The PANSS (Kay, Fiszbein, & Opler, 1987) is one of the most widely used measures of psychopathology in psychosis research, and was the only measure of negative symptom severity included in the EDEN battery. The PANSS is a 30-item instrument designed to measure a wide range of symptoms associated with schizophrenia. Symptom severity over the previous seven days is assessed by a trained rater on the basis of a semi-structured interview with the participant and the reports of professional carers or family members. Each symptom is rated on a 7- point scale from 1 (absent) to 7 (extreme) according to a set of symptom-specific anchoring criteria.
The PANSS items were originally grouped into three subscales: positive symptoms, negative symptoms and general psychopathology. However, it is now accepted that these a priori subscales are not an accurate reflection of the scales underlying factor structure (Kay, Opler, & Fiszbein, 2000). Numerous principle component analyses of the PANSS have been conducted, typically yielding four to seven factors, of which one corresponds to the negative symptoms construct (Fitzgerald et al., 2003). The negative symptoms factors identified by these analyses vary (Emsley, Rabinowitz, & Torreman, 2003; Wallwork, Fortgang, Hashimoto, Weinberger, & Dickinson, 2012), but none align with the original negative subscale. Indeed, it is now widely accepted that the PANSS negative subscale contains several items measuring symptoms that do not fall within the domain of negative symptoms (Kirkpatrick et al., 2006). As such, the negative subscale of the PANSS is an unsatisfactory tool for the assessment of negative symptom severity.
Due to the limitations of the original PANSS subscales, it is becoming increasingly common for studies using the PANSS to utilise a bespoke subscale structure based on a published factor model (Nicotra, Casu, Piras, & Marchese, 2015). However,
72
there remains much controversy surrounding the factor structure of the PANSS (Malaspina et al., 2014), making the choice of factor model to employ in using this strategy less than straightforward. Studies investigating the factor structure of the PANSS in schizophrenia have most commonly identified models with five factors, and the five-factor ‘pentagonal model’ (White, Harvey, Opler, & Lindenmayer, 1997) developed by the PANSS study group was included in the most recent PANSS manual (Kay et al., 2000). However, a subsequent independent study found that this model was an inadequate fit for data from a sample of 347 individuals diagnosed with schizophrenia (Fitzgerald et al., 2003). A recent attempt has been made to construct a ‘consensus’ five-factor model through identifying the most common item-factor assignments among 29 independent five-factor models (Wallwork et al., 2012). The resulting factor structure was found to be a good fit to data obtained from two independent samples from differing cultural backgrounds.
Such a consensus factor structure might be considered a suitable basis for the formation of a negative symptoms subscale for use in this thesis. However,
Wallwork et al.’s samples included only individuals with an established diagnosis of either schizophrenia or schizoaffective disorder and who were, on average, more than a decade older than the EDEN cohort. It cannot be assumed that a factor model confirmed in an older, diagnostically homogeneous sample can necessarily be successfully applied to an FEP cohort. Substantiating this assertion, Langeveld et al. (2013) examined the fit of five widely used PANSS factor models (including
Wallwork et al.’s consensus model) in a large FEP sample (n = 588) and found that none of the models tested met criteria for satisfactory model fit.
Use of PANSS symptom subscales based on an inadequate factor model may result in suboptimal sensitivity to change. As such, it is important to determine the best- fitting factor model for the population of interest when determining symptom subscales. Given a lack of a consensus regarding the optimum factor model of the PANSS in an FEP sample, the decision was taken to carry out a study to determine the factor structure of the PANSS in the EDEN cohort itself rather than choosing a published factor model. The factor model identified could then be used to determine the most suitable PANSS items to measure negative symptom severity for the
73
purposes of this thesis. This approach has been recognised as a valid means of ascertaining an appropriate subscale structure for the PANSS for the particular sample under investigation (Nicotra et al., 2015).
It should be noted that while none of the PANSS factor structures developed in schizophrenia samples were an adequate fit for Langeveld et al.’s FEP data, neither was the one factor structure developed in a sample with recent-onset psychosis. Thus their failure to confirm the published factor structures considered may represent a wider problem of lack of stability of PANSS factor structures across samples. A study that examined the goodness-of-fit of all previously published five-factor models of the PANSS in a sample of 5769 individuals diagnosed with schizophrenia failed to confirm the appropriateness of any of the models considered (van der Gaag et al., 2006a).
An important limitation of much work exploring the factor structure of the PANSS to date is the use of restrictive models that do not allow for the free estimation of cross-loadings, thereby restricting each item to load on only one factor. Some authors suggest that allowing free estimation of cross-loadings is necessary to
adequately reflect clinical reality and thus obtain satisfactory model fit (van der Gaag et al., 2006b; van den Oord et al., 2006). Following their failure to confirm any of the published five-factor models identified in the literature, van der Gaag et al. (2006b) used ten-fold cross-validation to develop a revised five-factor model. Ten- fold cross-validation involves randomly assigning participants to one of ten equally sized subsamples. Nine of these subsamples serve as training sets and the remaining subsample is used to test the validity of the resulting model. This process is then repeated with each of the subsamples in turn serving as the validation set.
Using this method, van der Gaag et al. demonstrated that a five-factor model can achieve good fit when items are permitted to load on more than one factor. Perhaps more importantly, they demonstrated the stability and clinical face-validity of such cross-loadings, indicating that they may be necessary due to some symptoms having multiple causes rather than certain PANSS items simply being ill-defined. The
74
negative symptom factor they identified was particularly stable; eight PANSS items – ‘blunted affect’ (N1)1, ‘emotional withdrawal’ (N2), ‘poor rapport’ (N3),
‘apathetic social withdrawal’ (N4), ‘lack of spontaneity and flow of conversation’ (N6), ‘motor retardation’ (G7), ‘uncooperativeness’ (G8) and ‘active social avoidance’ (G16) – loaded on the negative factor in all 10 cross-validations.
Van den Oord et al. (2006) also recognised the disadvantages of modelling the structure of the PANSS using restrictive models and thus used a combination of exploratory and confirmatory factor analysis to develop and assess the fit of a model that allowed items to load on multiple factors. The ‘Negative’ factor in the six-factor model they obtained was indicated by the items ‘blunted affect’ (N1), ‘poor rapport’ (N3), ‘motor retardation’ (G7) and ‘disturbance of volition’ (G13). However, the factor labelled ‘Withdrawn’ by the authors, indicated by ‘active social avoidance’ (G16), ‘emotional withdrawal’ (N2) and ‘apathetic social withdrawal’ (N4), could also be argued to reflect the negative symptoms construct.
For ease of comparison, the negative symptoms factors in van der Gaag et al. and van den Oord et al.’s models are presented alongside the negative factors from White et al.’s pentagonal model and Wallwork et al.’s consensus model in Table 3.1.
1Each PANSS items was labelled by the scale’s authors with a combination of a letter and a number.
The letter denotes which of the original subscales it formed part of (‘P’ for the positive subscale, ‘N’ for the negative subscale, and ‘G’ for the general psychopathology subscale).
75
Table 3.1. Summary of PANSS items assigned to the factor corresponding to the negative symptoms construct in four competing factor models.
PANSS Item White Wallwork Van der Gaag Van den Oord
N1 Blunted affect N N2 Emotional withdrawal W N3 Poor rapport N N4 Passive withdrawal W N6 Lack of spontaneity G5 Mannerisms and posturing G7 Motor retardation N G8 Uncooperativeness G13 Disturbance of volition N
G14 Poor impulse control G16 Active social
avoidance
W
‘ ’ = included in a single negative symptoms factor ‘N’ = included in van den Oord et al.’s ‘Negative’ factor ‘W’ = included in van den Oord et al.’s ‘Withdrawal’ factor
Note. Complete citations for the factor models compared are White et al. (1997), Wallwork et al. (2012), van der Gaag et al. (2006b) and van den Oord et al. (2006).
Several factor models were fitted to the data in the current study. Initially, exploratory factor analysis (EFA) was used to generate a factor model and
confirmatory factor analysis (CFA) used to test how well this model fitted the data. The advantage of this approach is that if it were possible to identify a factor model with adequate fit to the data using CFA then a single structural equation model incorporating both the measurement model for negative symptoms and longitudinal growth analyses would be able to be specified in the subsequent study. However, if (as was anticipated on the basis of the work by van der Gaag et al. and van den Oord et al. discussed above) it proved impossible to confirm the fit of the model suggested
76
by EFA using CFA, it was planned that exploratory structural equation modelling would be employed to determine the factor structure instead.
Exploratory structural equation modelling (ESEM) is a relatively new modelling technique (Asparouhov & Muthén, 2009) which combines advantages of both confirmatory and exploratory factor analysis (Marsh, Morin, Parker, & Kaur, 2014). Like EFA, ESEM does not require cross-loadings to be fixed at zero, allowing for the sort of complex factor models that van der Gaag et al. and van den Oord et al. argue are necessary to adequately reflect clinical reality and thus obtain satisfactory model fit. However, unlike EFA and in common with CFA, model fit indices can be obtained using ESEM, enabling the adequacy of the fit of the model to the data to be verified.