Abstract
Psychiatric classification is an enduring focus of debate. Despite the advances of the last quarter century, the taxonomy of mental disorder has yet to reach a state of professional consensus. Whether mental disorders should be represented as discrete categories or as continuous dimensions is one of the most fundamental points of contention. The prevailing systems of classification make some allowance for continuous variation, but assume that a categorical representation is generally appropriate: disorders are laid out as distinct types and diagnosed as present or absent. Critics of categorical diagnosis argue that it embodies a medicalized view of mental disorder and reifies disorders rather than recognizing their pragmatic character [1]. Some propose that a dimensional taxonomy should be introduced for at least some conditions [2].
As McHugh and Slavney argue [3], the categorical versus dimensional distinction runs deeply through psychiatric thinking. Traditionally, biological psychiatrists have taken a ‘disease perspective [which] rests on a logic that captures abnormalities within categories’ (p.15). Psychodynamic psychiatrists are more apt to take a dimensional perspective, in which disorder arises out of psychological vulnerabilities and environmental provocations that vary in degree rather than kind, a position dubbed the ‘continuum fallacy’ by one biomedical critic [4]. Compared to their medically trained colleagues, psychologists tend to favour dimensional models of individual differences [5, 6], and even to hold a prejudice against categories [7]. These views exist in tension in psychiatric research, where continuous measurement of syndromes sits uneasily alongside categorical diagnosis.
Categorical and dimensional views of psychopathology each have their advantages and disadvantages [8], and to some extent choosing between them can be a matter of theoretical preference. Dimensional variation exists within and between categories – at an observable level most psychological variation is continuous – and quantitative variation can be simplified into categorical distinctions. However, the choice between categories and dimensions at the level of latent structure is also an empirical matter, and resolving it has implications for psychiatric classification and the assessment and explanation of mental disorders. Instead of simply assuming that mental disorders are categorical or dimensional, this question should be a priority for empirical research.
Distinguishing between categorical and dimensional models of latent structure is a particularly knotty statistical problem. For example, latent dimensions can appear discontinuous when they are assessed in certain ways [9]. Several methods have been employed in psychiatric research, all of which have important limitations. Kendell for example, proposed that latent categories can be inferred when scores on symptom measures are bimodally distributed, so that a ‘point of rarity’ exists between the two modes [8]. Despite its intuitive appeal, bimodality is an unreliable criterion for assessing latent discontinuity [10]. When the observable manifestations of two latent categories overlap, their combined distribution is often unimodal. The excessive stringency of the ‘point of rarity’ criterion may explain Kendell's claim that ‘we have not yet established the existence of any disease entities within our territory’ (p.67).
Another popular method for psychiatric classification is cluster analysis, an ‘aggregative’ procedure for forming coherent groups of cases or clinical features. This diverse set of procedures also has a quite limited capacity to distinguish between categorical and dimensional models. In addition to having a poor record of correctly identifying latent structure in simulation studies [11], cluster analysis has a more basic limitation. Just as factor analysis invariably yields dimensions, cluster analysis invariably yields categories, whether or not underlying categories exist. Cluster analyses often group individuals in pragmatically useful ways, but it is unwise to mistake these groups for latent categories.
Mixture modelling is another widely used taxonomic procedure in psychiatric research [12], which ‘disaggregates’ observed distributions into multiple components. This method compares the statistical fit of models containing one (i.e. dimensional), two (i.e. categorical) or more components to these distributions. Although it is not subject to the same categorical bias as cluster analysis, this method is still problematic. Most importantly, it must make assumptions about the appropriate form of the latent distributions (e.g. normal). If these assumptions are incorrect then the results of the analysis may be invalid.
An alternative to these procedures that avoids some of their limitations is the family of taxonomic procedures developed by Paul Meehl and his colleagues [for procedural details see 13,14]. To distinguish these methods from other approaches to numerical taxonomy, Meehl coined the term ‘taxometrics’. Taxometric procedures examine the covariation among indicators (e.g. symptoms or test scores) of a latent variable such as a hypothesized mental disorder, seeking patterns that are diagnostic of latent categories (‘taxa’) or dimensions. A taxon, in this sense, is a nonarbitrary latent (‘genotypic’) category whose members differ qualitatively from nonmembers. Members of a taxon may nevertheless vary in apparently seamless ways in their observable (‘phenotypic’) characteristics, just as dogs differ in kind from cats but vary systematically in their attributes. Taxometric procedures make no strong assumptions about the form of the underlying distribution of taxon members, and have no inherent bias towards categorical or dimensional findings. A distinctive feature of taxometrics is the use of multiple independent procedures to assess whether a taxon exists and to estimate taxon parameters (e.g. prevalence). Conclusions are only drawn when findings are consistent across these procedures.
Taxometric procedures have received support from a small but growing body of simulation research, which shows that they are capable of discriminating between latent categories and dimensions with high accuracy given adequate sample size (< 300) and indicator validity [15]. This research also demonstrates that taxometric inference is quite robust under a variety of adverse psychometric conditions, and that taxometric procedures discriminate latent structure as well as or better than mixture and cluster analyses [16]. Although it has not been established that taxometric procedures are superior to all of their statistical alternatives, and they may turn out to have limitations in some circumstances, they clearly represent a promising new approach.
Taxometric methods have been employed in more than 60 substantive studies to date [17], with the pace of research accelerating in recent years. However, this research has yet to be disseminated widely outside the audience of specialists and is only beginning to cross the disciplinary divide from clinical psychology into psychiatry. Until very recently the volume of research was too small to offer a meaningful overview of psychopathology. A comprehensive review of the published taxometric evidence on the categorical versus dimensional status of mental disorders is timely.
Mood disorders
Taxometric studies have investigated mood disorders from three angles: the continuity versus discontinuity of general depression; of depressive subtypes; and of temperamental vulnerability. Whether depression differs in kind or only by degree from normal mood variations has long been a topic of theoretical debate within abnormal psychology [18, 19] and sociology and epidemiology [20], and taxometrics offers a rigorous means of assessing it. Five recent taxometric studies are very consistent, uniformly favouring the continuum position [21–25] in both clinical and non-clinical samples and across a variety of measurement instruments.
The evidence is equally consistent in subtyping studies. Taxometric analyses of several proposed nonendogenomorphic subtypes of major depression have all failed to support categorical models [26–28]. However, studies of melancholic or ‘nuclear’ subtypes, distinguished by high severity and classical endogenous features, consistently find evidence of a latent category [26, 27, 29, 30], although these findings have been somewhat controversial. This work speaks forcefully to an issue that has exercised academic psychiatrists for well over a century [31]. Although it is still rather sparse, taxometric research on the subject indicates that the dimensional and categorical views of depression each have merit, the former for general distress – Freud's ‘neurotic misery’ does indeed shade smoothly into ‘everyday unhappiness’ – and the latter for melancholia. Reconciling these distinct views to account for the emergence of a discrete subtype where no broader depressive ‘type’ exists may require a threshold model [32].
Taxometric research on temperamental vulnerabilities for mood disorders is in its infancy, but one large study [33] has found that hypomanic temperament, a proposed diathesis for bipolar disorder, is dimensional. No published studies of vulnerability to unipolar depression have appeared. Thus the current evidence inclines towards a continuum view of mood disorders, with the important exception of melancholia. Although this evidence is still thin, and largely neglects the bipolar spectrum, it suggests that proposals of discrete depressive subtypes should initially be met with skepticism, and that no underlying boundary separates the depressed and the non-depressed.
Anxiety disorders
Taxometric studies of anxiety disorders have only recently begun to appear. As with mood disorders, current evidence tends to favour dimensional models. A study of Posttraumatic stress disorder symptomatology in combat veterans supported a dimensional view [34]. Perhaps less surprisingly, a study of pathological worry in a student sample [35] also found no evidence of a latent category, and implies the continuity of generalized anxiety disorder. The only intimation of a discrete category in the anxiety disorder spectrum relates to social phobia, with one unpublished study [36] finding ambiguous evidence for a taxon of socially anxious individuals marked by extreme fears of public scrutiny. The existence of a taxon in this domain is rendered more plausible by taxometric evidence [37], replicated with mixture modelling [38], that inhibited temperament in childhood reflects a latent category. However, a finding that avoidant attachment style falls on a continuum in adults and infants [39, 40] also counts against this possibility. Any unqualified conclusions about the status of anxiety disorders are therefore premature, especially in the absence of research on obsessive-compulsive and panic disorders. The same can be said for proposed vulnerabilities for anxiety disorder, with the one study to date [41] supporting a dimensional model of anxiety sensitivity, a diathesis for panic disorder.
Eating disorders
Similarly limited is taxometric research on eating disorders, although with stronger evidence for at least some taxa. The categorical view of bulimia nervosa (BN) has been supported by three taxometric studies of undergraduate and clinical samples [42–44], but challenged by a more recent investigation [45]. The discreteness of anorexia nervosa (AN) has been supported by one of these [44]. One study [43] further indicated that the DSM-IV's restricting subtype of anorexia nervosa is discrete, but that the binge-eating/purging subtype belongs within the bulimia nervosa class. This finding implies that the DSM-IV draws the diagnostic boundary between AN and BN incorrectly, basing it on weight when the presence versus absence of bingeing and purging might be a more appropriate distinction.
Dissociative disorders
Two published taxometric studies of dissociative phenomena have appeared. The first [46], using a mixed sample of dissociative identity disorder (DID) cases and normal controls, found that depersonalization, derealization and dissociative amnesia were associated with a latent category. However, phenomena reflecting absorption and imaginative involvement were better modelled as continua. This analysis usefully parses dissociation into pathological and non-pathological components, and also indicated that DID might be overdiagnosed by the current diagnostic criteria. The second taxometric study [47], conducted in a general population survey, replicated the categorical status of pathological dissociation, estimated its point prevalence at a substantial 3.3%, and established that genetic contributions to it are negligible, consistent with traumatic aetiology. Although based on few studies, these findings are consistent and they matter for clinical assessment and practice. In a related vein, a taxometric study of normal adults has supported the categorical status of hypnotic susceptibility [48], a finding with possible implications for the study of dissociation given that hypnotisability is often claimed to play an aetiological role in dissociative conditions [49].
Personality disorders
The categorical versus dimensional status of DSMIV's axis II is a particularly important question. Dimensional models of personality disorders (PDs) are popular; there have been proposals to classify them dimensionally [2, 50, 51]. Many PDs are strongly associated with dimensions of normal personality such as the five-factor model [52], and most appear to shade imperceptibly into normal personality variation. Dimensional measures of PD symptomatology often predict clinical phenomena better than categorical diagnoses, and the troublingly high levels of comorbidity among some PDs might be understood better in terms of shared loadings on underlying dimensions of abnormal personality. Taxometric investigations of the latent structure of PDs are well suited to the task of testing the dimensional view.
Thus far only three PDs have undergone taxometric analysis, but the results are quite consistent [53]. The largest number of studies has addressed schizotypal PD and its childhood precursors [54–62]. These studies, conducted in normal and clinical samples and using a wide variety of questionnaire, interview and cognitive measures, overwhelmingly support categorical models of schizotypy, consistent with Meehl's influential theory of schizophrenia [63]. Almost without exception these studies find that schizotypal PD criteria and measures of perceptual aberration, magical ideation and social anhedonia all pick out a taxon whose prevalence is about 5% among normal adults. The taxon is detectable among at-risk children [55] and is stable from adolescence into middle age [62]. Although the schizotypal taxon is broader than DSM-IV schizotypal PD, the diagnosis appears to be a relatively severe subset of this taxon.
Three studies of antisocial PD, psychopathy, or their childhood antecedents also support a categorical model. The first study [64], using an adult offender sample, found evidence of a taxon defined by indicators of psychopathy, specifically chronic antisocial behaviour beginning in childhood. Criminality per se and the interpersonal and affective components of psychopathy were not categorical. This finding was replicated among offenders using measures of psychopathic tendencies and DSM-IV antisocial PD symptoms [65], and in a study of middle-school-age antisocial boys in a community sample [66]. By implication, antisocial PD is well understood as a latent category, identified better by antisocial conduct than by the less behavioural aspects of psychopathy, and is not equivalent to adult criminality.
The two published taxometric studies of borderline PD have both yielded dimensional conclusions [25, 67]. Should these findings be replicated it will suggest that PDs represent a mixture of latent categories and continua, and that any uniform preference for the categorical or dimensional classification of axis II misrepresents some disorders. Strong conclusions about the structure of PDs in general await taxometric studies of the seven PDs that have yet to be investigated.
Limitations
The research reviewed above indicates that taxometric research is helping to clarify the structure of several forms of psychopathology. However, it is important to acknowledge several limitations of this research, and to maintain a critical perspective towards it. Psychiatric researchers and practitioners are appropriately skeptical of exaggerated claims for the illuminating power of new statistical methodologies, and taxometric methods must prove rather than merely assert their reliability.
A first limitation is the still rather small quantity of taxometric research that has been conducted. Until more studies accumulate and findings are replicated across multiple samples, research groups and assessment instruments the robustness of these findings will remain unproven. Simulation research with taxometric procedures gives some reason for confidence in this regard, but it is still in its infancy and has barely begun to appear in mainstream statistical journals. On a related note, taxometric research has neglected several large regions of psychopathology. In particular, only one unpublished study of a psychotic disorder has been conducted, and no robust work on organic brain conditions such as dementias has appeared [17]. Similarly, child psychopathology has been almost completely neglected [68]. Taxometric researchers have generally failed to assess the structure of mental disorders in longitudinal perspective, an approach that might clarify how certain disorders emerge in the course of development and whether taxa are stable over time [69].
A second limitation is the uncertain relation of taxometric analysis to some alternative methods for testing between categorical and dimensional models. It is not yet clear, for instance, whether taxometric methods are more or less reliable guides to structural inference than other methods for comparing latent class and latent trait models [70, 71], or how their performance compares to widely used statistical procedures such as latent class analysis. Future research should establish the conditions under which particular methods are to be preferred and how they might be used in a complementary fashion. For example, although taxometric analysis may outperform cluster analysis in detecting the presence of latent categories, cluster analyses may sometimes be better at classifying individuals into these categories [72]. It is unlikely that any one method will be superior under all circumstances, and the findings of taxometric and other taxonomic methods will ultimately need to be pooled and integrated.
A third class of limitations is internal to taxometric methods themselves. Several authors have drawn attention to pitfalls involved when taxometric researchers employ dichotomous (e.g. present/absent) indicator variables [73], which may wrongly suggest that a taxon exists. Although procedures have been developed to guard against this possibility [74], and continuous indicators are now standard in taxometric studies, some early taxometric findings based on dichotomous indicators may be suspect. Similarly, taxometric studies have often been overly reliant on self-report scales, to the neglect of data drawn from clinical interviews or neuropsychological or biological tests. Not only is it desirable to employ a variety of data types in taxometric analyses, but selfratings may sometimes be particularly problematic. It has recently been shown [75] that raters' expectations about the latent structure of a phenomenon may influence their ratings in a way that may fool taxometric analyses. While it must be noted that this problem is also likely to affect other forms of taxonomic analysis, it implies that taxometric researchers must take greater care not to rely exclusively on self-report data, especially where raters are likely to have strong expectations about the nature of the phenomenon at issue
In sum, taxometric analysis remains a developing research methodology that has promise but is not without its problems and challenges. These have begun to receive critical attention [76], and the next decade is likely to see an increasing amount of critique and refinement of taxometric methodology.
Overview
Despite the limitations discussed above, the taxometric evidence that is now at hand allows some tentative conclusions about the structure of psychopathology. Most importantly, it offers no comfort to dogmatic positions on the issue of categorical versus dimensional models. Latent categories and dimensions both appear to be widely distributed. Dimensional models receive extensive support in the broad neurotic spectrum, predominating among the mood and anxiety disorders. However this support is qualified, as melancholic depression appears to be better understood as a latent category and there are intimations of latent categories in the domain of social anxiety and inhibition. Moreover, plausibly categorical conditions such as bipolar and obsessive-compulsive disorders have yet to be studied taxometrically. Discrete categories appear to exist among the eating disorders, although their boundaries do not appear to map perfectly onto DSM-IV disorders. Categorical models also enjoy more support in the domain of PDs than might have been anticipated, a finding that might give pause to advocates of dimensional classification.
Implications
If taxometric research is to go beyond botanizing, it must have implications for psychiatric theory and practice. These implications are important and varied, bearing on classification, assessment, aetiology and the theorizing of mental disorders.
Classification
The clearest implications of taxometric findings involve psychiatric classification. Most obviously, findings that support the existence of latent categories can be taken to vindicate categorical diagnosis, and dimensional findings to support the development of dimensional alternatives. Second, when evidence of a latent category is obtained, it can be used to refine the identification and description of the disorder. Taxometric procedures estimate the sample prevalence of a taxon, thus enabling a principled assessment of its community prevalence and allowing diagnostic thresholds to be calibrated to identify taxon members optimally. It might be determined, for example, that the cutoff on a disorder's symptom checklist is too liberal or too conservative, and that diagnostic accuracy is improved by adjusting it. In addition, taxometric methods estimate the validity of particular indicators of a taxon, so that weak or invalid symptoms can be removed or replaced.
However, the role of taxometrics in revising psychiatric classifications is not just a matter of winnowing out (valid) taxa from (invalid) dimensions. Kessler [77] has recently proposed a two-step process in which, if taxometric research supports a categorical model, the disorder's diagnostic criteria and prevalence are fixed by its internal structure. If a dimensional model is supported, a second step attempts to identify a diagnostic threshold by examining the association between the dimension and external, clinically relevant outcomes, in the same manner as diagnostic thresholds are set for hypertension and obesity. The implication of this proposal is that categorical and dimensional conditions both have a place in psychiatric classifications, the latter defined more in pragmatic or utilitarian terms. This review suggests that a psychiatric classification based exclusively on latent categories would omit many prominent forms of psychiatric suffering and disability. The development and revision of classification systems cannot be based solely on the detection of latent categories by taxometric and other statistical procedures, and pragmatic considerations will always have a role to play. Claims that taxometrics will ‘solve’ the classification problem in psychopathology [13] are clearly overstated.
Another classification-related implication of the categorical versus dimensional issue concerns comorbidity, the sometimes questionable diagnosis of multiple conditions in the same person. If taxometric evidence suggests that both codiagnosed conditions are taxa then it is reasonable to describe them as comorbid as they represent two distinct conditions [78]. If one or neither diagnosis corresponds to a taxon then it is inappropriate to refer to comorbidity. Instead the diagnoses are more properly thought of as correlated or overlapping aspects of a single underlying condition.
Assessment
Taxometric evidence also bears on psychiatric assessment, as the categorical versus dimensional status of a disorder fundamentally constrains how it should be measured [79]. If the disorder is dimensional, then it is a misrepresentation to assess it as categorically present or absent. When this is done, meaningful continuous variation is overlooked, resulting in a substantial loss of statistical power in research and of nuance in clinical formulation. In addition, the cut-point at which the dichotomous assessment decision is made is apt to be arbitrary. The situation is more complex when the disorder is categorical. Categorical assessment has a stronger justification in this case, but it is by no means mandatory. The existence of a taxon does not rule out the existence of meaningful variations in severity among taxon members or in resemblance to the category among non-members. Consequently it may sometimes be preferable to measure latent categories on continuous dimensions. When doing so, however, it is important not to forget that a discontinuity underlies the measurements, or that assigning individuals to categories may prove to be more valid.
Aetiology
The latent structure of a disorder represents the outcome of a pathological process. Taxometric findings should therefore have implications for aetiology. Different causal models are appropriate for latent categories and dimensions. If a mental disorder falls on a continuum, then it is likely to reflect a combination of many, relatively small contributing factors. An individual's position on the dimension will be determined by how many factors are present and to what degree. If a disorder corresponds to a taxon, however, some sort of dichotomous (i.e. present versus absent) causal factor must usually be invoked. If the difference between cases and non-cases is a difference in kind, then a causal model must explain what is responsible for this qualitative difference. Generally such models will propose some form of ‘specific aetiology’, which could be a major gene, a discrete biochemical abnormality, or a pathogenic environmental influence. Alternatively a threshold model could be proposed, in which people who exceed a critical value on a vulnerability continuum develop the disorder. In short, when taxometric studies point to the existence of a latent category, some kind of dichotomous causal factor will probably be required to account for it. Similarly, when a latent dimension is found, aetiological models that incorporate such dichotomous causes are likely to be mistaken: if a cause is categorically present or absent it is difficult to imagine how it might generate a continuous dimension.
Theory
Finally, the issue of categorical versus dimensional latent structure matters for psychiatric theory and the fundamental question of how mental disorders are to be conceptualized. To the extent that taxometric analyses support categorical models it is not unreasonable to consider some mental disorders to be entities of a sort. When a taxon is detected, the disorder would seem to have a nonarbitrary and at least somewhat ‘objective’ basis. What Jablensky [1] describes as the ‘reification fallacy’ seems less fallacious in this instance than it would for conditions that are dimensional.
‘Essentialism’ is perhaps more troublesome here than reification. When latent categories are mooted, people often infer a pathological ‘essence’, an identity-fixing attribute or microstructure that is shared by all category members and causes their disturbance. Commonly, this essential attribute is assumed to be biological, defining the mental disorder as a natural kind [80–82]. The concept of ‘disease’ is typically used and interpreted in just such an essentialist fashion [3, 83, 84]. This tendency to infer a natural, necessary and defining cause when a discrete category is observed is evident among lay people [85], but it represents a misconception. A taxon need not have a biological cause. As Meehl [13] notes: ‘the political taxon Trotskyist is a more tightly bound syndrome than any in DSM’ (p.274). Although biomedical and psychosocial approaches to psychiatry are aligned with categorical and dimensional views of mental disorder, respectively, it would be beneficial for psychiatric theorizing if the categorical versus dimensional issue were separated from the issue of causation. Categorical models do not imply that mental disorders are disease entities, nor do dimensional models pose sharp challenges to biomedical explanation. The categorical versus dimensional issue is better regarded as an empirical matter with a variety of implications, rather than as a theoretical assumption that carries explanatory baggage.
Conclusion
Taxometric research has made a promising start to the important task of clarifying the latent structure of psychopathology, although a great deal more research is needed to examine additional mental disorders, to replicate previous findings and to test the robustness of taxometric methods. Research conducted to date is encouraging in its consistency, and tends to support a pluralistic view of psychiatric classification. Some disorders appear to represent discrete categories, whereas others fall on a seamless continuum with psychological normality. Distinguishing between these alternatives has important implications for clinical practice and research. The process of revising psychiatric classification would do well to heed the findings of taxometric research. Psychiatric theorists might also profitably explore how categorical and dimensional understandings bear on the explanation and formulation of mental disorders.