Structure of Pathological Personality Traits through the Lens of the CAT-PD Model

Abstract

Personality pathology is increasingly conceptualized within hierarchical, dimensional trait models. The Comprehensive Assessment of Traits Relevant to Personality Disorders (CAT-PD; Simms et al., 2011) is a pathological-trait measure with potential to improve on currently prevailing instruments because it has wider content coverage; however, its domain-level structure, which is of scientific and clinical interest, is not established. In this study, we investigated the structure and construct validity of the CAT-PD’s domain level to facilitate wider use of the measure. We estimated five- and six-factor models with exploratory factor analysis in a pooled sample of 8 independent subsamples (N=3,987) and found that both models fit the data well; each had interpretable factors that were invariant across gender, sample type, and black/white racial groups; and the factors had good convergent validity with other measures of maladaptive traits, Big Five personality, and interpersonal problems. Our results support the validity of the CAT-PD for assessing multiple levels of the pathological trait hierarchy.

Psychology and psychiatry are undergoing a paradigm shift towards conceptualizing personality pathology within dimensional trait models (e.g., Cuthbert, 2020; Hopwood et al., 2018; Insel et al., 2010; Kotov et al., 2017). Accumulating evidence indicates that pathological personality traits are organized hierarchically, such that many narrow traits can be combined to form increasingly broad domains capturing more general maladaptive tendencies (Morey et al., 2013; Ringwald et al., 2019; Wright et al., 2012; Wright & Simms, 2014). Although a hierarchical personality trait structure is consistently found in factor-analytic research, the specific number and nature of factors varies across studies. The strongest consensus is on the five-factor level of the hierarchy including the trait domains of Negative Affectivity, Detachment, Disinhibition, Antagonism, and Psychoticism (Harkness et al., 2012; Widiger & Simonsen, 2005). These domains converge strongly with well-established structural models of personality including the Big Five traits (Gore & Widiger, 2013; O’Connor, 2005; Saulsman & Page, 2004) and interpersonal circumplex (Girard et al., 2017; Wright et al., 2012), as well as the more general structure of psychopathology (Kotov et al., 2017; Ringwald, Forbes, et al., 2021), suggesting that they may be a particularly useful level of analysis for researching cross-cutting psychiatric phenomena. Section III of the DSM-5 includes an instantiation of this domain model for diagnosing personality disorder, indicative of its growing influence on clinical practice.

Although this five-factor, dimensional trait model has gained wide acceptance, long-standing clinical theories posit pathology related to a sixth domain of Anankastia (e.g., Abraham, 1966; Freud, 1959). Anankastia encompasses features of perfectionism, rigid adherence to behavioral routines and viewpoints, overconcern with details and orderliness, hypercontrol, perseveration, and risk aversion (Clark & Krueger, 2010; Clark et al., 2021; Mulder et al., 2016). In contrast, five-factor models of personality pathology consider Anankastic features to reflect low (i.e., lack of) Disinhibition rather than a distinct trait (e.g., Krueger et al., 2012; Rojas & Widiger, 2017). Indeed, some key manifestations of personality pathology related to Anankastia (e.g., overconscientiousness, preoccupation with rules) are covered by the Disinhibition construct. At the same time, some Anankastic features are not captured well but Disinihibition (e.g., workaholism), suggesting it may be better conceptualized as a separate trait (e.g., Anderson & Sellbom, 2021; Bach et al., 2020; Clark et al., 2021; Clark & Ro, 2014; Sellbom et al., 2020; Stricker & Pietrowsky, 2021). As a testament to the contemporary clinical relevance of Anankastia, it is included in the World Health Organization’s ICD-11 dimensional trait model for diagnosing personality disorder, which has officially replaced the ICD-10 categorical system for use worldwide in all 194 WHO member states, including the U.S.

Despite the growing scientific and clinical importance of dimensional trait models for personality pathology, and the possibility of multiple, viable domain-level structures, most recent research has relied on a single measure of maladaptive traits, namely, the Personality Inventory for DSM-5 (PID-5; Krueger et al., 2012), which has both strengths and notable psychometric shortcomings. The PID-5 has been shown to have a stable factor structure (Somma et al., 2019) that is invariant across gender (Suzuki et al., 2018) and clinical/non-clinical samples (Bach et al., 2018); it has good convergent but poor discriminant validity (Al-Dajani et al., 2016); and it does not assess some aspects of personality pathology, such as workaholism, self-harm, and antisocial behavior, which may limit its utility for assessing consequential clinical problems (Few et al., 2013; Watters & Bagby, 2018; Yalch & Hopwood, 2016). The PID-5 scales, originally 37 in number but reduced to 25 following initial analyses (Krueger et al., 2011a; Krueger, et al., 2011b), have been found to load onto five higher order domains without a dedicated Anankastia factor (cf., Bach et al., 2017; Sellbom et al., 2020). Two scales that would plausibly form part of an Anankastia domain were, nonetheless, included in the final measure, but they loaded either weakly negatively (Rigid Perfectionism) or did not load (Perseveration) on Disinhibition, loading instead on Negative Affectivity (Clark & Watson, 2022). These limitations of the PID-5 notwithstanding, no other dimensional trait measure has yet received comparable empirical evaluation to be considered a sound alternative.

The Comprehensive Assessment of Traits Relevant to Personality Disorders (CAT-PD; Simms et al., 2011) is a measure of pathological traits that has the potential to address some of the PID-5 limitations. The CAT-PD consists of 33 lower order trait scales that overlap with nearly all of the PID-5’s 25 lower order trait scales, and includes ten scales that are not represented in the PID-5 (e.g., norm-violating behavior, self-harm). Previous work has found that the CAT-PD scales predict clinical outcomes (Yalch & Hopwood, 2016) and interpersonal problems (Williams & Simms, 2016) over and above the PID-5, which supports the conclusion that it covers aspects of personality pathology not covered by the PID-5.

The expanded content coverage at the lower order trait level gives the CAT-PD a clear advantage at predicting relevant criterion variables, but personality pathology is often studied and conceptualized at the higher order domain-level, which has not been firmly established in the CAT-PD. The CAT-PD was not designed to conform to a prespecified domain-level structure. Instead, it was developed to cover content from all major models of personality pathology, including those with a five-factor structure and those with Anankastia as a separate domain. Because of this bottom-up approach to scale development, the CAT-PD may reflect the five- and/or six domains specified by prominent maladaptive trait models. To date, the higher order structure of the CAT-PD has been inferred primarily by rational methods (e.g., Simms et al., 2011) or from joint factor analyses with other measures (Crego et al., 2018; Crego & Widiger, 2016; Wright & Simms, 2014). Results from joint factor analyses suggest that the CAT-PD’s higher order structure aligns with the five consensus domains, but there have been some discrepancies. In a large clinical sample, Wright and Simms (2014) conducted factor analysis of 88 scales from the CAT-PD, PID-5, and Big Five personality traits covering extensive personality and personality pathology content, and showed that a five-factor structure consistent with the consensus domains fit the data well. Although Crego and Widiger (2016) also found the five consensus domains in an Mturk sample with more limited construct coverage, they also found an additional “dependency” domain in two joint factor analyses of the CAT-PD, including one analysis with the PID-5 and another with the Five-factor Model of Personality Disorder scales (Widiger et al., 2012). Finally, in a joint factor analysis with Big Five personality traits in an Mturk sample, Crego and colleagues (2018) found four of the five consensus domains, with the psychoticism indicators split between two additional factors.

A strength of joint factor analyses is that this method arrives at solutions reflecting consensus constructs rather than the idiosyncrasies of a specific instrument. The remarkable consistency in personality pathology structure found across joint factor analyses despite using different measures, methods, and samples is a testament to the robustness of these consensus domains. Joint factor analyses are not appropriate, however, for determining the structure of a specific instrument such as the CAT-PD or whether it assesses the intended consensus constructs when used by itself. Only two factor-analytic studies have examined the CAT-PD structure independent of the scaffolding provided by other measures. Encouragingly, these studies have found similar results to the joint factor analyses. In a study of students and Mturk participants, Long and colleagues (2021) estimated an interpretable five-factor structure, but the model had poor global fit. Thimm (2020) examined the higher order structure of the Norwegian version of the CAT-PD in a student sample, and found a well-fitting, five-factor structure consistent with the expected domains. Both studies attempted to estimate models with more than five factors, but they either did not converge or did not provide the best fit; however, neither specifically considered a model with an Anankastia factor separate from Disinhibition. Although the CAT-PD scales were informed by the consensus five-factor domain structure (Simms, 2011), they were not explicitly developed to conform to that structure, and it is possible that Anankastia could be recovered. Findings from Long and colleagues hint at the possibility of an Anankastia domain separate from Disinhibition, as items tapping pathological constraint (e.g., Perfectionism, Workaholism) loaded primarily on Antagonism rather than negatively on Disinhibition. Furthermore, both studies on the CAT-PD higher order structure used non-clinical samples and only evaluated global fit, leaving open questions about the structure’s generalizability across a range of pathology and a need to evaluate other psychometric properties, such as (1) measurement invariance across different populations and (2) the construct validity of the factors.

Current Study

Given the importance of the domain level for studying personality pathology, and the need for comprehensive and valid measures of these domains, this study built on previous investigations to clarify and validate the higher order, domain level structure of the CAT-PD. We estimated the structure in a pooled sample from 8 independent studies (total N = 3,987) drawn from both clinical and non-clinical populations to maximize generalizability. The goal of these analyses was to determine whether the measure reflects a five-factor structure and/or a six-factor structure with a separate Anankastia domain. We also went beyond evaluating global fit to test the measurement invariance and construct validity of the empirically derived trait domains. We examined whether these domains converged with the predominant measure of pathological traits (i.e., PID-5) and took advantage of the strongly supported associations between hypothesized trait domains with Big Five personality traits and the interpersonal circumplex to enrich the nomological network of CAT-PD constructs and enable comparisons with previous work. Knowing which domain-level structure(s) can be estimated from the CAT-PD is necessary to determine what constructs from the major models of personality pathology it can validly assess in clinical and research settings as well as to facilitate its wider use.

Method

Most of our analyses were pre-registered, and all data and code needed to reproduce our results, as well as supplementary materials are available on the Open Science Framework (OSF; https://osf.io/3jazv/).We report herein all data exclusions and measures used in this study. Because we used existing data, the sample sizes for each subsample were determined by the original study aims, not our analyses.

Participants

Characteristics of the total sample and each subsample are presented in Table 1. Additional information about the individual study procedures is provided in the supplementary materials on OSF (Table S1). The pooled sample includes 3,987 participants drawn from eight independent subsamples that were administered the CAT-PD adaptive or static-form measure; the overall sample included 352 clinical and 3,635 non-clinical participants. Participants within each subsample were retained if they contributed complete or nearly complete CAT-PD data (i.e., > 90% of items).

Table 1.

Pooled Dataset Characteristics

% Race
Sample	Sample Type	n	% Female	M Age (SD)	Native American	Asian	Black	White	Other Race	Mixed Race	% Hispanica	Validators
1b	Clinical	211	62	44.09 (12.58)	-	1	22	64	-	12	9	PID-5, BFI, IIP-SC
2c	Clinical	141	62	28.34 (6.24)	1	2	12	77	4	5	3	NEO-PI-3, IIP-SC
2	Community	170	42	27.71 (6.50)	-	9	15	67	4	5	3	NEO-PI-3, IIP-SC
3	Community	299	71	46.50 (13.13)	1	-	48	46	-	3	4	PID-5, NEO-PI-3, FI-FFM
4	Undergraduate	396	57	18.74 (1.36)		12	7	81	-	-	5	IIP-SC
5	Undergraduate	374	69	19.20 (1.43)	1	-	4	78	10	4	11	BFI-2
6	Undergraduate	337	52	19.03 (1.17)	-	35	8	48	-	8	7	BFI, IIP-SC
7	MTurk	1511	46	35.84 (10.53)	1	7	9	79	2	3	9	-
8	MTurk	548	55	34.26 (10.76)	2	11	15	67	1.5	3	18	-
	Total:	3,987	54	31.56 (12.67)	1	9	13	70	2	4	9

Note.

^aCalculated separately from racial percentages;

^bCompleted the CAT-PD;

^cSample 2 consisted of participants selected to balance those in mental health treatment or not; PID-5 = Personality Inventory for the DSM-5; IIP-SC = Inventory of Interpersonal Problems Short Circumplex.

Measures

Internal consistencies for the CAT-PD and validation scales are summarized below, with full results reported in the supplementary materials on OSF (Table S2, S3).

Comprehensive Assessment of Traits Relevant to Personality Disorders (CAT-PD)

The CAT-PD (Simms et al., 2011) is a measure of 33 pathological personality traits. The CAT-PD can be administered in either its 216-item static form (CAT-PD-SF) or its adaptive form (CAT-PD-adaptive), which uses an IRT-based computerized-adaptive-testing algorithm to administer the most informative 188 items to participants from a larger pool of 1,366 items. Items on both measures were assessed using a 5-point response scale ranging from 1 (very untrue of me) to 5 (very true of me). Most participants in our pooled sample (n = 3,776; 95%) completed the CAT-PD-SF whereas a small number (n = 211, 5%) completed the CAT-PD-adaptive.

Trait scale scores were created by calculating the mean ratings of each scales’ items and then converting these means to T-scores using established norms for the CAT-PD (Simms et al., 2011). Converting all scores to T-scores enabled us to combine CAT-PD and CAT-PD-SF scores across subsamples. For the CAT-PD-SF, scale scores were first converted to z scores then converted to T-scores. For the CAT-PD-adaptive, scale scores in the form of thetas were created using an IRT model described by Simms and colleagues (2011). Descriptive statistics for the lower order trait scales are in the supplementary materials on OSF (Table S3). The average internal consistency (McDonald’s ⍵) of the trait scales was .85 (range = .80–.91).

Both forms of the CAT-PD include items designed to detect inconsistent responding, but these items were only administered to the two Mturk subsamples. In these two subsamples, the CAT-PD validity items were part of a larger series of items written to capture random responding, yea-saying, and impression management. Participants in these subsamples were excluded if more than two items from the broader suite of validity items were deemed invalid.

Validation Measures

Table 1 shows which samples were administered each validation measure.

Maladaptive Personality Traits.

The Personality Inventory for DSM-5 (PID-5; Krueger et al., 2012) is a 220-item self-report measure assessing the 25 maladaptive facets of Criterion B of the Alternative Model for Personality Disorders in DSM-5, Section III. The PID-5 uses a 4-point scale ranging from 0 (very false or often false) to 3 (very true or often true). Three facet scales were combined to assess each of the five maladaptive trait domains of Negative Affectivity (Emotional Lability, Anxiousness, Separation Insecurity), Detachment (Withdrawal, Anhedonia, Intimacy Avoidance), Disinhibition (Irresponsibility, Impulsivity, Distractibility), Antagonism (Manipulativeness, Deceitfulness, Grandiosity), and Psychoticism (Unusual Beliefs/Experiences, Eccentricity, Perceptual Dysregulation) (American Psychiatric Association, 2013). The PID-5 has been shown to have a relatively consistent structure and generally good convergence with normal-range Big Five traits as well as other maladaptive trait measures (Al-Dajani et al., 2016; Clark & Watson, 2022). The average internal consistency (McDonald’s ⍵) of the PID-5 trait scales was .94 (range = .91–.97).

Big Five Personality Traits.

Measures assessing the Big Five personality traits of Neuroticism, Extraversion, Conscientiousness, Agreeableness, and Openness were available in five subsamples. These measures and the participants that completed them are described in further detail below. To compare Big Five traits across measures, trait scales were standardized within samples. In samples in which two measures were completed, they were averaged before being standardized. The average internal consistency (McDonald’s ⍵) of the Big Five personality trait scales was .87 (range = .81–.93).

Big Five Inventory (BFI).

The 44-item BFI (John et al., 1991) and the 60-item BFI-2 (Soto & John, 2017) are self-report measures that assess the Big Five personality traits on a 5-point scale ranging from 1 (disagree strongly) to 5 (agree strongly). Both measures have been shown to have good convergence with other Big Five trait measures (John & Srivastava, 1999; Soto & John, 2017).

NEO-Personality Inventory-3 (NEO-PI-3).

The NEO-PI-3 (McCrae et al., 2005) is a well-validated and reliable 240-item self-report measure that assesses the Big Five personality traits on a 5-point scale ranging from 1 (strongly disagree) to 5 (strongly agree).

International Personality Item Pool-NEO-120 (IPIP-NEO-120).

The IPIP-NEO-120 (Johnson, 2014) is a 120-item shortened form of the International Personality Item Pool-NEO-300 that measures Big Five personality traits similarly to the NEO Personality Inventory. The IPIP-NEO-120 uses a 5-point scale ranging from 0 (very inaccurate) to 4 (very accurate). The IPIP-NEO-120 has been shown to have good convergent validity with the NEO-PI-R (Johnson, 2014).

Faceted Inventory of the Five factor Model (FI-FFM).

The FI-FFM (Watson, Nus, et al., 2019) is a 207-item self-report measure; it assesses 22 facets that were combined to measure the Big Five personality traits. Items were rated on a 5-point scale ranging from 1 (strongly disagree) to 5 (strongly agree). The FI-FFM has strong convergent and discriminant validity with other Big Five and maladaptive trait measures (Watson et al., 2019).

Interpersonal problems.

The Inventory of Interpersonal Problems—Short Circumplex Form (IIP-SC) is a 32-item self-report measure (Soldz et al., 1995) that assesses interpersonal functioning on a 5-point scale from 1 (not at all) to 5 (extremely). The items were combined into octant scales of four items each (Domineering, Vindictive, Cold, Socially Avoidant, Nonassertive, Exploitable, Overly Nurturant, and Intrusive). Octant scales were then combined to form broad dimensions of dominance, affiliation, and generalized interpersonal distress. Previous work has shown that the octant scales and higher order scales have good convergent validity related to pathological traits (Hopwood et al., 2008; Wright, et al., 2012). To compare scales across samples, scale scores were standardized within samples. Average internal consistency was .84 for the dominance scales (range = .77–.89), .85 for affiliation (range = .77–.92), and .84 for distress (range = .82–.86).

Analytic Plan

Unless otherwise noted, all aspects of our analytic plan were pre-registered. All analyses were conducted in Mplus Version 8.6 (Muthén & Muthén, 2020). Missing data were treated as missing at random and modeled with full information maximum likelihood estimation.

Structural analysis

To evaluate the higher order structure of the CAT-PD scales, we used exploratory factor analysis (EFA) to test an EFA model that aligned with the consensus five-factor structure and a six-factor model to evaluate whether an Anankastia domain could be recovered. For these analyses, we pooled all samples to maximize the generalizability of our results.

In addition to the number of factors to retain, the choice of rotation algorithm affects the interpretation of the retained factors. All EFA models with a given number of factors in the same data set will have equivalent fit regardless of rotation algorithm used, although some are better suited for certain analytic situations. Here we prioritized oblique target rotation (a.k.a. Procrustes rotation). Standard, mechanical rotation algorithms (e.g., Geomin, oblimax, promax) are designed to search for a simple-structure solution that minimizes the number and magnitude of cross-loadings. However, the higher order structure of personality pathology is not expected to conform to simple structure based on either theory or empirical work. There are numerous, large cross-loadings in nearly all factor-analytic work on personality pathology structure, which likely reflect the true, interstitial nature of many pathological features (Clark & Watson, 2019). For a highly complex true structure like this, mechanical rotation algorithms are at risk of finding inaccurate local solutions (Asparaouhov & Muthén, 2009). Target rotation avoids these shortcomings of mechanical algorithms under the conditions in our study. Namely, target rotation is ideal for when (1) there is theoretical precedence for a general structure, but the specific details are unknown, and (2) the indicators are complex (i.e., have multiple non-zero loadings) (Myers et al., 2015).

With target rotation, solutions are guided by user-selected criteria rather than the unrealistic criteria (for personality pathology) of simple structure. Anchor indicators for hypothesized factors are chosen based on theory/empirical precedence. Each anchor indicator is then targeted to have factor loadings near zero on all factors except on the factor for which it is a primary marker. This approach is different from confirmatory factor analysis in which indicators load onto one factor and cross-loadings are fixed to zero. Neither standard mechanical rotation algorithms nor target rotation are restricted to searching for simple structure or user-specified structure, respectively, but because target rotation explicitly searches for more complex solutions, this approach is more likely to arrive at an accurate solution when the true structure is complex. Indeed, simulation studies have shown that target rotation outperforms Geomin in terms of accuracy in such cases (Asparaouhov & Muthén, 2009; Myers et al., 2015), and the accuracy of target rotation solutions is robust to target error (i.e., targeting a factor loading to be zero when its true value is non-zero; Myers et al., 2013; 2015). Thus, target rotation balances confirmatory and exploratory approaches and accommodated our goal of evaluating two viable structures without imposing unrealistic assumptions of simple structure, while also allowing factors we did not hypothesize to be uncovered.

For the current study, anchor indicators were selected on the basis of the Wright and Simms’ (2014) joint factor analysis of the CAT-PD, PID-5, and NEO-PI-R that reported a five-factor solution. Our selection of anchor indicators was guided by a balance of empirical and rational considerations. We decided to use a joint factor analytic model for our empirical reference because one goal was to evaluate whether the CAT-PD reflects the consensus domain constructs that appear across trait measures. Of the few available joint factor analyses with the CAT-PD, we selected indicators from Wright and Simms’ model due to strengths of the study design noted in the introduction (clinical sample, superior construct coverage) and because the factor solutions had the best face validity (i.e., most interpretable).

Specifically, we used CAT-PD scales that loaded > |.60| on a primary factor and < |.40| on all other factors in this previous study as anchor indicators in our analyses. Non-anchor scales were allowed to load freely onto all factors. These pre-registered thresholds were selected to maximize conceptual clarity and discriminability between factors. Anger, Affective Lability, and Anxiousness scales were used as anchor indicators for Negative Affectivity; Social Withdrawal and Emotional Detachment for Detachment; Irresponsibility, Non-Perseverance, and Non-Planfulness for Disinhibition; Manipulativeness, Hostile Aggression, Domineering, Callousness, and Norm Violation for Antagonism; and Unusual Beliefs and Unusual Experiences for Psychoticism. The six-factor EFA model included a factor targeted to reflect Anankastia, which was not identified in Wright and Simms. Hence, we used CAT-PD scales with the strongest negative loadings on Disinhibition—Perfectionism and Workaholism—as anchors for the sixth, Anankastia factor. Otherwise, the same anchor indicators used in the five-factor model were used for factors one through five in the six-factor model.

Models were evaluated based on global model fit and factor interpretability. Because conventional model-fit thresholds were not developed for models with the number and complexity of indicators in our analyses, we set liberal, minimum thresholds for fit and prioritized interpretability of the factors. We considered models acceptable if they met the following criteria: (1) above conventional thresholds for acceptable global model fit according to at least one alternative fit index (i.e., root mean square error of approximation [RMSEA] < .06, comparative fit index [CFI] > .90, standardized root mean residual [SRMR] < .08; Hu & Bentler, 1999), and (2) all factors were marked by at least two indicators with primary loadings > |.30| and cross-loadings on all other factors at least |.20| lower (i.e., weaker) than the primary loading. We also evaluated the Bayesian information criterion (BIC) to compare models to one another, but this index was not used to determine acceptability. We also report the χ², Tucker-Lewis Index (TLI), and Akaike information criterion (AIC) for completeness, but we did not use these metrics to evaluate our models.

Measurement invariance

Measurement invariance of both models was tested using exploratory structural equation modeling (ESEM; Asparouhov & Muthén, 2009), which allowed us to compare the model fit between groups for the EFA models used to test our principal hypotheses. We examined whether the number of factors (configural invariance) and factor loadings (metric invariance) were invariant across three subgroupings. We tested measurement invariance across sample type (non-clinical/clinical) to ensure research in non-clinical samples can be generalized to clinical research and applications, and across gender (female/male) to ensure any observed gender differences in how personality pathology is expressed (e.g., Lynam & Widiger, 2007) are due to actual differences in the latent trait rather than artifacts of the instrument. Invariance analyses for race (black/white) were also included at the request of a reviewer due to recent evidence that the PID-5 maladaptive trait domains are not invariant across black and white Americans (Bagby et al., 2021). Invariance testing across racial groups was exploratory and was not pre-registered.

Construct validity

After determining the higher order CAT-PD structures, we evaluated the construct validity of the resulting factors by examining their bivariate correlations with multiple structural models that previous research has shown converge with the expected higher order domains; namely, the PID-5 as well as the Big Five traits and the interpersonal circumplex. These analyses were exploratory and were not pre-registered. The purpose of these analyses was not to test specific hypotheses about associations with external criteria, but rather to determine what content of interest to clinicians and researchers is covered by CAT-PD domains. Although exploratory, we had expectations for results that would align with previous work and broadly support the construct validity of the domains. For the PID-5, we expected positive correlations between CAT-PD domains and analogous PID-5 scales. For the Big Five measures, we expected negative correlations between CAT-PD domains and their normative personality trait variants: namely, Antagonism with Agreeableness, Detachment with Extraversion, and Disinhibition with Conscientiousness. For Negative Affectivity, we expected positive correlations with Neuroticism. We expected weaker correlations between Psychoticism and Openness given that their correspondence is tenuous in the broader literature (Watson et al., 2008; Widiger & Crego, 2019). There is far less research on relations between Anankastia and the Big Five compared to the five other maladaptive traits, but on the basis of conceptual grounds and some empirical work (Anissa et al., 2018; Ringwald et al., 2022), we expected Anankastia would be positively associated with Conscientiousness and negatively with Agreeableness. Inclusion of the interpersonal circumplex as a criterion measure was particularly pertinent to the CAT-PD’s construct validity given evidence that it predicts interpersonal problems over and above the PID-5 and therefore may capture unique, interpersonal content relevant to personality pathology (Yalch & Hopwood, 2016). Based on previous research (Wright et al. 2012), we expected all domains to be positively correlated with interpersonal distress, Detachment to be negatively correlated with dominance problems (i.e., more submissiveness problems), and Antagonism to be positively correlated with dominance problems and relatively less correlated with general distress.

For these analyses, we used observed domain scale scores by calculating the mean of the trait scales with primary loadings for the respective domain factor (i.e., loadings ≥ |.30| with no higher loadings on other factors). For scales with substantial cross-loadings, we used the scale to score the domain it had the highest loading on. We report results for observed scale scores in the manuscript to maximize relevance to applied research and clinical settings where scale scores are typically used, but we also report results using domain factor scores in the supplementary materials on OSF (Tables S4 and S5).

Robustness check with Geomin rotation

At the request of reviewers, we also estimated five- and six-factor EFA models using oblique Geomin rotation and compared their structural and construct validity to the models using targeted rotation. These analyses were exploratory and were not pre-registered. Because it could be argued that the target-rotation method based on empirical/theoretical precedence limits the ability to discover the “true” structure of the CAT-PD, these sensitivity analyses using a standard, mechanical-rotation algorithm allowed us to evaluate the robustness of factor solutions from the target-rotated models. Given that global fit is mathematically equivalent for all EFA models with the same number of factors–regardless of rotation algorithm–model solutions were compared based on factor interpretability and correlations with external criteria. Additionally, we indexed similarity in the factors obtained from different rotation methods using congruence coefficients for corresponding factors.

Results

All code and model output for our analyses, as well as descriptive statistics for all measures (Tables S3, S6), are provided in the supplementary materials available on OSF.

Structural analyses with targeted rotation

Both the five- and six-factor models met our pre-registered criteria for global fit and factor interpretability. Model fit statistics are provided in Table 2. The six-factor model had the smaller BIC and therefore fit the data better than the five-factor model. Factors were interpretable and consistent with the expected higher order domains in both models. The five-factor model had strong cross-loadings (i.e., > |.40|) than the six-factor model (6 vs. 1) and more scales without strong loadings (i.e., > |.40|) on any factor (7 vs. 3).

Table 2.

Fit Statistics from Exploratory Factor Analyses with Oblique Target Rotation

Model Fit Statistic	Five-factor Model	Six-factor Model
χ2(df)	7917.657 (373)	6209.219 (345)
RMSEA (90%CI)	.071 (.070, .073)	.065 (.064, .067)
CFI	.903	.925
TLI	.863	.885
SRMR	.029	.023
BIC	904010.574	901955.148
AIC	902620.308	900388.740

Note. RMSEA = Root Mean Square Error of Approximation, CFI = Comparative Fit Index, TLI = Tucker–Lewis index; SRMR = Standardized Root Mean Squared Residual, BIC = Bayesian Information Criterion; AIC = Akaike information criterion.

Factors were well-defined in both models with all marked by at least 3 primary factor loadings (i.e., > |.30|), except for Anankastia in the six-factor model. The only scale not to reach our threshold for a primary loading was Submissiveness in the six-factor model. Primary loadings were similar for analogous factors in the five and six-factor models. Factor loadings are shown in Tables 3 (five-factor model) and 4 (six-factor model). Most of these loadings were as expected, with a few notable exceptions. First, the primary loading of Self-Harm was on Psychoticism instead of Negative Affectivity in both models. Second, scales developed to assess oddity had primary loadings on Disinhibition instead of Psychoticism. Aside from these unexpected loadings, Disinhibition reflected the expected bipolar content in the five-factor model and these respective scales formed distinct Disinhibition and Anankastia factors in the six-factor model. In the five-factor model, Disinhibition had positive loadings of Non-Planfulness, Irresponsibility, Non-Perseverance, Cognitive Problems, Risk Taking, and Peculiarity and negative loadings of Perfectionism and Workaholism. In the six-factor model, Disinhibition was marked by Non-Planfulness, Cognitive Problems, Non-Perseverance, Fantasy Proneness, Irresponsibility, Peculiarity, and Risk Taking whereas Perfectionism and Workaholism became markers of Anankastia, which additionally had strong cross-loadings of Domineering, Rigidity, Fantasy Proneness, and (low) Irresponsibility and Exhibitionism.

Table 3.

Factor Loadings and Intercorrelations for the Five-factor Model.

CAT-PD Scales	NEG		DET		DIS		ANT		PSY
Anxiousness	.74	.91	.19	.02	.10	−.27	−.22	−.05	.18	.07
Affective Lability	.62	.81	.10	−.04	.30	.00	.02	.12	.06	−.05
Perfectionism	.52	.08	−.11	−.08	-.51	-.55	.30	.66	.10	.01
Anger	.52	.53	.12	.04	.13	−.03	.27	.38	−.03	−.13
Relationship Insecurity	.50	.68	.20	.08	.21	−.01	.10	.17	.12	.01
Depressiveness	.48	.79	.48	.31	.22	−.03	−.03	−.04	.05	−.03
Submissiveness	.37	.61	.05	−.05	.25	.03	−.06	.00	.31	.21
Fantasy Proneness	.36	.52	−.11	−.17	.27	.08	.05	.13	.28	.17
Mistrust	.36	.46	.30	.23	.08	−.01	.30	.34	.07	−.02
Health Anxiety	.35	.50	.16	.08	.02	−.15	−.01	.10	.35	.25
Social Withdrawal	.12	.43	.74	.62	.04	.00	.12	.00	.04	.01
Anhedonia	.21	.55	.66	.53	.16	.07	.16	.05	.08	.02
Exhibitionism	.17	−.06	-.57	-.48	.12	.12	.39	.55	.12	.03
Detachment	.11	.33	.48	.40	.13	.11	.20	.09	−.07	−.10
Romantic Disinterest	−.09	.13	.41	.36	−.04	−.01	.02	−.06	.23	.21
Non-Planfulness	.05	.36	−.14	−.18	.67	.58	.25	.10	.05	−.03
Irresponsibility	−.05	.43	.21	.13	.61	.54	.18	−.05	.15	.09
Non-Perseverance	.35	.70	.10	−.02	.56	.33	.08	.00	.04	−.05
Cognitive Problems	.41	.76	.02	−.11	.50	.23	−.05	−.05	.22	.11
Workaholism	.27	.00	−.05	−.02	-.40	-.39	.21	.46	.24	.17
Risk Taking	−.08	.04	−.24	−.20	.38	.42	.37	.31	.20	.12
Peculiarity	.30	.49	.06	−.01	.32	.17	.15	.15	.09	.01
Callousness	−.24	−.12	.36	.40	.07	.30	.70	.57	.16	.09
Domineering	.33	.03	−.17	−.11	−.13	−.08	.67	.88	.07	−.05
Manipulativeness	−.05	.06	.09	.13	.25	.36	.63	.55	.15	.05
Grandiosity	.02	−.10	−.06	.02	−.11	.03	.62	.73	.36	.24
Rigidity	.28	.14	.11	.13	−.11	−.06	.60	.72	.08	−.02
Rudeness	.14	.18	−.06	−.04	.29	.31	.58	.57	.05	−.05
Hostile Aggression	−.07	.02	.02	.07	.12	.23	.57	.57	.39	.27
Norm Violation	−.10	.09	−.04	−.02	.42	.48	.45	.33	.14	.06
Unusual Experiences	−.06	.22	−.01	−.01	.05	.03	.05	.12	.91	.77
Unusual Beliefs	−.15	−.02	−.09	−.04	−.05	.01	.19	.27	.81	.69
Self-Harm	.12	.43	.17	.10	.12	.00	-.03	.00	.60	.49

Factor Intercorrelations
NEG	—					NEG	—
DET	.21	—				DET	.30	—
DIS	.26	.02	—			DIS	.29	.23	—
ANT	.45	.04	.33	—		ANT	.34	.01	.25	—
PSY	.26	.09	.38	.48	—	PSY	.41	.19	.38	.53	—

Note. NEG = Negative Affectivity, DET = Detachment, DIS = Disinhibition, ANT = Antagonism, PSY = Psychoticism. Loadings ≥ .30 are bolded, primary loadings are bolded and underlined. Grey columns are target rotation results, white columns are Geomin rotation results.

As shown in Tables 3 and 4, factors in the six-factor model had stronger intercorrelations than in the five-factor model. Although the size of some correlations among the factors could be considered high, they were lower than PID-5 domain scale correlations in these samples (in the five-factor model, correlations between CAT-PD factors ranged from .00 to .53, with a median value of .30 and in the six-factor model they ranged from .00 to .52, with a median value of .30 whereas correlations between PID-5 scales ranged from .25 to .67, with a median value of .51), suggesting better discriminant validity for the CAT-PD factors than for PID-5 domain scales. In both models, Negative Affectivity and Psychoticism were the factors most strongly correlated with other domains. In the six-factor model, Anankastia was positively correlated with Disinhibition due to shared scale content (e.g., positive loading of Domineering) that may, in turn, reflect shared impairment. Their positive correlation also provides some evidence that Anankastia and Disinhibition are not simply opposite poles of the same domain, as expected by theories that posit a separate Anankastia factor.

Table 4.

Factor Loadings and Intercorrelations for the Six-factor Model.

CAT-PD Scales	NEG		DET		DIS		ANT		PSY		ANA
Affective Lability	.80	.90	−.06	−.13	.20	.04	.03	.05	.03	.00	.01	−.07
Anger	.73	.81	−.08	−.07	.02	−.17	.27	.27	−.08	−.08	.06	.02
Anxiousness	.67	.82	.16	.01	.10	.12	−.27	−.30	.14	.12	.19	.11
Depressiveness	.58	.74	.37	.28	.09	.04	−.01	−.10	.05	.02	−.05	−.08
Relationship Insecurity	.52	.62	.16	.08	.16	.12	.07	.07	.09	.05	.10	.03
Mistrust	.40	.51	.24	.22	.02	−.01	.27	.23	.04	.01	.11	.09
Health Anxiety	.35	.45	.12	.06	.00	.03	−.03	−.02	.32	.27	.12	.07
Social Withdrawal	.07	.23	.79	.71	−.03	.09	.11	−.05	.04	.00	.00	.02
Detachment	−.13	−.04	.69	.58	.19	.36	.11	.02	−.09	−.13	.15	.12
Anhedonia	.33	.47	.57	.52	.01	.01	.19	.07	.08	.04	−.10	−.09
Exhibitionism	.04	−.04	-.48	-.45	.25	.21	.28	.48	.08	.03	.30	.21
Romantic Disinterest	−.01	.08	.36	.36	−.12	−.07	.07	.00	.23	.20	−.09	−.06
Non-Planfulness	.10	.06	−.08	−.11	.63	.52	.27	.40	.06	−.06	−.13	−.24
Cognitive Problems	.20	.26	.19	.01	.60	.68	−.13	−.04	.20	.09	.14	−.01
Non-Perseverance	.27	.32	.20	.07	.56	.55	.04	.09	.04	−.06	.01	−.11
Fantasy Proneness	−.01	.04	.15	−.01	.47	.66	−.11	.00	.24	.14	.36	.22
Irresponsibility	.14	.16	.17	.16	.46	.34	.28	.33	.19	.07	-.30	-.37
Peculiarity	.07	.12	.24	.12	.41	.51	.05	.10	.06	−.02	.20	.10
Risk Taking	−.13	−.18	−.13	−.11	.41	.39	.35	.50	.19	.08	.06	−.03
Callousness	−.14	−.11	.34	.43	.00	−.01	.70	.70	.14	.06	.00	.02
Manipulativeness	.04	.04	.08	.15	.19	.12	.62	.70	.13	.04	.02	−.02
Hostile Aggression	.10	.11	−.07	.04	.04	−.04	.60	.70	.36	.27	.02	.00
Grandiosity	−.01	.01	−.05	.03	−.06	−.03	.54	.64	.30	.24	.28	.25
Rudeness	.17	.16	−.04	−.01	.28	.19	.53	.63	.02	−.06	.11	.05
Domineering	.17	.19	−.12	−.08	−.01	.01	.49	.58	−.01	−.03	.47	.42
Norm Violation	−.01	−.04	−.03	.02	.36	.26	.48	.59	.14	.04	−.08	−.14
Rigidity	.14	.20	.17	.18	−.03	.03	.45	.47	.02	−.01	.38	.35
Unusual Experiences	.00	.06	−.03	−.02	.02	.10	.11	.24	.86	.73	.05	−.01
Unusual Beliefs	−.10	−.07	−.11	−.05	−.07	.01	.22	.36	.76	.66	.09	.06
Self Harm	.25	.34	.08	.06	.01	.02	.04	.08	.58	.50	−.05	−.09
Perfectionism	.11	.19	.06	−.01	−.27	−.03	.03	.04	−.01	.03	.76	.73
Workaholism	−.05	.01	.10	.06	−.21	.01	.02	.04	.16	.17	.54	.52
Submissiveness	.24	.32	.13	.01	.29	.35	−.11	−.05	.28	.21	.14	.04

Factor Intercorrelations
	NEG	DET	DIS	ANT	PSY	ANA		NEG	DET	DIS	ANT	PSY	ANA
NEG	.—						NEG	.—
DET	.40	.—					DET	.52	.—
DIS	.63	.22	.—				DIS	.39	.18	.—
ANT	.37	.08	.42	.—			ANT	.25	.09	.32	.—
PSY	.34	.23	.38	.41	.—		PSY	.40	.27	.40	.47	.—
ANA	.11	−.05	−.03	.17	.12	.—	ANA	.27	.00	.19	.35	.30	—

Note. NEG = Negative Affectivity, DET = Detachment, DIS = Disinhibition, ANT = Antagonism, PSY = Psychoticism, ANA = Anankastia. Loadings ≥ .30 are in bold. Grey columns are target rotation results, white columns are Geomin rotation results.

Measurement invariance

After establishing acceptability of the five- and six-factor structures, we tested measurement invariance across gender (n_female = 2,154; n_male = 1,807), sample type (n_clinical = 352, n_non-clinical = 3,635), and race (n_black= 379; n_white = 2,789). In the first model, we estimated the same EFA models with target rotation separately in each group. This model constrained the number of factors to be equal across groups to test whether the same number of factors represent the CAT-PD higher order structure in different subpopulations (i.e., configural invariance). If this model had acceptable fit, we considered the structure to be invariant. Once configural invariance was established, we constrained the factor loadings to be equal across groups to test whether the higher order constructs were comparable in different subpopulations (i.e., metric invariance). We then compared the fit of this model to the configural model to determine metric invariance. Models were considered invariant across groups if the difference in CFI between the configural and metric models was < .01 (Cheung & Rensvold, 2002). We chose to compare fit using the CFI because the traditional χ² can be oversensitive to ignorable sources of ill fit in large samples such as those used in the current study, and changes in other alternative fit statistics (e.g., RMSEA) have been shown to be less robust to model parameters (Cheung & Rensvold, 2002).

Full model fit results are in Table 5. For all subgroupings considered, model fit for the configural model was good and the CFI difference between the configural and metric models was < .01, supporting measurement invariance across gender, sample type, and race.

Table 5.

Measurement Invariance Models

	Non-Clinical/Clinical		Female/Male		Black/White
Configural Model	Five-factor	Six-factor	Five-factor	Six-factor	Five-factor	Six-factor
χ2	8953.29	7018.62	8302.7	6647.74	7116.98	5739.86
RMSEA	.07	.07	.07	.07	.07	.07
SRMR	.03	.02	.03	.02	.03	.02
CFI	.90	.92	.90	.92	.90	.92
TLI	.86	.88	.86	.88	.86	.88
BIC	904123.55	902254.64	897158.42	895356.69	719799.96	718421.07
AIC	901343.02	899121.83	89438.78	892227.13	717121.06	715402.76
Metric Model	Five-factor	Six-factor	Five-factor	Six-factor	Five-factor	Six-factor
χ2	9204.12	7232.56	8637.98	6897.05	7187.85	5689.11
RMSEA	.07	.06	.07	.06	.07	.06
SRMR	.04	.03	.04	.03	.04	.03
CFI	.90	.92	.90	.92	.90	.92
TLI	.88	.90	.88	.90	.88	.91
BIC	903585.50	901502.38	896578.00	894649.30	719072.02	717477.90
AIC	901685.68	899388.67	89468.16	892537.79	717241.64	715441.45
CFI Δ	.001	.001	.003	.001	.001	.003

Note. RMSEA = Root Mean Square Error of Approximation, CFI = Comparative Fit Index, TLI = Tucker–Lewis index; SRMR = Standardized Root Mean Squared Residual, BIC = Bayesian Information Criterion; AIC = Akaike information criterion; CFI Δ = absolute value for the difference between nested configural and metric model CFIs. CFI Δ < .01 was considered evidence for measurement invariance.

Construct validity

Finally, to evaluate the construct validity of the five- and six-factor structures, we examined correlations with constructs that have well-established associations with the higher order trait domains. These results are shown in Tables 6 and 7. Convergence with expected constructs was generally good and comparable across models. Given the lack of previous research on Anankastia, there were few relevant measures available to test its construct validity and our expectations for its associations with the personality and psychopathology measures were primarily informed by the literature on obsessive-compulsive personality disorder (e.g., Morey et al., 2016; Samuel & Widiger, 2010).

Table 6.

Five-factor Model Correlations

	NEG		DET		DIS		ANT		PSY
Maladaptive Traits (n = 509)
PID-5
Negative Affectivity	.75	.74	.38	.33	.56	.48	.48	.49	.50	.50
Detachment	.57	.61	.76	.75	.46	.37	.40	.35	.37	.37
Disinhibition	.61	.63	.36	.33	.77	.77	.60	.52	.45	.45
Antagonism	.40	.38	.02	.06	.40	.49	.70	.68	.40	.40
Psychoticism	.56	.54	.26	.30	.54	.48	.51	.50	.63	.63
Big Five Traits (n = 1531)
Neuroticism	.73	.75	.49	.34	.42	.28	.30	.29	.37	.37
Extraversion	-.29	-.36	-.67	-.59	-.19	-.05	-.02	.05	-.14	-.14
Conscientiousness	-.34	-.42	-.29	-.23	-.68	-.60	-.33	-.18	-.28	-.28
Agreeableness	-.37	-.38	-.24	-.26	-.36	-.38	-.64	-.57	-.27	-.27
Openness	-.04	-.08	-.24	-.22	.03	.02	-.06	-.03	.03	.03
Interpersonal Circumplex (n = 1209)
IIP-SC
Dominance	.31	.28	-.14	-.13	.24	.30	.50	.53	.25	.25
Affiliation	.30	.27	-.18	-.25	.15	.13	.03	.11	.18	.18
General distress	.68	.69	.51	.48	.46	.35	.37	.38	.43	.43

Note. Grey columns are target rotation results, white columns are Geomin rotation results. External variables were not available in every sample, so correlations with different constructs were examined in subsets of the pooled sample; sample sizes are listed after the construct. Correlations ≥ .30 are bolded. Big Five traits were assessed with different measures including the Big Five Inventory, NEO Personality Inventory - 3, International Personality Item Pool-NEO-120, and Faceted Inventory of the Five-factor Model. Big Five personality trait scores were standardized before combining into the sample subset. NEG = Negative Affectivity, DET = Detachment, DIS = Disinhibition, ANT = Antagonism, PSY = Psychoticism, PID-5 = Personality Inventory for the DSM-5; IIP-SC = Inventory of Interpersonal Problems Short Circumplex.

Table 7.

Six-factor Model Correlations

	NEG		DET		DIS		ANT		PSY		ANA
Maladaptive Traits (n = 509)
PID-5
Negative Affectivity	.75	.75	.29	.40	.60	.61	.48	.48	.50	.50	.29	.06
Detachment	.59	.59	.74	.78	.47	.48	.40	.34	.37	.37	.18	-.03
Disinhibition	.61	.61	.29	.39	.78	.75	.60	.60	.45	.45	.13	-.18
Antagonism	.38	.38	-.04	.10	.45	.40	.70	.72	.40	.40	.24	.06
Psychoticism	.53	.53	.22	.32	.62	.63	.51	.51	.63	.63	.28	.08
Big Five Traits (n = 1531)
Neuroticism	.76	.76	.38	.43	.47	.50	.30	.26	.37	.37	.16	-.02
Extraversion	-.31	-.31	-.68	-.63	-.19	-.21	-.02	.06	-.14	-.14	.04	.15
Conscientiousness	-.37	-.37	-.23	-.29	-.62	-.59	-.33	-.33	-.28	-.28	.30	.53
Agreeableness	-.40	-.40	-.20	-.30	-.37	-.33	-.64	-.61	-.27	-.27	-.11	.05
Openness	-.08	-.08	-.26	-.24	.08	.09	-.06	-.01	.03	.03	.04	.07
Interpersonal Circumplex (n = 1209)
IIP-SC
Dominance	.30	.30	-.20	-.08	.28	.25	.50	.52	.25	.25	.20	.08
Affiliation	.24	.24	-.26	-.21	.19	.21	.03	.07	.18	.18	.15	.08
General distress	.64	.64	.45	.52	.52	.54	.37	.34	.43	.43	.24	.02

Note. Grey columns are target rotation results, white columns are Geomin rotation results. External variables were not available in every sample, so correlations with different constructs were examined in subsets of the pooled sample; sample sizes are listed after the construct. Correlations ≥ .30 are bolded. Big Five traits were assessed with different measures including the Big Five Inventory, NEO Personality Inventory - 3, International Personality Item Pool-NEO-120, and Faceted Inventory of the Five-factor Model. Big Five personality trait scores were standardized before combining into the sample subset. NEG = Negative Affectivity, DET = Detachment, DIS = Disinhibition, ANT = Antagonism, PSY = Psychoticism, ANA = Anankastia; PID-5 = Personality Inventory for the DSM-5; IIP-SC = Inventory of Interpersonal Problems Short Circumplex.

The CAT-PD domains showed good convergence with the PID-5. Most CAT-PD factors correlated > .30 with every PID-5 domain scale, indicating both instruments have somewhat problematic discriminant validity, but the strongest correlations clearly were with the corresponding domains (in the five-factor model, convergent correlations between corresponding domain scores ranged from .69 to .77, with a median value of .72; in the six-factor solution, convergent correlations ranged from .63 to .78, with a median value of .74). Anankastia was an exception to this pattern in that it had relatively low correlations with the PID-5 scales (i.e., all rs < .30).

There was good convergent and discriminant validity in relation to the Big Five personality traits. Antagonism, Detachment, Disinhibition, Negative Affectivity (from both the five- and six-factor models), and Anankastia correlated most strongly with the expected Big Five variants. Psychoticism correlated most strongly with Neuroticism (positively), Agreeableness (negatively), and Conscientiousness (negatively) instead of Openness.

Correlations with the IIP-SC showed most of the expected associations between the CAT-PD domains and the interpersonal circumplex. We present results for associations with the broad domains of dominance, affiliation, and general distress for ease of interpretation, but correlations with the octant scales are provided in the supplementary materials on OSF (Table S7). All domains except Anankastia were correlated with general distress (i.e., rs > .30) indicating that the CAT-PD effectively assesses the cross-cutting interpersonal impairment relevant to personality pathology. In both the five- and six-factor models, Antagonism was most strongly characterized by dominance problems and relatively low distress, in line with previous work. Contrary to expectations, Detachment was only weakly associated with dominance problems (rs = −.14/−.20).

Relative to the other traits, there was less empirical basis for hypothesizing correlations between the available validators and Anankastia or Psychoticism. No samples in our study had convergent measures of Anankastia, but we took advantage of one subsample (Sample 3; n = 299) that assessed more psychoticism-specific measures to explore the construct validity of Psychoticism further. Overall, Psychoticism from both models showed good convergence with measures of dissociation and schizotypal features, which is consistent with prior research (e.g., Watson et al., 2008). Even so, we are hesitant to overinterpret these results because they were from a single sample. These results are in the supplementary materials on OSF (Table S8a,b).

Robustness check with Geomin rotation

To evaluate robustness of the CAT-PD higher order structure, we compared the interpretability and construct validity of factors across target and Geomin rotation algorithms. Overall, factor solutions obtained with EFAs using Geomin rotation were highly similar to those from the EFAs using target rotation. Factor loadings from each rotation method are presented alongside one another in Tables 3 and 4. For the five-factor models, all factor congruences between analogous domains were ≥ .84 (range = .84 – .96) which is near or above the conventional threshold of .85 for considering factors to be the same construct (Fischer & Fontaine, 2011; Lorenzo-Seva & ten Berge, 2006). Complete factor congruence results are reported in the supplementary materials on OSF (Table S9). In the six-factor models, all factor congruences were ≥ .93 (range = .93 to .98). Of note at the individual factor-loading level, Self-Harm remained a primary indicator of Psychoticism but had a stronger, cross-loading on Negative Affectivity than was found in the target-rotated models. Also, Anankastia was marked by the same indicators with the addition of (low) Irresponsibility, and the cross-loadings of Fantasy Proneness and Exhibitionism attenuated to below |.30|.

Correlations with external criteria are presented alongside results for the target rotation models in Tables 6 and 7. The results were highly similar or identical (due to the having the same primary loadings) for scale scores based on either rotation algorithms, with the most notable differences found for Anankastia. The average difference in correlations with external variables was |.03| (five-factor) and |.05| (six-factor) with only three (six-factor, 14 including Anankastia) or four (five-factor) differences > |.10|. Although the factor loadings were very similar for Anankastia across rotation algorithms, due to Irresponsibility having a primary loading on Anankastia in the Geomin models (−.37) versus a strong cross-loading on Anankastia in the target rotated models (−.30) this meant that the observed domain score per the Geomin solution was calculated with this additional scale (given our heuristic of scoring with primary loadings). Including Irresponsibility in the domain score resulted in a stronger correlation with Conscientiousness and lower correlations with all PID-5 domains, Neuroticism, and interpersonal distress suggesting better convergence and discriminant validity.

Discussion

As conceptualization and diagnosis of personality pathology transitions to hierarchical, dimensional models, there is a need to test and refine measures for assessment of multiple trait levels. The lower order traits of the CAT-PD have shown promise to provide better construct coverage than predominant measures, and our results showed that the CAT-PD’s higher order structure aligns with prevailing dimensional models, which further reinforces the CAT-PD’s value for assessing multiple levels of personality pathology. Our study also points to areas in need of further inquiry in the ongoing development of the CAT-PD.

Five or six factors?

When pre-registering this study’s procedures, we anticipated being able to adjudicate clearly between a five and six-factor structure. However, with the exception of the expectably better model fit statistics in the six-factor model, the evidence did not obviously favor one model over the other for the CAT-PD lower order scales. Both structures were interpretable, fit the data well, were stable across rotation algorithms, and were invariant across the groupings we considered. Each structure had its own advantages and disadvantages. The six-factor model had slightly better fit according to the BIC and had fewer cross-loadings, whereas the five-factor model had lower factor intercorrelations. There were “orphan” indicators in both models, but this is not necessarily a weakness, given that these lower order scales have been shown to provide unique information about specific outcomes in previous research (e.g., Yalch & Hopwood, 2016). Correlations with the PID-5, Big Five personality traits, and the interpersonal circumplex indicated that the domains in both models captured constructs that are consistent with the consensus constructs in predominant dimensional trait models. The CAT-PD domains in both models had somewhat poor discriminant validity vis-à-vis the PID-5, but recent work examining each of these instruments has shown that the strong correlations among maladaptive traits reflects general personality dysfunction, suggesting that cross-trait correlations are, to some extent, substantive rather than artifactual (Morey et al., 2020).

Our analyses revealed that the CAT-PD higher order domain structures have several advantages over the PID-5. Factors from both CAT-PD models had somewhat lower intercorrelations compared to the PID-5 domain scales in our samples suggesting that the CAT-PD may have better discriminant validity at the domain level. We also found that the CAT-PD domain structures are invariant across Black and White Americans, whereas recent work has shown that the PID-5 domains are not (Bagby et al., 2021). Our results also show that the CAT-PD Anankastia scales have, at minimum, comparable structural and construct validity to Anankastia scales derived from the PID-5 in previous studies (Bach et al., 2017; Sellbom et al., 2019), with the Geomin rotation scored scale having stronger correlations with the criterion of Conscientiousness, and thus potentially better construct validity, than the PID-5 Anankastia scale (Sellbom et al., 2019). Based on the cumulation of evidence reviewed above, and in accordance with our preregistered criteria for model evaluation, it is our view that the five- and six-factor structures are psychometrically comparable and viable. However, many aspects of model selection are inherently subjective, therefore reasonable people could disagree with our conclusions depending on their priorities and preferences. Our interpretation of the results indicates that the CAT-PD may be adequate for research on the canonical five domains with a bipolar Disinhibition factor or up to six domains including Anankastia. These structures represent two different–but not incompatible–ways of conceptualizing personality pathology, as each may be useful for different applications. For instance, a five-factor model may be useful for studying the interface of pathological and normative personality traits, whereas separating Anankastia from Disinhibition could enable more precise, focused assessment capable of identifying pathology characterized by high Anankastia and high Disinhibition. Although such a trait profile may be uncommon, 2% of participants in our sample scored in the upper deciles (n = 84)—and 7% in the upper quintiles (n = 272)—of both Anankastia and Disinhibition, which is a large enough portion of the population to warrant scientific and clinical consideration. A person with high levels of anankastia and disinhibition may be, for example, extremely perfectionistic and meticulous at work but neglects their familial responsibilities and tends to be non-planful in their social life by forgetting get-togethers that are not work related. We see the ability of the CAT-PD to assess multiple domain-level structures as a strength in that it bridges various theoretical traditions and aligns with diagnostic models in both the DSM-5 Section III and ICD-11. Taken together, our results showing similar five and six factor domain models were consistent across rotation algorithms, invariant across race, gender, and sample type, and had good construct validity indicate that the CAT-PD is psychometrically sound and assesses a wide range of personality pathology content. Although the CAT-PD is a generally strong measure according to our evidence, in the following discussion, we consider how unexpected findings related to Disinhibition, and the relatively weaker validity evidence for Psychoticism and Anankastia, can serve as a blueprint for ongoing CAT-PD scale development and to highlight broader, fieldwide measurement issues.

Disinhibition

Our results showed the empirical organization of lower order CAT-PD traits into higher order domains largely mirrored the organization proposed by expert consensus in the original CAT-PD scale development paper, with the notable exception of scales intended to tap Psychoticism (i.e., Cognitive Problems, Peculiarity, and Fantasy Proneness) having primary loadings on Disinhibition (across all models except in the five-factor Geomin models where they had primary loadings on Negative Affectivity). The fact that the quantitatively derived structure differs slightly from the rationally derived structure underscores the importance of this study to inform how the CAT-PD is used and interpreted in practice. The unexpected loadings are not a fluke of our methods or samples given that they were present across rotation algorithms in this study, and nearly all previous factor analyses of the CAT-PD scales in other samples, using different methods, and combined with different instruments have also found that these three trait scales have primary loadings or strong, secondary loadings on Disinhibition and/or Negative Affectivity rather than Psychoticism factors (Crego et al., 2016; 2018; Long et al., 2021; Thimm et al., 2021; Wright & Simms, 2014).

It becomes clear by examining the item content for these scales how they could reflect distractability and poor executive functioning characteristic of disinhibition rather than oddity. For example, items used to score Cognitive Problems include “I often space out and lose track of what’s going on” and “I have a good memory for things I’ve done throughout the day,” whereas “I sometimes get lost in my daydreams,” and “I am sometimes so preoccupied with my own thoughts I don’t realize others are trying to speak to me” are used to score Fantasy Proneness. Similarly, the Peculiarity scale may be picking up on consequences of having low social inhibitions and not caring about one’s reputation, rather than schizotypy, with items such as “I am considered to be kind of eccentric.” Moreover, Disinhibition scored with these scales was most strongly positively correlated with PID-5 Disinhibition and Neuroticism, and negatively correlated with Conscientiousness which further reinforces that they are capturing disinhibited tendencies. Thus, we think that the bulk of evidence from the present study and others indicates that these scales reflect a blend of Disinhibition and Negative Affectivity despite initial expectations they would reflect Psychoticism. These findings also have implications for assessing the Psychoticism domain with the CAT-PD, which we turn to next.

Psychoticism

We found mixed support for the validity of the Psychoticism domain in the CAT-PD. These findings do not necessarily indicate a problem with the CAT-PD as a measure, given that there is generally less consensus on the nature of Psychoticism, how it maps onto normative personality (Chmielewski et al., 2014; Quilty et al., 2013; Watson et al., 2013), and its placement in dimensional models of personality pathology (Widiger & Crego, 2019). Indeed, the ICD-11 does not consider Psychoticism to be personality pathology, but rather considers it to reflect schizotypal disorder, which is in the “Schizophrenia or other psychotic disorders” chapter. As noted in the above section, traits hypothesized to be indicators of this domain appear to be markers of disinhibition. The composition of the Psychoticism factor across models in our study and its external correlations suggest that this domain taps uniquely psychotic features (i.e., convergence with PID-5 Psychoticism), but also seems to reflect general personality dysfunction. Psychoticism had primary factor loadings of Unusual Beliefs and Unusual Experiences, as expected, but Self-Harm also had a substantial loading on this factor. This may seem surprising on its face, but one thing that is shared among these traits is that they had the lowest endorsement suggesting they represent more severe dysfunction (descriptive statistics showing skews of the CAT-PD scales are in Table S3). Psychoticism also had the highest intercorrelations with other domains, which also suggests that it encompasses cross-cutting pathology. Additionally, Psychoticism correlated evenly with Neuroticism, Agreeableness, and Conscientiousness which are traits that correspond to general personality pathology (Ringwald et al., in press).

This evidence that the CAT-PD’s Psychoticism domain captures general dysfunction to a considerable extent suggests that the factor may have relatively weak discriminant validity that could be improved by adding lower severity Psychoticism items. It is also possible that Psychoticism is a more severe form of personality pathology and only very impaired individuals endorse problems in this domain; thus, low discriminability may be intrinsic to the construct rather than an artifact of the measure. At the heart of the issue is how to separate traits from level of dysfunction, which should be a central focus in the continued development of the CAT-PD in particular and dimensional models of personality pathology in general (Morey et al., 2020; Ro & Clark, 2009; Wright & Hopwood, 2021). Finally, it may be that psychoticism functions this way primarily as a self-report scale in samples with relatively little frank thought disorder and may function differently in samples with more severe and relevant psychopathology.

Anankastia

Although most dimensional models of personality pathology include features of pathological constraint, perfectionism, and rigidity, some conceptualize these features as low Disinhibition (e.g., PID-5), and others as a separate Anankastia domain with only minimal overlap with Disinhibition (e.g., ICD-11). We found evidence that the CAT-PD can assess both conceptualizations. Disinhibition formed a well-defined domain in the five-factor model with some evidence of bipolarity: Workaholism had a primary negative loading (target rotated = -.40/Geomin rotated = −.39) on the factor and the only other negative loading was a cross-loading of −.51/−.55 by Perfectionism, reinforcing how these important aspects of Anankastia can be folded into the Disinhibition domain. In the six-factor model, the lower order traits coherently divided into Anankastia and Disinhibition with only one notable cross-loading between them (Irresponsibility loaded 46 [target rotated]/.34 [Geomin rotated] on Disinhibition and −.30 [target rotated]/−.37 [Geomin rotated] on Anankastia).”

Correlations with other external variables further support that the CAT-PD captures a weakly bipolar Disinhibition and an Anankastia domain that is distinct from Disinhibition. Across the five- and six-factor target and Geomin-rotated models, CAT-PD Disinhibition converged with PID-5 Disinhibition and Conscientiousness. Both rotation algorithms arrived at a similar solution for the Anankastia factor (factor congruence = .96) and both had expected correlations with the available criterion (e.g., Conscientiousness). There were some notable differences in the nature of the factors from each solution, however, despite having a similar pattern of factor loadings. At the latent level, the Anankastia factor in the Geomin models had weaker intercorrelations with the other factors than it did in the targeted-rotated models indicating better discriminant validity. The observed Anankastia scale scored with the Irresponsibility facet per the Geomin model also had better convergent and discriminant validity than scoring it with just Perfectionism and Workaholism per the target-rotated models. Specifically, it had lower correlations with general personality pathology (i.e., PID-5 domains, Neuroticism, interpersonal distress) and a stronger positive correlation with Conscientiousness. Given the lack of external variables related to Anankastia to evaluate its construct validity adequately, future research is needed to test whether the CAT-PD Anankastia scale scored with Perfectionism, Workaholism, and Irresponsibility correlates highly with established Anankastia scales (e.g., Personality Inventory for the ICD-11; Oltmanns & Widiger, 2018; see also Clark et al., 2021, for preliminary ICD-11 scales created from more complete definitions of the model than earlier measures).

Despite these compelling findings, there was less evidence overall supporting the validity of Anankastia compared to the other CAT-PD domains. Anankastia had only two or three primary factor loadings (but several strong, theoretically consistent secondary loadings) and had unexpected positive loadings of Exhibitionism and Fantasy Proneness in the target-rotation models. One reason a more robust, well-defined Anankastia factor did not emerge may be because some aspects of the construct are not assessed by the CAT-PD, despite its relatively comprehensive content coverage. The CAT-PD includes items tapping perfectionism, workaholism, and rigidity about opinions but it lacks content related to risk aversion, perseveration, habit-governed behavior, and excessive orderliness, which also are considered core to this domain (Clark & Krueger, 2010; Clark et al., 2021; Mulder et al., 2016). A next step in CAT-PD scale development could be drawing from existing inventories of Anankastia and obsessive-compulsive personality traits (e.g., Clark et al., 2021; Oltmanns & Widiger, 2018; Samuel et al., 2012) to create items that fill these content gaps, and then testing whether a more robust Anankastia domain could be estimated.

Domain scale scoring

This study was not designed to identify optimal scoring for domain-level scales; however, our results offer a useful empirical reference for this purpose. Based on associations with external criterion, our results suggest that CAT-PD domain scales scored with the top-loading trait scales can effectively assess all six major personality pathology dimensions. Although the primary loadings for the targeted and Geomin-rotated models were very similar or identical, the Anankastia factor in the Geomin models had more primary loadings (3 vs. 2) which makes the scale more reliable, and it had stronger convergent and discriminant validity (as described above). For these reasons, we suggest using primary loadings from the six-factor Geomin models to score domains from the CAT-PD. Specifically, we suggest calculating the mean of the corresponding lower order scales, as was done in this study. We recognize that some scoring based on this recommendation may seem to lack face validity (e.g., Peculiarity for Disinhibition), but on the whole, our results support the construct validity of domains scored with these scales. At the same time, it is important to note that our recommendation is meant as tentative guidance for stimulating use of the CAT-PD rather than the definitive guideline for domain scoring. We strongly encourage more focused scale development research to improve performance of the CAT-PD, and hope that the specific areas for improvement identified by this paper can guide such efforts.

Limitations

This study is the most thorough evaluation of the higher order, domain-level structure of the CAT-PD to date, but it also has limitations. As noted above, despite the CAT-PD’s expanded coverage of personality pathology relative to other instruments, the Anankastia factor we found was not well-defined. Insufficient representation of maladaptive content from both poles of Big Five personality traits in maladaptive trait measures is not an issue only for the CAT-PD and Conscientiousness. For instance, most measures of maladaptive traits lack coverage of maladaptive high Agreeableness (e.g., overly deferent) and Extraversion (e.g., attention-seeking, maladaptively high positive affect and energy) despite evidence that features at these poles relate to personality pathology (Gore et al., 2012; Gore & Widiger, 2015; Lynam, 2012; Watson, Stanton, et al., 2019; Williams & Simms, 2018). It is an open question whether all maladaptive traits at the domain-level are best conceptualized and measured as bipolar or unipolar dimensions or whether polarity depends on the trait; thus, more research is needed to determine whether the CAT-PD would benefit from additional items assessing underrepresented poles or if an asymmetry of bipolar and unipolar dimensions reflects the natural structure of personality pathology. A challenge facing modeling bipolar factors is the shared impairment that generally creates positive manifolds in the correlations among maladaptive scales. This can be circumvented with separate modeling of impairment in a distress or dysfunction factor, as has long been done in the IIP-SC (Alden et al., 1990), but has not been attempted in any comprehensive way in a published five- or six-factor pathological trait inventory.

We did not screen for non-credible, biased responding in our study; therefore, our results may have been affected by under- and/or over-reporting of personality pathology or acquiescence. Consistent with the possibility of bias, research has shown that people self-report more (Oltmanns et al., 2013) or less pathology than informants (Carnovale et al., 2019; Sleep et al., 2019) with discrepancies increasing with level of pathology (Carnovale et al., 2019), and that scales designed to detect biased responding can distinguish groups of individuals who report significantly more or less pathology than “credible” reports (McGee et al., 2016; Sellbom & Bagby, 2008). However, it is impossible to know the extent to which these systematic differences reported in previous work are artifactual or substantive. Thus, to avoid losing potentially valid responses, we did not exclude participants in the pooled sample. Because both under- and over-reporting can attenuate correlations with external variables (Burchett & Ben-Porath, 2010; Dhillon et al., 2017), it was a conservative decision not to screen participants, as our validity results would be strengthened if we removed influence by biased responding. Researchers using the CAT-PD in future research must consider carefully the potential for self-report bias, as is true for any other measure or method, and make their own determination of whether and, if so, how to screen responses.

A strength of our study is that we investigated a pooled sample representing a diversity of sample types, study designs, and a wide range of personality pathology to provide robust and generalizable results. However, all samples were collected in the United States and the majority of the pooled sample was white. To further improve the generalizability of evidence for the CAT-PD’s structure, its psychometrics should be tested across a broader range of geographic, cultural, and racial groups.

The next step towards using the CAT-PD in research, and eventually clinical applications, is to develop reliable, standardized domain-level scales to accompany those for the 33 lower order traits. This would enable the CAT-PD to be used for assessing multiple levels of personality pathology and promote more direct integration with contemporary, dimensional trait models.

Conclusion

Our study adds to the literature showing that the CAT-PD is a valid and uniquely comprehensive measure of personality pathology from its lower order traits to higher order domains. The CAT-PD may be used to assess multiple conceptualizations of maladaptive traits and can thus play an integral role in progressing the fields of psychology and psychiatry towards hierarchical, dimensional models of personality pathology.