Thesis
Research Master Psychology
A Network Model of Comorbidity
in Borderline Personality Disorder
by
Phillip Lino von Klipstein
Student ID 10620818
1. Supervisor: Denny Borsboom
2. Supervisor: Arnoud Arntz
Abstract
Comorbidity in patients with borderline personality disorder (BPD) is very high and linked to impaired functioning. Such comorbidity has been interpreted both as a product of and a challenge to classic disorder conceptualizations. That is, some researchers have proposed that symptom overlap between disorder definitions is an important source of comorbidity, while others have proposed that high comorbidity challenges the notion that disorders constitute empirically distinguishable phenomena. The recently emerged network approach to comorbidity allows us to address both of these perspectives and provides a unique symptom-level approach to comorbidity. The network approach conceptualizes psychopathology as a network of causally related symptoms, where disorders are constituted by symptom clusters and comorbidity arises from symptom relations between these clusters. We employed network analysis to symptom-level data on BPD and highly comorbid Axis-I disorders in a sample consisting largely of BPD patients (N = 773). The results clearly showed strong, ramified relations between symptoms of the same disorder, while relations between symptoms of different disorders were generally sparse and weak, resulting in distinct disorder clusters with few and mostly weak relations between them. These results support the disorder concepts as empirically distinguishable phenomena. Conclusions about the role of symptom overlap were impeded by methodological concerns. We explore the role of single symptoms for BPD comorbidity and discuss implications for comorbidity theories.
Introduction
Borderline personality disorder (BPD) is defined by such severe symptoms as self-mutilating behavior and suicide attempts (American Psychiatric Association, 2000). While BPD by itself thus inflicts a high burden on the individual, it seldom stands alone. Zimmerman and Mattia (1999) found that 70% of their BPD sample were diagnosed with three or more additional Axis I disorders. In a recent review of BPD comorbidity Links and Eynan (2013) thus concluded that “Perhaps the most characteristic of BPD is the finding of multiple Axis I disorders (three or more) comorbid with the BPD diagnosis” (p. 544). Such Axis I comorbidity largely accounts for functional impairment and treatment contact of BPD patients (Coid et al., 2009; Lenzenweger, Lane, Loranger, & Kessler, 2007) and thereby contributes significantly to the burden weighing on them. The scale and impact of comorbidity in BPD make it an essential aspect of understanding this disorder and helping the ones suffering from it.
To understand comorbidity one needs to distinguish between co-occurrence and covariation, as comorbidity may refer to either one. Co-occurrence refers to the simultaneous occurrence of two disorders in the same individual that may transpire without any relation between the disorders. Covariation in contrast refers to a systematic relation between the disorders, where having one disorder increases the likelihood of having a second disorder. Covariation is of higher theoretical interest to research as it implies a systematic relation between the disorders.
Researchers have suggested a number of possible interpretations of the covariation between mental disorders. One line of reasoning is that covariation may not actually exist in nature, but instead may be a methodological artifact of diagnostic classification systems (e.g., Maj, 2005; Neale & Kendler, 1995). More specifically, such reasoning points out that there is considerable overlap between symptoms of different disorders, which may be the source of covariation. While many have argued that symptom overlap likely does not account for most of the systematic covariation of disorders (e.g., Bleich, Koslowsky, Dolev, & Lerer, 1997; Franklin & Zimmerman, 2001; Kessler, DuPont, Berglund, & Wittchen, 1999), recent efforts by Cramer, Waldorp, van der Maas, and Borsboom (2010) as well as Borsboom, Cramer, and Schnittmann (2011) have demonstrated the importance of such symptoms in the covariation of disorders. The first goal of present study was to assess the role of overlapping symptoms in the comorbidity of BPD.
Another approach to interpreting high covariation between disorders suggests that it points towards the existence of common causes or higher order factors of psychopathology (e.g., Andrews et al., 2009; Kraemer, Shrout, & Rubio-Stipec, 2007; Krueger & Markon, 2006). For example, James and Taylor (2008) suggested a hierarchical factor structure incorporating two broad factors (termed internalizing and externalizing) to explain covariation of BPD and a number of other disorders. A third approach suggests that the high covariation between disorders challenges their validity (e.g., Kendell & Jablensky, 2003; Maj, 2005). It argues that high covariation of disorders conflicts with the assumption that disorders represent distinct entities. Thereby, current disorder classification might be ill-defined and research should redefine disorders in terms of separate entities that are identifiable by necessary and sufficient mechanisms. While the latter two approaches differ in their proposals on how to further our concepts of mental disorders, they correspond in their attempt to explain the occurrence of psychopathological symptoms in terms of underlying factors (van Loo, Romeijn, de Jonge, & Schoevers, 2013); an approach that reflects the psychometric framework that underlies the conceptualization of mental disorders: the latent variable model.
The latent variable model assumes the existence of a latent variable (i.e., a disorder) that underlies the occurrence of a set of symptoms. It holds that the latent variable causes symptoms to arise, and thus symptoms in turn can be used to measure the latent variable. Underlying this model is the assumption that symptoms are locally independent, that is, symptoms are independent given the latent variable (Lord & Novick, 1968). Thereby, the model does not allow direct causal links between symptoms, which has been the cause for strong criticism (e.g., Borsboom, 2005; 2008; Borsboom, Mellenbergh, & van Heerden, 2003). Taking a look at single symptoms illustrates the problem with the assumption of local independence. For example, sleep disturbances and fatigue are both symptoms of major depressive disorder according to the Diagnostic and Statistical Manual of Mental Disorders (4th
ed.; DSM-IV; American Psychiatric Association, 2000). Under the latent variable model the covariation of these symptoms is explained entirely by the latent variable. However, it seems logical that there is a direct causal link – not sleeping makes you tired. Analogically, for BPD it seems unlikely that there is no direct causal influence of patients’ emotional instability and impulsivity on the stability of their personal relationships. As past comorbidity research on BPD is based on the latent variable model it focused entirely on investigating covariation on a disorder-level and thereby fails to capture such symptom-to-symptom relations. Capturing them was the second goal of the present study.
While from a strict psychometric perspective the disorder concept used in the current classification systems (i.e., the latent variable model) is problematic, it might still be a legitimate concept in terms of clinical utility. Under such an approach one might conceptualize disorders as an empirical phenomenon, namely sets of typically co-occurring symptoms. The legitimacy of disorders defined in this fashion depends on the extent that they are empirically distinguishable from other disorders. More specifically, such symptom-sets should form clusters with strong relations between symptoms within the same cluster but few and weaker relations between symptoms of different clusters. Strong covariation between disorders, as it can be found in BPD, constitutes a serious challenge to such a disorder concept as it suggests that disorders are not easily empirically distinguishable. The third goal of the present study was to assess whether the symptom-set of BPD forms a cluster that is distinguishable from non-BPD symptoms.
With their introduction of a network approach to psychopathology Cramer et al. (2010) provided methods to address the goals outlined above (also see Borsboom et al., 2011; Borsboom & Cramer, 2013). In this approach psychopathology is conceptualized as a network of causally related symptoms, where mental disorders are formed by clusters of interrelated symptoms. Thereby, comorbidity is no longer understood as relations between disorders, but as direct relations between symptoms of different symptom clusters. Symptom overlap plays an important role in explaining comorbidity under the network approach, because symptoms included in two disorders provide a direct causal link between them, as they send out and receive effects from both of them. Thereby, such symptoms were termed bridge symptoms. Symptom networks can be estimated as Markov Random Fields (Epskamp, Maris, Waldorp, & Borsboom, in press), that provide estimation of the symptom relations with high efficiency and specificity (van Borkulo et al., 2014). Graphical representations of the network – where symptoms are represented as nodes and relations between symptoms as edges between nodes – provide an interpretable visualization of the data structure. They allow an assessment of whether symptoms of the same disorder form a cluster and provide insight into the role of single symptoms in connecting other symptoms or symptom-sets.
Using this approach, the present study implemented a network analysis of DSM-IV BPD symptoms and symptoms of Axis-I disorders highly comorbid with BPD. Specifically the included disorders were generalized anxiety disorder (GAD), post-traumatic stress disorder (PTSD), major depressive episode (MDE), alcohol-related disorders, and substance-related disorders (these disorders show especially high associations with BPD; Coid et al.,
2009; Grant, Chou, Goldstein, Huang, & Stinson, 2008; Lenzenweger et al., 2007). The resulting network was analyzed to explore the character of BPD comorbidity regarding three central aspects: How is BPD symptomology related to highly comorbid disorders on a symptom-level? Do symptoms of the same disorder form distinguishable clusters? What role does content overlap between symptoms play in BPD comorbidity?
Method Sample
The present study used a combined sample comprised of already existing datasets that could provide data on the Borderline Personality Disorder Severity Index, 4th
version (BPDSI-IV; Arntz et al., 2003) and symptom-level ratings on the Structure Clinical Interview for DSM-IV Axis-I Disorders (SCID-I; First, Spitzer, Williams, & Gibbon, 2002). Data from three sources was included: 230 subjects stem from a validation study (Giesen-Bloo, Wachters, Schouten, & Arntz, 2010) of the BPDSI-IV (Arntz et al., 2003); 446 subjects were acquired from a large international multicenter randomized controlled trial of schema therapy for BPD (Wetzelaer et al., 2014); 97 subjects stem from unpublished data acquired at the Riagg Maastricht as part of their routine outcome measurement. For the latter two a BPD diagnosis or sub-threshold BPD was an inclusion criterion, while the first sample included groups with BPD, Cluster C personality disorder, and Axis-I disorders, as well as a group of healthy participants. Exclusion criteria of the schema therapy trial did not apply, as we used data from the screening phase. Exclusion criteria for the Riagg sample were primary diagnoses of bipolar or psychotic disorders. For the divers exclusion criteria used for the different groups in the Giesen-Bloo et al. (2010) study please refer to the original publication.
Measures
BPDSI-IV. The BPDSI-IV is a semi-structured interview comprised of 70 items that
are based on the nine BPD criteria defined in the DSM-IV. Interviewers rate the frequency of specific manifestations of the criteria on an 11-point scale from “never” to “daily”. The BPDSI-IV has shown good reliability and validity in multiple languages and countries (Arntz et al., 2003; Giesen-Bloo et al., 2010; Kröger, Vonau, Kliem, Roepke, & Kosfelder, 2013; Leppänen, Lindeman, & Arntz, 2013). The nine subscales representing the nine DSM-IV criteria were included in the network.
SCID-I. The SCID-I is a semi-structured interview for making DSM-IV Axis I
diagnoses (First et al., 2002). Interviewers rate the presence of DSM-IV symptoms in patients on a 4-point scale (? - not enough information, 1 - absent/false, 2 - subthreshold, 3 - threshold/true). The SCID-I has shown moderate to excellent inter-rater reliability (Lobbestael, Leurgans, & Arntz, 2011). All items that represent DSM-IV symptom criteria or
screening criteria were included in the network. In an effort to only display acute symptoms, SCID-I items were only considered present if symptoms were rated as “present in the last month” (as indicated by chronology specifier items). Ratings on the SCID-I follow a screening method, where ratings for all symptoms of a disorder are only made if certain screening items are rated positively. If screening items are rated negatively, items are skipped and assumed to be absent. Such skipping can further occur among alcohol/substance items, where depending on the amount consumed by the patient, only abuse items or only dependence items are rated. If screening for alcohol/substance was positive and only dependence items were rated, we considered abuse items to be missing. For all other skipped items we assumed “absent” ratings.
Data Analysis
Missing data. Missing data on the SCID-I resulted from “not enough information”
ratings, ordinary missingness, and skipping of abuse items in the presence of dependence items. Missing values were imputed through multiple imputation by a chained equations technique (van Buuren & Oudshoorn, 2000). In this technique, missing values on a specific variable are imputed from a conditional distribution of that variable conditioned on other variables in the data. The technique was implemented because it can handle ordinal and binary data. Imputation was conducted using the R-package mi (Yu-Sung Su, 2011). On account of the ordinal or binary character of SCID-I variables, imputation models were proportional odds logistic regression or Bayesian logistic regression models. Models for each variable incorporated those other variables that were also part of the network and had a correlation with the variable in question of .1 or higher (Bouhlila & Sellaouti, 2013). Ten datasets with imputed values where created this way.
Network estimation. The network model used for the analysis of BPDSI and SCID-I
data was a Markov Random Field (Epskamp et al., in press), where edges between nodes represent partial association. Currently, there are two ways to estimate Markov Random Fields: the Gaussian Random Field (GRF) for data assumed to be normally distributed, and the Ising Model for binary data. To avoid loosing information through dichotomization, a GRF was employed. As SCID-I data is clearly of ordinal character, the estimation of a GRF
for the present data should ideally be based on polychoric and polyserial correlations. However, this was not feasible due to sparsity of the cross tables (i.e., for some symptom pairs there were empty cells in the cross table of observations). Therefore, the GRF was instead based on Pearson-correlations after non-paranormal transformation. To ensure that this approach did not bias results due to the violation of underlying assumptions, an Ising Model was also applied after dichotomizing data. Dichotomization was achieved by setting sub-threshold ratings on the SCID-I to absent ratings and by median split on the BPDSI scales. If the GRF did not bias results, the Ising Model should constitute a subset of the GRF. More specifically, the Ising model should display fewer relations between symptoms because dichotomizing data costs power, but it should not show different relations. For both network estimations model parameters were estimated using the least absolute shrinkage and selection operator (Tibshirani, 1996) with the extended bayesian information criterion (J. Chen & Chen, 2008) for model selection. To further investigate the role of single symptoms in the network, a centrality analysis was conducted. Analyses were conducted using the R-packages
qraph (Epskamp, Cramer, & Waldorp, 2012) and IsingFit (van Borkulo & Epskamp, 2014).
Graphical representations were created using the Fruchterman-Reingold algorithm (Fruchterman & Reingold, 1991).
Results Sample Characteristics
See Table 1 for basic demographics and prevalence rates. Our sample was mostly female (81.1%), on average 33.2 years old, and consisted largely of BPD patients (82.9%). See Table 2 for association measures between diagnoses of disorders included in the network.
Multiple Imputation
After 260 iterations of the mi algorism the sampling distribution of imputed values failed to converge using an 𝑅 statistic of 1.1 as criterion (Gelman, Carlin, Stern, & Rubin, 2004). Failed convergence suggests considerable remaining variation in the imputed values between datasets. However, when estimating networks separately for the imputed datasets, edge weights correlated very highly for both GRFs (Mr = .97, minr = .95) and Ising models (Mr = .91, minr = .86), suggesting that results are stable across imputed datasets. This remained the case when correlating GRF edge weights of variables with large numbers of missings (i.e., alcohol/substance abuse items), Mr = .89, minr = .80. Thereby, a pooled correlation matrix was created by calculating the mean correlations from the imputed datasets
Characteristics full sample BPD patients
N 773 641 Age (mean ± SD) 33.2 ± 10.2 32.5 ± 9.28 Gender (% female) 81.8 86.8 no diagnosis (%) 9.7 0 BPD (%) 82.9 100 MDE (%) 31.4 35.6 GAD (%) 11.4 12.8 PTSD (%) 20.4 24.0 Alcohol abuse (%) 10.3 11.4 Alcohol dependence (%) 4.9 6.8 Substance abuse (%) 2.6 3.0 Substance dependence (%) 7.6 9.0 Table 1
Basic demographics and prevalence rates for full sample and borderline personality disorder (BPD) patients
Note. MDE – major depressive episode; GAD – generalized anxiety disorder; PTSD –
using Fisher’s z-transformation. Correlations matrices in the imputed datasets constituted Pearson correlations after non-paranormal transformation. Standard deviations of pooled correlations were moderate at their largest (maxSD = .11), further demonstrating the stability of results between imputations. The pooled correlation matrix was used as the basis for the central GRF network.
Imputation of missing data on the basis of available data generally serves to strengthen existing relations in the data. To assess whether this may have influenced results, we also estimated networks employing listwise and pairwise deletion of missing data (see Appendix A). Results show that patterns did not differ markedly between methods of handling missing data. Interpretations were based on the GRF of imputed data.
The Network
Figure 1 shows the GRF network. See Table 3 for a legend providing symptom descriptions. The network clearly displays symptoms of the same disorder forming distinct clusters. In these disorder clusters symptoms are well connected among each other and connections within the disorder are markedly stronger than connection to symptoms of other disorders. Connections between symptoms of different disorders are much more sparse with some disorder pairs showing no connected symptoms (e.g., PTSD and GAD, Alcohol
alcohol substance
BPD PTSD GAD MDE abuse dependence abuse dependence
BPD – 8.34 2.61 3.35 2.10 5.00 1.50 8.49 PTSD .54 – 0.82 2.28 0.53 0.75 0.98 1.99 GAD .29 .05 – 1.70 0.41 1.21 0.81 2.19 MDE .43 .34 .20 – 1.05 1.58 1.87 0.89 alcohol abuse .22 -.15 -.18 .04 – 13.49 0.19 0.89 dependence .41 -.04 .04 .17 .64 – 3.28 3.00 substance abuse .27 .13 .07 .21 -.16 .28 – 9.58 dependence .50 .25 .24 .11 .04 .32 .57 – Table 2
Tetrachoric correlations and adjusted odds-ratios between diagnoses of disorders included in the nework
Note. Values above the diagonal represent adjusted odds-ratio from logistic regressions
controlling for the other diagnoses included in the network. Values below the diagonal represent tetrachoric correlations. MDE – major depressive episode; GAD – generalized anxiety disorder; PTSD – post-traumatic stress disorder.
abuse/dependence and depression) and often only a few symptoms per disorder connected to symptoms of other disorders.
On the basis of the imputed datasets 10 Ising Models were estimated. Correlations between edge weights of the Ising Models and the pooled GRF were high, Mr = .82, minr = .81. Further, comparing the Ising Model with the lowest correlation (see Appendix B) to the GRF clearly shows their structural compatibility, while the Ising Model shows fewer relations. Thus, one may justifiably conclude that the Ising Model constitutes a subset of the GRF. F103 F104 F105 F106 F107 F108 F109 F110 F111 F113 F114 F115 F116 F117 F118 F119 F121 F122 F123 F124 F125 F135 F136 F138 F139 F140 F141 F142 F143 A1 A2 A3 A6 A9 A12 A13 A16 A19 alcS E2 E3 E4 E5 E7 E8 E9 E10 E11 E12 E13 E14 subS E36 E44 E52 E60 E68 E76 E84 E148 E156 E164 E172 BPD1 BPD2 BPD3 BPD4 BPD5 BPD6 BPD7 BPD8 BPD9
Alcohol
BPD
GAD
MDE
PTSD
Substance
Alcohol
BPD
GAD
MDE
PTSD
Substance
Figure 1. Gaussian Random Field of Borderline Personality Disorder (BPD) symptoms and
symptoms of related Axis-I disorders, based on pooled correlation matrix. See Table 3 for symptom legend. Alcohol – alcohol-related disorders; GAD – generalized anxiety disorder; MDE – major depressive episode; PTSD – post-traumatic stress disorder; Substance – substance-related disorders.
Borderline personality disorder
BPD1 abandonment BPD2 interpersonal relationships BPD3 identity BPD4 impulsivity BPD5 parasuicidal behavior BPD6 affective instability BPD7 emptiness BPD8 outbursts of anger
BPD9 dissociation or paranoid ideation
Post-traumatic stress disorder
F103 experience of a trauma-like event F104 re-experiencing the trauma-like event or
being upset when reminded of event F105 event involved physical threat to self/others F106 intense fear during the event
F107 recurring recollections of event F108 recurring dreams of event F109 acting/feeling like event recurs F110 distress when exposed to event-cues F111 physiological reactivity to event-cues F113 avoiding thoughts/feelings/conversations
associated with event
F114 avoiding activities/places/people associated with event
F115 inability to recollect aspects of event
F116 diminished interest/participation in activities F117 feeling detached from others
F118 restricted range of affect F119 sense of a foreshortened future F121 sleeping problems
F122 irritability/outbursts of anger F123 concentration problems F124 hypervigilance
F125 exaggerated startle
Generalized anxiety disorder
F135 chronic worrying
F136 no control over worrying F138 restlessness F139 easily fatigued F140 concentration problems F141 irritability F142 muscle tension F143 sleeping problems
Major depressive episode
A1 depressed mood
A2 diminished interest or pleasure A3 weight change
A6 sleeping problems
A9 psychomotor agitation/retardation A12 fatigue
A13 feeling of worthlessness/guilt A16 concentration problems A19 suicidal ideation
Alcohol-related
alcS alcohol screening
Abuse
E2 failure to fulfill major obligations E3 recurrent use in dangerous situations E4 alcohol-related legal problems
E5 recurrent use despite alcohol-related social problems
Dependence
E7 consuming more/longer than intended E8 desire/unsuccessful effort to stop
E9 much time spend obtaining/recovering from alcohol
E10 reducing other activities
E11 continued use despite knowledge of harm E12 tolerance
E13 withdrawal a: physical symptoms
E14 withdrawal b: use to counteract withdrawal
Substance-related
subS substance screening
Dependence
E36 consuming more/longer than intended E44 desire/unsuccessful effort to stop
E52 much time spend obtaining/recovering from alcohol
E60 reducing other activities
E68 continued use despite knowledge of harm E76 tolerance
E84 withdrawal
Abuse
E148 failure to fulfill major obligations E156 recurrent use in dangerous situations E164 substance-related legal problems
E172 recurrent use despite substance-related social problems
Table 3
Description of variables included in the network
The Role of Single Symptoms in the Network
Figure 2 displays centrality measures for all nodes in the network. See Table 3 for a legend providing symptom descriptions. The strength of a node in the network constitutes the sum of edge weights connected to that node. The closeness of a node is a measure of distance to all other nodes in terms of path length, where higher closeness indicates shorter path lengths. The betweenness of a node indicates the proportion of shortest paths between pairs of nodes that include the node in question.
Looking at criteria of BPD, four criteria play a pivotal role in connecting to other disorders: BPD2 interpersonal relationships with both ties to PTSD and substance clusters, BPD4 impulsivity with unsurprising connections to substance and alcohol symptoms (impulsivity scale includes items on substance and alcohol use), BPD5 parasuicidal behavior with an also content related connection to A19 suicidal ideation and the strongest connection to the GAD cluster over F135 chronic worrying, and BPD9 dissociation or paranoid ideation with diffuse relations to PTSD and MDE symptoms. Within the BPD cluster BPD6 affective
instability displays the strongest relations. Among PTSD symptoms the screening item F103 ever experienced a trauma displays the strongest relations to symptoms of other disorders.
Further connections to other disorders consist of relations of F119 sense of a foreshortened
future, F122 irritability/outbursts of anger, and F123 concentration problems to MDE, and
diffuse relations to BPD9 dissociation or paranoid ideation. GAD symptoms show the weakest relations to symptoms of other disorders. These relations mostly consist of multiple weak relations to BPD symptoms, with only F135 chronic worrying displaying a somewhat stronger connection to BPD5 parasuicidal behavior. MDE symptoms show a rather diffuse pattern of relations to BPD symptoms with A19 suicidal ideation and A1 depressed mood displaying the strongest relations. Further, A6 sleeping problems and A16 concentration
problems have connections to mentioned PTSD symptoms and some weak relations to GAD
symptoms exist mostly through A9 psychomotor agitation/retardation. Relations to symptoms of other disorders for alcohol- and substance-related symptoms almost exclusively consist of relations of the screening items subS and alcS, relations between alcohol and substance symptoms, and relations to BPD4 impulsivity.
Betweenness Closeness Strength F103 F104 F105 F106 F107 F108 F109 F110 F111 F113 F114 F115 F116 F117 F118 F119 F121 F122 F123 F124 F125 F135 F136 F138 F139 F140 F141 F142 F143A1 A2 A3 A6 A9 A12 A13 A16 A19 alcSE2 E3 E4 E5 E7 E8 E9 E10 E11 E12 E13 E14 subSE36 E44 E52 E60 E68 E76 E84 E148 E156 E164 E172 BPD1 BPD2 BPD3 BPD4 BPD5 BPD6 BPD7 BPD8 BPD9 −2 0 2 −2 0 2 −2 0 2
Figure 2. Centrality measures for Gaussian Random Field displayed in Figure 1. See Table
Power, Restriction of Range, Skipping
The pattern of low between disorder symptom relations is surprising in light of the high comorbidity of the included disorders and past network analyses including disorders of the network finding more connected patterns. Further, some symptom pairs with obvious content overlap were completely unrelated, for example BPD8 outbursts of anger in BPD and F122 irritability/outbursts of anger in PTSD were unrelated, and F143 sleeping problems and F140 concentration problems in GAD were unrelated to their counterparts in PTSD and MDE. To better understand these results, we investigated whether low power, restriction of range, or the skipping structure of the SCID-I may have contributed to the emerged pattern.
To investigate whether low power may have been a problem we firstly looked at networks of disorder pairs (see Appendix C for pairs PTSD, MDE-GAD, and BPD-MDE). These networks show that between disorder relations remain sparse even for smaller networks with higher power. Further, assuming that our sampling was unbiased, we artificially doubled our sample size by copying the dataset to simulate whether more data would have led to markedly more between disorder relations. The resulting network shows that this was not the case (see Appendix D).
While contemporary comorbidity research is usually conducted using community sample to avoid restriction of range, the largely clinical sample in the present study allows for the possibility that results were biased by restriction of range. To assess whether this was the case, networks were estimated for datasets excluding healthy participants and artificially adding healthy participants (see Appendix E). Excluding the 75 healthy participants from analysis only led to a marginal reduction in between disorder relations. Artificially adding 3000 healthy participants (without any symptom present) in turn only led to a marginal increase in between disorder relations.
To investigate whether the skipping structure of the SCID-I may have biased the network we used a different dataset of arguably similar data that does not involve skipping and artificially imposed a skipping structure. The data consisted of a large sample (N = 2165) of personality disorder patients rated on the Structure Clinical Interview for DSM-IV Axis-II Personality Disorders (SCID-II; First, Gibbon, Williams, Spitzer, & Benjamin, 1997), which is currently used in a different network project. For every personality disorder a random screening item was created with a 30% chance of a negative screening for every subject. If the screening was negative “absent” ratings were imposed on all symptoms of the relevant
disorder. Looking at the change in edge weights between the original network and the network after this procedure showed that within disorder relations increased, whereas between disorder relations decreased (see Appendix F). These results clearly substantiate the presumption that the skipping structure may have influenced results.
Discussion
In the present study we employed network analysis to provide a symptom-‐level perspective on comorbidity in BPD. The resulting network represents an estimate of partial relations between symptoms of BPD, MDE, GAD, PTSD, alcohol-‐related disorders, and substance-‐related disorders. The results clearly showed strong, ramified relations between symptoms of the same disorder, while relations between symptoms of different disorders were generally sparse and weak, resulting in distinct disorder clusters with few and mostly weak relations between them. We elaborate on the role of symptom overlap, the identifiability of disorders as distinguishable symptom clusters, and implications for comorbidity theories. Further, we explore the role of specific symptoms in the network.
Opportunities for comparison with previous research are limited, as, while the idea of a network model of comorbidity has gained traction in the literature (Eaton, 2015), there has only been one previous empirical effort to estimate a comorbidity network (Cramer et al., 2010). In contrast to the present findings, Cramer et al. found symptoms of GAD and MDE to be strongly related with each other. This difference in symptom relations was mirrored by the relations on the disorder-‐level, where the association between MDE and GAD was much lower in our sample than in Cramer et al.’s sample (i.e., the National Comorbidity Survey Replication; Kessler et al., 2008). In light of the low association on the disorder-‐level the absence of relations on the symptom-‐level is not surprising. Other disorder relations in our sample show substantial associations of BPD with other disorders, but many weaker associations between the included Axis-‐I disorders. Comparing disorder associations to estimates from epidemiological literature shows that associations of BPD are comparable with previous estimates (Coid et al., 2009; Grant et al., 2008; Lenzenweger et al., 2007; Westphal et al., 2013; Zimmerman & Mattia, 1999), while associations between Axis-‐I disorders were often lower than previous estimates1 (Compton, Thomas, Stinson, & Grant, 2007; Hasin, Stinson, Ogburn, & Grant, 2007; Kessler, Chiu, Demler, & Walters, 2005; Scott, McGee, Oakley Browne, & Wells, 2006). This pattern suggests the possibility that raters may have made Axis-‐I
1 Note that the meaning of such comparison is limited as estimates generally vary a lot due to the use of different
measures, different variables controlled for, and different reference periods (e.g., 12-month vs. lifetime diagnoses). Further estimates stem almost exclusively from community samples as contemporary comorbidity studies avoid clinical samples to avoid selection bias, while older studies often do not report association statistics.
diagnoses in a satisficing manner (i.e., making one diagnosis leads to disregarding others). However, prevalence rates of Axis-‐I disorders among BPD patient in our sample did not fall outside the range of prevalence rates found in other BPD samples (Coid et al., 2009; Grant et al., 2008; Lenzenweger et al., 2007; McGlashan, Grilo, & Skodol, 2000; Pagura et al., 2010; Westphal et al., 2013; Zimmerman & Mattia, 1999). Thereby, such a bias seems unlikely. In conclusion, while somewhat weak associations among Axis-‐I disorder limit the ability of our network to qualify comorbidity between Axis-‐I disorders, such limitations do not apply to associations with BPD, which are the focus of this paper.
Regarding symptom overlap, our results mostly showed weak or absent relations between symptoms with apparent content overlap (e.g., sleeping problems in MDE, PTSD, and GAD; relations of BPD’s parasuicidal behavior and impulsivity are exceptions). The interpretation of these results is complicated by the use of screening questions in the SCID-‐I and the thereby induced skipping structure. More specifically, the assessment of each disorder starts with a screening question; only if this screening question is rated positively (i.e., symptom present) the other symptoms are assessed; in case of a negative rating, the other symptoms are skipped and considered absent. Thereby, symptoms can only be rated as present in the context of a positive screening and should interpreted accordingly. For example, MDE’s sleeping problems can only be rated as present if previously one of the screening symptoms depressed mood or diminished
interest/pleasure was rated as present. Therefore, it should be interpreted as sleeping problems in the context of depressed mood and/or diminished interest/pleasure. Such a
contextualized interpretation of symptoms ultimately reduces the content overlap between symptoms with different contexts to an unknown degree. For example, adopting a contextualized interpretation, the overlap between sleeping problems in the
context of chronic worrying (GAD) and sleeping problems in the context of depressed mood and/or diminished interest/pleasure (MDE) becomes unclear. Consequently, our results
cannot be adequately used to explore the role of content overlap in BPD comorbidity. High comorbidity poses a challenge to whether disorders offer clinical utility by describing empirically distinguishable phenomena. The exceptionally high comorbidity found for BPD (Links & Eynan, 2013) makes this challenge especially applicable to BPD. Our results show clearly distinguishable symptom clusters that correspond accurately with the DSM-‐IV disorder concepts. However, we demonstrated that the skipping
structure of the SCID-‐I likely contributed to this pattern by inducing common sources of variation for symptoms of the same disorder (see Appendix F to see the influence of imposing a skipping structure with random screening items on a network of personality disorder symptomology). While the present study therefore does not constitute an immensely critical test of disorder concepts, the clarity with which disorder clusters emerged in the network is notable in light of past challenges. These findings demonstrate that even where associations on the disorder-‐level are high (here most notably between BPD and PTSD), disorders may assert as empirically distinguishable phenomena.
Our results on content overlap and distinguishability of symptom clusters can inform the discussion about comorbidity theories, insofar as theories put different emphasis on these issues. Latent variable and network models of comorbidity cannot be distinguished on a statistical level, as they have statistically equivalent representations (Epskamp et al., in press). However, they suggest different causal processes as the origin of comorbidity. These differences in the proposed mechanisms put different emphasis on certain data patterns, which may provide some indication on the plausibility of the models. In the latent variable approach comorbidity arises from processes on the disorder-‐level. The latent disorders cause symptom expression and relations between symptoms are a spurious expression of the latent disorders and their relation. Such latent variables directly imply the existence of clusters in a symptom network, thereby corresponding with the pattern found in our network.
In the network approach symptom relations are not spurious, as they represent direct (although potentially mediated) causal effects between symptoms. If the same symptom is part of two different disorder concepts, it becomes a prime candidate for explaining comorbidity between the disorders, because such a symptom receives and sends out effects to symptoms of both disorders, providing a direct causal link between them. These so-‐called bridge symptoms are an important mechanism to explain comorbidity in the network approach. However, symptoms with apparent content overlap often show little to no relation in our network (BPD’s parasuicidal behavior and
impulsivity are exceptions). This pattern needs to be interpreted considering the
influence of the skipping structure reducing content overlap, as well as potential power issues in light of some low associations on the disorder-‐level. Yet, many relations between symptoms without any content overlap overcame these restrictions. Further,
consider that Cramer et al. (2010) have shown data patterns in line with bridge symptoms despite a skipping structure. Thereby, the rather consistent absence of relations between symptoms with content overlap questions the notion that bridge symptoms are central mechanisms of comorbidity.
Further, the relatively low associations between disorders with considerable symptom overlap constitute a problem for the network approach (most notably MDE and GAD, sharing the symptoms sleeping problems, concentration problems, and fatigue). Under the network approach such overlap constitutes bridge symptoms, which provide a direct causal link between disorders and drive comorbidity. In light of normal disorder prevalence rates and no indication that such symptoms showed low activity2, the low associations between disorders seems to suggest that it is not these symptoms that drive comorbidity. While our results are somewhat at odds with the concept of bridge symptoms, they by no means falsify the network approach, as comorbidity may also arise from causal symptom-‐relations that do not involve bridge symptoms.
Our results further allow us to evaluate the role of specific symptoms within the psychopathology covered in the network. Note that the method of network estimation we employed has very high specificity (van Borkulo et al., 2014), so that present relations in the network with high certainty represent relations in the real world. The absence of relations, however, should not be over interpreted, as both lacking power and the skipping structure may explain them. The following will address the role of single symptoms within the BPD cluster and for BPD comorbidity: In the BPD cluster
affective instability and dissociation/paranoid ideation are the most central symptoms.
While affective instability is generally considered to be a hallmark symptom of BPD (Hooley, Cole, & Gironde, 2012), the strong relations of dissociation/paranoid ideation somewhat surprisingly suggest a quite prominent role. The strong relation to
parasuicidal behavior further substantiates this role, suggesting that
dissociation/paranoid ideation may play a substantial role in the arguably most harmful
symptom of BPD. This is in line with past studies that suggest a direct link between dissociation and self-‐mutilating behavior (Brodsky, Cloitre, & Dulit, 1995; Maaranen et al., 2005; Zlotnick et al., 1996). Also, a number of relations between
dissociation/paranoid ideation and PTSD-‐symptoms suggest a significant role of this
2 E.g., MDE’s sleeping problems, fatigue, and concentration problems showed prevalence rates of 30.8%,
symptom in the comorbidity with PTSD. This finding does not surprise, as the occurrence of dissociative symptoms in PTSD patients has long been recognized (Armour, Karstoft, & Richardson, 2014; Bremner et al., 1992; van der Hart, van Ochten, van Son, Steele, & Lensvelt-‐Mulders, 2008) – so much so that the DSM-‐V includes a dissociative subtype of PTSD (American Psychiatric Association, 2013). Further, the network suggests that the experience of a trauma-‐like event plays an important role in comorbidity between PTSD and BPD, which corresponds to accounts that view trauma as an important etiological influence in BPD (Ball & Links, 2009; Bradley, Jenei, & Westen, 2005; Vermetten & Spiegel, 2014). More specifically, the connection between
experiencing a trauma-‐like event and BPD-‐symptom interpersonal relationships is
consistent with literature showing that traumata negatively affect interpersonal relationships mediated by traumata’s effect on core beliefs (e.g., Biruski, Ajdukovic, & Stanic, 2014).
Relations to the MDE cluster most prominently feature the relation between BPD’s parasuicidal behavior and MDE‘s suicidal ideation, which is likely explained by content overlap, as both measures cover suicide plans and attempts, while parasuicidal behavior further includes automutilation (Arntz et al., 2003; First et al., 2002). Further, BPD’s (fear of) abandonment displayed a connection to MDE’s depressed mood, which may be explained by findings showing that fear of abandonment is associated with increased negative emotional reactions in the face of social rejection (Buckley, Winkel, & Leary, 2004). BPD’s emptiness further shows interesting relations. Although emptiness intuitively seems to have content overlap with MDE’s diminished interest/pleasure, these symptoms showed no relation in the network. While this absence alone should not be overinterpreted, the fact that emptiness instead displayed relations to MDE’s depressed
mood and psychomotor agitation/retardation may indicate that the missing relation is
not a power issue and hint at a qualitative difference between feelings of emptiness in BPD and diminished interest/pleasure in MDE.
The only substantial connection to GAD constituted of the relation between BPD’s
parasuicidal behavior and MDE’s chronic worrying, which corresponds to past findings
showing that suicidal ideation is associated with GAD (Weisberg, 2009). The central connection of BPD to substance-‐use disorders was exhibited by relations between BPD’s
impulsivity and the screening items for substance and alcohol use. These relations are
well accepted as consequences of impulsivity that the BPDSI uses indicators of substance use to measure impulsivity (Arntz et al., 2003). Further relations between BPD’s interpersonal relationships and substance screening, as well as between BPD’s
impulsivity and social problems in substance abuse indicate the importance for social
processes for the comorbidity with substance-‐related disorders. This assertion corresponds to research indicating that BPD patients self-‐medicate in response to social problems (Kruedelbach, McCormick, & Schulz, 1993; Porter, 2008). Further, studies have shown that positive social experiences can serve protecting against substance use (Hussong, Hicks, Levy, & Curran, 2001; Shadur, Hussong, & Haroon, 2015).
The results of our study are most notably limited by the skipping structure of the SCID-‐I, which impedes the evaluation of the role of content overlap and to some degree contributed to the formation of distinguishable symptom clusters. Further, the composition of our sample, consisting mostly of BPD patients, limits generalizability. However, we demonstrated that neither restriction of range, nor limited power significantly influenced the results. Also, the sub-‐sample provided by Wetzelaer et al. (2014) includes data from various international locations.
These limitations directly translate into implications for future research: Firstly, future studies should aim to investigate the structure of psychopathology on a symptom-‐level while refraining from using measures that rely on a skipping structure, as such measures may bias results towards the structure of traditional classification systems. Secondly, the particular composition of our sample notably differs from past studies that challenge the distinguishability of the disorders included in the network. Specifically, both past network estimations showing GAD and MDE symptoms to be strongly intertwined (Borsboom & Cramer, 2013; Cramer et al., 2010) and epidemiological studies finding high comorbidity rates of BPD and among Axis-‐I disorders (e.g., Coid et al., 2009; Grant et al., 2008; Kessler et al., 2005; Scott et al., 2006) are based on community samples. This difference led us to speculate about the possibility that symptom expression at the pathological level might organize differently than at the sub-‐clinical or healthy level. That is, while symptom expression among healthy individuals may show divers relations, symptoms expression at pathological levels may crystallize and become clustered. One could for example imagine that the formation of habits strengthens certain reaction patterns to the extent that alternative reactions are strongly reduced. However, one can just as easily find arguments for the
opposite pattern. That is, relations between symptoms are likely moderated by personal resources, so that patients, who are likely to have fewer resources, should exhibit stronger relations between disorders. Evidence for such moderation can easily be found for symptoms in our network. For example, the relation between negative mood and alcohol use is moderated by social support (Reimuller, Shadur, & Hussong, 2011). These considerations illustrate that differences in the structure of symptom relations between populations are very much conceivable and may provide important insights into the nature of psychopathology. Traditional comorbidity studies provide very little insight into this matter, as they largely restrict themselves to community samples to avoid selection bias and restriction of range. We suggest that future studies use network models to illuminate differences between populations and we provided some initial evidence suggesting that network estimations are rather robust against restriction of range.
Further, through our analysis of single symptoms in the network we were able to qualify their role in the psychopathology of BPD. We identified symptoms that play a central role in the larger system of psychopathology (e.g., dissociation in BPD) and found symptom relations that may constitute causal links between disorders. As our analysis is of an exploratory nature these findings call for further research and validation. While caution is thus warranted, our network in specific and psychopathology networks in general make important contributions to identifying targets for interventions and advancing our understanding of the etiology of psychopathology.
The present study for the first time provides a symptom-‐level perspective on comorbidity in BPD. We provided insights into the relative role of BPD symptoms within the disorder and for connections to Axis-‐I disorders. Further, the present study provides support for the notion that BPD and highly comorbid Axis-‐I disorders constitute empirically distinguishable phenomena despite of high associations between them. While we could not illuminate the role of symptom overlap in BPD comorbidity due to methodological issues, we demonstrated that network analysis is a fitting approach to further this issue. Altogether, the present study serves as a demonstration of the innovation that network analysis and its symptom-‐level perspective can bring to comorbidity research.
References
American Psychiatric Association. (2000). Diagnostic and statistical manual of mental disorders (revised 4th ed.) Washincton, DC: American Psychiatric Association.
American Psychiatric Association. (2013). Diagnostic and Statistical Manual of Mental Disorders (5th ed.). Washincton, DC: American Psychiatric Association.
Andrews, G., Goldberg, D. P., Krueger, R. F., Carpenter, W. T., Hyman, S. E., Sachdev, P., & Pine, D. S. (2009). Exploring the feasibility of a meta-structure for DSM-V and ICD-11: could it improve utility and validity? Psychological Medicine, 39(12), 1993–2000.
Armour, C., Karstoft, K.-I., & Richardson, J. D. (2014). The co-occurrence of PTSD and dissociation: differentiating severe PTSD from dissociative-PTSD. Social Psychiatry and
Psychiatric Epidemiology, 49(8), 1297–1306.
Arntz, A., van den Hoorn, M., Cornelis, J., Verheul, R., van den Bosch, W. M. C., & de Bie, A. J. H. T. (2003). Reliability and Validity of the Borderline Personality Disorder Severity Index. Journal of Personality Disorders, 17(1), 45–59.
Ball, J. S., & Links, P. S. (2009). Borderline personality disorder and childhood trauma: Evidence for a causal relationship. Current Psychiatry Reports, 11(1), 63–68.
Biruski, D. C., Ajdukovic, D., & Stanic, A. L. (2014). When the world collapses: changed worldview and social reconstruction in a traumatized community. European Journal of
Psychotraumatology, 5.
Bleich, A., Koslowsky, M., Dolev, A., & Lerer, B. (1997). Post-traumatic stress disorder and depression. An analysis of comorbidity. The British Journal of Psychiatry, 170(5), 479– 482.
Borsboom, D. (2005). Measuring the Mind: Conceptual Issues in Contemporary Psychometrics. Cambridge University Press.
Borsboom, D. (2008). Psychometric perspectives on diagnostic systems. Journal of Clinical
Psychology, 64(9), 1089–1108.
Borsboom, D., & Cramer, A. O. J. (2013). Network Analysis: An Integrative Approach to the Structure of Psychopathology. Annual Review of Clinical Psychology, 9(1), 91–121. Borsboom, D., Cramer, A., & Schmittmann, V. D. (2011). The small world of
psychopathology. PLoS ONE, 6(11), 1–11.
Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2003). The theoretical status of latent variables. Psychological Review, 110(2), 203–219.
Bouhlila, D. S., & Sellaouti, F. (2013). Multiple imputation using chained equations for missing data in TIMSS: a case study. Large-Scale Assessments in Education, 1(4), 1–33. Bradley, R., Jenei, J., & Westen, D. (2005). Etiology of borderline personality disorder:
disentangling the contributions of intercorrelated antecedents.
and Mental Disease, 193(1), 24–31.
(1992). Dissociation and posttraumatic stress disorder in Vietnam combat veterans.
American Journal of Psychiatry, 149(3), 328–332.
Brodsky, B. S., Cloitre, M., & Dulit, R. A. (1995). Relationship of dissociation to self-mutilation and childhood abuse in borderline personality disorder. The American Journal
of Psychiatry, 152(12), 1788.
Buckley, K. E., Winkel, R. E., & Leary, M. R. (2004). Reactions to acceptance and rejection: Effects of level and sequence of relational evaluation. Journal of Experimental Social
Psychology, 40(1), 14–28.
Chen, J., & Chen, Z. (2008). Extended Bayesian information criteria for model selection with large model spaces. Biometrika, 95(3), 759–771. http://doi.org/10.1093/biomet/asn034 Coid, J., Yang, M., Bebbington, P., Moran, P., Brugha, T., Jenkins, R., et al. (2009).
Borderline personality disorder: health service use and social functioning among a national household population. Psychological Medicine, 39(10), 1721–1731.
Compton, W. M., Thomas, Y. F., Stinson, F. S., & Grant, B. F. (2007). Prevalence, Correlates, Disability, and Comorbidity of DSM-IV Drug Abuse and Dependence in the United States: Results From the National Epidemiologic Survey on Alcohol and Related Conditions. Archives of General Psychiatry, 64(5), 566–576.
Cramer, A. O. J., Waldorp, L. J., van der Maas, H. L. J., & Borsboom, D. (2010). Comorbidity: A network perspective. Behavioral and Brain Sciences, 33(2-3), 137–150. Eaton, N. R. (2015). Latent variable and network models of comorbidity: toward an
empirically derived nosology. Social Psychiatry and Psychiatric Epidemiology, 50(6), 845–849.
Epskamp, S., Cramer, A., & Waldorp, L. J. (2012). Qgraph: Network visualizations of relationships in psychometric data. Journal of Statistical Software, 48(4), 1–18.
Epskamp, S., Maris, G., Waldorp, L. J., & Borsboom, D. (in press). Network Psychometrics. To appear in: P. Irwing, D. Hughes, & T. Booth (Eds.), Handbook of Psychometrics. New York: Wiley.
First, M. B., Gibbon, M., Williams, J. B., Spitzer, R. L., & Benjamin, L. S. (1997). Structured Clinical Interview for DSM-IV Axis II Personality Disorders (SCID-II). Washington D.C.: American Psychiatric Press.
First, M. B., Spitzer, R. L., Williams, J., & Gibbon, M. (2002). Structured clinical interview for DSM-IV-TR (SCID-I)-research version. New York: Biometrics Research.
Franklin, C. L., & Zimmerman, M. (2001). Posttraumatic Stress Disorder and Major Depressive Disorder: Investigating the Role of Overlapping Symptoms in Diagnostic Comorbidity. The Journal of Nervous and Mental Disease, 189(8), 548–551.
Fruchterman, T., & Reingold, E. M. (1991). Graph drawing by force-directed placement.
Software: Practice and Experience, 21(11), 1129–1164.
Gelman, A., Carlin, J., Stern, H. S., & Rubin, D. B. (2004). Bayesian Data Analysis (2nd ed.). London: Chapman & Hall/CRC.