatively correlated change between Machiavellianism and life satisfaction with a practically significant effect size of r = −0.39. Our model implies an additional increase in life satisfaction over two years of d = 0.28 associated with being one standard deviation below the average on the linear time slope for Machiavellianism. This finding suggests that maturation in terms of decreasing Machiavellianism is part of a desirable developmental pattern.

The desirability of maturation in terms of decreasing Machiavellianism and psychopathy in young adulthood seems clear. For an individual, more mature socioemotional personality characteristics should lead to better social relationships (Miao et al., 2019; Muris et al., 2017; Vize et al., 2018) and, in turn, to an increased quality of life. From a societal perspective, reduced aggression and delinquency should especially benefit those people (Muris et al., 2017) and organizations (Ellen et al., 2021) who are negatively affected by the behavior of less mature individuals.

Based on our findings, we can derive relevant questions for future research investigating the mechanisms of maturation. First, what could be the specific mechanism underlying the decrease in Machiavellianism and psychopathy? Second, what factors are driving individual differences in maturation? And third, why do Machiavellianism and life satisfaction partially develop in tandem?

If one decides to draw on the social investment principle, it provides a very general answer to the first question—namely that young adults invest effort into fulfilling normative social roles and that these investments are targeted towards specific behaviors that are part of relevant aspects of personality. But this answer lacks a lot of detail. What are the exact challenges in young adulthood and which of them are most critical for maturation? And what are the corresponding aspects of personality on which mastering these challenges depends? Investigating these questions could start with cross-sectional research on identifying the most important challenges of young adulthood. Short-term longitudinal studies (over several weeks or months) could then be conducted to examine whether experiencing or overcoming any of those challenges is associated with change in relevant aspects of personality.

Second, a good understanding of the factors underlying individual differences in maturation can point to effective starting points for interventions. From the perspective of the social investment principle, one might examine individual differences in the amount of effort that young adults invest into mastering social roles. In case that individual differences in social investment can be identified as underlying maturation, what causes them? Relevant variables could include individual differences in exploration behavior, emotion regulation, metacognition, and learning from mistakes (Keith & Frese, 2005; Keith & Wolff, 2015). Individual differences could also be rooted in young adults’ environments, such as in the focus of a particular degree program (Krishnan, 2008). After identification of relevant factors, experimental studies could attempt to manipulate them in order to examine their role for maturation.

Third, coupled development between life satisfaction and Machiavellianism suggests that maturation may drive or sustain happiness in the long term. This could imply that the mastery of age-related challenges is related to the experience of higher positive affect and/or lower negative affect as compared to failure in responding adequately to those challenges. However, some choices—such as doing a demanding internship—may lower satisfaction in the short term while increasing it in the long run. Future research should include multiple outcomes to be able to examine potentially diverging consequences on different time frames.

The present study focused on the broad personality traits Machiavellianism, psychopathy, and narcissism as these traits have significant historical importance in the literature on antagonistic personality traits. However, and as discussed in the introduction, their conceptualizations reflect several aspects per trait (i.e., manipulativeness, callousness, impulsivity, risk-seeking, etc.). In future research, it might be helpful to employ pure measures of these aspects (called “themes” in Bader et al., 2021) and to control for a global tendency towards antagonism using bifactor models. In that way, the interplay between life challenges and narrow aspects of personality can be examined in more detail and with higher specificity.

Finally, genetically informative study designs are required to estimate the proportion of individual differences in maturation that can be attributed to genetic factors. A moderate proportion of variance due to genes would emphasize the potential of individuals and/or societies to influence the presence and the extent of maturation. Previous research suggests that only a part of maturation can be attributed to genetic factors (Briley & Tucker-Drob, 2014; McGue et al., 1993; Roberts & Nickel, 2021).

With four waves of measurement and a preregistered model-building procedure, the present study was able to identify a best-fitting growth function separately for each of the four traits. This allowed us to attempt to detect nonlinear change, which can happen when an initial increase later turns into a decrease. Another strength of the present study was its ability to expose potential uniqueness in the development of the investigated traits. All traits were assessed with full-length measurement instruments.

A final strength of the current study might be seen in our effort to conduct state-of-the-art data analysis. More specifically, we specified measurement models for each trait in which we accounted for the categorical response format of all indicators, and we tested and specified longitudinal measurement invariance. We conducted model comparisons instead of Wald tests to examine the presence of a random effect of the linear time slope and used equivalence testing for a “no change” hypothesis. In case of multiple comparisons, we controlled the rate of false discoveries. Monte Carlo simulation studies were used to estimate the statistical power for detecting mean-level change or the absence of it. Lastly, we provide scripts and data for our study.

The present study has several limitations. The first limitation concerns the homogenous composition of our sample. It consisted almost entirely of students (approx. 94% students at T2, see Table S1), almost half of which studied psychology. The majority (74%) of the participants identified as female and almost all participants (88%) were between 18 and 25 years old. On top of that, men (and individuals high in psychopathy) were more likely to drop out of the study. As a consequence, we cannot generalize our findings to all students and certainly not to the entire population of young adults in Germany.

Second, the duration of our study was limited to two years. Length of study is an important factor influencing growth curve reliability (Rast & Hofer, 2014). If traits develop linearly, longer study durations imply larger accumulated effect sizes which could accentuate individual differences in trait development. This limitation (together with the homogenous sample) has most likely reduced the statistical power to detect reliable individual differences in the developmental paths for psychopathy, narcissistic rivalry, and narcissistic admiration (H3b-d) as well as correlated change (H4-H6).

Third, the present research documented substantial contextual effects for Machiavellianism, psychopathy, and narcissistic rivalry. More specifically, the final models imply that older participants entered the study with higher levels on all of these traits. In the case of narcissistic rivalry, the final model also implies that older participants were on developmental paths with a more positive linear slope than those of younger participants. We did not gather enough information to be able to examine potential explanations for these contextual effects. On the one hand, these contextual effects may reflect a cohort effect in the sense that older students differ somehow from younger students. Older students may represent a population of individuals that has already worked for a few years before they started studying (and maybe some of them study part-time). On the other hand, these contextual effects may reflect a selection effect in the sense that those students who are still studying at a higher age may be those who had more difficulties with making progress in their studies or who were confronted with external obstacles that prevented them from quickly completing their studies.

The ability to generalize our results is limited by several aspects of study design that negatively affect statistical power—most importantly, sample size, attrition, and study duration. Due to the complexity of our analyses and the interplay of the various parameters involved, we had difficulties estimating statistical power for detecting a random effect around the linear time slopes (H3) and, depending on that, statistical power for detecting correlated change (H4-H6).

We have full near confidence in our results pertaining to H1 because these results were only based on correlations and the corresponding Wald-test statistics (values range from 12.76 to 16.24) yielded high values. Based on the reported power simulation results, we have high confidence in the reliability of the decrease in Machiavellianism (H2a). Based on the Wald-test statistic (6.96) for the variance around the linear time slope and the χ²-test statistic for the corresponding model comparison (27.10, df = 1), we also have high confidence in our results concerning individual differences in change in Machiavellianism, as well as their covariance with individual differences in change in life satisfaction (Wald test = −3.72). In summary, we think that all hypothesis tests involving just Machiavellianism (and life satisfaction) appear quite reliable—i.e., that Machiavellianism decreases, that individuals differ in their development, and that change in Machiavellianism correlates inversely with change in life satisfaction.

In contrast, our evidence involving narcissistic rivalry and narcissistic admiration appears very weak. We had extremely low power for detecting the absence of change in narcissistic admiration (H2d). We found an inverted U-shaped developmental trajectory for narcissistic rivalry with a Wald-test statistic of −2.77. However, in combination with a positive linear slope (Wald test = 3.02) and substantial cohort effects, this finding might be specific to this sample. The evidence involving psychopathy is modest for H2b (Wald test = −2.40) and very weak for all other hypotheses. Overall, we cannot distinguish whether we rejected H3 to H6 due to a lack of statistical power or due to the absence of the expected effects/non-effects. In models with random linear time slopes (despite rejection of H3), we found substantial covariances between most pairs of traits (Table S8) but we refrain from interpreting them due to the rejection of H3.

Besides statistical power, another important constraint on generality is imposed by the high proportion of women in our sample (74% of the entire sample; 75% at T1, and slightly more at T2-T4 due to selective attrition). Gender is a non-trivial factor in antisocial behavior and its development. Average levels of antagonistic traits such as Machiavellianism (Götz et al., 2020), psychopathy (Tuvblad et al., 2016), and narcissism (Chopik & Grimm, 2019; Weidmann et al., 2023; Wetzel et al., 2020) have consistently been found to be higher in men than in women (Hartung et al., 2022). In contrast, gender differences in the patterns of mean-level change are much less evident and harder to detect. Some studies suggest that developmental paths run mostly parallel throughout adulthood (e.g., Hartung et al., 2022), with a peak around the same age (in levels of Machiavellianism in adolescence; Götz et al., 2020). Other studies suggest that women show a more desirable developmental pattern, either in terms of lower increases (in callous affect; Klimstra et al., 2020), earlier declines (in willfulness; Chopik & Grimm, 2019), or steeper declines (in Machiavellianism and narcissism, respectively; Götz et al., 2020; Zettler et al., 2021).

In light of this evidence, we see two ways how the gender imbalance in our sample may have affected our results: First, assuming that gender differences in developmental paths are negligible, the bias due to our imbalanced sample composition should be minimal. Alternatively, assuming that women, on average, develop towards maturity more consistently than men, the estimated effect sizes for mean-level decreases (H2a-c) should be inflated by some degree. In a sample with a balance of men and women, mean-level decreases could be expected to be weaker.

On average, young adult students appear to become more mature over time in terms of decreasing Machiavellianism and decreasing psychopathy. Individuals differ substantially from each other with regard to the presence or extent of maturation in Machiavellianism. Decreasing Machiavellianism coincides with stable or increasing life satisfaction and appears desirable.

The development of narcissism seems to be more complex. It is possible that narcissistic rivalry changes directions in young adults who are mostly female students—an initial increase could be followed by a later decrease. While the average developmental path on narcissistic admiration over two years appears to be relatively flat, our study design was unable to produce conclusive evidence.

The present study’s results are somewhat in line with expectations based on the maturity principle. Inversely correlated change between Machiavellianism and life satisfaction contributes new information in support of the adaptiveness of maturation. The exact timing of mean-level change, its underlying mechanisms, as well as the factors driving it remain areas for future research.

This article was presented as a talk at the 16th meeting of the section Differential Psychology and Psychological Assessment of the German Psychological Society (Ulm, September 12-15, 2021) and at the 52nd Congress of the German Psychological Society (Hildesheim, September 10-15, 2022).

Contributed to conception and design: EW; contributed to acquisition of data: EW; contributed to analysis and interpretation of data: CW, EW; drafted and/or revised the article: CW, EW; approved the submitted version for publication: CW, EW.

This research was supported by grants from the German Research Foundation (WE 5586/2-1) and the Elite Program for Postdocs of the Baden-Württemberg Stiftung and the Young Scholar Fund of the University of Konstanz to EW. EW was also supported by the research-training group Statistical Modeling in Psychology (GRK 2277) funded by the German Research Foundation. We acknowledge support by the Open Access Publication Fund of Magdeburg University.

The authors declare that no competing interests exist. Eunike Wetzel is an associate editor at Collabra: Psychology. She was not involved in the review process of this article.

An extensive preregistration (https://osf.io/98wqf/), the analysis scripts (https://osf.io/p8f67/), and a reduced data set with age groups and item-level data for all measures of Machiavellianism, psychopathy, and narcissism (https://osf.io/j5ub7/) are available on the project page (https://osf.io/5947v/) on the Open Science Framework. We can only provide a reduced data set because of confidentiality concerns (i.e., especially older participants in our sample might be re-identifiable in the full data set).