Recent studies of disease burden have demonstrated the importance of psychological disorders. For instance, depression was the fourth leading cause of disease burden, accounting for 4.4% of total disability adjusted life years in the world in 2000 [1]. The 12-item General Health Questionnaire (GHQ-12) has been widely used in many countries for detecting psychological morbidity. Some major national studies such as the British Household Panel Survey (BHPS) also employ this instrument [2]. Calibration of this instrument may therefore contribute significantly to a large community of researchers.

While the longer versions of the GHQ are normally considered multidimensional, the GHQ-12 is often regarded as measuring only a single dimension of psychological health. For example, Corti [3] analyzed the GHQ-12 data in the BHPS and maintained that the high Cronbach's alpha value indicated the unidimensionality of this instrument. However, several authors suggested that the GHQ-12 contained two or three clinically meaningful factors. Using principal component analysis, Politi et al. [4] identified two factors: general dysphoria and social dysfunction. Andrich and van Schoubroeck [5] suggested that the positively worded items formed one factor and the negatively worded items formed another. Graetz [6], Martin [7] and Worsely and Gribbin [8] proposed three different 3-factor models. In a multi-centre study, although considerable between-centre variation was found, the final solution tended to have either two or three factors [9].

Using confirmatory factor analysis (CFA) to analyze the BHPS data, Cheung [10] compared various models and found that the 3-factor model proposed by Graetz [6] gave the best fit. The factors are anxiety and depression (4 items), social dysfunction (6 items), and loss of confidence (2 items). In a study of employees in New Zealand, Kalliath et al [11] also employed CFA to compare various models. They also found that Graetz's 3-factor model gave better goodness-of-fit than the others. However, they maintained that none of the models they examined gave a sufficient level of goodness-of-fit. Hence they modified the instrument to propose a short (8-item) version of GHQ. In a study of college students and young adolescents in Australia, French and Tait [12] found that Graetz's model not only fitted the data better than other models, but also satisfactorily achieved some fit indices targets such as Comparative Fit Index > 0.95. In a study of a rural population in Australia [13], the model of Worsely and Gribbin fitted best and that of Graetz was second best.

While the structure of the GHQ-12 has been studied using factor analysis methods, the construct validity and usefulness of those resulting factors are not often tested. The question is whether the additional information provided by the 2 or 3 factors, if they exist, is clinically useful. In other words, will multiple scores be more useful than a total single score in helping us to understand respondents' health status?

The purpose of this study was therefore two-fold. First, we aimed to compare the previously proposed models of the GHQ-12 in an oriental population and identify the best-fitting one. It was not our objective to assess their absolute level of fit or to derive new model or version of the GHQ. Second, we aimed to assess whether the factors identified relate to clinical and health-related quality of life variables in different ways.

Figure 1 displays the standardized factor loadings and between-factor correlation of model IIIB. The factor loadings ranged between 0.72 and 0.90. The three factors were strongly correlated. The correlation between factor 1 (Anxiety and Depression) and factor 2 (Social Dysfunction) was 0.89. The correlation between factor 2 and factor 3 (Loss of Confidence) was 0.83. That between factor 1 and 3 was 0.90. These strong correlations suggest that even if there were in fact three factors, in practice it may be very difficult to discern them.

Several previous confirmatory factor analyses found that the 3-factor model of Graetz gave better fit to survey data from Australia [12], Britain [10] and New Zealand [11]. In this study we examined the issue in an Asian population in Singapore, whose members are mainly ethnic Chinese. All three goodness-of-fit indices employed, namely AIC, CFI and RMSEA, agreed that the 3-factor model of Graetz out-performed the other five models. The CFI value was 0.935. Conventionally, a CFI of 0.90 or larger is taken as evidence of sufficient fit. A more stringent criterion of CFI larger than 0.95 has recently been proposed and debated [27, 28]. The RMSEA also indicated that even the best-fitting model did not fit well, using the cut-off of 0.08 as a criterion. However, our aim is to compare the models rather than to modify the instrument. So for our purpose it is the comparison of the goodness-of-fit of the six models that matters, not the absolute values of the fit indices. We consider the "correctness" and "usefulness" of a model two fairly separate issues. Although the goodness-of-fit of Graetz's model was limited, we proceeded to examine the factor scores in relation to external criteria in order to reach a conclusion about the usefulness of the model.

The one-dimensional model was the worst according to all three goodness-of-fit indices.

The three factors in the model proposed by Graetz were found to be strongly correlated with each other, with correlation coefficients in the neighborhood of 0.8 to 0.9. Such strong correlations suggest that even if there were indeed three different factors, in practice it is quite difficult to differentiate them. The study of French and Tait [12] also showed strong correlation between the factors, which led the authors to recommend that it may be prudent to use the overall score rather than overinterpret the factors within the GHQ-12. We examined the three factor scores and the overall GHQ-12 score in relation to clinical diagnoses. The four scores behaved in fairly similar ways. Although the Loss of Confidence scale was significantly different between patients with and without panic disorder while the other three scales did not show significant differences between the two groups of patients, the difference was only about 0.4 SD. This is smaller than a recommended threshold (0.5 SD) corresponding to minimal clinically important differences for health states questionnaires [29]. We also examined the association between the three GHQ scores and the Beck Anxiety Inventory, a clinical impression score, and the 8 scales of the SF-36. The three factors were associated with the clinical and HRQoL variables to similar degrees.

Two limitations of the study should be noted. Firstly, the sample size was somewhat small for confirmatory factor analysis. Secondly, the participants were clinical cases. This homogeneity might have made it more difficult to detect variations in GHQ-12 scores. We believe that the question about the relative plausibility of various factor models have been sufficiently answered by this and several previous studies [10–12]. Nevertheless, future studies of non-clinical participants based on larger sample sizes will be helpful to further assess the practical usefulness of the factors of the GHQ-12.

Several studies, including the present one, have found that Graetz's 3-factor model of the GHQ-12 is more plausible than other models. However, the factors were strongly correlated and difficult to discern. Our analysis of the three GHQ scores in relation to clinical variables and aspects of health-related quality of life did not appear to be more informative than analysis of a single overall GHQ-12 score. As such, from a pragmatic point of view we consider it acceptable to use this instrument as a one-dimensional measure. Unless one has specific questions that are best answered by a subset of the three factors, there is no need to consider the multi-dimensionality.