A twin study of normative personality and DSM-IV personality disorder criterion counts; evidence for separate genetic influences

Abstract

Background

Both normative personality and DSM-IV personality disorders have been found to be heritable. However, there is limited knowledge about the extent to which the genetic and environmental influences underlying DSM Personality disorders are shared with those of normative personality.

Method

A large population-based sample of adult twins was assessed for DSM-IV personality disorder criteria with structured interviews at two waves spanning a 10-year interval. At the second assessment wave participants also completed the Big Five Inventory, a self-report instrument assessing the five-factor normative personality model. The proportion of genetic and environmental liability unique to the individual personality disorder measures, and hence not shared with the five Big Five Inventory domains was estimated by means of multivariate Cholesky twin decompositions.

Results

The median percentage of genetic liability to the ten DSM-IV personality disorders assessed at wave 1 that was not shared with the Big 5 domains was 64%, whereas for the six personality disorders that were assessed concurrently at wave 2 the median was 39%. Conversely, the median proportion of unique environmental liability in the personality disorders across the two waves was 97% and 96% respectively.

Conclusion

Our results indicate that a moderate to sizable proportion of the genetic influence underlying DSM-IV personality disorders is unshared with the domain constructs of the Big 5 model of normative personality. Caution should be exercised in assuming that normative personality measures can serve as proxies for DSM personality disorders when investigating the etiology of these disorders.

Introduction

Models of normative personality strive to achieve the most parsimonious way of describing individual differences in characteristic patterns of thinking, feeling and behaving. Consensus has converged on a model with five dimensional factors that provided an adequate representation of normative personality (1). According to the Big 5, the main features of normative personality can be summarized by scores on the five primary domains of, Extraversion, Agreeableness, Conscientiousness, Neuroticism and Openness to experience (2).

According to the Diagnostic and Statistical Manual of Mental Disorders (DSM) (3, 4), personality disorders constitute an enduring pattern of inner experience and behavior that deviates markedly from the expectations of the individual’s culture, and is manifested in at least two of the domains; cognition, affectivity, interpersonal functioning or impulse control. The fourth and fifth edition of DSM (DSM-IV and DSM-5 respectively) list criteria sets for the same ten distinct and categorical personality disorders, where diagnosis requires a specific number of criteria to be endorsed. Numerous studies have concluded that DSM personality disorders can be characterized with the five-factor model of personality both conceptually and empirically (5, 6). While the DSM constructs constitutes the personality disorder measures most widely used both by clinicians and researchers, 18 different dimensional models of pathological personality have been published (7). Of the more widely used is self-report instrument Dimensional Assessment of Personality Pathology (8). Psychometric studies comparing the factor structure of the Dimensional Assessment of Personality Pathology to that of normative personality have found strong similarities with respect to the number of underlying domains and their content (9). The notable exceptions being a lack evidence for an “Openness to experience” dimension in pathological personality (10), and a psychoticism dimension in normative personality (11).

Normative personality traits were among the first psychological phenotypes to be studied using genetically informative samples, such as twins, and it is well established that these traits are moderately heritable, with genetic influences accounting for approximately 40–60% of individual differences across the Big 5 domains (12, 13). Only more recently have personality disorders been investigated using genetically informative samples, and results have demonstrated that the heritability of personality disorders as defined by the DSM criteria are similar in magnitude to normative personality (14, 15). Similarities in the extent of genetic influences have fueled speculation that largely the same etiological factors may underlie both normal personality and personality disorders (16). Shared genetic influences are plausible, given the large number of studies that have found overlapping genetic influences between neuroticism and axis-I disorders such as mood and anxiety disorders (17), and the substantial comorbidity typically observed between axis-I disorders and personality disorders (18) attributable to shared genetic risk factors (19). However, to our knowledge, no study has had the appropriate data to directly assess the extent of overlapping genetic etiology in normative personality and DSM personality disorders. The only study to investigate the extent of shared genetic influences between normative personality and a wide set of pathological personality domains (as indexed by the Dimensional Assessment of Personality Pathology inventory) was performed by Jang & Livesley (16), and concluded that there was evidence in favor of a common broad-based genetic architecture. Two studies have estimated the amount of genetic liability to Borderline personality disorder that was unique to this construct, and not shared with the five domains of normative personality. Distel et al analyzed a large twin sample, and found no unique genetic liability in the Personality Assessment Inventory–Borderline Features scale (20). More recently, Kendler, Myers (21) estimated genetic correlations between the Big 5 domains and four Dimensional Assessment of Personality Pathology subscales that were judged to assess the core components of Borderline personality disorder. They found substantial genetic correlations between Borderline personality disorder and Neuroticism, Conscientiousness and Agreeableness. In summary, there are strong phenotypic associations between the Big 5 domains and pathological personality traits, and the limited empirical evidence available suggests that this association may be largely due to shared genetic influences.

The aim of this paper is two-fold; First, to assess the phenotypic similarity between normative and pathological personality, and second, to investigate the extent to which genetic and environmental influences underlying individual differences in normative personality account for symptom variance across all ten personality disorders in DSM-IV.

Methods

Participants

Data for the present study were drawn from two waves of a longitudinal population-based study of mental disorders in Norwegian twins, in which a sample was recruited from the Norwegian Institute of Public Health Twin Panel (22, 23). The first wave of data collection was carried out between 1999 and 2004, at which time the DSM-IV Axis I and Axis II psychiatric disorders were assessed in 2,801 adult twins (44% of those eligible) born between 1967 and 1979. The sample consisted of 2,793 twins with valid data for DSM-IV personality disorders; 220 monozygotic (MZ) male, 117 dizygotic (DZ) male, 449 monozygotic female, 259 dizygotic female and 340 dizygotic opposite sex twin pairs in addition to 23 single twins.

The second wave of data collection was conducted in 2010 and 2011, and to limit the length of interviews, and thus maximize participation, the twins were reassessed only on a subset of the disorders from wave 1. Of the twins that participated in the first wave, 17 had withdrawn their consent to participate in further research, 14 had unknown addresses and 12 had died, leaving 2,758 eligible twins that were invited to participate in a follow up study. After two written reminders and a final telephone contact to non-responders, 2284 twins were interviewed in wave 2 (82.8% of the eligible). The distribution of zygosity groups for the pairs with complete personality disorder data at wave 2 were 154 monozygotic male, 76 dizygotic male, 358 monozygotic female, 179 dizygotic female and 219 dizygotic opposite sex, comprising 986 twin pairs and 312 single twins who participated in the personality disorder interviews.

Zygosity was determined by a combination of questionnaire items and genotyping, and the misclassification rate has been estimated to be less than 1.0%, an error rate unlikely to be a source of bias.

Measures

In both waves personality disorders were assessed using a Norwegian version of the comprehensive Structured Interview for DSM-IV Personality (SIDP-IV) (24). The specific DSM-IV criterion associated with each set of questions is rated using the following scoring format: 0 = “not present”, 1 = “sub-threshold”, 2 = “present”, and 3 = “strongly present”. Behaviors, cognitions, and feelings that were prominent for most of the past 5 years are thought to be representative of an individual’s long-term personality. At wave 1 all 10 the DSM-IV personality disorders were assessed, whereas at wave 2 only six personality disorders (two from each DSM-IV cluster) were reassessed; paranoid, schizotypal, antisocial, borderline, avoidant and obsessive-compulsive.

In wave 1, all but 231 (8.3%) of the interviews were conducted face-to-face, while the remaining were obtained by telephone. At wave 2, all interviews were conducted over the telephone. Interviewers at both waves were mainly senior clinical psychology graduate students or experienced psychiatric nurses, while some were clinical psychologists. Each twin in a pair was interviewed by a different interviewer.

The endorsement rates for the individual personality disorder criteria were in general too low for twin models to be fitted to DSM derived categorical personality disorder diagnostic status. We therefore adopted a dimensional approach, in which we analyzed variables defined as the counts of positively endorsed criteria for each personality disorder. To improve statistical power, we treated criteria endorsed at the subclinical level (i.e. SIDP criteria scored 1 or greater) as being positive. Finally, to lessen the impact of empty cells in the twin contingency tables during model estimation, symptom counts above 3 for each of the personality disorder variables were collapsed. For all personality disorders, this resulted in variable values ranging from 0 to 3. In the first publications from wave 1, (25), we investigated whether; i) subthreshold endorsement of individual criteria on the SIDP interview (i.e. a score of 1) should be considered qualitatively different from scores at a clinical level (rated 2 or 3), and; ii) whether the subthreshold count of endorse criteria (a count less than the clinical threshold as given in the DSM for the various PDs), should be treated qualitatively differently to scores above the clinical threshold. Results from Multiple Threshold tests supported the notion that scores below and above the clinical threshold, both for individual criteria and their counts, represent different levels of severity on the same liability dimension.

The inter-scorer reliability of the SIDP interview was assessed at both waves. At wave 1, 70 interviews were recorded and scored also by a second interviewer, while at wave 2, 95 interviews were recorded and scored by two additional interviewers. Intraclass (ICC) and polychoric correlation between subthreshold PD counts as judged by the different reviewers were calculated. At wave 1 ICC across the PDs ranged from 0.81 to 0.96, (polychoric correlations{0.80–0.99}), while at wave 2 ICC ranged from 0.68 to 0.85 (polychor correlations {0.81–091}).

Normative personality was assessed by the Big Five Inventory (26), a self-report instrument completed by participants at wave 2. The Big Five Inventory is a self-report instrument developed to measure the five prominent domains of normative personality, and consists of 44 items each scored on a five-point scale. Extraversion is represented by 8 items (α=0.85), Agreeableness by 9 (α=0.71), Conscientiousness by 9 (α=0.75), Neuroticism by 8 (α=0.84), and Openness by 10 (α=0.79). The ordinal response options on these items were summed for each of the five domains, resulting in variables that were reasonably normally distributed, and in all subsequent analyses the Big Five Inventory variables were treated as continuous variables.

Statistical Analyses

In order to determine the degree of phenotypic association between the five Big Five Inventory domain sum scores and the subthreshold DSM-IV personality disorder criterion counts, polyserial correlations were estimated. Polyserial correlations are well suited to quantify the association between a continuous and an ordinal variable, and are less prone than Pearson correlation to underestimate this association if the ordinal variable is skewed or contains few categories (27).

The extent of shared genetic variance underlying normative and pathological personality was investigated using a series of multivariate twin models. Twin models allow the variance of an observed phenotype to be partitioned into three sources. The influence of additive genetic factors (A) can be inferred by the extent to which the correlation between MZ twins are twice as large as the correlation between DZ twins. Common environmental influences (C) are those that can be inferred if the correlation between MZ twins is equal in magnitude to the correlation between DZ twins. Any remaining variance in the phenotypes that cannot be accounted for by A or C influences is attributed to a unique environmental component (E), representing factors that contribute to making individuals within both MZ and DZ twin pairs dissimilar. Analogous to the way in which the variance in a phenotype can be partitioned into A, C and E, the covariance between variables can be decomposed similarly using a multivariate twin model. The extent of genetic and environmental overlap between the Big 5 domains and each of the DSM-IV personality disorders was estimated by fitting a series of 6-variate Cholesky twin decompositions to the five Big Five Inventory domains and each of the ten personality disorders measured at wave 1, and the six personality disorders measured at wave 2. The Cholesky decomposition is one of the most widely employed multivariate twin analyses, and contains as many latent A, C and E factors as there are observed variables (28). The first five factors orthogonally contribute to the variance of a given personality disorder that is shared with the Big 5 domains, while the last factor contributes variance that is unique to each personality disorder. Due to the large number of twin pairs required to estimate sex-specific effects, Cholesky path coefficients were constrained to be equal across sex, but separate thresholds and means were estimated for males and females, as there are systematic differences in the mean levels of Big 5 traits and endorsement of personality disorder criteria across sex that may otherwise add an unwanted confound to the interpretation of results from the biometric analyses if not taken into account.

The best fitting models were selected based on the lowest value for Akaike’s Information Criterion, a fit statistic that jointly expresses the parsimony and explanatory power of a model (29).

For each personality disorder, we report the total genetic variance, the proportion that is unique, and the proportion that is shared with each of the Big 5 domain constructs. Genetic and environmental correlations were also calculated and reported. The genetic correlation quantifies the extent to which the genetic variance in two phenotypes is shared.

All statistical analyses were performed in R 3.1.2 (30), and twin analyses were carried out using the free R based OpenMx structural equation package (31), an R extension developed to analyze twin and family data. Model parameters were estimated by means of full information maximum likelihood, an approach that makes use of all observed data. If missing data can be considered to be missing at random, this method returns asymptotically accurate parameter estimates.

Results

Phenotypic correlations

Phenotypic correlations between the Big 5 domains and the personality disorder symptom counts assessed as wave 1 and wave 2, are given in figures 1a and 1b respectively. At wave 1, the Big 5 domain construct with the largest absolute valued median correlation across all ten personality disorders was Neuroticism, with a median correlation (r_m) of .21 (range .08 to .36). This was followed by Conscientiousness (r_m = −.15, range −.28 to −.05), Agreeableness (r_m = −.14, range −0.24 to −.01), Extraversion (r_m = −.13, range −.54 to .15) and Openness to experience (r_m = .08, range −.18 to .15).

Figure 1.

a. Polyserial phenotypic correlations between Big Five Inventory sum scores and personality disorder symptom count variables at Wave 1.

b. Polyserial phenotypic correlations between Big Five Inventory sum score and personality disorder symptom count variables at Wave 2.

For the wave 2 data, the Big 5 domain with the largest absolute valued median correlation across the six personality disorders assessed concurrently was again Neuroticism, r_m = .34, (range.13 to .51), followed in decreasing order by Agreeableness r_m =−.20 (range −.26 to −.09), Conscientiousness, r_m = −.19 (range −.31 to −.06), Extraversion r_m = −.16 (range −.61 to .01) and Openness to experience r_m = .09 (range −.16 to .14).

Heritability, unique to each personality disorder and shared with Big Five Inventory

For all the twin decompositions, dropping all the common environmental parameters (C) resulted in a more parsimonious solution compared with the full ACE or the CE models, as indicated by a lower Akaike’s Information Criterion value. Results from the Cholesky AE models are summarized in table 1. For the wave 1 data, the percentage of genetic variance in the personality disorder traits not shared with the Big 5 domains ranged from 22 % (Avoidant) to 79 % (Schizotypal), with a median of 64%. Conversely, the percentage of unique environmental variance in the personality disorders not shared with the Big 5 domains raged from 89% (Avoidant) to 99% (Schizoid), with a median of 97%.

Table 1.

Heritability (a²) and proportion of unique environmental variance (e²) within each personality disorder, including the proportions of genetic and environmental variance explained by the Big 5 personality domains.

	Genetic Effects			Individual Environmental Effects
	a2	% Shared with Big Five Inventory	% Unique	e2	% Shared with Big Five Inventory	% Unique
Wave 1
Paranoid	0.20	41.5	58.5	0.80	3.5	96.5
Schizoid	0.27	46.8	53.2	0.73	1.2	98.8
Schizotypal	0.27	21.2	78.8	0.73	5.5	94.5
Antisocial	0.41	30.7	69.3	0.59	3.4	96.6
Borderline	0.36	48.3	51.7	0.64	2.3	97.7
Histrionic	0.32	32.6	67.4	0.68	3.4	96.6
Narcissitic	0.24	33.2	66.8	0.76	1.3	98.7
Avoidant	0.35	78.4	21.6	0.65	10.6	89.4
Dependent	0.30	39.3	60.7	0.70	3.7	96.3
Obsessive-Compulsve	0.26	23.7	76.3	0.74	1.4	98.6
Wave 2
Paranoid	0.19	79.4	20.6	0.81	3.2	96.8
Schizotypal	0.29	47.3	52.7	0.71	4.3	95.7
Antisocial	0.37	55.1	44.9	0.63	4.3	95.7
Borderline	0.32	61.1	38.9	0.68	10.8	89.2
Avoidant	0.28	81.8	18.2	0.72	20.9	79.1
Obsessive-Compulsve	0.22	42.1	57.9	0.78	2.4	97.6

Across the six personality disorders assessed at wave 2, the percentage of genetic variance not shared with the Big 5 domains ranged from 18% (Avoidant) to 58% (Obsessive), with median equal to 42%. The percentage of unique environmental variance at wave 2 ranged from 79% (Avoidant) to 98% (Obsessive), with median 96%. On average, the percentage of genetic variance that was unique to the six personality disorder traits assessed at both waves was 59% at wave 1 personality disorder, and 39% at wave 2, when the personality disorders and Big 5 were measured concurrently.

To further facilitate comparisons across the personality disorder traits, additive genetic and environmental variance unique to each personality disorder and the portion shared with the five factors of normative personality is also presented in the form of stacked bar-charts in figures 2a and 2b.

Figure 2.

a. Stacked bar-plot displaying, for each personality disorder, the proportion of genetic and individual specific environmental variance at wave 1 that is shared with the Big Five Inventory factors, and unique to each personality disorder.

b. Stacked bar-plot displaying, for each personality disorder, the proportion of genetic and individual specific environmental variance at wave 2 that is shared with the BFI factors, and unique to each personality disorder.

Genetic correlations

Genetic correlations between the domains of normative personality and the personality disorder traits are presented in figures 3a and 3b. For the wave 1 data, the Big 5 domain with the highest absolute valued median genetic correlation across the ten personality disorder traits was Agreeableness (r=−0.40), followed by Conscientiousness (r=−0.38), Neuroticism (r=0.36), Openness (r=0.20) and Extraversion (r=−0.18). At wave 2, the median genetic correlations across the six personality disorders were, in order of decreasing absolute magnitude, Neuroticism (r=0.56) Conscientiousness (r=−0.54), Agreeableness (r=−0.46), Extraversion (r=−0.28) and Openness (r=0.19).

Figure 3.

a. Genetic correlations between Big Five Inventory sum score and the personality disorder symptom count variables at Wave 1.

b. Genetic correlations between BFI sum score and the personality disorder symptom count variables at Wave 2.

Discussion

Phenotypic correlations

The pattern of correlations between DSM-IV personality disorder criterion counts and the Big 5 domains were largely consistent with previous meta-analyses (5, 32). The Big 5 domain construct of Neuroticism had the strongest association with personality disorder, and like Samuel et al., we observed the highest correlations between neuroticism and borderline, avoidant and dependent personality disorder criterion counts, while the lowest correlations were found for antisocial, narcissistic, histrionic and obsessive. The same pattern was also evident for the wave 2 data. Of the Big 5 domains, Openness to experience displayed the weakest association with personality disorder criterion counts, a finding consistent with results from both previous meta-analyses of DSM personality disorder and normative personality, where no significant correlations were reported between personality disorders and Openness to experience (5, 32). The weak association is also consistent with the lack of an Openness factor reported by psychometric studies of the Dimensional Assessment of Personality Pathology instrument, arguably due to a lack of indicators of Openness (10).

While we found correlations between the Big 5 domains and the personality disorder criterion counts to be higher when assessed concurrently at wave 2, overall the difference was modest, considering that up to ten years separated the waves. A reduction in the strength of association over longer intervals of time is to be expected, for while a high level of temporal stability is commonly reported for normative personality traits (33), the stability for personality disorders is typically found to be lower (34). Any age-specific genetic or transient environmental influence operating at wave 1 will not contribute to shared genetic or environmental variance across the timepoints.

Shared and unique genetic variance

For the personality disorders assessed at wave 2, the average genetic variance unique to the personality disorder criterion counts was 38.9%, suggesting a moderate influence of genetic factors specific to DSM-IV personality disorders. While the present study is the first to estimate the proportion of genetic liability in DSM personality disorders that is shared with normative personality, the limited evidence available from analyses of data based on dimensional models of pathological personality suggest that only modest genetic liability is specific to pathological personality (19, 35). Borderline is the only personality disorder for which the shared genetic variance with the five factor model of personality has been studied more extensively, but not using the DSM criteria. In a large extended twin sample, Distel et al. found that all genetic variance in Personality Assessment Inventory–Borderline Features scale was shared with normative personality, as measured by NEO-FFI (20). In contrast, our twin Cholesky decomposition results indicated that approximately 39% of the genetic variance was unique to DSM-IV Borderline personality disorder criterion count, and hence not shared with the Big 5 domains. Further differences between our results and those of Distel et al. were evident in the genetic correlations between Borderline personality disorder and the normative personality domains. Overall, the genetic correlations between Borderline personality disorder and the Big 5 domains were lower in our sample, and this was especially pronounced for agreeableness and extraversion. While Distel et al. observed genetic correlations of .81 with agreeableness and .62 with extraversion, the absolute valued estimates in our sample were .44 and .25 respectively.

This difference could in part be due to the measures of normative personality used. The Big 5 (26) and the five-factor model (FFM) (33) are taxonomies of personality traits derived through factor analysis, both positing that individual differences can be attributed variability on five broad domains. However, while the Big 5 is rooted in the lexical approach, and based on the investigation of descriptive terms embedded in natural language, the five-factor model is based on the analyses of questionnaire data. The associated measurement instruments, the BFI for the Big 5, and the NEO PI-R for the five-factor model, may therefore differ in genetic or environmental correlation with personality disorders.

The genetic correlations between normative personality and Borderline personality disorder in the present sample are more in agreement with those reported by Kendler et al, where the correlations between Big 5 domains and the core features of Borderline personality disorder as measured by the Dimensional Assessment of Personality Pathology were in decreasing order; neuroticism conscientiousness and agreeableness (21). The modest genetic correlations found in our sample between Openness and all personality disorder criteria counts are consistent with those found between the 18 Dimensional Assessment of Personality Pathology-BQ subscales and the NEO-FFI domains (16).

Across measurement waves, the genetic correlations were largely similar, but the proportion of unique genetic variance in personality disorder traits was approximately 50% lower at wave 2. The lower level of non-shared genetic influences at wave 2 is most consistent with the influence of time specific genetic effects in one or both traits, and is somewhat at odds with the longitudinal stability observed in the phenotypic correlations. While the mean level of personality disorder symptoms is known to decline over time (36), the limited empirical evidence suggests that the underlying genetic influences are relatively stable. For example, Bornovalova et al. found no time specific genetic influences on Borderline personality disorder across four waves spanning the ages 14 to 24 years (37). In conclusion, our results indicate that the etiology underlying DSM-IV personality disorders is not well captured by Big 5 normative personality measures. This is in contrast to PID-5-Norwegian Brief Form (PID-5-NBF) a short-form of the PID-5 (38), dimensional model of personality pathology designed to cover all the maladaptive trait features of DSM-IV-TR personality disorders. A recent publication based on the current sample found PID-5-NBF at an aggregate level to tap the same genetic risk factors as the DSM-5 section II classification for most of the personality disorders (39). Therefore, while Wright et al. have reported an overlap in both phenotypic and genetic correlations between normative personality and PID-5 (35), we believe a reasonable conclusion that follows from the results presented in the present paper and that of Reichborn-Kjennerud et al., is that PID-5 is a better trait representation of DSM personality disorders than the Big Five Inventory.

In conclusion, our results suggest that while the observed association between DSM personality disorder criteria and normative personality is largely due to common genetic rather than environmental influences, a substantial proportion of the genetic risk underlying the endorsement of personality disorder criteria appears not to be shared with normative personality.

Limitations

The interpretation of results presented in this study should be considered in the light of several possible limitations. First, due to the low prevalence of endorsed criteria, we were unable to analyze categorical personality disorder diagnoses. In previous publications we have examined whether the personality disorder criterion count variables are in accordance with an underlying continuous liability to increasing levels of endorsements of the personality disorder criteria, and found this assumption to be satisfied empirically (25). Second, the sample consists of Norwegian twins in a fairly limited age range of adulthood (30–44), and the results may therefore not generalize to other populations. Third, only a subset of DSM-IV personality disorder traits was assessed at wave 2, and we were therefore unable to replicate wave 1 results for all disorders. The median summaries of non-overlapping genetic variance differences may therefore in part be due to four personality disorders not being assessed at wave 2. However, the results for the six personality disorder symptom counts assessed at both waves were very similar. Fourth, while there was evidence for selective attrition from wave 1 to wave 2, this was modest. The full-information maximum likelihood estimation approach employed in the twin analyses is robust against biases due to common types of missing data (40), so the attrition is unlikely to affect the estimates from our analyses. Therefore, we believe that attrition should not impact estimates from our analyses. A final limitation concerns the lack of more explicit modeling of sex-differences. Sex-limited twin models of ordinal data require very large samples to attain sufficient power. However, previous twin studies have failed to find either quantitative or qualitative gender differences for DSM-IV personality disorders and personality traits (13, 14).