Resumen

Antecedentes: El General Health Questionnaire (GHQ) de 28 ítems está construido a partir del análisis de componentes principales del GHQ-60 (escoge 28 ítems que se agrupan en cuatro subescalas). Lógicamente, la utilización en población española de la traducción del GHQ-28 desarrollado en población inglesa, debe presentar peores valores predictivos.

Método: Analizamos la estructura factorial de la versión de 60 ítems del GHQ, a partir de los resultados obtenidos en una muestra de población general de un núcleo urbano próximo a Madrid. Posteriormente realizamos todo el proceso de construcción del GHQ-28 para población general española utilizando la misma metodología empleada en la versión original inglesa.

Resultados: En este trabajo se proponen dos versiones más cortas del GHQ: una con 6 escalas y 30 ítems y ,una segunda, con cuatro escalas y 28 ítems. Ambas son comparadas con las versiones inglesas y, la última, con la versión mejicana del GHQ de 28 ítems.

Conclusiones: con este trabajo presentamos una adaptación real del GHQ-28 en población general española.

Palabras clave: Cuestionario de Salud General, GHQ, GHQ-28, versión española del GHQ-28, estructura factorial del GHQ-60, versión escalar, estudio comunitario, población general española, análisis factorial, estadística, tests psicológicos, cuestionarios.

Abstract

Background: The 28-item General Health Questionnaire (GHQ) is constructed on the basis of a principal components analysis of the GHQ-60 (28 items selected and grouped under four subscales). When used on a Spanish population, a translation of the GHQ-28 developed for an English population logically yields worse predictive values.

Methods: Data from a community study conducted in a town near Madrid was used to explore the factor structure of the 60-item version of the GHQ. Using the methodology employed in the original English version, a 28-item version of the GHQ was constructed based on the general Spanish population.

Results: Two shorter versions were proposed: one with 6 scales and 30 items, and the other with four scales and 28 items. These and the 28-item Mexican versions were then compared to the English version.

Conclusions: The resulting GHQ-28 was a successful adaptation for use on the Spanish population.

Key words: General Health Questionnaire, GHQ-28, spanish version of GHQ-28, factor structure of GHQ-60, scaled version, community study, general spanish population, factor analysis, statistical, psychological tests, questionnaires.

The late 70s saw the emergence of an approach which relied on standardised methods of diagnosis and which Dohrenwend and Dohrenwend (1982) termed "third-generation studies". At this same point in time, coinciding with the introduction of two-phase sampling methodology for psychiatric epidemiological research, the first structured interview, the Clinical Interview Schedule (CIS) (Goldberg et al. 1970), and the first non-symptomatic "psychiatric case" detection scale purpose-designed as a screening instrument, the General Health Questionnaire (GHQ) (Goldberg, 1972), were both published.

The conjunction of these two types of instruments, one designed to detect symptoms or signs in the population capable of rendering a case potentially suspect, and the other derived from clinical psychiatric examination and designed to arrive at an accurate and reliable diagnosis, has enabled two-phase sampling to be used to study large-sized populations in a reliable manner. In the first phase, the study population is classified (usually by means of a survey) in terms of the probability of the presence or absence of psychiatric morbidity. In the second, a variable proportion of probable cases and non-cases must be simultaneously examined by a psychiatrist in order to assess the definitive diagnosis.

At present, the GHQ is the most widely used instrument for detecting non-psychotic psychiatric "cases", whether in the general population or among patients attending a given general practitioner's practice. This questionnaire exists both as a 60-item version and in the form of shorter versions (comprising 30, 28 and 12 items). In terms of validity, reliability and prediction coefficient, the 60-item version has outperformed its shorter counterparts (Goldberg, 1972), yet thanks to their brevity, the 12- and, above all, the 30-item versions have seen wide use in general population studies.
The 28-item version (GHQ-28) developed by Goldberg and Hillier (1979) is constructed on a basis unlike that of the other versions. Whereas the 30- and 12-item versions contain a selection of items that retain a similar discriminative power, the GHQ-28 is based on an analysis of the principal components of the GHQ-60, with 28 items then being chosen and grouped under four subscales.
Over the last ten years, use of the GHQ-28 has come to assume growing importance in epidemiological studies, a trend reflected in the increasing interest in adapting this instrument to different populations and languages. It is succinct, simple to use and yields comparable results in general population and primary-care settings. Furthermore, as mentioned above, it derives from an approach totally different to that of the remaining General Health Questionnaire versions, and apart from providing an overall assessment, also contains four scales that furnish additional information with a breakdown by symptom area.

Seva et al. (1992) reported the predictive validation data for the GHQ-28, using the Spanish translation of the English GHQ-28 in a general Spanish urban population such as ours, with a sample of 117 patients and the same instrument as in the original validation, the CIS. In Spain there was a clear fall-off in specificity and a rise in the percentage of misclassified subjects vis-à-vis the original English version, namely, 21.3% versus 14.2% for the same cut-off point of 5/6. This same loss of specificity and increase in the number of misclassified subjects (17% for a cut-off point of 5/6) had previously been observed by us in a sample of 100 general medicine patients, in which the CIS was again used as the external validation criterion (Lobo et al.1986). Although raising the cut-off point (6/7) places the percentages on a par with the English version, it nevertheless results in a marked decline in sensitivity. Indeed, what is reflected here is the different predictive performance of a questionnaire constructed on the basis of an English population versus the selfsame questionnaire transferred to a Spanish population, in both community studies and primary-care settings.

The important study by Medina-Mora et al.(1983) is noteworthy for having developed a scaled 28-item, as well as a 30-item, version based on a factor analytic study of the results obtained with the GHQ-60 in a Mexican population. The scaled 28-item version, constructed on the basis of this Spanish-speaking population with the same methodology as the original, shows only 67% of the items selected by the British authors.

Lastly, Vázquez-Barquero (1988) analysed the factor structure of the GHQ-60 in a Spanish rural population sample. This study reported differences between the structure identified in his factor analysis and that described by Goldberg and Hillier and by Medina-Mora et al. An overall comparison of the data obtained in these three studies is not possible, since Vázquez-Barquero et al. did not draw up a scaled 28-item version and used a different technique to retain items with greatest loading on each factor (Catell scree test). Nevertheless, in their principal components analysis with varimax rotation they identified six factors with a different item composition, in which no equivalent was found for the "Somatic symptoms" factor described by Goldberg and Hillier.

The present study therefore sought to replicate the entire process of construction of the GHQ-28 in a general Spanish population, using the same methodology as the original. Available data indicated that the same items as those comprising the English version would not be obtained. For study purposes, we used the Spanish version of the GHQ-60, drawn up on the basis of the 140-item English version (Muñoz et al., 1978,1979).

METHOD

Data-collection

This study relied on data generated during the first phase of field work undertaken for a project entitled, "Estudio de validación predictiva de diferentes versiones del GHQ en población general urbana" (Predictive validation study of different versions of the GHQ in a general urban population), and sponsored by the Spanish Health Research Fund (Fondo de Investigación Sanitaria -FIS). We used all the GHQ-60 questionnaires completed in the first phase of the above study as the basis for comparison and development of our 28-item version of the GHQ.

The designated study population was that of Tres Cantos, a new town created just under 30 years ago. The electoral roll of 01/01/93 shows the total population as 19,563. Broken down by sex, this gives 9,680 men and 9,883 women, with a male: female ratio of 0.98. Table 1 shows the distribution by age group and sex for the 15-65 age range used to delimit our study population.

Table 1. Distribution of the population of Tres Cantos (age-range 15-65 years), by age and sex

In order to obtain the necessary sample, the municipal electoral office was asked to draw up a random list of 5% of the population stratified by age and sex, along with an adequate, also randomly selected list of replacements for each stratum, to be used in the event of possible refusal. The sample so obtained comprised 654 subjects, age-range 15-65 years, made up of 334 women (51.5%) and 320 men (48.9%). Table 2 shows the distribution of the sample by age and sex cohort. All subjects duly completed the GHQ-60 questionnaire; this was self-administered during a home interview conducted by purpose-trained personnel, and was also used to gather other socio-demographic and general health data.

Table 2. Demographic characteristics of the sample (N=654)

Table 2 shows the demographic characteristics of the sample. Mean age was 34.8 years, with a standard deviation of 11.96. The age and sex balance of the sample meant that the mean age proved to be similar, if not practically equal, across the sexes, i.e., 34.71 years with a standard deviation of 11.83 for women, and 34.89 years with a standard deviation of 11.96 for men. When broken down by marital status, the most numerous group was made up of "Married" subjects, and the least represented groups were "Separated" (0.3%) and "Divorced" (0.2%), with "Widowed" (1.2%) being similarly very reduced. In all, 35.6% of the sample had a university education, with solely 1.7% of subjects being "Illiterate" or "Sub-primary". While certain differences between the sexes were in evidence, this divergence proved widest in the case of university graduates, with men accounting for 42% and women, 29.3%. The largest occupational group was "Active, in full-time employment" (71.6% of men and 38.4% of women), followed by that of "Housewife", composed totally of women. The low percentage of unemployed (4.1%), with figures far below those for the Madrid population as a whole, can be explained by the fact that this is a newly settled urban population, which has largely opted to move house in order to be near the workplace. Tres Cantos lies between the so-called "Madrid Technology Park", consisting of a group of high-technology companies lining its southern access point, and a sizeable industrial estate located at its northern access point. Furthermore, the town has tended to be a population-catchment area for teaching and other categories of staff employed at the nearby Autonomous University of Madrid.

Data-analysis

Factor structure of the GHQ-60

Given that our aim was to construct a real Spanish-language adaptation of the GHQ-28 in the general Spanish population, in the process of constructing the GHQ-28 we had to follow the selfsame step-by-step methodology as that used by the authors when they developed and drew up the original version, rather than rely on a mere translation of the same items obtained for the English population.

In order to determine whether it was possible to develop a shorter, balanced version of the GHQ, a principal components analysis was performed, using GHQ-60 responses with Likert scoring (0-1-2-3). The unrotated solution was first examined and varimax rotation was then applied. The number of factors to be retained and rotated was determined by the criteria used by Goldberg and Hillier (1979).

Comparison between the results obtained by us on the one hand, and the factor structures of the original English and the Mexican versions of the GHQ-28 and data obtained for the general rural Spanish population on the other.

The GHQ-28 items were disembedded from the GHQ-60. The results of the four analyses were then inspected and compared.

RESULTS

Factor- structure of the GHQ-60

In the principal component analysis, 14 components with an eigenvalue greater than or equal to 1.00 were obtained, accounting for 59.9% of total variance, but only the first 6 could be readily conceptualised. These factors accounted for 43.6% of the total variance (see Table 3).

Table 3. Unrotated principal components analysis of GHQ-60

The first principal component was a general unipolar factor. Subsequent components were generally bipolar, with fewer items loading on each successive factor. As with the original version, when these factors were subjected to a varimax rotation, they proved even easier to conceptualise, due to the more equal distribution of variance. More items were loaded on each. By taking the 5 items with the highest loads on each factor (all with loads of over 0.50), it proved possible to construct 6 five-question scales in which no item loaded significantly on any but the assigned scale (see Table 4).

Table 4. Factor structure of the GHQ-30 version. Varimax rotation of the 6-factor solution, accounting for 43.6% of variance

A new analysis retaining 4 factors was now performed, thereby allowing for each factor to be described with a larger number of items, i.e., the 7 with the highest loads per factor. We then proceeded to construct four scales with seven items each (28 items in all) relating to the four factors accounting for 37.8% of total variance.

Individual analysis of the items showed that all but one loaded more than 0.48 on the chosen factor. This sole exception was item 50, which was also the only item to register a greater load on any scale other than that chosen. It loaded more on the first factor than on factor 4, where it was chosen as the seventh highest load overall (see Table 5).

Table 5. Items of greatest factor load after varimax rotation of first four factors (accounting for 37.8% of variance).

In similar circumstances, the authors of the original version (Goldberg and Hillier, 1979) opted for greater homogeneity and excluded item 49, which in their study loaded equally on two scales (B and D). They thus obtained a version in which no item loaded significantly on any other scale. Other authors have adopted a similar strategy. In their Mexican population study, Medina-Mora et al. (1983) found that in the previous stage (varimax rotation of the six-factor solution) their item 36 loaded among the first five items on two scales (C and D). Accordingly, they replaced this in one of the scales ( D) with item 28, which had approximately the same load.

We found no factor-4 item that was similar to 50, did not load significantly on any other scale and could thus be used to replace it, as Goldberg and Hillier had done in a similar situation. It thus became necessary to choose a solution that had 4 scales of equal size (7 items each), or alternatively, one that -for the sake of greater homogeneity- dispensed with 4 scales of equal length and instead had one with six and another with eight items.

Given the importance of this question, we decided to pursue both possible avenues of research and leave analysis of the results for in-depth discussion at the end. The chosen 28 items were thus subjected to a second analysis, using varimax rotation with these alone to observe the final solution. This step was performed twice:

1. The first time, item 50 was included among the chosen 28 comprising factor 4 (as the item with the seventh highest load on this factor);
2. The second time, with the aim of achieving greater homogeneity, item 50 (which showed a heavier load on a factor other than that from which it was selected) was excluded from factor 4, and item 17 selected instead to form part of the 28 on which this analysis was based (as the item with the eighth highest load on the first factor).

1. Final solution with item 50: GHQ-28 with four scales of 7 items each

On repeating the analysis with the twenty-eight selected items, the four factors identified accounted for 50.7% of the corresponding variance (see Table 6). All items showed a heavy load on their pre-assigned factors except for item 50 ("lost confidence in oneself"), which continued to register a greater load on a factor other than the original. This item, which had been chosen in the conceptualised factor as Severe depression for the purposes of analysis, replicated its previous behaviour pattern and continued loading more on the same alternative factor (conceptualised as Anxiety and Insomnia). In essence, the situation was similar to that observed previously.

Table 6. Factor structure of the 28-item GHQ version (varimax rotation of the 4-factor solution, accounting for 50.7% of variance ). Solution with four scales of 7 items, including item 50

As observed in the first unrotated analysis, a general factor accounted for 22.9% of the total variance, thereby rendering it inevitable that the various scales would not be pure measures of the four factors. Table 7 shows the mean factor loads for each scale on each factor.

Table 7. Mean factor loads of the 4 scales on the four factors. Solution with four scales of 7 items, including item 50 (50.7% of variance on these factors: varimax rotation of the 28-item questionnaire)