The five-factor model of the Positive and Negative Syndrome Scale II: A ten-fold cross-validation of a revised model

Abstract

Objective

The lack of fit of 25 previously published five-factor models for the PANSS items, can be due to the statistics used. The purpose of this study was to use a ‘new’ statistical method to develop and confirm an improved five-factor model. The improved model is both complex and stable. Complex means that symptoms can have multiple factor loadings, because they have multiple causes, not because they are ill defined. Stable means that the complex structure is found repeatedly in validations.

Methods

A ten-fold cross-validation (10 CV) was applied on a large data set (N = 5769) to achieve an improved factor model for the PANSS items. The advantages of 10 CV are minimal effect of sample characteristics and the ability to investigate the stability of items loading on multiple factors.

Results

The results show that twenty-five items contributed to the same factor all ten validations with one item showing a consistent loading on two factors. Three items were contributing to the same factor nine out of ten validations, and two items were contributing to the same factor six to eight times. The resulting five-factor model covers all thirty items of the PANSS, subdivided in the factors: positive symptoms, negative symptoms, disorganization, excitement, and emotional distress. The five-factor model has a satisfactory goodness-of-fit (Comparative Fit Index = .905; Root Mean Square Error of Approximation = .052).

Conclusions

The five-factor model developed in this study is an improvement above previously published models as it represents a complex factor model and is more stable.

Introduction

Despite being the most widely used instrument to assess the symptoms of schizophrenia, the Positive and Negative Symptom Scale (PANSS) does not have a satisfactory five-factor model. None of 25 previously published five-factor models could be confirmed using confirmatory factor analysis (CFA; van der Gaag et al., 2006).

Three possible explanations have been put forward to explain this lack of fit:

• Different subgroups of schizophrenia patients have different symptom structures;

• The multiple loading items are ill-defined and should be discarded or rewritten;

• The multiple loading items are well defined, but have multiple causes.

The first explanation seems not very likely. As was discussed in the companion article, medication use, gender, age and cognitive decline do not seem to influence the factor structure. Stage of illness and the duration of illness have shown only some minor variation in the model. Furthermore, the subsamples in this study that varied widely in age, severity, inpatient–outpatient status, and duration of illness, were tested for heterogeneity and were found to be homogeneous. That finding is a replication of White et al. (1997). This means that despite the fact that different subsamples vary, the structure of the symptoms on the PANSS is the same for all patients with schizophrenia.

The second explanation of ill defined items that try to cover different dimensions in one rating cannot be ruled out as an explanation. The instability of the item G12 (lack of judgment and insight) can be due to the fact that the item tries to rate general social judgment, insight into the disorder and acceptance of treatment in a single rating (White et al., 1997). When these three dimensions are not highly correlated in schizophrenia patients, the item will be interpreted in different ways by different raters. When an item loads on three or more factors, most researchers have removed the item as ill defined or not interpretable. This decision might harm the model, when the symptom is not ill defined, but has multiple causes.

This third explanation of symptoms having multiple causes, has not been researched yet and can be investigated by using ten-fold cross-validation (10 CV) (Breckler, 1990, MacCallum et al., 1990). This statistical method is ‘new’ as it has not been applied before in modeling the symptoms of schizophrenia. The advantage of 10 CV is that the developed final model is minimally dependent on sample characteristics, while exploratory factor analysis (EFA) is notoriously dependent on sample characteristics. A second advantage is that 10 CV can investigate how stable items are represented over multiple validations. A high degree of stability predicts that the model will probably be confirmed again in a future CFA (Efron and Tibshirani, 1993).