Network Analysis of Triarchic Psychopathy Measure (TriPM) and Other Self Report Measures of Psychopathy

ABSTRACT

Introduction

Network analysis is a statistical tool for understanding interrelations between symptoms and the relative importance of symptoms in a disorder. This is especially appealing to psychopathy researchers, as network analysis may provide some insights that bring the field closer to resolving debates about psychopathy‐relevant features.

Methods

In the present study (N = 612), we first conducted network analysis on the Triarchic Psychopathy Measure (TriPM) and subsequently conducted a scale‐level analysis of psychopathic traits from seven psychopathy inventories.

Results

In the TriPM network, we found that items from Meanness and Disinhibition were among the most central. Meanness items assessing deficits in empathy were highly influential nodes in the network, but also demonstrated significant topological overlap. Boldness items were less central but still important to the TriPM network. Scales indexing impulse control problems, affective deficits, and interpersonal dominance were among the most influential in the domain‐level network. This included measures of callousness, impulsiveness, and need for stimulation, and grandiosity and manipulativeness.

Conclusion

Despite concerns about the application of network analysis to cross‐sectional data, our results support the relevance of boldness, meanness, and disinhibition to the conceptualization and measurement of psychopathy.

1. Introduction

After close to 100 years of psychopathy research, there is little agreement about the most important characteristic(s) of psychopathy, although some consistent themes have emerged across the literature (Skeem et al. 2011). Affective deficits (i.e., lack of empathy/callousness, impaired social relatedness) and disconstraint (i.e., irresponsibility, impulsive behavior, and recklessness) are widely accepted as core psychopathic traits and are measured to varying degrees by all existing psychopathy instruments (Drislane et al. 2014). Fearlessness (i.e., low harm avoidance, anxiety), by contrast, often arises as a key correlate of psychopathy in empirical studies across multiple paradigms (Lykken 1957; Patrick et al. 1993) but receives mixed support as a core feature. Some self‐report measures of psychopathy explicitly measure tendencies toward fearlessness or calmness under conditions that would otherwise provoke anxiety, such as the Fearless Dominance (FD; Benning et al. 2003) factor of the Psychopathic Personality Inventory (PPI; Lilienfeld and Andrews 1996) and the Boldness scale of the Triarchic Psychopathy Measure (TriPM; Patrick 2010). By contrast, other psychopathy measures place less emphasis on fearlessness, such as the PCL‐R (Hare 2003), which only indirectly assesses lack of anxiety on the interpersonal facet. The relevance of fearlessness or boldness to psychopathy is also debated. Meta‐analytic studies (Miller and Lynam 2012; Sleep et al. 2019) have found that PPI‐FD and TriPM Boldness exhibit small correlations with antisocial behavior or other externalizing problems along with strong positive associations with adaptive functioning. Thus, some authors have concluded that boldness is not a necessary condition for psychopathy and is better conceptualized as a diagnostic specifier (Lynam and Miller 2012).

Another salient point of contention among psychopathy researchers is whether antisocial behavior is a core psychopathic trait or incidental correlate. Structural studies of the PCL‐R have concluded that antisocial behavior is an important and potentially central feature of psychopathy (Hare and Neumann 2005). The PCL‐R is a consistent predictor of violent and non‐violent offenses both within an institution and after release (Leistico et al. 2008), thus supporting the notion that psychopathy is fundamentally intertangled with antisocial behavior. Some researchers (e.g., Cooke and Michie 2001; Skeem and Cooke 2010) have argued that the PCL‐R overemphasizes the role of antisocial behavior relative to affective or interpersonal traits. The content of the PCL‐R, which contains several items that expressly reference overt criminal behavior, also makes it inappropriate for use in non‐forensic or correctional settings, which limits investigation into non‐criminal forms of psychopathy (e.g., subclinical or corporate psychopathy, “successful psychopathy;” Benning et al. 2018). For this reason, many measures of psychopathy (e.g., PPI, TriPM) focus on dimensional personality constructs rather than behavioral indicators of criminality. Nonetheless, there is no current consensus as to whether criminal behaviors are central to psychopathy itself or are a manifested outcome of a specific trait constellation (Hare and Neumann 2010; Skeem and Cooke 2010).

Discrepancies between psychopathy measures, as well as their respective representation of various psychopathic traits, inevitably affect the conceptualization of psychopathy as a construct. The prevailing statistical method used to uncover the structure of psychopathy is the common factor model, such that symptoms of psychopathy are assumed to be manifestations of latent trait dimensions (Markus and Borsboom 2013). Latent traits then help to explain the co‐occurrence of certain symptoms within an individual, and extreme high or low standing on one or more traits may influence maladaptive behavioral patterns. However, lumping symptoms within a trait may disregard symptom‐specific etiologies, relative impacts on functioning, and causal interactions between symptoms (e.g., Fried and Nesse 2015). As an alternative to the hierarchical method, network analysis emphasizes the nature of particular symptoms and assumes them to be psychologically meaningful and distinct, while interactions between symptoms could reveal causal pathways to psychopathology (Borsboom and Cramer 2013). Certain symptoms can also exert disproportionate causal influence on other symptoms and serve as central, unifying factors, helping to explain observed correlations among other symptoms. Network analyses have demonstrated utility in understanding the multivariate dependencies within personality constructs at a more granular level (Constantini et al. 2017). This approach offers a complement to that of more traditional techniques examining latent constructs (e.g., exploratory factor analysis [EFA]) but provides an alternative vantage point to understand and communicate associations in psychopathy that may not be identifiable in EFA (Borsboom et al. 2021).

Personality research, the present study included, often involves cross‐sectional data that do not allow for this type of causal inference. For our purposes, network analysis is a useful exploratory paradigm for describing the associations between psychopathic traits and generating new hypotheses. This may provide initial evidence for the centrality of certain psychopathic traits over others. Furthermore, it may lead to the generation of new hypotheses regarding the associations between psychopathy symptoms that can inform longitudinal and experimental studies. For example, if antisociality is central to psychopathy, it should be highly influential in the endorsement of other psychopathic traits and serve as a conduit for the relations between other traits. Network analysis correctly identifies that endorsement of certain symptoms tends to covary, just like other approaches, although the explanation for this covariation may differ. The benefit of a network analysis approach is to identify how an item (symptom) may influence the relations between other items. While the triarchic model conceives of psychopathy from a latent trait framework (Patrick et al. 2009; Patrick and Drislane 2015), it is plausible that specific symptoms or indicators may be particularly useful for indicating levels of an underlying construct (e.g., meanness), which could be assessed in a manner such as EFA, while also linking that symptom with indicators indexing other psychopathic traits (e.g., boldness and disinhibition).

Studies examining the network structure of the PCL‐R have consistently identified lack of empathy and shallow affect as the most central items across US forensic samples (Preszler et al. 2018; Trupp et al. 2021; Verschuere et al. 2018), although irresponsibility and parasitic lifestyle were more central in an inpatient Dutch sample (Verschuere et al. 2018). Likewise, lack of empathy (callousness) was replicated as the keystone symptom of psychopathy among adolescent offenders using the Psychopathy Checklist‐Youth Version (PCL‐YV; McCuish et al. 2019). Similar results have been found using the Levenson Self‐Report Psychopathy Scale (LSRP) in community (Tsang and Salekin 2019) and forensic samples (Wang et al. 2022; Zhang et al. 2022). Taken together, these studies point to the importance of empathy deficits within the psychopathy network, with relative agreement on the centrality of specific items, and suggest that initial affective deficits bidirectionally influence the development of other psychopathy symptoms. However, disinhibitory traits appear to be more important to the psychopathy network when assessed using alternative measures. For example, in the psychopathic personality inventory‐short form (PPI‐SF), impulsive nonconformity (Item 22) is most central, and social deviance items (e.g., Item 28) were centrally placed on a network model of the Self‐Report Psychopathy Scale‐II (Tsang and Salekin 2019). Some evidence also points to the centrality of impulsivity items (e.g., Unreliable) on the Comprehensive Assessment of Psychopathic Personality Self‐Report (Zhang et al. 2022). Thus, disinhibitory tendencies or deficits in empathy may be more or less central to the construct depending on how psychopathy is assessed.

To further explore key questions about which traits are most central to the psychopathy construct, it is also necessary to look across disparate measures rather than within a specific instrument. Tsang and Salekin (2019) addressed this in a combined network analysis on subscales from multiple measures (LSRP, PPI‐SF, SRP‐II, Personality Assessment Inventory—Antisocial Features Scale [PAI‐ANT; Morey 2007]) by treating subscale scores as nodes (i.e., observed variables) within the network. In contrast to PCL‐R networks, the domain‐level model revealed that scales assessing stimulus seeking (PAI‐ANT; Sokolowski et al. 2020), social deviance (SRP‐II), and interpersonal affective traits (LSRP) were all among the most central to the psychopathy network. Interestingly, Fearlessness (PPI‐SF) was associated with many other psychopathic traits (high betweenness), whereas other FD subscales (Social Potency, Stress Immunity) formed their own cluster with Interpersonal/Affective (SRP‐II) and Coldheartedness (PPI‐SF) subscales. This may indicate that much of boldness is ancillary—rather than central—to other psychopathic traits, or that features of meanness (e.g., Coldheartedness) act as a bridge connecting the interpersonal/affective and behavioral deviance features of psychopathy.

The centrality of psychopathic trait domains in Tsang and Salekin (2019) yielded a combination of sensation seeking, social deviance, and interpersonal affective traits similar to the triarchic conceptualization of psychopathy (Patrick et al. 2009). In the triarchic model, psychopathy consists of three phenotypic components reflecting separate neurobehavioral systems: disinhibition, meanness, boldness. Disinhibition manifests as general externalizing proneness including inhibitory control deficits, irresponsibility, lack of planfulness, and foresight. Meanness consists of a constellation of features related to reduced empathy, exploitativeness, cruelty, and lack of close attachments with others. Finally, boldness is characterized by emotional stability, resilience to stress, social efficacy, and sensation seeking. The triarchic model was originally formulated to resolve discrepancies between different psychopathy conceptualizations and inventories. As such, this model may be particularly useful for comparing the relative importance of these trait constructs within the overall network of psychopathy using a range of instruments, which each reflect somewhat different content and theoretical conceptualizations.

1.1. Present Study

To date, no study has investigated the psychopathy network from the framework of the triarchic model, and the present study is in part an exploratory endeavor that aims to capture the symptom‐level interrelations between psychopathic traits through the lens of the triarchic model. The present study examined the network structure of psychopathy, first through an item‐level network of the TriPM and second through a scale level analysis of psychopathic traits assessed by seven psychopathy inventories. Although we conduct null hypothesis tests to determine whether individual items or scales are more central than others, conclusions about the dynamic associations between psychopathic traits cannot be inferred from the cross‐sectional data. Nonetheless, some expectations for the results can be made based on previous research and the known relations between scales. Based on the centrality of items indexing callousness/lack of empathy found in previous studies, we expected that items measuring meanness would be most central to the network, thus indicating that meanness is most influential in the broader psychopathy network. Likewise, meanness and disinhibition were expected to have stronger and closer associations than with indices of boldness. Boldness nodes were expected to be less central to the network and have modest associations with meanness nodes.

2. Methods

2.1. Participants

The present study used pre‐existing data collected by Drislane et al. 2014 from undergraduate students in the Southeastern United States (M age = 18.8, SD = 1.65; 56.2% female; 76.2% White). Importantly, no exploratory factor analyses have been published using the current data. Self‐report data for the present study were collected in two waves, with 197 participants responding in person using pencil and paper while the remaining 453 participants responded online. The order in which questionnaires were presented to participants was randomized to reduce the effects of respondent fatigue. Further, there were no differences in scale scores between participants who responded electronically versus in person, nor were there any significant differences between waves in age, race, or gender (ps = 0.37–0.97). More detailed demographics of the sample and psychometric properties of the measures are described in greater depth in Drislane et al. 2014. Participants were excluded for demonstrating > 25% missingness on all self‐report items (n = 19), or for inconsistent response patterns on the PPI (≥ 3 SD above the mean on PPI VRIN scale; n = 13), resulting in a final sample of 618 in the original paper. ¹ In the present study, an additional six participants were excluded for demonstrating > 25% missingness on all TriPM items, leaving a final sample of 612 participants.

2.2. Measures

2.2.1. Triarchic Psychopathy Measure

The Triarchic Psychopathy Measure (TriPM; Patrick 2010) is a 58‐item inventory with subscale scores on Boldness, Meanness, and Disinhibition. The Meanness (19 items) and Disinhibition (20 items) scales were taken from the Externalizing Spectrum Inventory (ESI; Krueger et al. 2007) with TriPM Meanness items reflecting ESI callous‐aggression and TriPM Disinhibition reflecting general externalizing on the ESI. The Boldness (19 items) scale was developed independently (Patrick et al. 2019) and reflects social efficacy, emotional stability, and venturesomeness. Participants rated each item on a 4‐point Likert scale: 3 (true), 2 (somewhat true), 1 (somewhat false), 0 (false). Internal consistencies as measured by Cronbach's alpha were acceptably high for all subscale scores (αs = 0.79–0.83; Drislane et al. 2014).

2.2.2. Psychopathic Personality Inventory

The Psychopathic Personality Inventory (PPI; Lilienfeld and Andrews 1996; Lilienfeld and Widows 2005) is a 187‐item inventory yielding scores on eight subscales: Alienation, Carefree Nonplanfulness, Coldheartedness, Fearlessness, Impulsive Nonconformity, Machiavellian Egocentricity, Social Potency, and Stress Immunity. For the purposes of the present study, we opted to use the two‐factor structure of the PPI (Lilienfeld and Widows 2005) along with Coldheartedness as nodes in the domain‐level network. Social Potency, Stress Immunity, and Fearlessness served as indicators of the Fearless Dominance factor (FD), and Machiavellian Egocentricity, Impulsive Nonconformity, Alienation, and Carefree Nonplanfullness served as indicators of the self‐centered impulsivity factor (SCI). Coldheartedness was treated as an independent subscale as it does not load substantially onto either factor (Benning et al. 2003). Participants responded to each item on a four‐point Likert scale: 4 (true), 3 (somewhat true), 2 (somewhat false), 1 (false). Internal consistencies of individual PPI subscales were all acceptably high (αs = 0.79–0.88; Drislane et al. 2014).

2.2.3. Self‐Report Psychopathy Scale‐III

The Self‐Report Psychopathy Scale‐III (SRP‐III; Paulhus et al. 2009) is a 64‐item inventory developed to measure PCL‐R facets via self‐report in community samples. Subscales from the SRP‐III, consisting of Interpersonal Manipulation, Callous Affect, Erratic Lifestyle, and Criminal Tendencies (16 items each), served as individual nodes in domain‐level network analysis. Participants rated the extent to which they agreed with each item on a five‐point Likert scale (5 = agree strongly, 4 = agree, 3 = neither agree nor disagree, 2 = disagree, 1 = disagree strongly). Internal consistency of the SRP‐III subscales was high, with coefficient alpha ranging from 0.76–0.86 (Drislane et al. 2014).

2.2.4. Levenson's Self‐Report Psychopathy Scale

Levenson's Self‐Report Psychopathy Scale (LSRP; Levenson et al. 1995) is a 26‐item inventory originally intended to assess primary (16 items) and secondary (10 items) subdimensions of psychopathy. However, more recent factor analytic studies suggest that a three‐factor structure provides a superior fit to the data and better distinguishes between interpersonal and affective traits that are conflated in the two‐factor model (Sellbom 2011). Therefore, we opted to use the three‐factor LSRP scores consisting of Egocentricity, Antisocial, and Callousness as nodes in the domain‐level network. Participants responded to items on a four‐point scale (5 = agree strongly, 4 = agree somewhat, 2 = disagree somewhat, 1 = disagree strongly). Internal consistency was high for the Egocentricity subscale (α = 0.81) and marginal for the Callousness (α = 0.64) and Antisocial subscales (α = 0.63).

2.2.5. Youth Psychopathic Traits Inventory

The Youth Psychopathic Traits Inventory (YPI; Andershed et al. 2002) is a 50‐item self‐report questionnaire developed to assess behavioral, interpersonal, and affective features of psychopathic traits in adolescents. The measure consists of 10 subscales (Dishonest Charm, Grandiosity, Lying, Manipulation, Callousness, Unemotionality, Remorselessness, Impulsivity, Thrill‐Seeking, and Irresponsibility) that are combined into three factors: Grandiose Manipulative, Callous‐Unemotional, and Impulsive Irresponsible. In the present study, we used these three‐factor scores as nodes in the domain‐level network. Participants indicated the frequency of a behavior on a four‐point scale (4 = Applies very well to 1 = Does not apply at all). Internal consistency of the YPI factors was high, with coefficient alpha ranging from 0.82 to 0.90 (Drislane et al. 2014).

2.2.6. Antisocial Process Screening Device (APSD)

The Antisocial Process Screening Device (APSD; Frick and Hare 2001) is a 20‐item behavioral rating inventory modeled after the PCL‐R to screen for psychopathic traits in adolescents. The APSD has 13 subscales assessing different aspects of antisocial and aggressive behavior. The APSD also yields broader factor scores of Callous/Unemotionality (i.e., lack of remorse and restricted affect), Impulsivity/Conduct Problems (i.e., proneness to boredom, risk taking), and Narcissism (i.e., egocentric manipulativeness). Factor scores were included as nodes in the domain‐level network. Participants responded to items on a three‐point scale (2 = Definitely true, 1 = Sometimes true, 0 = Not at all true). Internal consistency was marginal for the APSD factors (αs = 0.51–0.67; Drislane et al. 2014).

2.2.7. Inventory of Callous‐Unemotional Traits

The Inventory of Callous‐Unemotional Traits (ICU, Frick 2004) is a 24‐item inventory developed to expand coverage of callous‐unemotional traits from APSD items. It is designed to measure precursors to psychopathy in children and adolescents. The ICU renders facet scores of Unemotional, Callous, and Uncaring. However, in contrast to the other measures in the domain‐level network, we opted to use the total score as a single node. This decision was made based on a lack of consensus regarding the appropriate factor structure of the ICU, such that five (Feilhauer et al. 2012), three (Ezpeleta et al. 2013), and two (Houghton et al. 2013) factor solutions are often found. Likewise, the Uncaring factor contains all reverse‐coded items, which suggests this factor may reflect method variance rather than a distinct latent construct. Therefore, we opted to use a total ICU score which has been previously demonstrated to have higher internal consistency than subscale scores along with positive correlations with externalizing outcomes (Fanti et al. 2009; Kimonis et al. 2008). Participants responded to items on a four‐point scale: 3 (Definitely true) to 0 (Not at all true). ICU total scores demonstrated good internal consistency (α = 0.84; Drislane et al. 2014).

2.3. Statistical Procedure

Network models reflect a network of nodes and edges. Observed variables (e.g., items) are represented by nodes, and the relationship between two nodes is represented by edges. In graphical representations, thicker edges indicate stronger connections between two nodes. Centrality indices (strength, closeness, betweenness) are then used to examine the overall structural importance of nodes within the network. Strength equals the sum of absolute partial correlation coefficients, and indicates how strongly a node is directly connected to other nodes in the network. Expected influence (EI) is like strength in that it measures direct connections to other nodes, but EI uses raw, rather than absolute, edge weights. Thus, nodes with no negative edges will have equal EI and strength values. Closeness, measured by the inverse of the sum of shortest paths between a node and all other nodes, indicates how strongly a node is indirectly connected to other nodes. Betweenness, an index of the number of shortest paths on which a node sits between two other nodes, indicates the importance of a node in connecting other nodes. Compared to other centrality indices that reflect the position of a node, EI reflects node importance in maintaining the network and is preferable to other indices in networks with negative edge weights where higher values on all nodes indicate higher psychopathology (Robinaugh et al. 2016). Negative edge weights are to be expected in a network analysis of a multidimensional construct containing items or scales that have contrasting associations with external criteria—even when all indicators are positively keyed (Patrick et al. 2009; Patrick and Drislane 2015). Although we provide the results for all centrality indices in figures, it is worth noting that EI is the most interpretable indicator of node importance in both networks due to the presence of negative edges.

All analyses in the present study were conducted using the statistical application R (version 4.2.1). Following from previous studies (e.g., Verschuere et al. 2018; Wang et al. 2022) and tutorials on network analysis with personality traits (Constantini et al. 2017), an item‐level (TriPM) network model was estimated using the qgraph package (version 1.9.2; Epskamp et al. 2012). We used Spearman's rank correlation coefficients ² to account for potential non‐normality of TriPM items and the large number of nodes in the network. The same estimation method was again used to estimate a domain‐level psychopathy network using scales or factor scores for the TriPM, PPI, SRP‐III, YPI, APSD, LSRP, and ICU.

2.3.1. Edge Weight Accuracy

Non‐parametric bootstrapped confidence intervals (CI), with 2500 bootstrapped samples, were estimated to examine the accuracy of associations between nodes (edges) in the model (Epskamp et al. 2018). This allows for an assessment of variability in edge weights, such that smaller bootstrapped CIs yield more accurate, and thus more interpretable, edge weight estimates. Since edge weights directly affect centrality, small bootstrapped CIs also make it easier to interpret the accuracy of centrality indices. Importantly, bootstrapped CIs containing zero do not indicate that an edge weight is statistically indistinguishable from zero. Instead, edge weight accuracy only reflects the confidence with which the size of associations between nodes can be interpreted.

2.3.2. Centrality Stability

The qgraph package was used to estimate four indices of centrality: strength, closeness, betweenness, and expected influence (Epskamp et al. 2018). Prior to interpretation of the centrality indices, within‐sample centrality stability was assessed with subset bootstrapping in the bootnet package (Epskamp 2015). Summaries of these results for both networks are available in Tables S1–S11 of the Supporting Information (osf.io/8kfe3). Stability is quantified by the correlation stability coefficient (C‐S coefficient), which reflects the maximum number of cases that can be dropped from a sample while maintaining a correlation of 0.70 (95% CI) between centrality indices from the original sample and those estimated from a subset of the data (Epskamp et al. 2018). Epskamp et al. (2018) recommend that a C‐S coefficient of at least 0.25 (but preferably greater than 0.50) is needed to interpret centrality indices.

2.3.3. Centrality Difference Test

Centrality indices (strength, closeness, betweenness, expected influence) reflect the overall importance of a single node within a network. High standing on one centrality index does not necessarily confer high standing on all other indices, and observed variables in a network will have different values for each of these indices. However, whether the centrality of one item is statistically different from that of another item cannot be inferred by comparing their respective numerical values. As such, the node with the highest strength may be statistically indistinguishable from the node with the lowest strength, despite an ostensible numerical difference in strength estimates. Therefore, a bootstrapped difference test with 2500 bootstrap samples was conducted in the bootnet package providing a null hypothesis test of the difference in centrality between two nodes. Of note, there is currently no solution for multiple testing in network analysis (Epskamp et al. 2018).

2.3.4. Addressing Topological Overlap

One concern in network analysis of personality measures is that of topological overlap between nodes (Fried and Cramer 2017). Topological overlap occurs when two nodes are connected with each other and share similar connections with other nodes in the network (i.e., two nodes appear to have nearly identical nomological networks). If kept separate, these nodes can have artificially inflated centrality indices due to their connection with each other. Topological overlap is likely when conducting a network analysis on a personality scale, because redundant items are sometimes retained due to their importance to the internal consistency of the scale (for discussion of this in relation to the TriPM see Patrick et al. 2021). We addressed topological overlap in both networks using the goldbricker function in the networktools R package (Jones 2025). The goldbricker function compares two nodes on every possible combination of correlations within a network and calculates a percentage of correlations that are significantly different for each pair of nodes. Nodes with lower percentages of significantly different correlations exhibit a greater degree of topological overlap. A threshold is set to a specific proportion of different correlations that are too low, and pairs below this threshold can be combined into principal components using the net_reduce function.

We followed this procedure in both networks, setting the threshold for redundant nodes to 50% (i.e., node pairs with fewer than 50% significantly different correlations). Table 1 provides a summary. In the item‐level network, 19‐item pairs were identified as meeting this criterion, and eight were identified in the domain‐level network. We then examined the content of the overlapping node pairs to evaluate whether they are derived from the same facet within a scale or measure the same psychopathic construct, which would lend further support for redundancy. Only items 29 (Empathy; “I don't see any point in worrying if what I do hurts someone else.”) and 24 (Theft; “I have taken money from someone's purse or wallet without asking.”) were identified as redundant nodes but were not from the same facet within a TriPM scale. Node pairs were then reduced into principal components. This resulted in four principal components (CU‐ICU; Cal‐Ego; CALU‐CA; IRR‐Imp) in the domain network and 11 principal components in the item‐level network. Notably, not all node pairs could be reduced, as some nodes were redundant with multiple items (see Table 1). This reduced set of indicators (47 nodes for the TriPM and 16 nodes for psychopathy scales) was then submitted to network analysis.

TABLE 1.

Pairs of items/scales with greater than 50% overlap.

Item (scale; facet)	Item (scale; facet)	%
52 (Meanness; Empathy) a	2 (Meanness; Empathy) a	0.25 a
51 (Disinhibition; Problematic Impulsivity) a	49 (Disinhibition; Irresponsibility) a	0.25 a
30 (Disinhibition; Dependability) a	5 (Disinhibition; Dependability) a	0.32 a
29 (Meanness; Empathy) a	24 (Disinhibition; Theft) a	0.38 a
33 (Meanness; Empathy) a	11(Meanness; Empathy) a	0.39 a
55 (Meanness; Empathy) a	36 (Meanness; Empathy) a	0.39 a
45 (Meanness; Excitement Seeking) a	6 (Meanness; Excitement Seeking) a	0.43 a
41 (Boldness; Dominance) a	13 (Boldness; Dominance) a	0.43 a
29 (Meanness; Empathy)	26 (Meanness; Relational Aggression)	0.43
39 (Meanness; Honesty)	33 (Meanness; Empathy)	0.43
42 (Meanness; Relational Aggression) a	23 (Meanness; Relational Aggression) a	0.45 a
39 (Meanness; Honesty)	11 (Meanness; Empathy)	0.46
24 (Disinhibition; Theft)	12 (Disinhibition; Irresponsibility)	0.46
26 (Meanness; Relational Aggression)	23 (Meanness; Relational Aggression)	0.46
42 (Meanness; Relational Aggression)	36 (Meanness; Empathy)	0.46
20 (Meanness; Empathy) a	8 (Meanness; Empathy) a	0.48 a
57 (Boldness; Persuasiveness) a	19 (Boldness; Persuasiveness) a	0.48 a
34 (Disinhibition; Fraud)	24 (Disinhibition; Theft)	0.48
48 (Meanness; Empathy)	36 (Meanness; Empathy)	0.48
Scale (Measure)	Scale (Measure)	%
Callous‐Unemotionality (APSD)	ICU	0.22 a
Callousness (LSRP)	Egocentricity (LSRP)	0.33 a
Callousness (LSRP)	Meanness (TriPM)	0.39
Egocentricity (LSRP)	Narcissism (APSD)	0.39
Callous‐Unemotional (APSD)	Callous Affect (SRP‐III)	0.39 a
Egocentricity (LSRP)	Meanness (TriPM)	0.44
Impulsive Irresponsible (YPI)	Impulsivity/Conduct Problems (APSD)	0.44 a
Callous Affect (SRP‐III)	Interpersonal Manipulation (SRP‐III)	0.44

Note: % = The percentage of correlations with other nodes in the network that were statistically significantly different from each other.

Abbreviations: APSD, Antisocial Process Screening Device; ICU, Inventory of Callous‐Unemotional Traits; LSRP, Levenson's Self Report Psychopathy Scale; PPI, Psychopathic Personality Inventory; SRP‐III, Self Report Psychopathy Scale; TriPM, Triarchic Psychopathy Measure; YPI, Youth Psychopathic Traits Inventory.

^a Item/scale pairs that were combined into a principal component.

Creating principal components of certain nodes did not result in large changes to the centrality indices or the overall interpretation of our results. Nonetheless, they did improve the stability of the centrality indices and are ultimately more parsimonious and thus are presented in the main Results section of the present paper. For interested readers, we provide network models, at the item and domain level, that were created prior to these reductions in Figures S10–S21 and Tables S12–S23 of our Supporting Information on the OSF project page (osf.io/8kfe3).

3. Results

3.1. Item‐Level TriPM Network

Figure 1 displays the network structure of the TriPM network. The strength of associations between TriPM items corresponds to edge thickness, with positive associations depicted in blue and negative associations depicted in red. Notably, negative edge weights appear between multiple Disinhibition and Boldness items. This is to be expected at the item level considering the item content (e.g., item 16 [“I have a hard time making things turn out the way I want”] and 27 [“People often abuse my trust”]) and the expected null relation between TriPM Boldness and Disinhibition (Patrick 2010). Results for all interpretable centrality indices are provided, while results for strength and EI are reported below. We encourage readers to focus on EI due to negative edges (Robinaugh et al. 2016). The locations of nodes within the network cannot be interpreted directly from the network visualization (Epskamp et al. 2018). As such, any reference to location within this paper is made purely to orient the reader to the graphics.

FIGURE 1.

Network structure of the TriPM. This figure depicts the network structure estimated for TriPM items. Edge thickness corresponds to the strength of associations between nodes. Red edges indicate negative associations, and blue edges indicate positive associations.

Plots for the within‐sample stability of the centrality indices for the TriPM can be found in Figure S1 of the Supporting Information (osf.io/8kfe3). Strength and expected influence display high stability across the bootstrapped subsamples, while closeness and betweenness show more marked decreases in stability with smaller subsamples. This is perhaps unsurprising given the relatively modest sample size (N = 614). Stability as depicted in Figure S1 was supported by the C‐S coefficients. Strength (C‐S = 0.44), expected influence (EI; C‐S = 0.44), and closeness (C‐S = 0.28) all had C‐S coefficients above the lowest recommended value of 0.25 for interpretation. Betweenness (C‐S = 0.21) was below this threshold and therefore deemed uninterpretable.

Figure 2 displays the centrality statistics for strength, EI, and closeness of the TriPM items. The principal component of items 42 (Relational Aggression; “I sometimes insult people on purpose to get a reaction from them”) and 23 (Relational Aggression; “I enjoy pushing people around sometimes”) had the highest strength, followed by the principal component of 29 (Empathy) and 24 (Theft) and that of items 11 (Empathy; “I sympathize with others' problems”) and 33 (Empathy; “I am sensitive to the feelings of others”). Item 34 (Conning; “I have conned people to get money from them”) demonstrated similar strength. Items 3 (Impatient Urgency; “I often act on immediate needs”), 1 (Optimism; “I'm optimistic more often than not”), and 54 (Social Assurance; “I never worry about making a fool of myself with others”) had the lowest strength and are depicted graphically as relatively isolated from other items with few strong connections.

FIGURE 2.

Centrality indices estimated for TriPM items. Centrality indices for the TriPM network model. Higher values indicate greater centrality.

The results for EI closely resemble those for strength. Components 42–23 (Relational Aggression), 29–24 (Empathy‐Theft), and item 34 (Conning) were highly influential in the network when interpreting their EI scores. Close inspection of Figure 2 reveals that most items among the highest in EI were from the Meanness scale including components 33–11 (Empathy), 45 (Excitement Seeking; “Things are more fun if a little danger is involved”), 6 (Excitement Seeking; “I would enjoy being in a high speed chase”), 20 (Empathy; “It doesn't bother me to see someone else in pain”), and 8 (Empathy; “I don't mind if someone I dislike gets hurt”). Among Boldness items with the highest EI's were 10 (Courage; “I get scared easily”), 44 (Social Assurance; “It's easy to embarrass me”), and component 57 (Persuasiveness; “I'm not very good at influencing people”) and 19 (Persuasiveness; “I have a knack for influencing people”). Disinhibition items high in influence included items 37 (Problematic Impulsivity; “I get in trouble for not considering the consequences of my actions”), 15 (Problematic Impulsivity; “I jump into things without thinking”), and component 51 (Problematic Impulsivity; “Others have told me they are concerned about my lack of self‐control”) and 49 (Irresponsibility; “I have lost a friend because of irresponsible things I've done”). Of note, four items appeared to be especially low in influence: 27 (Alienation; “People often abuse my trust”), 21 (Planful Control; “I have good control over myself”), 1 (Optimism; “I'm optimistic more often than not”), and 3 (Impatient Urgency; “I often act on immediate needs”).

The results of bootstrapped difference tests, with black boxes indicating significant differences in item strength and EI are displayed in Figures S2–S4 of the Supporting Information. Many of the items were not significantly different from each other in terms of strength (Figure S2) or closeness (Figure S4). Components 42–23 (Relational Aggression), 29–24 (Empathy‐Theft), 33–11 (Empathy), and item 34 (Empathy) exhibited significantly higher influence than most other items. By contrast, 3 (Impatient Urgency), 1 (Optimism), 21 (Planful Control), 27 (Alienation), 16 (Optimism), and 46 (Impatient Urgency) demonstrated significantly lower EI than all other items. These nodes were not significantly different from each other.

3.2. Domain‐Level Network

The domain‐level network structure is presented in Figure 3. Centrality stability for the domain‐level psychopathy network was similar to that of the TriPM item‐level network. As is shown in Figure S5 of the Supporting Information, only a moderate proportion of the sample could be dropped to maintain a correlation of 0.7 between the centrality indices in the original and subsamples. Based on the C‐S coefficients, all of the centrality indices were interpretable. Expected Influence (C‐S _EI  = 0.75) and strength (C‐S _strength  = 0.67) demonstrated the highest stability and were well above the recommended cut‐off of 0.50, although closeness (C‐S _closeness  = 0.44) and betweenness (C‐S _betweenness  = 0.44) were above the lower limit for interpretability of 0.25. Like the item‐level network, we predominantly interpret EI but provide the results for the rest of the indices.

FIGURE 3.

Domain‐level network structure for seven inventories of psychopathic traits. This figure depicts the domain‐level network structure. Edge thickness corresponds to the strength of associations between nodes. Red edges indicate negative associations, and blue edges indicate positive associations. APSD, Antisocial Process Screening Device; ICU, Inventory of Callous‐Unemotional Traits; LSRP, Levenson's Self Report Psychopathy Scale; PPI, Psychopathic Personality Inventory; SRP‐III, Self Report Psychopathy Scale; TriPM, Triarchic Psychopathy Measure; YPI, Youth Psychopathic Traits Inventory.

Centrality indices for the domain‐level network are shown in Figure 4. Component Callous‐Unemotional (YPI) and Callous Affect (SRP) were the highest in strength, followed by Erratic Lifestyle (SRP), Interpersonal Manipulation (APSD), Grandiose Manipulative (YPI), Fearless Dominance (PPI), and Meanness (TriPM). Inspection of the bootstrapped difference tests in Figures S6–S9 of the Supporting Information suggests that differences in centrality between the scales were trivial. Only three scales displayed significantly lower strength than most other scales. These were Criminal Tendencies (SRP), Narcissism (APSD), and Antisocial (LSRP). Criminal Tendencies showed the lowest strength overall but was no different from Narcissism or Antisocial. Closeness scores were also quite consistent throughout the domain‐level network. The Callous‐Unemotional (YPI)‐Callous Affect (SRP) component was among the highest in closeness, along with Self‐Centered Impulsivity (PPI), Meanness (TriPM), Disinhibition (TriPM), Interpersonal Manipulation (SRP), and Erratic Lifestyle (SRP).

FIGURE 4.

Centrality Indices estimated for psychopathy domains. Centrality indices for the domain‐level network model. Higher values indicate greater centrality.

EI rendered clearer indications of the most central scales within the network. Component Callous‐Unemotional‐Callous Affect (YPI‐SRP) and scales Erratic Lifestyle (SRP), Interpersonal Manipulation (APSD), Grandiose Manipulative (YPI), Meanness (TriPM) were significantly more influential than most other scales. The Self‐Centered Impulsivity (PPI) and Callousness‐Eccentricity (LSRP) nodes had similar expected influence. Self‐Centered Impulsivity (PPI) was only lower than Callous‐Unemotional‐Callous Affect (YPI‐SRP), while Callousness‐Eccentricity (LSRP) was additionally lower in influence than Interpersonal Manipulation (APSD), Erratic Lifestyle (SRP), and CALU‐CA (YPI‐SRP). Coldheartedness (PPI) and Antisocial (LSRP) demonstrated significantly lower EI than all other scales in the domain network. Although Boldness (TriPM) showed higher EI than these two scales, it was lower in influence than most other scales, albeit no different than Narcissism (APSD) and Criminal Tendencies (SRP).

4. Discussion

The present study was an exploration of the network structure of psychopathy, first as operationalized by the TriPM and subsequently as operationalized by multiple measures of psychopathic traits. We expected that Meanness items would be most central to the TriPM, which was strongly supported. Components from the Relational Aggression (42–23), Empathy (33–11), and Excitement Seeking facets of the Meanness scale were among the most influential. Along with Disinhibition item 34 (Conning), these nodes had the highest expected influence and were statistically distinguished from most other items in the bootstrapped difference test. The prevalence of Meanness items among the most influential nodes in the TriPM network parallels early clinical descriptions of “lovelessness” and “guiltlessness” (McCord and McCord 1964) as core to psychopathy and contemporary, basic trait perspectives implicating low agreeableness (Miller and Lynam 2015). Our results are also comparable to the centrality of Callous/Lack of Empathy in the PCL‐R (e.g., Verschuere et al. 2018) and antagonism from the LSRP (e.g., Tsang and Salekin 2019) networks. Thus, despite concerns regarding the replicability of psychological networks (Forbes et al. 2017), there is growing evidence for the replicability of traits central to psychopathy networks across measures and populations.

Contrary to hypotheses, Boldness and Disinhibition items had similar influence as other nodes in the network, as is indicated in the bootstrapped difference tests. Parity across TriPM items, and the interconnectedness of Meanness, Disinhibition, and Boldness, may support the mutual relevance of all three constructs to the conceptualization of psychopathy. Nonetheless, Meanness items and components were disproportionately represented among highly influential items in the network, even though they were not significantly more influential than those from Boldness and Disinhibition. Items low in EI may be important to refining the conceptualization and measurement of psychopathy at the facet level. Notably, items 27 (Alienation), 21 (Planful Control), and 3 (Optimism) were especially low in EI and may be less important in influencing the endorsement of other items. Alienation and Optimism are less classically characteristic of psychopathy compared to, for example, lack of empathy, irresponsibility, or dominance. While they are clearly related to items in their respective scales and thus psychometrically important to maintaining internal consistency, their lack of centrality may challenge whether they contain construct relevant variance.

The domain‐level network, in contrast to the TriPM item‐level network, provided a clearer picture and appears to map closely onto the triarchic model. The three triarchic constructs were well represented among scales highest in EI including CALU‐CA (YPI‐APSD), Erratic Lifestyle (SRP), Interpersonal Manipulative (SRP), Grandiose Manipulative (YPI), Meanness (TriPM), and Self‐Centered Impulsivity (PPI). These findings are consistent with Tsang and Salekin's (2019) domain‐level psychopathy network, which revealed stimulus seeking, social deviance, and primary psychopathy as most central in a college sample. Scales central to the domain‐level network appeared to represent impulse control (i.e., Erratic Lifestyle, Self‐Centered Impulsivity) and affective traits (i.e., Callous Affect, Callous‐Unemotional) along with features of interpersonal dominance (Interpersonal Manipulative, Grandiose Manipulative). Thus, at the domain level, a three‐factor structure of psychopathy appeared more evident.

One of the purposes of this study was to investigate the centrality of antisocial and criminal behavior to the conceptualization of psychopathy. Previous network structures of the PCL‐R have found mixed evidence, with many items on the PCL‐R Antisocial facet falling between low and moderate centrality across all indices (Preszler et al. 2018; Trupp et al. 2021; Verschuere et al. 2018). In our study, Criminal Tendencies (SRP) and Antisocial (LSRP) were among the lowest betweenness and closeness scores, although these were not significantly different from most other nodes in the networks. Results for strength and expected influence were starker by comparison. Aside from Coldheartedness (PPI), Antisocial (LSRP) was significantly lower in expected influence compared to all other scales. Criminal Tendencies was among the lowest in EI, as it was more central than only Coldheartedness and Antisocial and equivalent to Boldness (TriPM) and Narcissism (APSD). Similar results were found for strength, where Criminal Tendencies was the lowest in overall strength, albeit no different from Narcissism or Antisocial. The characteristics of our sample potentially influenced such low centrality, as Criminal Tendencies may have a low rate of endorsement among a college sample. However, these results may challenge the claim that criminal behavior is a keystone characteristic of psychopathy, especially in non‐forensic populations.

Evidence for the centrality of boldness was less clear. Boldness demonstrated especially low influence, while FD appeared to be influential although lower in EI compared to other scales like Erratic Lifestyle. As such, our results may run in contrast to the notion that boldness could indeed give rise to other psychopathic traits (Lykken 1957). However, interpersonal features of psychopathy as measured by Interpersonal Manipulation (SRP) and Grandiose Manipulative (YPI), which reflect boldness to some degree (Drislane et al. 2014), were among the most central to the domain network. Thus, Boldness and its related scales were less cohesive than expected, potentially reflecting the facet level differences in influence that were observed at the item level.

Finally, an unexpected outcome of the present study was the topological overlap of LSRP scales Callousness and Egocentricity. We opted to use the three‐factor structure, as opposed to the traditional two‐factor structure, due to evidence of superior fit and indications that the three‐factor structure better distinguishes between interpersonal and affective traits (Sellbom 2011). In our data, only 33% of the correlations between Callousness and Egocentricity and other measures demonstrate statistically significant differences from each other, which may challenge whether these scales are truly indexing different constructs. Future studies should determine whether this implies further refinement of the LSRP factor structure or item content is needed.

5. Limitations

There are multiple limitations to the present study. Our use of a cross‐sectional study design meant that causal inferences regarding potential symptoms interactions cannot be made from our results. Due to our use of cross‐sectional data, our results should be viewed as descriptive. Nonetheless, we hope that future studies will test our findings in longitudinal and experimental data. Furthermore, our convenience sample of college students is a limitation, as college students are not representative of the general population and may have influenced the centrality of certain items and scales. Nevertheless, important information can still be gleaned from this sample. An undergraduate sample was used for the present study since the purpose of the study was to understand how relationships between psychopathic trait dimensions are related in individuals with normal range personality. The scales used in the present study were developed for use in non‐clinical settings and provide information across the spectrum of severity for traits assessed similar to that of non‐pathological personality instruments (Walton et al. 2008). Finally, most previous network analyses conducted on psychopathic traits use forensic populations, yet there is evidence to suggest that the psychopathy network is somewhat different in non‐criminal settings (Tsang and Salekin 2019). Thus, using an undergraduate sample may inform how psychopathic traits are expressed in the absence of antisocial behavior (Lilienfeld 1998).

Sample size was another important limitation, as our sample was small in comparison to those used in other network analyses. Currently, there is no established method for determining sufficient power in network analysis (Epskamp et al. 2018), and we suspect sample size was the source of instability in some centrality indices. As such, the results of our study should be interpreted with caution, and future studies should attempt to replicate the TriPM and domain‐level networks in larger, more representative samples. Our selection of scales in the domain‐level network potentially contributed to disparate findings as well. In our study, we opted to use scale scores for the TriPM, SRP‐III and factor scores for the LSRP, PPI, YPI, and APSD, and total scores for the ICU. Other researchers have made different decisions, such as using scale scores from the PPI (Tsang and Salekin 2019).

Finally, there are limitations to network analysis that may affect the generalizability of our results. Generally, our results are in agreement with previous studies on the centrality of specific trait dimensions, although the absolute centrality of particular items may be subject to variation (Forbes et al. 2017). Furthermore, betweenness and closeness rely on the assumption of shortest paths, which assumes that transactions between related nodes will take a direct path (Bringmann et al. 2019). However, this is often not the case with psychological networks whose nodes have many indirect relations with negative edges. There is ongoing debate as to whether these limitations obscure the interpretability of centrality indices in the presence of an underlying latent structure (Hallquist et al. 2021). Simulation studies have demonstrated that betweenness and closeness are vulnerable to spurious correlations while strength is statistically equivalent to factor loadings (Christensen and Golino 2021; Hallquist et al. 2021). In this view, nodes that are higher in strength are simply more consistent indicators of an underlying latent trait. For example, Hallquist et al. (2021) demonstrated that strength increases for nodes that contain more factor‐related variance. However, others have demonstrated that network models and latent variable models are only generally equivalent, while subsets of these models lead to different predictions (van Bork et al. 2021). Furthermore, statistical equivalency may be unimportant due to differing interpretations and conceptualizations of the models, such that the theoretical implications vary between each method (Christensen and Golino 2021; van Bork et al. 2021).

Nonetheless, the centrality indices used in this study should be interpreted with caution due to the collinearity of nodes in both the item‐level and domain‐level networks, as the known correlations between items and scales make it difficult to claim one node is more influential to the network than any others (Bringmann and Eronen 2018). We attempted to address the issue of topological overlap in our analyses by creating principal components of items/scales that were effectively redundant. Regularized partial correlation matrices for the networks along with bootnet summaries are available in Tables S1–S11 of the Supporting Information (osf.io/8kfe3). However, topological overlap is still a threat to the results presented in this paper, especially in the item‐level network where components of Meanness items were often among the most central. This is likely caused by the overrepresentation of items and scales assessing deficits in empathy, both in the TriPM (more than half of the Meanness items are Empathy items) and the larger psychopathy literature. As such, it is plausible that the centrality of Meanness found in our networks and in other works (e.g., Preszler et al. 2018; Trupp et al. 2021; Verschuere et al. 2018) is caused by statistical redundancy. Indeed, at least two of the self‐report measures used in this study are derived from the same conceptual model of psychopathy and thus index highly similar traits (i.e., APSD and SRP‐III are modeled from the PCL‐R). While an important consideration for interpreting our results, we suspect that this may reflect broader problems in psychopathy research, especially since new self‐report measures are validated based on their pattern of association with existing measures, which further reifies the emphasis on callousness items as essential to the measurement of psychopathy.

6. Conclusion

Despite these limitations, our findings may have important implications for future research refining the measurement and conceptualization of psychopathy. For example, certain boldness characteristics appear to be especially central to the psychopathy network (i.e., interpersonal dominance) whilst others are less influential (i.e., optimism). Furthermore, investigation of the psychopathy network may eventually offer a data‐driven approach to treatment target selection in psychopathy. From a network perspective, central traits serve as intervention targets, and the similar centrality across trait domains suggests that intervening on any trait should effectively disrupt the psychopathy network. Presently, this is beyond our abilities to test here, and future studies using longitudinal designs and more diverse sample characteristics will be better equipped to explicitly examine the dynamic associations among psychopathy symptoms. Nonetheless, there is growing support for the high centrality of meanness characteristics to the conceptualization of psychopathy (e.g., Verschuere et al. 2018; Tsang and Salekin 2019), but not to the exclusion of disinhibited or boldness‐related features.

Author Contributions

Ky Bray contributed to formal statistical analyses, data visualizations, and original drafting of the current manuscript. Dr. Laura E. Drislane was responsible for data curation and supervision. All authors contributed equally to conceptualization, reviewing, and editing.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: jopy70047‐sup‐0001‐Supinfo.zip.