INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Epidemiological research on asthma focuses on comparisons of prevalence between populations or between groups. Definition of asthma in these studies seeks an operative definition based on the same well-standardized method in all the populations to be compared (1). The European Community Respiratory Health Survey (ECRHS) in adults and the International Study on Asthma and Allergy in Children (ISAAC) measured international variations in asthma prevalence using symptoms questionnaires to define asthma (2, 3). In the ECRHS, as well as in the ISAAC, notable differences in the frequency of asthma by country were observed. The use of the same questionnaire and definitions suggests that these differences are real. However, a good replicability of the diagnostic tool (i.e., the same protocol and questionnaire) did not exclude international cross-cultural differences in the subjective recognition and report of symptoms.

Some of the questions of the ECRHS questionnaire had already been validated against bronchial hyperresponsiveness in three of the countries (4). Nevertheless, bronchial hyperresponsiveness is not a true "gold standard" for asthma (5). A full clinical examination would be a better gold standard (6), but it is not feasible in studies in large populations or in international surveys, where there could be regional variations in clinical methods and interpretations.

An alternative method to investigate differences in the validity of the definition of asthma between countries, in the absence of a true gold standard, is an analytical approach using confirmatory factor analysis (7). This approach provides a measure of the internal consistency of the way symptoms are interrelated in different countries, on the basis that if symptoms were interrelated in the same way, they presumably measure the same thing. This approach has been used in psychological medicine (8, 9), and recently to assess the Gulf War syndrome (10).

Factor analysis is a technique that attempts to identify groups of variables, from a large set of variables, with a strong correlation among variables within a group, but weak correlations to those variables outside the group. These groups of variables are called factors, latent variables, or dimensions. Factors, then, although not directly observable, are estimated from the measurement of observable variables. Factor structure is how variables are related according to factors. In confirmatory factor analysis, the factor structure is prespecified and the purpose of the modeling is to test how well the data fit this prespecified structure. Judgment of the fit of the structure in confirmatory analysis is straightforward and various measures of fit are used. If structure was different between populations, confirmatory analysis specifies which variables contributed to the differences between countries.

We assume that a similar reporting of symptoms among the different populations in the ECRHS study would be consistent with a similar interrelationship of symptoms. We aimed to identify a factor structure underlying the pattern of interrelationships among the respiratory symptoms, and to test if this factor structure was common to all countries. Thus, the final objective was to test if the same factors were observed in all countries.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Subjects

The protocol for the ECRHS has been described elsewhere (2, 11). Briefly, participating centers selected an area defined by preexisting administrative boundaries, with a population of at least 150,000 individuals. An up-to-date sampling frame was used to select randomly at least 1,500 men and 1,500 women, aged 20 to 44 yr. In Stage I, subjects were sent a questionnaire enquiring about respiratory symptoms. A 20% random sample of subjects was selected to take part in Stage II, in which they were invited to answer a more detailed administered questionnaire, and to take part in blood tests, skin tests, assessment of lung function by spirometry, and airway challenge with methacholine. The present study included subjects randomly selected in Stage II. Of 43 participating centers, we included data from 35 centers in 15 countries (Table 1). The remaining centers had not fully checked and edited their data, or have a small sample size. Response to Stage II varied from 20% in France to 85% in Sweden. The overall response rate, with complete data for all the questions, was 60% (n = 16,635). The Institutional Review Board of the participating centers approved the study protocol and participants gave informed written consent.

The Questionnaire

Questions on symptoms and medical history were adapted from a preexisting questionnaire (i.e., the IUALTD [12]). A version of this questionnaire was first developed in England (13) with subsequent studies in other countries (4, 14). The questions were tested for comprehensibility and translated, with back translation into English. The results from the questionnaire has been reported elsewhere (11), including respiratory symptoms, questions on asthma diagnosis, and asthma treatment.

Analysis

The first step was to build up the factor structure by exploratory factor analysis (15). Exploratory factor analysis is a statistical technique used to identify whether the covariances between a set of observed variables (i.e., symptoms) can be explained by a few latent, unobserved variables (i.e., factors). The structure is specified by parameters such as factor loadings (i.e., the load of each symptom in a factor), the covariances between factors, error terms of factor loadings, and the covariance between error terms. The structure was built up with the United Kingdom data set, since the questionnaire was developed in the United Kingdom. Criteria of goodness of fit have been used to set up the factor structure. A factor was defined by symptoms with a factor loading higher than 0.40, and the rest of the symptoms in a factor were given a factor loading of zero. Factors were named with a term appropriate for describing the content of the group of symptoms.

To assess whether the factor structure was coherent, the independent associations between the factors and the following outcomes: bronchial responsiveness, atopy, and smoking were measured by using the odds ratio from logistic regression models. A multivariate model was built up separately for bronchial hyperresponsiveness, atopy, and tobacco smoking as the outcome variables. The "exposure" variables were the factors. On the basis of the factor structure, a value for each factor for each subject was assigned.

Subsequently, a confirmatory analysis of the factor structure specified for the United Kingdom was assessed for all countries, through paired comparisons of each country's data with the data from the United Kingdom. This confirmatory factor analysis tested whether the prespecified structure for the United Kingdom provides an adequate explanation of how the observed variables were mutually related in the different countries (i.e., how close the prespecified correlations are to those observed). The paired comparisons were made in several steps, increasing the number of prespecified parameters in the factor structure at each step. In the first and least restrictive comparison, we tested if all countries had the same number of factors composed of the same variables. In a second step, factor loadings and factor covariances were forced to be equal to the United Kingdom. In the final and most restrictive step, all parameters of the prespecified factor structure were forced to be equal. The goodness of fit of each model indicates whether the prespecified structure was appropriate for the different countries.

To judge the goodness of fit of models in confirmatory analysis we used "goodness of fit indices" that have been developed in confirmatory factor analysis, including the nonnormed fit index (NNFI) and the comparative fit index (CFI), which reflect fit independently of sample size. Since both tests resulted in similar results only NNFI tests are reported. A value of 0.9 or more is considered an adequate fit (15). The EQS modeling package (16) was used.

Comparison between models in consecutive steps could also be assessed by the likelihood ratio test (15), which has a ² distribution. A limitation of this test in large sample sizes, such as in our study, is the presence of significant results from non-relevant departures between the two nested models. However, we used this test since it identifies whether some individual parameter accounts for the majority of the ² difference between models (i.e., which questions on symptoms contributed to the differences between countries).

Since geographical variations in questions using the term asthma could reflect patterns of health services utilization and medical criteria (17), we repeated exploratory and confirmatory factor analysis after excluding questions using the term asthma. Finally, an alternative method was used to assess the geographical variation in the internal consistency of symptoms, consisting of assessing the homogeneity of the odds ratios from the association between bronchial hyperresponsiveness, atopy, and smoking and the factors, using metaanalytical techniques.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Selection of the Factor Structure: How Symptoms Were Organized

Prevalence of symptoms and response rates by country have been presented elsewhere (8). In exploratory factor analysis, a structure involving four factors, accounting for 58% of the variance, was first selected. These four factors were composed of variables relating to (1) wheezing and shortness of breath, (2) cough, (3) phlegm, and (4) asthma. Symptoms were clearly separated among these four factors (Table 2).

All countries achieved the criterion of goodness of fit for the four factor models, indicating the same factor form (number of factors and composition of these factors) in comparison with the United Kingdom (all NNFI tests > 0.91). However, the correlation between the factors representing wheezing/ shortness of breath and asthma (correlation coefficient, r = 0.75), and cough and phlegm (r = 0.61), were high, which suggests that a structure with fewer factors may also be adequate to describe the data. A structure with three factors (including cough and phlegm as a single factor) achieved a smaller variance (53.5%) and the criteria of goodness of fit also reached the standard cutoff point of 0.9 in the United Kingdom, but not in The Netherlands (NNFI = 0.89) or Spain (NNFI = 0.88). A factor structure with only two factors (one with variables of wheezing/shortness of breath and asthma and the other with cough and phlegm) did not reach the goodness of fit value in the United Kingdom, and the explained variance was much lower (42%).

On the basis of the above exploratory analysis the factor structure selected was that composed of four factors. Each of the factors included the variables with a factor loading > 0.40 in Table 2, with the exception of Q6 (attack of cough at night) that was also included in Factor 2 (cough) due to clinical coherence, although not reaching this value. The structure with the best fit also incorporated covariances between paired error terms of some variables (Q1-Q101, Q1-Q102, Q8-Q801, Q10-Q1001, Q13-Q1301). These covariances probably incorporated the lack of error independence due to the ordering of questions. The four factors, adjusted for one another, were independently associated with smoking. In addition, the factors wheezing/shortness of breath, asthma, and cough were independently associated with bronchial hyperresponsiveness; and the factors wheezing/ shortness of breath and asthma were independently associated with atopy. These associations were found both for the United Kingdom and in all the countries (Table 3).

Comparison of the Factor Structure between Countries

In the confirmatory factor analysis of a structure specifying not only the same form but also the factor loadings and the factor covariances, all countries showed an adequate fit, except Spain (Table 4). The prespecified parameters (i.e., constraints) that account for the difference in Spain refer to the question on asthma treatment (Q1306) and, second, to the question on current asthma (Q1305). Goodness of fit in Spain improved after relaxing the constraint on Q1306. The comparison of likelihoods between the model with the same form and the more constrained model showed significant differences in other countries not detected by the goodness of fit index. Most of these differences were in factor loadings of questions on asthma as shown in Table 4. In addition, there was a difference in Italy in Q101 on wheezing, a difference in The Netherlands in Q2 on tightness in the chest, and differences in Switzerland, Germany, and Spain in Q1001 on chronic phlegm. However, these differences had a minor impact since they did not affect the goodness of fit.

In a more rigorous comparison forcing all parameters to be equal to the United Kingdom (factor loading and residual variances, as well as covariances), all English-speaking countries, as well as Norway, Sweden, France, and Switzerland, had reasonable fit whereas Iceland, Germany, Belgium, The Netherlands, Italy, and Spain showed an inadequate fit (Table 4). Most of the constraints that had to be relaxed to achieve an acceptable fit were related to the variables that composed the factor "asthma." After relaxing constraints of loadings and residual variance on questions related to asthma, differences among countries disappeared, except in Iceland and Italy. Italy differed on the error variance of the question about wheezing with shortness of breath, and Iceland differed on the error variance of questions about shortness of breath at night and chronic phlegm.

After the exclusion of questions about asthma, respiratory symptoms were organized on the basis of the same factors as in the preceding structure (with the obvious exception of the factor on asthma), and the same form was observed for all countries. However, as shown in Table 4, Italy (due to the question on wheezing), The Netherlands (due to questions on shortness of breath) and Germany, Iceland, and Spain (due to questions on phlegm) showed differences in the more restricted factor structure. Finally, the odds ratios of the association between bronchial hyperresponsiveness, atopy, and smoking with the factors were homogeneous between the countries (all homogeneity tests > 0.1).

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The ECRHS is the first study able to measure geographical variations in respiratory symptoms between a large number of countries not limited by the use of different questionnaires or protocols. We found that the four-factor model was the optimal model for explaining symptom correlations in all countries. The same groups of symptoms (i.e., the same latent variables) were identified in all countries, which shows a good internal consistency and suggests that the experience of reporting of symptoms was not affected by cross-cultural differences such as a different meaning of the terms used to define the symptoms in different languages.

In the more constrained analysis (when the structure that fixed how symptoms were interrelated was completely prespecified), the English-speaking countries and some other countries such as the Scandinavian countries, France, and Switzerland showed no differences. This suggests that there was both a similarity in type of symptom clusters identified, and that the relationship between symptoms was also quantitatively similar in the different countries. The exact structure of symptoms was not replicated in Belgium, Germany, The Netherlands, Iceland, Italy, and Spain. Differences were mostly centered on questions using the term asthma. Differences in questions using the term asthma may reflect variations in its reporting, use and structure of related health services, how the label of asthma is used by doctors (17), and how treatment and symptoms are correlated (18). The fact that English-speaking countries showed an identical structure might indicate that report of symptoms, even those including the term asthma, is identical in all these countries, and therefore that definitions of asthma relying on questions that included the term asthma are exactly replicated between these countries. However, the lack of an identical replication of the factor structure in all countries, mainly because of differences in questions on asthma, suggests using with caution definitions of asthma based on a single question, particularly when doing international comparisons. Any definition including questions with the term asthma should combine various questions, precisely the strategy used in ECRHS analysis (11).

Answers to questions on wheezing, shortness of breath, cough, and phlegm are not influenced by the diagnostic and treatment practices. Only a few symptoms in a few countries (wheezing in Italy, nocturnal attack of shortness of breath in Iceland) showed a significant difference in the way that they related to the other symptoms, when compared with the same relations in the United Kingdom. The analysis carried out after excluding questions on asthma provides similar results on how the respiratory symptoms, other than asthma, were organized. Therefore, in general, respiratory symptoms, apart from asthma, were reported in a similar way between countries, and hence can be used for international comparisons.

A secondary finding is that the best factor structure specified in the present analysis included four factors. A three-factor structure did not fit with data in The Netherlands and Spain. In addition, the four factors were associated independently with smoking. Factors for wheezing/shortness of breath and asthma had an opposite association with smoking, and this supports the separation of these two factors in epidemiological studies. In adults, wheezing is separated from asthma in other epidemiological studies (19). Factors for cough and phlegm had an opposite association with bronchial responsiveness, which also supports the separation of these two factors. Some occupational studies have differentiated the effects on chronic cough from those on chronic phlegm (20). Although the factors wheezing/shortness of breath and asthma and phlegm and cough were highly correlated, the factor structure chosen suggests that in the general population, respiratory symptoms may be grouped in four different dimensions, beyond the two classic respiratory syndromes most commonly recognized in the clinical setting (i.e., asthma, chronic bronchitis). The homogeneity of odds ratios for the four factors among all countries reinforces a similar geographical grouping of symptoms and suggests that combinations of symptoms reduced the measurement variance in a similar way in all countries. Finally, the question on coughing at night was not well represented in any factor, which suggests that this symptom seems to be something different than usual cough, particularly in women, among whom it is common (11).

However, factor analysis is a statistical technique that has some limitations. The fact that symptoms cluster together does not mean that they are equally important, and some of the symptoms may perform much better than others in a single population. Therefore, we do not recommend routinely grouping symptoms into four factors on the basis on the present results, without considering whether this is appropriate. In addition, it might be argued that there was no a priori reason to define the structure in the country it was created. However, if we followed a different strategy, such as building up the structure with data from all countries, we obtain the same results. Factor analysis using dichotomous variables is limited in the calculation of the correlation matrix of the observed variables. An alternative is to calculate tetrachoric correlations instead of Pearson correlations. However, we were unable to do so, since the EQS program required larger sample sizes for calculation of the tetrachoric correlations. The use of Pearson correlations would underestimate the "true" correlations for dichotomous variables, but this underestimation is invariant between countries and did not affect the comparisons between countries.

We have used factor analysis to assess cross-cultural validity of the reporting of respiratory symptoms, and cross-cultural variations in the derived definitions of respiratory diseases such as asthma, in an international study that uses the same tools and protocol. We conclude that the fact that the symptom groups behave in the same way in all countries makes it possible to perform international comparisons of respiratory diseases defined on groups of symptoms without an underlying error due to cross-cultural and language differences. This has important implications for the interpretation of findings from the ECRHS and ISSAC. In addition, we conclude that the fact that an identical structure was not replicated in all countries was due to questions including the term asthma, and this suggests using with caution definitions of asthma based only on questions including the term asthma. Any definition including questions with the term asthma should combine various questions. The other respiratory symptoms are similarly reported, in spite of idiomatic and cross-cultural differences, suggesting they can be used in international studies. Overall, the present analysis implies that international differences in the prevalence of respiratory symptoms, or asthma defined through a combination of questions, are not affected by errors due to cross-cultural variations, and are indeed due to actual variations.