Latent structure analysis of DSM-IV borderline personality disorder criteria

Latent Structure Analysis of DSM-IV Borderline Personality Disorder Criteria Andrea Fossati, Cesare Maffei, Maria Bagnato, Deborah Donati, Caterina Namia, and Liliana Novella The aim of this study was to evaluate the structure of DSM-IV borderline personality disorder (BPD) criteria. The study group consisted of 564 consecutively admitted inpatients and outpatients. BPD criteria discriminatory power was tested by using corrected i t e m - t o t o t a l and item-to-diagnosis correlations. Weighted least-squares (WLS) confirmatory factor analysis (CFA) was used to assess the fit of DSM-IV BPD unidimensional model. The categorial model of BPD was tested by exploratory latent class analysis (LCA). Item analysis suggested a hierarchy in BPD criteria discrimina-

tory power, even if with different rank order with respect to the DSM-IV model. CFA showed a unifactorial structure with congeneric items as the best fitting model for DSM-IV BPD criteria (X2 = 18.89, d f = 27, P > .87), LCA showed evidence for three latent classes; heterogeneity was observed only among subjects falling below DSM-IV diagnostic threshold for BPD. These results support the categorial model of BPD, even if with several differences with respect to

ORDERLINE PERSONALITY DISORDER (BPD) is one of the most studied personality disorders (PDs). 1,2 DSM-III3 and DSM-III-R4 described BPD as a unidimensional, categorial PD. It should be stressed that unidimensionality does not mean homogeneity. A unidimensional model of BPD simply means that all the diagnostic criteria measure, and belong to, a single diagnostic entity (namely, BPD), even if with different prevalence and diagnostic efficiency. None of the criteria can be considered as pathognomonic, i.e., as the necessary and sufficient condition for diagnosing the disorder. With respect to diagnostic criteria, the unidimensional model hypothesizes a certain degree of variability (i.e., heterogeneity') within BPD subjects, but not the existence of distinct subpopulations of BPD. Moreover, DSM-III-R considered BPD, as well as several other mental disorders, as a diagnostic class with clinical features which make it clearly recognizable and establish clear boundaries with respect to other axis I and II diagnoses (i.e., as a categorial construct). However, this model of BPD raised much contro-

versy. Some studies based on exploratory factor analysis (EFA) or cluster analysis questioned the assumption of unidimensionality of BPD, suggesting the presence of three to four subsets of DSMIII-R BPD criteria. 57 These data suggested the presence of clinical heterogeneity within the BPD criteria set, as well as within BPD patients. 8 However, the structure of BPD criteria subsets could not be completely replicated between different studies due to different BPD assessment, sample size and characteristics, and statistical analyses. Moreover, none of these studies statistically assessed the goodness-of-fit of a specific BPD multidimensional model (i.e., distinct subsyndromes of BPD), as well as its superiority to the DSM-III-R unidimensional model. A second criticism concerns the categorial model of BPD adopted by DSM-III-R. Several authors questioned the validity of the categorial model of PDs, suggesting a shift towards dimensional conceptualization and assessment. 9-13 Few studies were performed specifically to test the hypothesis of BPD categorial structurel4,15; their results were contrasting, 13 because they did not support the categorial hypothesis of BPD. At the same time, they did not find clear evidence of a dimensional model. Many taxometric techniques were used, ranging from cluster analysis to admixture and maximum covariation analysis, u,13 Latent class analysis (LCA), a statistical method particularly useful in detecting latent taxons when categorial manifest variables (such as DSM-III-R PD criteria) are used, 16was never applied to BPD criteria. 13 It should be stressed that these studies were highly relevant, because they showed that the categorial model of necessary and sufficient conditions was not optimal for BPD, and raised the

B

From the Medical Psychology and Psychotherapy Unit, lstituto Scientifico Ospedale San Raffaele, Department of Neuropsychiatric Sciences, University of Milan School of Medicine, Milano; and Institute of Psychology, University of Urbino, Urbino, Italy. Presented in part at the 5th International Congress on the Disorders of Personality, June 25-27, 1997, Vancouver, BC, Canada. Address reprint requests to Professor Cesare Maffei, Medical Psychology and Psychotherapy Unit, Istituto Scientifico Ospedale San Raffaele-DSNP, via Stamira d'Ancona 20, 1-20127 Milano, Italy. Copyright © 1999 by W.B. Saunders Company 0010-440X/99/4001-0003510.00/0 72

DSM-IV, Copyright © 1999 by W.B. Saunders Company

Comprehensive Psychiatry, Vol. 40, No. 1 (January/February), 1999: pp 72-79

LATENT STRUCTURE OF BORDERLINE PD

scientific inquiry on this topic to a higher level. Their results suggested the need for further studies, based on different statistical techniques, and played a major role in the aims and design of this study. With the introduction of the DSM-IV, t7 the debate on these topics raised to a deeper level. Despite the suggestion of adopting a dimensional model, ]1,]3 DSM-IV maintained a categorial model for BPD, as well as for other PDs. According to criticisms to DSM-III-R polythetic format l] and psychometric studies of the diagnostic efficiency of the individual PD criteria, 18 polythetic PD diagnoses were slightly modified in DSM-IV by entering a hierarchy in PD criteria. Research findings played a major role in entering a new BPD diagnostic feature (stress-related paranoid ideation or severe dissociative symptoms). Starting from these considerations, the present study was performed to provide the following: (1) an analysis of the diagnostic efficiency of the individual DSM-IV BPD criteria; (2) a test of the fit of DSM-IV unidimensional model of BPD, and its superiority to other proposed models; and (3) an evaluation of the presence and number of latent taxons underlying the DSM-IV BPD criteria. METHOD The study group consisted of 564 subjects consecutively admitted from January 1995 to May 1996 to the Medical Psychology and Psychotherapy Unit of the Scientific Institute H San Raffaele of Milan, Italy. None of these subjects met any of the following exclusion criteria: (1) DSM-IV axis I diagnosis of schizophrenia, schizoaffective disorder, delusional disorder, or delirium, dementia, amnestic, and cognitive disorder not otherwise specified (NOS); (2) IQ -< 75; or (3) education level lower than elementary school. Two hundred thirty-nine (42.4%) subjects were male and 325 (57.6%) female; mean age was 29.92 (SD = 8.50) years. Three hundred sixty-eight (65.2%) were inpatients and 194 (34.8%) outpatients. Four hundred eighteen (74.2%) subjects received at least one DSM-IV axis I diagnosis; most frequently diagnosed DSM-IV axis I disorders were anxiety disorders (n = 178, 31.6%), eating disorders (n = 93, 16.5%), mood disorders (n = 63, 11.2%), substance abuse/dependence disorders (n = 59, 10.5%), and brief/NOS psychotic disorder (n = 35, 6.2%). Twenty-six subjects (4.6%) received other axis I diagnoses (e.g., paraphilias, sleep disorders, etc.). The cumulative frequency and percentage of subjects with specific DSM-IV axis I diagnoses exceeded the frequency and percentage of subjects with at least one DSM-IV axis I diagnosis because of multiple axis I diagnoses. No significant difference was observed between inpatients and outpatients with respect to demographic variables; as expected, inpatients showed a significantly higher frequency of axis I diagnoses (n = 331, 89.9%) than outpatients (n = 87, 44.4%): Yates-corrected X2 = 135.986, d f = 1, P < .001. With

73

respect to BPD diagnosis, no significant difference was observed between inpatients (n = 63, 17.1%) and outpatients (n = 37, 18.9%): Yates-corrected ×2 = 0.164, d f = 1, P > .60. After a complete description of the study to the subjects, written informed consent was obtained. DSM-IV BPD criteria and diagnosis were assessed using Structured Clinical Interview for DSM-IV axis II Personality Disorders, Version 2.0 (SCID-II), 19 a semis~uctured interview designed to diagnose the DSM-IV PDs. SCID-II was preceeded by the administration of its self-report screening questionnaire (PQ). SCID-II was administered by eight trained raters blind to the hypothesis of this study, in the context of patient routine diagnostic assessment. Both SCID-II and PQ were translated into Italian; the adequacy of the Italian translation was checked through backversions by a professional English mother-tongue translator. Subjects with axis I diagnoses were administered SCID-II at acute symptom remission by expert, trained raters. SCID-II interrater reliability was evaluated in a subsample composed by the first 231 consecutively admitted inpatients and outpatients. A pairwise interviewer-observer design was used to assess interrater reliability. Raters were paired randomly. Observers were explicitly not allowed to interfere with the interview. Both raters sat in on the interviews. Each rater served approximately equally as interviewer and observer. SCID-II was rated independently by interviewer and observer. Shrout and Fleiss intraclass correlation coefficient 1.1 (ICC) 2° was used to evaluate interrater reliability of SCID-II PD dimensional diagnoses, while interrater reliability of dichotomously scored DSM-IV BPD criteria and categorial PD diagnoses was assessed by computing Cohen K. DSM-IV BPD diagnosis proved to have excellent interrater reliability (ICC = .952, K = .909). Also, the individual DSM-IV BPD diagnostic criteria showed adequate interrater reliability (median K = .868; Table 1). All other DSM-IV PD diagnoses presented satisfactory interrater reliability coefficients (median ICC = .937, minimum = .901 [Depressive PD], maximum = .982 [antisocial PD]; median K = .912, minimum = .651 [Depressive PD], maximum = .981 [Narcissistic PD]). However, it should be considered that both SCID-II format and pairwise interview design could have spuriously increased the agreement between raters. The presence of significant association between DSM-IV BPD and other PDs was assessed using phi coefficient. Nominal alpha level was controlled by using the Bonferroni procedure (.05/11 = .0045). DSM-IV BPD criteria diagnostic accuracy was assessed by computing item-total (no. of criteria) point-biserial correlation coefficient (crpbi) and item-diagnosis phi coefficient (cqb), both corrected for overlap. 21,22 Within each set of item-total and item-diagnosis comparisons, nominal alpha level was stabilized using Bonferroni correction (.05/9 = .0056). Correlations (qb coefficients) between DSM-IV BPD criteria and other DSM-IV PD diagnoses were computed to evaluate BPD criteria discriminatory power. Sensitivity, specificity, efficiency (i.e., total probability of making a correct statement about the presence or absence of a particular disease; Youden J* was used as efficiency measure), and positive (PPP) and negative (NPP) predictive power of the individual BPD criteria could not be computed from standard

*Youden J = (sensitivity + specificity) - 1.

74

FOSSATI ET AL

Table 1. Item Analysis

DSM-IVBPD Criteria Frantic efforts

IRR K

BR (N)

.915 .165 (93) Unstable .951 .170 relationships (96) Identity .868 .200 disturbance (113) Impulsivity .883 .252 (142) Suicidal .873 .165 behavior (93) Affective .739 .216 instability (122) Feelings of .866 .163 emptiness (92) Inappropriate .861 .271 anger (153) Paranoid .711 .122 ideation (69)

Item-Cluster Item-PD~, BPD ~, Median Median (minimum/ (minimum/ c~pbi C~ maximum) maximum) .54 .41 .72 .68 .70 .67 .64 .51 .53 .43 .63 .54 .61 .56 .61

.44

.54 .45

.04 (-0.8/`14) .01 (-.13/`19) -.04 (-.16/.23) -.03 (-.13/.30) .05 (-.03/,16) -.01 (-.10/`19) -.03 (-.06./.16) .04 (-.09/.25) .03 (-.09/`13)

.11 (.04/`14) .12 (.10/.19) .22 (.16/`23) .18 (.18/.30) .13 (.05/`16) .17 (.12/.19) .13 (.08/`14) .17 (.12/`25) .12 (.09/.13)

NOTE. BPD criteria are listed in DSM-IV order. Abbreviations: IRR, interrater reliability (based on 231 observations); BR, base rate; crpbi, item-total corrected point-biserial r; c 4, item-diagnosis corrected ~ coefficient.

formulas, based on contingency tables, because of several violations of the assumptions underlying these statistics (e.g., presence of item-diagnosis overlap, lack of independency among BPD criteria, lack of a "gold standard" for BPD diagnosis, etc.).23 As suggested by several authors, 23-25LCA was used to evaluate BPD criteria diagnostic accuracy. The formulas for multiple latent classes were used to derive BPD criteria sensitivity, specificity, efficiency, PPP, and NPP from latent class conditional probabilities. 25 The hypothesis that one factor was sufficient to explain BPD criteria covariance was tested using confirmatory factor analysis (CFA). CFA is a covariance structure analysis that has several advantages over ordinary EFA. While EFA often relies on a sort of "shot-gun empiricism," CFA has a strong hypothesis-testing approach: starting from a theoretical model, CFA allows to preconstruct a specific mathematical model that explains the interrelationships among both observed and latent variables (i.e., factors), as well as to test its goodness-of-fit and its superiority to other alternative models. Moreover, different from EFA, CFA requires that the factorial model meets the necessary and sufficient conditions for obtaining a unique factorial solution (i.e., model identification). Several procedures are available to test these conditions. 26 Given the dichotomous assessment of DSM-IV BPD criteria, a weighted least-square (WLS) CFA was performed by using the tetrachoric correlation matrix as input matrix. 26 Covariance matrix of estimated tetrachoric correlations was used as weight matrix. According to the hypothesis of this study and the DSM-IV model of BPD, a unidimensional model with congeneric (i.e., linearly related, with no additional constrain) items was built and tested against the following alternative models: (1) unidimensional with parallel items (i.e., items with equal true score and error variance) (compatible with

DSM-III-R model); (2) unidimensional with tan-equivalent items (i.e., items with equal true score variance, but with different error variance) (compatible with DSM-III-R model); (3) three dimensional (uncertainty about the self and interpersonal difficulties, affect and mood regulation, and impulsivity) with orthogonal factors, derived from the Clarkin et al.7 exploratory study; (4) three dimensional (identity, affect, and impulse clusters) with orthogonal factors, derived from the Hurt et al. s study based on cluster analysis of PDs; and (5) four dimensional (uncertainty about self and interpersonal difficulties, affect and mood regulation, anger, and impulsivity) with orthogonal factors, derived from the Clarkin et al.7 study. Published factor loadings were used as starting points for models derived from the Clarkin et al. 7 study. The identification of the BPD model based on DSM-IV was assessed using t and three-factor rules. Model goodness-of-fit was evaluated using WLS asymptotic ×2 statistic. A significant (i.e., P < .05) goodness-of-fit X2 m e a n s that the corresponding factorial model should be rejected, because it does not reproduce adequately the correlations among the observed variables. At the opposite, a nonsignificant (i.e., P > .10) goodness-of-fit X2 shows that the corresponding factorial model should be accepted, because it has an acceptable fit to the data. Tetrachoric correlation matrix (and its asymptotic covariance matrix of correlation estimates) and CFA were performed using, respectively, PRELIS 27 and LISREL 7. 28 The categorial model of BPD hypothesizes the existence of natural clusters of subjects, i.e., nonartifactual groups in which subjects belonging to one group are maximally similar to each other and dissimilar to subjects belonging to the other group(s). Multivariate mixture analysis is a class of cluster analysis techniques designed to uncover, rather than to impose, the presence of natural groups in multivariate data, and to allow statistical inference from samples to populations. 16,29,3° When the data are dichotomous, as BPD characteristics, LCA is the appropriate multivariate mixture analysis technique. 16.29,30 In LCA, the observed contingencies among several dichotomous variables are explained by assuming the population is a mixture of latent (i.e., not directly observed and measured) classes within each of which the variables are independently distributed (in other words, it is assumed that subjects come from a mixture of multivariate Bernoulli distributions). Exploratory LCA was used to identify the number of latent classes underlying DSM-IV BPD criteria. Best-fitting model was identified by using improvement in likelihood X2 statistic (L 2) and bayesian information criterion (BIC). A significant (i.e., P < .05) goodness-of-fit L 2 means that the corresponding LCA model should be rejected. At the opposite, a nonsignificant (i.e., P > . 10) goodness-of-fit L 2 shows that the corresponding LCA model should be accepted, because it has an acceptable fit to the data. The L 2 difference between nested models (e.g., one class v two classes, two classes v three classes, etc.) was used to assess the significance of the incremental fit. On the other hand, according to BIC, the model with the smallest value is the best one. 29 LCA subject classification was generated using individual latent class membership probabilities. LCA and DSM-IV classifications of BPD were then compared. The potential confounding role of patient severity on clustering was assessed testing the association between latent classes and, respectively, inpatient/ outpatient status and presence of any axis I disorder. LCAG 31 and 1EM32 computer programs were used to perform LCA.

LATENT STRUCTURE OF BORDERLINE PD

75

RESULTS According to SCID-II, 370 subjects (65.6%) received at least one DSM-IV PD diagnosis; mean number of PD diagnoses was 1.14 (SD = 1.11). Mean number of DSM-IV BPD criteria was 1.71 (SD = 2.37); 280 subjects (49.6%) received no BPD criteria. DSM-IV BPD was diagnosed in 100 subjects (17.7%). Among these, 39 subjects (39%) received a "pure" BPD diagnosis, while 61 (61%) received one or more additional PD diagnoses (range, one to four). Significant, even if weak, positive associations were observed between BPD and passive-aggressive (d~ = .22), antisocial (+ = .19), histrionic (~b = .19), and narcissistic (+ = .18) PDs (all P < .0045). DSM-IV BPD diagnosis showed negative significant correlations with avoidant (~b = - . 1 9 ) and obsessive-compulsive (+ = - . 1 3 ) PDs (all P < .0045). Mean number of PD codiagnoses in subjects with DSM-IV BPD was 1.14 (SD = 1.18). Mean number of BPD criteria among subjects with BPD diagnosis was 6.21 (SD = 1.32); only seven subjects could be considered as "prototypical" cases (i.e., meeting all nine diagnostic criteria) of DSM-IV BPD. Item analysis results are listed in Table 1. All item-total and item-diagnosis corrected correlation coefficients were large and significant (all P < .0056), as well as larger than the correlation coefficients between BPD criteria and other DSM-IV PD diagnoses. However, as expected, differences in diagnostic efficiency were observed. In particular, considering ~b coefficients corrected for criterion-diagnosis overlap, DSM-IV BPD criteria should be ranked as follows: (1) unstable relationships; (2) identity disturbance; (3) feelings of emptiness; (4) affective instability; (5) impulsivity; (6) paranoid ideation; (7) inappropriate anger;

(8) suicidal behavior; and (9) frantic efforts. It should be noted that the rank order of BPD criteria ~b coefficients was independent from the rank order of the interrater reliability (K) coefficients: Spearman r = .067 P > .80. Sensitivity, specificity, efficiency (Youden J), PPP, and NPP of BPD criteria are listed in Table 2. As shown, unstable relationships and identity disturbance appeared as the most relevant diagnostic criteria. Moreover, the data listed in Tables 1 and 2 suggested a small heterogeneity of BPD subjects with respect to these two criteria; rather, they appeared as the two main BPD characteristics. For instance, when the conjoint presence of unstable relationships and identity disturbance was considered, the sensitivity, specificity, efficiency, PPP, and NPP were as follows: .736, .998, .758, .985, and .952. On the other hand, frantic efforts to avoid abandonment was one of the criteria provided with the worst diagnostic accuracy. All tetrachoric correlation coefficients among BPD criteria were large, positive, and significant, even after the nominal alpha level Bonferoni correction (.05/36 = .0014). Median tetrachoric correlation was .62 (minimum = .48, maximum = .82). CFA results are shown in Table 3. The BPD model derived from DSM-IV (unidimensional/ congeneric items) showed adequate fit, and was clearly more valid than other alternative models. Considering LCA, the unidimensional model with one latent class did not fit adequately (L 2 = 1484.122, d f = 502, P < .001). Adding a second class significantly improved the model (L z difference = 1,025.993, d f = 10, P < .001) and adequately fitted the data (L 2 = 458.129, d f = 492, P > .80). The model was improved further (L 2 difference = 96.594, d f = 10, P < .001) when a

Table 2. Item Analysis: Diagnostic Efficiency Statistics BR-BPD(N)

Sensitivity

Specificity

Youden J

PPP

NPP

Frantic efforts Unstable relationships

.560 (56) .750 (75)

Identity disturbance Impulsivity Suicidal behavior Affective instability Feelings of emptiness Inappropriate anger Paranoid ideation

.820 .810 .570 .760 .650 .800 .480

.543 .809 .878 .789 .534 .784 .663 .777 .461

.914 .925 .934 .863 .913 .899 .938 .840 .947

.457 .766 .812 .652 .447 .684 .601 .617 .408

.551 .785 .722 .527 .544 .602 .672 .484 .626

.833 .925 .949 .912 .830 .912 .873 .907 .812

(82) (81) (57) (76) (65) (80) (48)

NOTE. Derived from LCA best fitting model (see Table 3) - BR-BPD = Base rate of the individual criteria in subjects (N = 100) with a DSM-IV BPD diagnosis (frequencies are between brackets) - PPP = Positive Predictive Power - NPP = Negative Predictive Power Youden J = (Sensitivity + Specificity) - 1,

76

FOSSATI ET AL Table 3. DSM-IV BPD Criteria: WLS CFA Model

X2

Unidimensional/congenericity Unidimensional/tau-equivalence Unidimensional/parallelism Three dimensions (uncertainty a b o u t the self, affect regulation, impulsivity)/orthogonal factors Three dimensions (identity disorder, affective instability, impulsivity)/ orthogonal factors Four dimensions (uncertainty about the self, affect regulation, anger, impulsivity)/orthogonal factors

df

P

18.89 80.30 80.30

27 35 43

.874 <.001 <.001

1858.31

24

<.001

2889.53

25

<.001

2713.85

24

<.001

Factor Loadings (completely standardized solution) Rest-Fitting Model (BPD items) Frantic efforts Unstable relationships Identity disturbance Impulsivity Suicidal behavior Affective instability Feelings of emptiness Inappropriate anger Paranoid ideation

.718 .903 .889 .818 .703 .807 .788 .784 .726

third latent class was added (goodness-of-fit: L 2 = 361.535, df = 482, P > .90). No significant improvement was observed when a fourth latent class was added (L 2 difference = 14.475, df = 10, P>.10). BIC reached its minimum value (3962.849) in correspondence of the three-class model, indicating that this model is the best fitting one. According to these results, DSM-IV BPD criteria multivariate distribution could be considered as a mixture of two, or more likely, three multivariate bernoulli distributions. Conditional probabilities for LCA best-fitting model are shown in Table 4. LCA conditional probabilities fall between zero and one, and represent the probability of belonging to a latent class given the presence of a BPD criterion (e.g., P(frantic efforts/class 1) = .543). In a sense, LCA conditional probabilities are analogous to factor loadings in conventional factor analysis. As shown, all DSM-IV BPD criteria "loaded" on class 1 (with a less clear contribution of criterion 9). Class 2 consisted of zero "correlated" variables, and class 3 was characterized by "impulsivity" (.375) and "inappropriate anger" (.450) BPD criteria. In other words, even if BPD appeared as a distinct latent class, a dimensional gradient was observed for

impulsivity and anger discontrol conditional probabilities. Individual LCA membership probabilities were computed for each subject. Each subject was ascribed to a given class according to his/her highest membership probability (for instance, a subject with latent class membership probabilities of .75, .05, and .20 was ascribed to class 1). Unweighted (raw) means and SDs of BPD criteria were computed for each class. Class 1 was mainly composed of subjects (N = 91, % = 16.2) scoring above DSM-IV diagnostic threshold for BPD (mean = 6.32, SD = 1.34). Classes 2 and 3 were composed of, respectively, 316 (56.0%) and 157 (27.8%) subjects, showing no (class 2) or few (class 3) BPD criteria. In particular, class 3 consisted of non-BPD subjects characterized by marked impulse and anger discontrol. As expected, a highly significant difference was observed between latent classes 1 and 3 when BPD criteria dispersion matrices were compared: Box M = 80.118, X 2 = 76.740, df = 45, P < .005 (latent class 2 dispersion matrix could not be included in the analysis because singular). Ten subjects diagnosed as having BPD according to DSM-IV threshold were misclassified by LCA as belonging to class 3, while only one subject diagnosed as non-BPD according to DSM-IV was classified in class 1 by LCA. When LCA classes 2 and 3 were grouped to form a non-BPD class, the agreement between DSM-IV and LCA classificaTable 4. LCA Three-Class Model

DSM-IVBPD Items Frantic efforts Unstable relationships Identity disturbance Impulsivity Suicidal behavior Affective instability Feelings of emptiness Inappropriate anger Paranoid ideation No. of subjects (%)* Mean no. (SD) of BPD items

Latent Class 1 LatentClass 2 LatentClass3 Conditional Conditional Conditional Probabilities Probabilities Probabilities .543

.017

.224

.809 .878 .789 .534 .784

.000 .023 .019 .012 .030

.129 .151 .375 .237 .242

.663 .777 .461

.015 .016 .011

.158 .450 .138

91 (16.2) 6.32 (1.34)

316 (56.0) 0.12 (0.33)

157 (27.8) 2.25 (1.11)

*Subjects were ascribed to latent classes according to their highest individual membership probability. Means and SDs of BPD criteria are raw means (and SDs) of the 9 items.

LATENT STRUCTURE OF BORDERLINE PD

tions was almost perfect (Cohen K = .931 P < .001). It should be noted that the agreement between LCA and DSM-IV classifications of BPD was also substantial for the less fitting two latent class model (K = .794 P < .001). No significant association was observed between latent classes and, respectively, inpatient/outpatient status (×2 _ 0.143, df = 2, P > .90) and any axis I disorder (×2 = 2.091, df = 2, P > .35). DISCUSSION Despite previous negative findings, 5,7 this study supported the hypothesis of BPD as a unidimensional construct. Item analysis results confirmed a differential diagnostic efficiency of BPD criteria. However, the rank order of BPD diagnostic criteria observed in this study was substantially different from the one proposed by DSM-IV. Despite the fact that intolerance to aloneness was recently proposed as the main characteristic of BPD subjects, 33 this study showed that interpersonal difficulties and identity problems were the most typical features of BPD, while frantic efforts to avoid abandonment was among the worst-performing BPD criteria. This result was in agreement with previous studies based on DSM-III-R criteria, 34,35 which evidenced the high diagnostic efficiency of unstable relationships (and, to a lesser degree, of identity disturbance), as well as the low discriminant efficiency of intolerance to abandonment. In our opinion, these data should not be considered as evidence that intolerance to abandonment (and aloneness) is not an important characteristic of BPD subjects; rather, they seemed to suggest that, in agreement with other observations, 34,35 abandonment fears are not central as relationship instability and identity disturbance to BPD psychopathology. According to these results, a revision of DSM-IV BPD criteria hierarchical listing should be proposed. On the other hand, BPD intolerance to abandonment could not be expressed mainly by frantic, but goal-oriented, efforts to avoid it. Abandonment may trigger several disorganized, rather than goal-oriented, reactions in BPD subjects, such as an increase in impulsivity, severe mood deflection, etc. Moreover, DSM-IV criterion does not specify how close and significant should the relationship be for the subject to cause severe reactions to abandonment in BPD. In fact, subjects with BPD diagnosis usually have a pattern of unstable and intense relationships, but these hardly have the

77

same importance. Devaluation could help BPD subjects to overcome abandonment fears. LCA results (as well as dipersion matrix analysis) gave evidence of a categorial structure of BPD. BPD (i.e., latent class 1) appeared as a distinct, cohesive class of subjects, not only quantitatively (i.e., no. of diagnostic criteria), but also qualitatively different from the other latent classes. However, evidence was found for a dimensional distribution of some temperamentally based BPD characteristics (namely, impulsivity and inappropriate anger). In other words, according to our results, BPD should be considered as a definite personality disorder, deeply intertwined with dimensionally distributed temperamental characteristics. Since identity disturbance and unstable relationships were relevant diagnostic markers of BPD, the probability of finding a BPD subject without either of these two characteristics (or a non-BPD subject with either of these two characteristics) is low, even if a variability of BPD criteria was observed also in BPD subjects. Thus, despite better reliability of BPD dimensional assessment, 36 the categorial BPD diagnosis seemed more correspondent to the actual BPD structure. However, the evidence for a dimensional distribution of anger discontrol and impulsivity strongly stressed the need for a thorough investigation of the ways BPD interact with, and stem from, these temperamental dimensions. The strong agreement between LCA and DSM-IV models of BPD seemed to show the adequacy, or at least the empirical reproducibility, of the DSM-IV diagnostic threshold for BPD. It was also interesting to observe roughly the same average number of BPD diagnostic features (n = 6) in subjects diagnosed as BPD according to DSM-IV cut-off threshold score and in subjects belonging to latent class 1. According to diagnostic features, DSM-IV BPD should be described as a unidimensional, categorial PD. This does not mean that subjects with BPD could not be subtyped using variables external to diagnostic criteria. Diagnostic homogeneity does not imply the absence of natural subgroups of BPD subjects when additional characteristics relevant to personality description and development are considered. In fact, BPD heterogeneity was shown when temperamental factors 37 or variables related to developmental history were considered. 38 This is a well-known phenomenon in medicine, not only in psychiatry; for instance, if we focus our attention

FOSSATI ET AL

78

only on diabetes manifest symptoms, diabetes appears as a unitary syndrome. But if laboratory findings, as C-peptide and insulin plasma levels are added, then two types of diabetes can be identified, with a different etiopathogenesis, course, and treatment. The same might be true for BPD; for instance, different subtypes of BPD could need different therapeutic approaches, and show different clinical course and response to treatment. However, more research on subtyping BPD according to external variables is needed before drawing any conclusion on this topic. Finally, the following caveats should be stressed: (1) It is possible that a method-by-data interaction, 39 related to SCID-II structure, could have played a role in producing these results, including the lower comorbidity between BPD and the other D S M - I V PDs. SCID-II does not reorganize the PD symptomatology and retains the organization of D S M - I V PD criteria. These characteristics are likely to inflate the internal consistency and distinctiveness of the respective diagnostic criteria, and decrease the occurrence of multiple diagnoses. In other words, our findings supporting the hypothesis that D S M - I V BPD has a unidimensional, categorial structure could have been somehow influenced by SCID-II format. On the other hand, using raters blind to the aim of the study should have decreased the risk of artefactual results. However, this meth-

odological problem related to SCID-II format strongly stresses the need for further studies based on different interviews (i.e., interviews that reorganize PD symptomatology in different areas o f dysfunction). (2) CFA allowed us to test the unidimensional model of BPD against only some alternative factorial models of BPD. Even if it performed better, there could be other multidimensional models of BPD that could be superior. This is to say that more research is needed before accepting this unidimensional model of BPD. (3) Exploratory L C A was used in this study, and it should be stressed that exploratory L C A suffers from the same undeterminacy problems of EFA. A replication study, based on confirmatory LCA, should be performed before accepting these results. (4) The presence of a large group of inpatients in this sample could have influenced clustering by severity. Unfortunately, the sample size of inpatient and outpatient subgroups was insufficient to perform separate CFA and L C A analyses. However, the lack of a significant association between latent class membership and inpatient/outpatient status seemed to show that L C A results were not significantly biased by patient severity. ACKNOWLEDGMENT

The authors thank Marco Battaglia, M.D. for his useful suggestions.

REFERENCES

1. Blashfield R, McElroy R. The 1985 journal literature on personality disorders. Compr Psychiatry 1987;28:536-546. 2. WidigerTA, Trull TJ. Borderline and narcissistic personality disorders. In: Adams H, Sutker P (eds): Comprehensive Handbook of Psychopathology,Ed. 2. New York, NY: Plenum Press, 1993;371-394. 3. American PsychiatricAssociation. Diagnostic and Statistical Manual of Mental Disorders, Ed. 3. Washington, DC: American Psychiatric Press, 1980. 4. American PsychiatricAssociation.Diagnostic and Statistical Manual of Mental Disorders, Ed. 3. Rev. Washington, DC: American Psychiatric Press, 1987. 5. Hurt S, Clarkin J, Widiger T, Fyer M, Sullivan T, Stone M, et al. Evaluation of DSM-III decision rules for case detection using joint conditional probability structures. J Personal Disord 1990;4:121-130. 6. Livesley WJ, Schroeder ML. Dimensions of personality disorder: The DSM-III-R cluster B diagnoses. J Nerv Ment Dis 1991;179:320-328. 7. Clarkin JF, Hull JW, Hurt SW. Factor structure of borderline personality disorder criteria. J Personal Disord 1993;7:137143. 8. Trull TJ, McCrae RR. A five-factorperspective on personality disorder research. In: Costa PT, Widiger TA (eds): Person-

ality Disorders and the Five-Factor Model of Personality,Ed. 2. Washington,DC: AmericanPsychologicalAssociation, 1994;5971. 9. Eysenck H. The definition of personality disorders and the criteria appropriate for their description. J Personal Disord 1987;1:211-219. 10. Livesley WJ. Classifying personality disorders: ideal types, prototypes, or dimensions? J Personal Disord 1991;5:5259. 11. Widiger TA. Personality disorder dimensional models proposed for DSM-IV.J Personal Disord 1991;5:386-398. 12. Oldham JM, SkodolAE, Kellman HD, Hyler SE, Rosnick L, Davis M. Disagnosis of DSM-III personality disorders by two structured interviews: patterns of comorbidity.Am J Psychiatry 1992; 149:213-220. 13. WidigerTA, FrancesAJ. Towarda dimensionalmodel for the personality disorders. In: Costa PT, Widiger TA (eds): Personality Disorders and the Five-Factor Model of Personality, Ed. 2. Washington, DC: American Psychological Association, 1994;19-39. 14. Numberg HG, Hurt SW, Feldman A, Shu R. Evaluation of diagnostic criteria for borderline personality disorder. Am J Psychiatry 1988;145:1280-1284. 15. Trull TJ, Widiger TA, Guthrie P. The categorical versus

LATENT STRUCTURE OF BORDERLINE PD

dimensional status of borderline personality disorder. J Abnorm Psychol 1990;99:40-48. 16. Everitt BS. Cluster Analysis. Ed. 3. London, England: Edward Arnold, 1993. 17. American Psychiatric Association: Diagnostic and Statistical Manual of Mental Disorders, Ed. 4. Washington, DC: American Psychiatric Press, 1994. 18. Gunderson JG, Zanarini MC, Kisiel CL. Borderline personality disorder: a review of data on DSM-III-R descriptions. J Personal Disord 1991;5:340-352. 19. First MB, Spitzer RL, Gibbon M, Janet BV, Benjamin L. Structured Clinical Interview for DSM-IV Axis II Personality Disorders (SCID-II), Version 2.0. New York, NY; Biometric Research Department, New York State Psychiatric Institute, 1994. 20. Shrout PE, Fleiss JL. Intraclass correlation: Uses in assessing rater reliability. Psychol Bull 1979;86:420-428. 21. Henrysson S. Correction of item-total correlations in item analysis. Psychometrika 1963;28:211-218. 22. Guilford JP, Fruchter B. Fundamental Statistics in Psychology and Education, Ed. 6. New York, NY: McGraw-Hill, 1978. 23. Rindskopf D, Rindskopf W. The value of latent class analysis in medical diagnosis. Stat Med 1986;5:21-27. 24. Young MA. Evaluating diagnostic criteria: a latent class paradigm. J Psychiatr Res 1983;17:285-196. 25. Uebersax JS, Grove WM. Latent class analysis of diagnostic agreement. Stat Med 1990;9:559-572. 26. Bollen KA. Structural Equations with Latent Variables. New York, NY: Wiley, 1989. 27. J6reskog KG, Srrbom D. PRELIS: A Preprocessor for LISREL. Mooresville, IN: Scientific Software, Inc., 1988. 28. Jrreskog KG, Srrbom D. LISREL 7: User's Reference Guide. Mooresville, IN: Scientific Software, Inc., 1989. 29. Clogg CC. Latent class models. In: Arminger G, Clogg CC, Sobel ME (eds): Handbook of Statistical Modeling for the Social and Behavioral Sciences. New York, NY: Plenum Press, 1995;311-359.

79

30. McLachlan GJ, Basford KE. Mixture Models. Inference and Applications to Clustering. New York, NY: Marcel Dekker, 1988. 31. Hagenaars J, Luijkx R. LCAG. A Program to Estimate Latent Class Models and other Loglinear Models with Latent Variables with and without Missing Data. Tilburg, The Netherlands: Tilburg University, Department of Sociology, 1990. 32. Vermunt JK. 1EM: Log-linear and Event History Analysis with Missing Data using the EM Algorithm. Tilburg, The Netherlands: Tilburg University, Department of Sociology, 1993. 33. Gunderson JG. The borderline patient's intolerance of aloneness: insecure attachment and therapist availability. Am J Psychiatry 1996;153:752-758. 34. Clarkin J, Widiger T, Frances A, Hurt S, Gilmore M. Prototypic typology and the borderline personality disorder. J Abnorm Psychol 1983;92:263-275. 35. Nurnberg H, Hurt S, Feldman A, Suh R. Efficient diagnosis of borderline personality disorder. J Personal Disord 1987;1:307-315. 36. Pilkonis PA, Heape CL, Proietti JM, Clark SW, McDavid JD, Pitts TE. The reliability and validity of two structured diagnostic interviews for personality disorders. Arch Gen Psychiatry 1995;52:1025-1033. 37. Akiskal HS. Subaffective disorders: Dysthymic, cyclothymic, and bipolar II disorders in the "borderline" realm. Psychiatr Clin North Am 1981;4:25-46. 38. Paris J, Zweig-Frank H. Parental bonding in borderline personality disorder. In: Paris J (ed): Borderline Personality Disorder. Etiology and Treatment. Washington, DC: American Psychiatric Press, 1993;147-166. 39. Zimmerman M. Diagnosing personality disorder. A review of issues and research methods. Arch Gen Psychiatry 1994;51:225-245.