Testing Genetic and Environmental Associations between Personality Disorders and Cocaine Use: A Population-Based Twin Study

Abstract

Aims

To investigate the association between all ten DSM-IV personality disorders (PDs) and cocaine use. Our aim was to determine which PD traits are more strongly linked phenotypically and genetically to cocaine use.

Design

Stepwise multiple and Least Absolute Shrinkage and Selection Operator (LASSO) regressions were used to identify which PD traits best predict cocaine use. Biometrical genetic twin models were then fitted to estimate the size of genetic and environmental associations between the significant PD traits and cocaine use.

Setting

Cross-sectional data were obtained in a face-to-face interview between 1999–2004 as part of a population-based study of mental health.

Participants

Subjects were 1,419 twins (μ_age = 28.2 years, range = 19–36) from the Norwegian Institute of Public Health Twin Panel with complete PD and cannabis data.

Measurements

Lifetime cocaine use and criteria for all 10 DSM-IV PDs.

Findings

In the multiple regression, Antisocial (OR = 4.24, 95%CI = 2.66–6.86) and Borderline (OR = 2.19, 95%CI = 1.35–3.57) PD traits were the only significant predictors of cocaine use. In the LASSO regression, Antisocial, Borderline, and Histrionic were significant predictors of cocaine use. In the twin modelling, Antisocial and Borderline PD traits each explained 72% and 25% of the total genetic risks in cocaine use respectively. Genetic risks in Histrionic were not significantly related to cocaine use. After removing criteria referencing substance use, Antisocial explained 65% of the total genetic variance in cocaine use, whereas Borderline explained 4%.

Conclusions

Among all ten PD traits, Antisocial is the strongest correlate of cocaine use, for which the association is driven largely by common genetic risks.

Introduction

In 2013, 601,000 persons in the US reported using cocaine for the first time within the previous 12 months (1). Although the economic, physiological and psychiatric consequences of cocaine use and misuse are well documented (2–4), the psychiatric antecedents are not well understood. Commensurate with research-based prevention programs that targeting communities and individuals at high risk of substance use, research has begun linking personality disorders (PDs) to licit and illicit substance use (5–12). Approximately 13–15% of adults have at least one PD (13, 14). However, a comprehensive investigation of the genetic and environmental pathways linking pathological personality to lifetime cocaine use has never been reported.

Among the ten DSM-IV PDs (15) consistently linked to cocaine use, Borderline and Antisocial have received the strongest attention and empirical support (16–22). Both of these PDs are associated with increased prevalence of substance use disorders (8–10, 12). While twin studies provide compelling evidence that PDs and cocaine use are heritable (23–29), genetically informative data have not been available until now to elucidate the genetic and environmental sources of comorbidity between all ten PDs and cocaine use.

To our knowledge, no study to date has modelled the association between all ten DSM-IV (15) Axis-II PD traits and cocaine use. It is plausible that within a more integrated model, a number of PD traits may correlate with the liability to cocaine use. We address these gaps by focusing on two aims. The first aim will identify which DSM-IV PD traits best predicts the liability to cocaine use. The second aim is to use genetically informative models to identify and obtain estimates of the genetic and environmental factors that influence the relationship between PD traits and cocaine use.

Methods

Sample

Participants came from the Norwegian National Institute of Public Health Twin Panel (30) comprising twins born from 1967–1979 who were identified through the Norwegian National Medical Birth Registry established in 1967. Data were collected in an interview study between 1999 and 2004, which assessed DSM-IV Axis I and Axis II disorders (15). Of the 3,221 twin pairs eligible for the interview study, there were 1,391 complete pairs (43.2%) and 19 single twins (0.6% pairwise), totalling 2,801 twins who participated (43.4%) (63% female, μ_age = 28.2 years, range = 19–36). All interviewers were advanced psychology students or psychiatric nurses, who received standardised training, and were supervised during data collection. Written informed consent was obtained from all subjects who received stipends of $35 for participation. Ethical approval came from The Norwegian Data Inspectorate and the Regional Ethical Committee.

Measures

Personality Disorders

The DSM-IV (15) personality disorders were assessed using a Norwegian version of the Structured Interview for DSM-IV Personality Disorders (SIDP-IV) (31) comprising: Schizotypal (9 criteria); Schizoid (7 criteria); Paranoid (7 criteria; conduct disorder before age 15 criterion not included); Histrionic (8 criteria); Borderline (9 criteria); Obsessive-Compulsive (8 criteria); Dependent (8 criteria), Avoidant (7 criteria); Narcissistic (9 criteria); and Antisocial (7 criteria). The SIDP-IV used non-pejorative questions organised into topical sections rather than by individual PD, thereby improving the flow of the interview. The SIDP-IV interview was conducted after the Composite International Diagnostic Interview (CIDI) (15, 32) to enable interviewers to distinguish long-standing behaviours from temporary states resulting from Axis I disorders. Each criterion was scored on a 4-point scale (absent, subthreshold, present, or strongly present), which was then dichotomized (0 = absent, 1 = ≥sub-threshold), and summed for each PD. However, because very few subjects endorsed most criteria, each PD sum score was recoded onto a 3-point ordinal scale (0 criteria, 1–2 criteria, ≥3 criteria) (see Table S1 for variable distributions). We have previously tested the validity of this approach by examining the fit of the multiple threshold model in order to determine whether the number of endorsed criteria reflects differences in severity along a single, normally distributed continuum of liability. In brief, this was done using the twin data and testing the fit of the bivariate normal distribution. This assumption is supported for all ten PDs examined (33–35).

Cocaine Use

As part of a Norwegian version of the Composite International Diagnostic Interview, (15, 32), all twins were asked ‘Have you ever taken cocaine’ (No = 1,362, Yes = 57). Cocaine use was 5.2% and 3.5% for males and females respectively, while the average age of most frequent cocaine use was 21.6 years (SD = 3.2, range = 15–28).

Complete PD trait scores were available from 2,793 twins respectively. Data on cocaine use were available from 1,419 twins.

Statistical Analyses

Overview

To identify which PD traits best predict lifetime cocaine use, our first aim applies linear regressions in which all 10 DSM-IV PD traits together with sex and age are entered as predictors of cocaine. We next validate this method using a Least Absolute Shrinkage and Selection Operator (LASSO) regression. Any PDs traits which are significantly predictive of cocaine use are then brought forward and biometrical twin models are fitted to estimate the proportion of genetic and environmental risks shared between the PD traits and cocaine use.

Regressions

We fitted logistic regressions using the generalised linear model glm() function in R_3.1.1 (36) to determine how well each PD trait predicts cocaine use. Our rationale for presenting univariate results is to illustrate the strength of each predictor when other PDs are not taken into account. This was followed by a multiple regression in which cocaine use was regressed onto all ten PDs, again using the glm() function with forward and backward selection in R_3.1.1 (36). We applied this stepwise method to determine which subset of PD traits best predicts cocaine use. All regressions included sex and age as covariates, and in order to account for non-independence introduced by twin data, the standard errors were corrected for clustering.

Least Absolute Shrinkage and Selection Operator (LASSO) regression

Given the widely recognised limitations of stepwise regression (37) we sought to replicate the findings by fitting a 10-fold cross-validated Least Absolute Shrinkage and Selection Operator (LASSO), or penalised regression, using the cv.glmnet() function (38) in R_3.1.1 (36). This method works by adding an L1-penalization term $(\lambda {\Vert \beta \Vert }_1=\lambda {\sum }_{j=1}^p|\beta |)$ to the regression equation (39), where larger λ values correspond to the shrinking of more regression coefficients to 0. We ran the cv.glmnet() function 1,000 times and averaged the error curves. The λ with the smallest Binomial error deviance between predicted and actual observations was then fitted to a final penalised regression again using the glmnet() function (38) in R_3.1.1 (36) to identify the subset of PD traits predictive of cocaine use.

Twin Analyses

PD traits found to predict cocaine use were then brought forward into bivariate twin analyses. All twin models were fitted to the raw ordinal data using Full Information Maximum Likelihood using the OpenMx_2.0 software package (40) in R_3.1.1 (36). This approach assumes that the observed ordinal categories within each variable are an imprecise measure of a latent normal distribution of liability and that this liability distribution has one or more threshold values. Thresholds can be conceived of as cut-points along a standard normal distribution that relate category frequencies to cumulative probabilities indicating increasing levels of risk. All thresholds were adjusted for the fixed effects age and sex.

By exploiting the expected and observed genetic and environmental correlations between monozygotic (MZ) and dizygotic (DZ) twin pairs, standard bivariate biometrical genetic methods (41) were used to estimate the size and significance of the genetic and environmental correlations between each significant PD and the CU and CUD outcomes. The Classical Twin Design assumes that covariance between MZ and DZ twin pairs can be decomposed into additive (A) genetic, shared environmental (C), and non-shared or unique (E) environmental components or risks. Because MZ twin pairs are genetically identical, compared to DZ twin pairs who share on average half of their genes, the expected twin pair correlations for the genetic (A) effects are 1.0 and 0.5 respectively. An important assumption is that common environments (C) are equal in MZ and DZ twin pairs and because non-shared environments (E) are by definition uncorrelated, E must also reflect measurement error. In order to determine the best fitting model, a fully saturated (A+C+E) model was used as a reference to compare models in which the shared environmental, and genetic parameters were dropped to zero.

Computational demands increase with increasing numbers of ordinal variables. Therefore, only bivariate predictors that explained significant genetic, or environmental covariance with cocaine use were included in the multivariate twin analyses. As illustrated in Figure 1, a popular multivariate analysis is the Cholesky triangular decomposition (41), whereby the first variable is assumed to be caused by an unobserved, or latent genetic factor (A₁), which explains variance in the remaining observed variables. Variance in the next observed variable is assumed to be caused by a second latent genetic factor (A₂), which is excluded from explaining variance in the first observed variable. We specified this decomposition for each source of latent risk (A, C, and E). A fully saturated A+C+E Cholesky was used as a reference to compare sub-models in which the C and A parameters were fixed to zero. In both the bivariate and multivariate twin analyses, model comparisons were evaluated using the Akaike Information Criterion (42), which provides a balance between model complexity, and model or data misfit.

Figure 1.

Pathway diagram illustrating a bivariate Cholesky Decomposition for decomposing the genetic and environmental variance-covariance between a personality disorder trait and cocaine use.

Note: A₁ and A₂ denote the unobserved or latent additive genetic risk factors responsible for variation in the observed personality disorder trait and cocaine use respectively. The model assumes an underlying multivariate normal liability, with latent factor means of zero (not shown) and variances of 1 (double-headed arrows). The latent factor contributions are estimated via the pathway coefficients (a₁₁–a₃₃). Latent shared (C_1–3) and non-shared (E_1–3) environmental sources of variance are not shown.

Results

Regressions

With the exception of Schizoid and Obsessive Compulsive, all PD traits in the univariate logistic regressions had significant partial odd ratios greater than one for cocaine use (see Table 1). In the multiple regression with forward and backward selection, the partial regression coefficients and corresponding odds ratios for Antisocial and Borderline PD traits were significantly, and positively predictive of cocaine use. Age was significant such that younger subjects reported more frequent cocaine use. Standard errors corrected for family clustering in the multiple regression are shown in Table S2.

Table 1.

Univariate, multiple stepwise¹, and Least Absolute Shrinkage and Selection Operator (LASSO) regression coefficients.

	Cocaine Use
	Logistic regressions		Multiple regression1		LASSO2
	OR	(95%CI)	OR	(95%CI)
Sex			-	-
Age at interview			0.89	(0.82–0.97)
DSM-IV PD traits:
Paranoid	1.88	(1.25–2.78)	-	-	-
Schizoid	1.10	(0.59–1.93)	0.58	(0.29–1.06)	-
Schizotypal	1.69	(1.02–2.69)	-	-	-
Antisocial	6.05	(3.95–9.45)	4.24	(2.66–6.86)	0.08
Borderline	3.73	(2.49–5.71)	2.19	(1.35–3.57)	0.02
Histrionic	2.50	(1.69–3.73)	-	-	0.01
Narcissistic	1.84	(1.23–2.75)	-	-	-
Avoidant	1.29	(1.87–1.89)	-	-	-
Dependent	1.67	(1.09–2.51)	-	-	-
Obsessive Compulsive	1.25	(0.83–1.89)	-	-	-

Notes: Univariate and multiple stepwise linear regression coefficients (including odds ratio and 95% confidence interval) represent the amount of change in the measures of cocaine use for every one unit of change in the ordinal personality disorder trait scores.

¹Predictors in the multiple stepwise linear regression are based on the best fitting solution with forward and backward selection. Significant predictors are bolded.

²λ = 0.006.

LASSO

As shown in Table 1, LASSO identified three PD traits significantly predictive of cocaine use: Antisocial; Borderline; and Histrionic PD traits.

Bivariate Twin Analyses

The three PDs that were significant predictors in the multiple linear regression or LASSO were then examined in separate bivariate twin-analyses. In each instance, an additive genetic model in which the shared environmental risks were removed provided the most parsimonious fit to the data. See Table S3 for all model comparisons.

As shown in Table 2, the phenotypic (r_P) correlations between the three PD traits and cocaine use ranged from moderate (+0.33) to large (+0.63). Unlike the genetic (r_A) and environmental (r_E) correlations between Borderline and cocaine use, which approximated the level of phenotypic association, the r_A between Antisocial PD and cocaine use was very high (+0.83). Among the three PD traits, the phenotypic correlation between Histrionic and cocaine was the smallest, with a slightly higher genetic correlation.

Table 2.

Phenotypic (r_P), additive genetic (r_A) and environmental (r_E) correlations between each of the significant personality disorder traits and cocaine use, along with standardized proportions of genetic and environmental variance in cocaine use (including 95% CIs) explained by each trait.

				Proportions of variance in cocaine use
	Correlations			Genetic variance				Environmental variance
DSM-IV PD traits:	rP	rA	rE	Shared		Unique		Shared		Unique
Antisocial	0.63	0.85	0.43	41%	(15–73)	16%	(00–43)	8%	(03–25)	35%	(15–62)
Borderline	0.50	0.51	0.50	13%	(01–39)	38%	(09–61)	12%	(01–33)	36%	(15–66)
Histrionic	0.33	0.44	0.27	10%	(00–36)	40%	(05–67)	4%	(00–16)	47%	(23–79)
Antisocial (trimmed)	0.51	0.82	0.26	36%	(11–70)	18%	(00–48)	3%	(00–17)	43%	(20–71)
Borderline (trimmed)	0.32	0.22	0.40	2%	(00–20)	47%	(15–71)	8%	(00–27)	43%	(19–75)

Notes: Results based on the most parsimonious additive genetic (A) + unique environment (E) bivariate model; trimmed = Borderline trait excluded ‘Failure to conform to social norms with respect to lawful behavior as indicated by repeatedly performing acts that are grounds for arrest’, Antisocial trait excluded ‘Impulsivity in at least two areas that are potentially self-damaging (e.g., spending, sex, substance abuse, reckless driving, binge eating)’.

The Histrionic, Borderline and Antisocial PD traits explained 10%, 13% and 52% of the total variance in cocaine use respectively. Effectively, 72% (41/(41+16)) of the total genetic risks in cocaine use were shared with the Antisocial PD trait. In comparison, 25% (13/(13+38)) of the total genetic risks in cocaine use are shared with the Borderline PD trait. Confidence intervals surrounding these genetic covariance estimates were significant for Antisocial PD, and only marginally for significant the Borderline PD trait. For the Histrionic PD trait, despite modest genetic and environmental correlations, the confidence intervals on both the genetic and environmental variance components shared with cocaine use spanned zero.

The Antisocial and Borderline PD trait scores each included criteria referencing substance use or substance use related problems. Therefore, the bivariate analyses were repeated after removing the ‘Failure to conform to social norms with respect to lawful behaviour as indicated by repeatedly performing acts that are grounds for arrest’ and ‘Impulsivity in at least two areas that are potentially self-damaging (e.g., spending, sex, substance abuse, reckless driving, binge eating)’ criteria from Antisocial and Borderline respectively (see Table S1 for variable distributions). There were reductions in the phenotypic, genetic and environmental correlations between ‘drug-free' Antisocial and Borderline PD traits and cocaine use (see Table 2). Correspondingly, Antisocial PD explained 67% (36/(36+18)) of the total genetic variance in cocaine use and remained significant. In contrast, the proportion of total genetic variance in cocaine use explained by the Borderline PD trait dropped to 4% (2/(2+47)) and was no longer significant.

Multivariate Twin Analyses

The Antisocial and Borderline PD traits (with substance use criteria retained) were next entered into a multivariate Cholesky decomposition to predict cocaine use (see Figure S1). There was no evidence that shared environmental (C) risk factors were responsible for any covariance between these measures. These C factors could, therefore, be dropped from the model without any significant change in the −2×log-likelihood (see Table S4). Consequently, additive genetic (A) risk factors best explained the familial aggregation between all three variables. Moreover, additive genetic risk factors unique to cocaine use (see a₃₃ Figure S1) could be dropped from the model without any significant deterioration in model fit (Δ−2LL=0.66, Δdf=1, p=0.42).

As shown in Table 3, the genetic correlation between Antisocial and Borderline PD traits was high, whereas the non-shared environmental correlation was more modest. The genetic correlation between Antisocial PD trait and cocaine use was also significantly higher than the Borderline-cocaine correlation; a model in which these the PD-cocaine genetic correlations were constrained to equal (r_A = 0.73) produced an albeit marginal significant deterioration in fit (Δ−2LL=3.94, Δdf=1, p<0.047). In terms of the non-shared environmental risk factor correlations, the Antisocial and Borderline PD traits were each modestly and comparably associated with cocaine use. As a final check, the variable order was reversed before re-fitting the Cholesky decomposition. We obtained an identical pattern of results, including the same pattern of additive genetic, and nonshared environmental correlations.

Table 3.

Additive genetic (below diagonal) and non-shared environmental (above diagonal) correlations between the latent factors for Antisocial and Borderline personality disorder traits and cocaine use.

	1.	2.	3.
1. Antisocial	1	0.46	0.45
2. Borderline	0.72	1	0.47
3. Cocaine	0.83	0.48	1

Discussion

To our knowledge, this is the first study integrating all ten DSM-IV personality disorder traits within a genetically informative design to predict cocaine use. Multiple regression, LASSO, and twin analyses revealed that Antisocial followed by Borderline PD traits were the strongest phenotypic and genetic correlates of cocaine use. The multivariate twin analysis further illustrated that the risk of cocaine use is significantly linked to correlated genetic risk factors mostly through Antisocial, followed by Borderline PD traits. Although a proportion of the shared genetic risks were driven by PD criteria referencing substance use, when these criteria were excluded, Antisocial still explained two-thirds of the genetic risks in cocaine use. Non-shared environmental risks in Antisocial and Borderline PD were only modestly linked to cocaine use.

Antisocial (43) and Borderline (18) PDs have been linked to cocaine use disorder. Our findings are commensurate with results based on the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC) data showing an approximately three-fold increase in the number of cocaine users who have a diagnosed PD, for whom the probability of transitioning to cocaine dependence is very high (12). The strength of our findings is: (i) the demonstration of the relative importance of Antisocial PD vis-à-vis all other PD traits, even after removing criteria referencing drug use; and (ii) the very high genetic and modest non-shared environmental correlations between Antisocial and Borderline PD traits and lifetime cocaine use.

Recently, we have shown that Antisocial and Borderline PD traits also predict the phenotypic and genetic risks in alcohol and cannabis use and misuse (CUD) criteria (44, 45). Indeed, the correlations between these two PDs is best explained by common and longitudinally stable genetic risk factors (46). Despite cannabis and cocaine having very distinct pharmacological properties, including routes of administration, we speculate that the genetic and environmental risks in Antisocial and Borderline PD traits are indexing the sequelae of behavioural and psychological processes linked to the initiation and maintenance of substance use and misuse in general. This is consistent with PDs as significant indicators of the spectrum of externalising disorders, which is highly heritable (47) and characterised by conduct disorder, licit and illicit substance use and misuse (48, 49), as well as Antisocial and Borderline PDs (50–52). Borderline PD also predicts, at least phenotypically, the liability to facets within the spectrum of internalising disorders (50). In summary, not only do Antisocial and Borderline PD criteria predict cocaine use, the same criteria index the broader genetic liability to substance use, and other correlated psychiatric disorders. Although our modelling is consistent with correlated, non-causal mechanisms underpinning associations between PD traits and cocaine use, we emphasise that we have not tested alternative, causal hypotheses, which may otherwise have direct clinical implications. Such analyses were beyond the scope of this manuscript.

Limitations

Our results should be interpreted in the context of four potential limitations.

First, there was some sample attrition from the original birth registry through to the interview study. In longitudinal studies, attrition reduces statistical power, but will only introduce bias if it is non-random with respect to critical dependent variables (53). Multiple lines of evidence have indicated that the sample is broadly representative with respect to the key areas of interest whereby demographic, and not psychiatric and substance use measures significantly predicted cooperation across assessment waves (53). No psychiatric variables predicted cooperation assessed during an earlier study in 1998. Instead, the strongest predictors of participation were sex, zygosity, and education. Based on examination of 45 variables potentially predictive of cooperation from a richer 1998 survey, including 22 indicators of mental health, only 2 of 45 variables – age and zygosity – significantly predicted cooperation at the interview study (1999–2004), whereas none of the psychiatric variables predicted cooperation. Using the 1998 data, we also fitted standard twin models to 25 variables (including proxies for all ten PDs, five Axis I psychiatric disorders, and alcohol abuse) to determine if results differed between non-subjects and subjects in the interview study. None of the parameters differed significantly. So while some attrition bias remains possible, our use of Full Information Maximum Likelihood is robust to missing data when missingness is random or predicted by other variables in the analyses (54), which means that attrition was unlikely to have biased our results.

Second, there were 91 complete and 164 incomplete (singletons) opposite-sex DZ twin pairs with cocaine use data. Consequently, the lack of statistical power precluded investigating sex differences in the additive genetic risk factors (41).

Third, the study relied on Norwegian adults. Consequently, variation and replication of our results are required to determine if they generalise to different age groups and sample populations.

Finally, administration of the substance use items was contingent upon the response to, “Are you prepared to speak openly about this subject?” Substance use was higher among cooperative twins whose co-twin was unprepared to speak openly about his/her history of substance use. The Antisocial and Borderline bivariate analyses were, therefore, re-run, wherein cocaine use was contingent upon ‘speaking openly’ (see Figure S2). There were declines in the phenotypic (+0.63 to +0.44) and genetic (+0.85 to +0.73) correlations between Antisocial PD criteria and cocaine use. For Borderline PD criteria, the phenotypic correlation declined (+0.50 to +0.45), whereas the genetic correlation (+0.51 to +0.54) increased marginally. We, therefore, conclude that this contingency had minimal impact on the results.

Conclusion

Among the ten DSM-IV PDs, Antisocial followed by Borderline PD traits are the strongest phenotypic and genetic correlates with lifetime cocaine use. Associations between these PD traits and cocaine use are driven largely by common genetic risk factors. Future twin analyses will test the direction of causation between these PD traits and cocaine use.