Discriminating Borderline Disorder From Other Personality Disorders Cluster Analysis of the Diagnostic Interview for Borderlines Joseph Barrash; Jerome Kroll, MD;

Kathleen

Carey, RN, MS; Lloyd Sines,

\s=b\ The statistical technique of cluster analysis was applied to

hospitalized patients' scores on the 29 statements of the Diagnostic Interview for Borderlines. We found that this statis-

252

tical treatment could reliably differentiate borderline disorder from other personality disorders. Two subtypes similar to Spitzer's schizotypal and unstable subtypes emerged. (Arch Gen Psychiatry 1983;40:1297-1302)

previous articles, reported the reliability of the Diagnostic Interview for Bor¬ In derlines (DIB)1 the validity the concept of we

the results of studies

on

as well as of borderline personality disorder (BPD).2 3 We concluded that the DIB demonstrated satisfactory levels of interrater reliability for research and clinical purposes. We further concluded that the consensus of studies, including our own, showing that persons with BPD can be distinguished from DSM-III axis I disorders across a variety of instruments (DIB,46 Perry and Klerman checklist,7 Sheehy checklist,8 DSM-III, and Minnesota Multiphasic Personality Inven¬ tory) lent validity to the borderline construct. The present article addresses one of the persistent questions relating to the concept of BPD: can it be differentiated from other personality disorders? This question is addressed by multivariate covariation analysis of the 29 DIB statements of 252 psychiatric inpatients. The DIB, developed by Gunderson and Kolb, is a semistructured interview containing 29 items (statements) de¬ scribing five areas of functioning and/or psychopathology that most workers have associated with borderline condi¬ tions. The ratings on the 29 statements are reduced to a

Accepted for publication Feb 14, 1983. From the Department of Psychology, University of Iowa, Iowa City (Mr Barrash); and the Department of Psychiatry, University of Minnesota Medical School, Minneapolis (Drs Kroll and Sines). Ms Carey is a private consultant in St Paul.

Reprint requests to Box 393 Mayo, University of Minnesota Hospitals, Minneapolis, MN 55455 (Dr Kroll).

PhD

scaled score (0 to 2) for each section. The summation (0 to 10) of these section scores provides the basis of a borderline diagnosis; a score of 7 or greater is interpreted as indicating borderline disorder. SUBJECTS AND METHODS Almost half the sample (117 patients) consisted of consecutive admissions to the University of Minnesota, Minneapolis, inpatient psychiatric service; 88 additional patients on the same service were added to the study during a two-year period. This latter group represented the most recent admissions to the wards on the days when the DIB interviewers were available. Most were interviewed within two days of admission. There were no other selection criteria used for the 88 patients. The last cohort of our sample consisted of 47 patients hospitalized in a public mental hospital near

Cambridge, England.3

All patients were diagnosed by DSM-III criteria, the initial 117 Minnesota and 47 British patients by one of us (J.K.) and the 88 additional Minnesota patients by the psychiatric residents on the service in collaboration with their attending psychiatrists. Several interviewers were involved with the data collection: a psychiatric resident, a clinical nursing specialist, a first-year graduate student in clinical psychology, and a mental health worker. All interviewers were trained in accordance with Gunderson's instructions, and acceptable interrater reliability (.71) was established. The DSM-III diagnoses and DIB scores were deter¬ mined independently. Cluster analysis of the DIB data was used for the determination of borderline patients. A cluster analysis sorts subjects (or varia¬ bles) into groups in which they are most similar to others in that group and most dissimilar to subjects (or variables) in all other groups. The definitions of similarity and dissimilarity are deter¬ mined by the particular method ofcluster analysis chosen.9 The use of cluster analysis methods requires that certain questions regard¬ ing the handling of the data that will affect the outcome of the analysis be conscientiously examined.10 The first question to be answered in this case was whether to cluster on the basis of covariation among five variables (ie, a subject's DIB section totals) or among the 29 statements that constitute the DIB. Pilot analyses with both sets demonstrated the ability of the 29-statement profiles to form clusters at least as intuitively meaningful as

Downloaded From: http://archpsyc.jamanetwork.com/pdfaccess.ashx?url=/data/journals/psych/12379/ on 01/16/2017

clusters from section totals, while forming at lower levels of error. (Error is the total of the sum of squared deviations for each cluster in an analysis.) The decision to use correlated or orthogonal variables can have a profound effect on results; if unweighted correlated variables are used, the factors most frequently repre¬ sented will make the greatest contributions to the formation of the clusters. Since Gunderson's scoring system was designed to give equal weight to the five factors presumed to be measured by the five sections, statements from a given section were weighted according to the number of statements in that section. The effect of weighting by section was to maintain equal contributions to the diagnostic process (ie, clustering) by the five areas. In other pilot analyses, the results of a hierarchical clustering technique, Ward's" hierarchical grouping method, were compared with those obtained by a nonhierarchical technique, MacQueen's convergent K-means.1213 The nonhierarchical approach formed more valid clusters, the criterion being DSM-III diagnosis. This is assumed to be due to the noniterative procedure of Ward's; once subjects have been paired together, they may not be split at later stages of clustering. Because of this feature, random error will have a greater impact than in a nonhierarchical analysis. After these analyses, it was decided to use MacQueen's conver¬ gent K-means,lzl3 and the two cluster analyses to be reported were then performed. The first analysis clustered all 252 subjects. The second analysis clustered the 49 patients (14 men, 35 women) with a DSM-III diagnosis of personality disorder who had no axis I diagnosis. The rationale for the second analysis was the expecta¬ tion that the clustering rule (that within-group profiles have maximal dissimilarity to all those outside the group) would have the confounding effect of pushing all personality-disordered sub¬ jects closer together when considered together with organic, psychotic, and affectively disordered patients.

RESULTS DIB Scores

Forty-eight (19%) of the 252 patients were diagnosed as bor¬ derline by the DIB, including 17 men (16.3%) and 31 women (20.9%). These patients are referred to as DIB positive. A break¬ down of the DIB-positive patients by their DSM-III diagnoses is presented in Table 1. According to DSM-III criteria, there were 18 borderlines in the sample, 13 (72.2%) of whom were DIB posi¬ tive. Thirty-one patients were diagnosed as having personality disorders other than borderline, 13 (41.9%) of whom were indi¬ cated as borderline by the DIB. Two patients were diagnosed as having chronic factitious disorder; both were DIB positive. Twenty other patients scored as borderline on the DIB. By DSM-III diagnostic criteria, four had a schizophrenic disorder; three, functional psychosis; three, affective disorder; three, organic mental disorder; three, substance abuse; two, anorexia nervosa; one, adjustment disorder; and one, explosive disorder. Fifteen percent of the DSM-III nonborderlines were diagnosed as bor¬ derlines by the DIB. In the second analysis, when only patients with personality disorders were considered, 41.9% of the personal¬ ity-disordered patients who did not meet the DSM-III borderline criteria were DIB positive. Clusters The first analysis clustered patients into eight groups, a number found to balance successfully the opposing requirements of dis¬ crimination between borderline and nonborderline and parsimony. Two of the clusters included 15 of the 18 borderlines, and were regarded as borderline clusters. That is, the patients in these clusters were diagnosed as having a BPD on the basis of their profile (the pattern of scores on the 29 statements). Hereafter, such patients are referred to as cluster positive. In addition to the 15 DSM-III borderlines, 25 other patients were cluster positive. Ten of these had other personality disorders; four, an organic mental disorder; four, affective disorder; one, schizoaffective disor¬ der; one, paranoid state; one, agoraphobia; one, conduct disorder; and one, alcohol abuse; both patients with chronic factitious disorder had borderline profiles. The composition of the borderline clusters is reported in Table 1. The borderline clusters correctly indicated 83.3% of the DSM-

Table 1.—DSM-III

Diagnoses of Patients Diagnosed as

Borderline

by DIB* and Clustering

DIB Positive

Cluster Positive

18

13

15

other than borderline Chronic factitious disorder Organic mental disorder

31

13

10

Affective disorder

77

Schizophrenic disorder Other functional psychosis Disorder of adjustment, anxiety, eating, Impulse control, or

46

Diagnosis Borderline personality disorder Personality disorders DSM-III

substance abuse Total

28

4t

2*

13

37

48

252

40

*DIB indicates Diagnostic Interview for Borderlines. tTwo have a schizoaffective disorder. ß has a schizoaffective disorder.

DSM-III Borderline DSM-III Nonborderline

Personality-Disordered All Patients

(

60 .8

1

=

252)

40

* 20 o

Patients

(N 49) =

1B. .9? 10

UÀ

Hü

1*2*3 4 5 6 7 8

1f 2* 3 4 5 6 7

Cluster

Cluster

Fig 1.—Cluster analysis of DSM-III borderline and nonborderline patients. Asterisk indicates that members of these clusters are considered borderline; dagger, members of this cluster are consid¬ ered type 1 borderline; double dagger, members of this cluster are considered type 2 borderline.

only 10.7% of patients with other diagnoses diagnosed as borderline. A graphic representation of the clusters' composition is shown in Fig 1, left. The sensitivity of diagnosis by clustering was. 83, compared with 72 for diagnosis by the DIB. The specificity of profile-based diagnosis was .89, com¬ pared with .85 for the DIB. Among patients with a personality disorder other than borderline, 32.3% were diagnosed as bor¬ derline by clustering. Diagnostic agreement with DSM-III was improved by the application of clustering techniques to the DIB data. Agreement between the DIB and DSM-III included 13 borderline diagnoses and 199 nonborderline diagnoses. The coefficient between DIB and DSM-III diagnoses was .32. In the 40 cases when disagree¬ ment occurred, the DIB accounted for 35 (87.5%) of the borderline diagnoses. Agreement between DSM-III and cluster-based diag¬

/// borderlines, while were

.

included 15 borderlines and 209 nonborderlines; the coef¬ ficient was .46. Again, DSM-III criteria were more restrictive when disagreement occurred (28 cases), only three (10.7%) of these patients being diagnosed as borderline by the DSM-III criteria. The second analysis, of the 49 patients with personality disor¬ ders, resulted in seven clusters. As expected, the second analysis improved the discrimination between borderline and other person¬ ality disorders. Two of the seven clusters were considered bor¬ derline clusters because they included 13 (72.2%) of the borderlines and only two (6.5%) of the patients with a personality disorder noses

Downloaded From: http://archpsyc.jamanetwork.com/pdfaccess.ashx?url=/data/journals/psych/12379/ on 01/16/2017

Table 2.—Mean

Ratings on

29 Statements for Borderlines and Nonborderlines Mean

Statements

Ratings*

Type 1 Borderline

Type 2 Borderline

Other PersonalityDisordered Patients

(N 10)

(N 5)

(N 35)

=

I. Social adaptation 1. School/work achievement

(as Diagnosed by Clustering)

=

=

ft (Differences Between Two Types of Borderlines)

0.7

1.2

1.0

1.0

0.6

0.9

NS NS

1.8§ 2.0II

0.9