Journal of Personality Disorders, 19(6), 597–623, 2005 2005 The Guilford Press
HIGHER-ORDER DIMENSIONS OF PERSONALITY DISORDER: HIERARCHICAL STRUCTURE AND RELATIONSHIPS WITH THE FIVE-FACTOR MODEL, THE INTERPERSONAL CIRCLE, AND PSYCHOPATHY Ronald Blackburn, PhD, Caroline Logan, DPhil, Stanley J. D. Renwick, PhD, and John P. Donnelly, ClinPsyD
Two studies examined the higher-order factor structure of DSM-IV personality disorders using the International Personality Disorder Examination in male forensic psychiatric patients. In Study 1 (N = 168), exploratory factor analysis at the level of individual personality disorder criteria indicated nine primary factors. Exploratory and confirmatory factor analyses of these first-order factors supported a hierarchical structure in which two of three second-order factors covaried to yield a third-order factor. The two resulting superordinate factors were labelled Anxious-Inhibited and Acting Out. In Study 2 (N = 160), we used exploratory and confirmatory factor analyses to test hypotheses of common dimensions underlying these superordinate factors of personality disorder and superordinate factors of the five-factor model of personality, dimensions of the interpersonal circle, and psychopathy. Of three common factors, one combined Anxious-Inhibited disorders, “neurotic introversion,” and hostile-submission. The other two factors of Acting Out/ psychopathy and antagonism/hostile-dominance covaried to yield a superordinate factor. Possible substrates underlying two superordinate dimensions common to normal and abnormal personality were identified in the theoretical literature.
Successive editions of the Diagnostic and Statistical Manual of Mental Disorders (DSM) have promoted a categorical approach to the classification of personality disorders; however, research has clearly established the
From the Division of Clinical Psychology, University of Liverpool, UK (R. B., C. L.); The Duchess of Kent’s Psychiatric Hospital, Catterick Garrison, UK (S. J. D. R.); The State Hospital, Scotland, UK (J. D. P.). This research was supported by grants from the former Special Hospitals Service Authority, Ashworth Hospital Authority, and the State Hospital. We are grateful to Dr. Paul Barrett, Ms. Anne Beadle, and Ms. Moira F. Scott for their help in data collection and analysis. Correspondence concerning this article should be addressed to Ronald Blackburn, Division of Clinical Psychology, University of Liverpool, The Whelan Building, Quadrangle, Brownlow Hill, Liverpool L69 3GB, UK; E-mail:
[email protected]
597
598
BLACKBURN ET AL.
greater utility of a dimensional representation of these disorders both in assessment and conceptualization (Livesley, 2001; Widiger & Frances, 2002). Although the categorical model has some advantages in clinical communication, it treats abnormalities of personality as discrete entities, and misleadingly implies clear boundaries between disorders and between normality and abnormality. A dimensional model, in contrast, treats personality disorders as measurable continua differing only quantitatively from the normal range of functioning. In this model, the problem of diagnostic overlap, or comorbidity, is simply a consequence of the hierarchical structuring of personality dispositions in which traits lower in the hierarchy are correlated and give rise to broader dimensions at the higher level. Identifying these fundamental higher-order dimensions more readily permits the translation of personality disorders into the dimensions of personality identified in theory and research. Somewhat paradoxically, the DSM also recognizes a hierarchical structure of personality disorder. Classes of disorder are made up of traits assumed to be correlated with each other, while the clustering of disorders into A (paranoid, schizoid, schizotypal), B (antisocial, borderline, histrionic, narcissistic), and C (avoidant, dependent, obsessive-compulsive) groupings assumes that disorders within clusters are more similar to each other than those of other clusters. Research has not, however, consistently supported this proposed hierarchical structure at either the first-order (trait) or second-order (cluster) levels. The grouping of traits into the classes proposed by the classification rests on clinical tradition rather than empirical findings (Frances, 1980), and relevant research remains limited. In a study meeting the taxonomic assumptions of the DSM classification, Morey (1988) conducted a cluster analysis of DSM-III-R personality disorder criteria as recorded by clinicians on a checklist. The 11 clusters identified corresponded “relatively closely” to the DSM classes, but many criteria failed to belong to expected classes. For example, of ten criteria in the borderline cluster, four were from other disorders, and most schizoid and schizotypal criteria fell in the same cluster. Arntz (1999) reported a confirmatory factor analysis of DSMIII-R criteria assessed by semistructured interview which also supported the DSM classification. Most criteria loaded factors as predicted, although several correlated more highly with other factors. Factor analyses using self-report measures, however, provide less support for the DSM classification either in terms of relationships between criteria or number of factors identified (e.g., Hyler et al., 1990). Blais and Norman (1997) also found that the majority of DSM-IV criteria failed to discriminate clearly between classes, most correlating with more than one category. The available data therefore question whether personality disorder criteria are optimally organized by the categories proposed in the DSM classification. As is acknowledged in DSM-IV, the A, B, and C clustering has not been well validated. Although Morey’s (1988) analysis provided support for the DSM classes, the hierarchical relationships of his 11 clusters gave rise to
HIGHER-ORDER DIMENSIONS
599
two rather than three higher-order clusters. Morey labelled these “anxious rumination” and “behavioral acting out.” Other multivariate studies have produced variable findings, partly reflecting differences in populations, assessment methods, and analytic procedures (Widiger & Frances, 2002). Nevertheless, exploratory factor analyses have yielded some consistencies, the most common finding being that four second-order dimensions underlie the DSM-III and DSM-III-R classifications (e. g., Blackburn & Coid, 1998; Kass, Skodol, Charles, Spitzer, & Williams, 1985; Mulder & Joyce, 1997). Mulder and Joyce described these as “the four A’s”, antisocial (paranoid, antisocial, borderline, histrionic, narcissistic, passive-aggressive), asocial (schizoid, schizotypal), asthenic (dependent, avoidant), and anankastic (obsessive-compulsive). This structuring seems inconsistent with either the DSM three clusters or Morey’s two higher-order clusters. The antisocial factor corresponds to Morey’s “acting out” cluster, but the remaining three are merged in his “anxious rumination” cluster. In a hierarchical analysis, the number of dimensions identified will vary according to where the hierarchy is “sliced” (Harkness, 1992), and none of these factor analyses derived oblique solutions that might reveal correlations between factors. It is therefore possible that two or three higher-order factors might represent a better fit to the data than the four second-order factors commonly found. Moreover, apart from Morey’s analysis, which began with the individual DSM criteria, these studies examined relationships between personality disorders as defined by the DSM classification. Therefore, higher-order structure of the DSM classes might partly reflect inappropriate placement of some of the criteria at the first-order level. This paper reports two studies in a forensic psychiatric population. In the first, we examine the hierarchical structure of the DSM-IV personality disorders. We investigate the primary factor structure of the personality disorder criteria using exploratory factor analysis to identify homogeneous classes. We then determine the second-order and higher-order structure of these first-order factors through exploratory factor analysis with oblique rotation, and test the adequacy of the structure indicated through confirmatory factor analysis. In the second study, we explore the meaning of the emerging dimensions of personality disorder by examining their relationship to established higher-order dimensions of personality.
STUDY 1 METHOD PARTICIPANTS The sample comprised 168 male forensic psychiatric patients detained in two of the four high security psychiatric hospitals in Britain, who completed assessments in a survey of personality disorder and psychopathol-
600
BLACKBURN ET AL.
ogy (Blackburn, Logan, Donnelly, & Renwick, 2003). Their mean age was 37.02 years (SD = 10.01) and mean length of detention 80.52 months (SD = 68.82). All were detained under mental health legislation because of perceived dangerousness. We attempted to obtain a representative sample, but the final sample (67% of those approached) is biased toward clinically stable patients who could participate in lengthy interviews and testing. Primary clinical diagnosis agreed at the patient’s most recent multidisciplinary case conference was personality disorder for 67 (40%) patients and mental illness (mainly schizophrenia) for 101 (60%) patients; however, formal assessment using structured diagnostic measures indicated a wide range of Axis I disorders and extensive Axis I-Axis II comorbidity in both clinical groups (see Blackburn et al., 2003).
INSTRUMENTS International Personality Disorder Examination (IPDE). The IPDE (World Health Organisation, 1995) yields categorical diagnoses of the 10 DSM-IV Axis II disorders and also dimensional scores (the sum of criteria in each category) from a semistructured interview and file information. IPDE interviews were conducted by a single interviewer at each hospital. A definite categorical diagnosis of at least one personality disorder was obtained by 114 (68%) patients, and a probable (subthreshold) diagnosis by 142 (85%) patients. The most prevalent definite diagnoses were antisocial (51%), narcissistic (16%), and borderline (11%), least prevalent being dependent (2%) and schizotypal (1%). However, current analyses are based on dimensional scores. We assessed inter-rater reliability from independent ratings of videotaped interviews (n = 10). The mean intra-class correlation for dimensional scores across the ten disorders was .77, ranging from .66 (dependent) to .92 (schizoid). Mean internal consistency (coefficient alpha) was .76, and most measures demonstrated acceptable homogeneity in exceeding the minimum requirement of an alpha of .70, except for schizotypal (.63) and borderline (.64). PROCEDURE Interviewers received training in the IPDE from staff involved in the development of the instrument. The research was approved by local ethics committees and patients gave written consent to participate. Statistical Analyses. We carried out three sets of analyses. We first examined the dimensional structure of the 93 individual DSM-IV criteria (including 15 conduct disorder criteria) by means of principal axis factor analysis followed by rotation to an oblique solution using Promax. This analysis extracts the common factors in the correlation matrix by analyzing only shared variance, and hence eliminates unique and error variance from factors. Although the subject-variable ratio is below the 5 : 1 ratio generally considered a minimum for factor analysis, the sample size is sufficient to ensure reasonably stable correlations between variables. We then
HIGHER-ORDER DIMENSIONS
601
estimated scores on the resulting primary factors by the regression method, and examined the second-order factor structure, again using principal axis factor analysis. We rotated factors to an oblique solution using Promax to determine correlations between second-order factors and hence the possibility of higher-order (i.e., third-order) factors. Finally, we tested the adequacy of the dimensional structure indicated by the exploratory analysis by means of confirmatory factor analysis (CFA), analyzing the covariance matrix with maximum likelihood estimation procedures using EQS 6.1 (Bentler, 1995). This entails an exploratory rather than strictly confirmatory use of CFA; however, model construction and model testing are on a continuum (Bentler, 1995; Gerbing & Hamilton, 1996), and as Anderson and Gerbing (1988) observed, rather than being a strict dichotomy, “the distinction in practice between exploratory and confirmatory analysis can be thought of as that of an ordered progression” (p. 412). Goodness of Fit Indices. Goodness of fit (GOF) in CFA is tested statistically by χ2. A significant χ2 indicates statistical discrepancy between the model and sample covariance matrices, but χ2 is usually supplemented by other GOF indices evaluated by conventional cutoff values rather than significance tests. We used three indices recommended by Hu and Bentler (1999) as sensitive to model misspecifications. The Tucker-Lewis (or nonnormed fit) Index (TLI) and the Comparative Fit Index (CFI) are measures of incremental fit comparing the proposed and null (no common factors) models, values of .90 or greater conventionally indicating acceptable fit. The Root Mean Square Error of Approximation (RMSEA), a measure of absolute fit, is the average residual correlation, values of up to .08 indicating acceptable fit. Also included was the normed chi-square (χ2/df), an index of parsimony with acceptable values ranging from 1.0 to 2.0 (Hair, Anderson, Tatham, & Black, 1998).
RESULTS AND DISCUSSION PRIMARY FACTOR ANALYSIS We first examined the factorability of the matrix of correlations between the 93 items (DSM-IV criteria) using two commonly used tests (Hair, Anderson, Tatham, & Black, 1998). The Bartlett test of sphericity yielded a highly significant value, 8522.18 (4278), p
