Statistically derived symptom dimensions or clusters

Factor analysis has been applied to psychiatric rating scales since the 1960s.¹²³ Essentially, factor analysis and related methods reduce the covariation of the primary data matrix to covariances of small numbers of latent factors which account for the interrelationships among the primary variables and explain a proportion of their variance. Based on a relatively small number of input variables (SANS/SAPS scores), a three-factor structure was proposed by Liddle¹²⁴ and replicated by other investigators.^{125, 126, 127} In this model, negative symptoms load on a single factor of 'psychomotor poverty', while positive symptoms split into a delusions-and-hallucinations factor ('reality distortion') and a thought-and-speech disorder factor ('disorganization'). The model has been shown to be longitudinally stable¹²⁸ and replicable in non-European populations.^{129, 130} It was incorporated in DSM-IV³ and tested in field trials.¹³¹

Results of factor-analysis of symptomatology depend strongly on the content of the clinical rating scales used as input. Studies using the SANS and SAPS result in different solutions from those produced by the PANSS,¹³² BPRS,¹³³ or OPCRIT,¹³⁴ including a general neurotic syndrome factor;¹³⁵ excitement and depression;^{136, 137} paranoid, first-rank delusions and first-rank hallucinations;¹³⁸ premorbid adjustment deficits factor;¹³⁹ and autistic preoccupation factor.¹⁴⁰ In a large sample of probands with schizophrenia, McGrath et al.¹⁴¹ identified five factors (positive, negative, disorganized, affective, and early onset/developmental), which were associated with risks of psychoses and affective disorders in relatives. In a series of factor analyses based on an expanded list of 64 psychopathological symptoms, Cuesta and Peralta¹⁴² concluded that a hierarchical 10-dimensional model provided the best fit on statistical and clinical grounds. Factor solutions, therefore, are not unique and the question 'how many factors parsimoniously describe the symptomatology of schizophrenia?' can only be answered in the context of a particular selection of symptoms and measurement methods. Therefore, factor-analytical studies suggesting 'established' dimensions or syndromes of schizophrenia should be viewed with caution, considering the diversity of clinical populations and the limitations of the instruments used to generate the input data.

Whereas factor analysis groups variables, cluster analysis groups individuals on the basis of maximum shared characteristics. Farmer et al.¹⁴³ identified two clusters into which patients with schizophrenia could be fitted, based on their scores on a checklist of 20 symptom and history items: one characterized by good premorbid adjustment, later onset, and well organized delusions, and another including early onset, poor premorbid functioning, incoherent speech, bizarre behaviour, and family history of schizophrenia. However, using the Positive and Negative Syndrome Scale,¹³² Dollfus et al.¹⁴⁴ obtained four quite different distinct clusters, corresponding to positive, negative, disorganized and mixed symptomatology. Thus, cluster analysis is as dependent on the selection of input variables as factor analysis.

Latent class analysis (LCA) assumes the existence of a finite number of mutually exclusive and jointly exhaustive groups of individuals, within which person characteristics, for example, responses to symptom items, are: (a) determined by class membership and (b) locally independent.¹⁴⁵ A latent class typology of schizophrenia, proposed by Sham et al.,¹⁴⁶ using data on 447 patients with nonaffective psychoses, suggested the existence of three subgroups: a 'neurodevelopmental' subtype resembling the hebephrenic form of the disorder (poor premorbid adjustment, early onset, prominent negative, and disorganized features); a 'paranoid' subtype (less severe, better outcome); and a 'schizoaffective' subtype (dysphoric symptoms). In an epidemiological sample of 343 probands with schizophrenia and affective disorders, Kendler et al.¹⁴⁷ found six latent classes, broadly corresponding to the nosological groups of 'Kraepelinian' schizophrenia; major depression, schizophreniform disorder; schizoaffective disorder (manic), schizoaffective disorder (depressed), and hebephrenia. Increased risk for schizophrenia spectrum disorders was found among the relatives of subjects assigned to the schizophrenia and schizophreniform classes, while increased risk for affective disorders was only found in the relatives of patients assigned to the major depression and schizoaffective (depressed) classes. Similar results, using a combination of principal component (factor) analysis and LCA in an epidemiologically ascertained sample of 387 patients with psychoses, have been reported by Murray et al.¹⁴⁸

In contrast to conventional LCA, a special form of latent structure analysis, the grade of membership (GoM) model, allows individuals to be members of more than one class and represents the latent groups as 'fuzzy sets',^{149, 150} where individuals can be members of more than one set. The GoM model simultaneously extracts from the data matrix a number of latent 'pure types' and assigns each individual a set of numerical weights quantifying the degree to which that individual resembles each one of the identified pure types. When applied to the symptom profiles of 1065 cases in the WHO International Pilot Study of Schizophrenia,¹⁵¹ the method identified eight pure types of which five were related to schizophrenia, two to affective disorders and one to patients in remission, all showing significant associations with course and outcome variables used as external validators.

Top of page

Date: 2016-04-22; view: 704