This paper presents the findings of a Medical Research Council (UK)-funded study to estimate the population prevalence of CF and CFS in an ethnically diverse sample. In this study, we tested whether there is a consistently higher prevalence of CF and CFS in specific cultural and ethnic groups and whether variations in prevalence can be explained by social adversity (social strain and perceived discrimination), social support, physical inactivity, anxiety and depression.

Our hypothesis was as follows: (1) CF and CFS show differing patterns of prevalence across ethnic groups, and this difference is independent of variations by age, sex and socioeconomic status; and (2) prevalence variations may be explained by variations in psychosocial risk factors and physical inactivity, such that (a) social adversity such as social strain, low social support and perceived discrimination account for a higher prevalence of CFS; (b) physical inactivity accounts for a higher prevalence of CFS; and (c) anxiety and depression account for a higher prevalence of CFS.

The Health Survey for England (HSE) comprises a series of annual surveys beginning in 1991. It was designed to provide regular information on the nation's health. The detailed methods used in the Ethnic Minority Psychiatric Illness Rates in the Community (EMPIRIC) study to sample from HSE have been published 27 28 29 . Therefore, in this paper, we briefly set out the sample methods. The EMPIRIC national population survey took place in 2000 and sampled White British participants from the 1998 Health Survey for England (HSE98) 26 . The 1999 Health Survey for England (HSE99) (the ninth in the series) was the first to increase the representation of ethnic minority groups, boosting the sample of adults from Black Caribbean, Indian, Pakistani, Bangladeshi and Irish communities. So, for the EMPIRIC survey, people from these ethnic groups were sampled starting from HSE99 27 (see Figure 1). The EMPIRIC survey of 4,281 people ages 16 to 74 years included White British (n = 837), White Irish (n = 733), Black Caribbean (n = 694), Bangladeshi (n = 650), Indian (n = 643) and Pakistani (n = 724) people 27 28 . Response rates were highest among the White (71%) and Irish (72%) groups and lowest among the Indian group (62%). Weighting in EMPIRIC analyses ensured that the sample was representative of the population samples in the HSE annual surveys by removing different probabilities of selection. The EMPIRIC study included structured and validated questions about demographic variables and measures of social adversity and measures of anxiety and depression 29 .

A question asking about fatigue in the Revised Clinical Interview Schedule (CIS-R) has previously been used to operationally define CF 7 and to validate more extensive measures of CF and CFS 3 30 (see Appendix). Therefore, to define CF and CFS, we used the fatigue question and other stem questions from the CIS-R (these questions were distinct from measures of anxiety and depression symptoms). The questions asked about the following symptoms: (1) getting tired and lacking energy and whether there were any reasons for this (such as physical illness), (2) having any problems with concentrating or noticing any problems with forgetting things, (3) having problems with falling asleep or with getting back to sleep, (4) any sort of ache or pain or being troubled by any sort of discomfort such as headache or indigestion.

Participants also rated the duration of symptoms. Using these items, we defined CF as the presence of unexplained fatigue for a period of 6 months. We defined a CFS-like syndrome as closely as possible to the international criteria for CFS 21 . Although the literature does at times make a distinction between CFS-like syndrome and CFS, this is not consistently followed, for example, the published studies on which the methods in this paper were adopted 3 7 31 . Therefore, in this paper, we make explicit the study methods and that the operational definition of CFS is compatible with what has been called by some a CFS-like syndrome. We counted as CFS those individuals with all four of the above symptoms, each lasting for at least 6 months. Our definition of at least four symptoms meets the threshold criteria, as CFS is usually diagnosed on the basis of four of eight symptoms, including fatigue 21 . Sensitivity analyses showed that when a fatigue syndrome was defined using fatigue and only two additional symptoms rather than three, the findings from the analyses were identical to the findings for CF. This supported the notion that our definition of CFS did identify a group that was distinct from the CF group.

The EMPIRIC study (2000) 28 29 was undertaken by following up and oversampling people from HSE98 and HSE99 26 27 . We obtained the individual identifier codes that linked the two data sets, which permitted us to search for physical illnesses reported in HSE98 and HSE99. We were able to identify diseases that might exclude a diagnosis of CFS (31 of 108 (28.7%) participants met the criteria for CFS): cancer, diabetes, epilepsy, arthritis or fibrositis and infectious or parasitic disease. However, we did not know the duration of these physical diseases, whether they were in remission or whether they were causing the fatigue at the time of the HSE survey. Although one analytic approach might be to exclude all those individuals with these physical illnesses, given the lack of information about duration, degree of fatigue and whether it was being treated, this approach might introduce a significant selection bias. We therefore included those participants with the specified conditions in our sample and adjusted for all these illnesses in all of our logistic regression analyses where CFS was an outcome. This adjustment was part of the operational definition; therefore, we do not show the unadjusted analyses. This approach provided estimates of risk independent of physical illnesses that might cause fatigue. This method was supported by sensitivity analyses, which showed that the odds ratios following our analyses did not differ whether we included or excluded people with these physical illnesses. Therefore, using these operational definitions of CF and CFS, linked closely to carefully defined analytic plans, we tested the hypotheses by analysis of data collected in the EMPIRIC study 29 .

The EMPIRIC study included the CIS-R, which asks about 14 symptoms in the week preceding a structured interview 32 . Anxiety and depression subscales of the CIS-R were used in the analyses (both are scored on a scale from 0 to 4). These were distinct from the questions about fatigue. The CIS-R shows good face content and construct validity in culturally diverse populations 28 29 32 .

Social class was classified by the head of household's occupation recoded into a 'manual' or 'nonmanual' group. Other measures of socioeconomic position were educational qualifications (recoded into no qualifications, degree or secondary school qualifications, or foreign qualifications) and employment status (employed or unemployed, retired or other economically inactive status). Social support was measured by the Close Persons Questionnaire 31 . Scores were classified using usual validated thresholds into 'low', 'medium' or 'high' levels of confiding or emotional support (seven items), practical support (four items) and negative aspects of close relationships (four items) 31 .

Chronic social strain was measured by questions from the Whitehall II study 15 33 . The five-point Likert scale responses for each domain were dichotomized: (1) problems with relatives and problems with finances (always, often, sometimes and seldom or never) and (2) problems with essentials, problems with housing and problems with neighbourhood (few or no problems and many or great problems) 33 . These were summed to give a total score (on a scale from 0 to 4), where a higher score indicated more frequent (for relatives and finances) or greater social strain. This total score (from 0 to 4) was recoded into tertiles.

Perceived discrimination questions asked about insults in the preceding 12 months; unfair treatment at work regarding promotions; and refusals of employment; and assaults in relation to race, religion or ethnic group. Responses of 'Yes' or 'No' (score of 1 or 0, respectively) were summed to produce a total discrimination score from 0 to 4. These questions have been used in previous studies of perceived discrimination and health 34 .

Physical activity levels were measured in the HSE98 and HSE99, so they preceded the measures of CF and CFS in the EMPIRIC study (measured in 2000). The intensity of physical activity was assessed by an aggregate measure of the physical demands of people's jobs and four other types of activity: housework, gardening and 'do-it-yourself' jobs around the house (including painting, decorating, repairs), walking, sports participation and exercise. The 'intensity level' of activity was already derived in the original HSE data from the estimated energy costs of each activity: vigorous, moderate, light and inactive 26 27 . These were recoded into two groups: vigorous or moderate activity versus no activity or light activity. The number of days in the preceding 4 weeks during which subjects had participated in at least a moderate level of activity for at least 30 minutes daily was the recommended level of healthy activity at the time of the original HSE surveys 35 36 . The total number of active days that each informant reported was divided by four to produce an (average) weekly figure of <1 day, 1 to 4 days or 5 or more days weekly, so that the policy and clinical implications for recommended activity might more easily be discerned as activity in an average week. For each activity variable, the highest activity levels (intensity and frequency) were used as the reference group.

The EMPIRIC study piloted and then used a structured questionnaire that was available in five languages (Hindi, Gujarati, Punjabi, Urdu and Bengali) 27 . A letter introduced the study before the survey. Gender and language matching were used whenever requested.

All analyses used the survey commands in Stata version 10 software (StataCorp, College Station, TX, USA) and were weighted for probability sampling and nonresponse (in HSE99 and EMPIRIC) using established methods with these data to represent population estimates 27 29 . Unweighted descriptive data are presented, but all regression models were weighted. Differences between weighted means were assessed using weighted regression, with an overall P value reported. Logistic regression models were built using likelihood ratio (LR) tests to justify inclusion and then retention of variables in the models. Gender did not appear to contribute significantly to the models, but, given its importance, it was retained in all regression models.

The unadjusted associations of demographic factors, including ethnic groups, with CF and CFS are presented in Table 1. Further analyses then investigated univariate associations of social strain, perceived discrimination, social support, physical inactivity and anxiety and depression variables with CFS (Table 2). Table 3 sets out the relationships between these variables and ethnic groups.

Table 4 sets out the nonparametric correlations between demographic and risk factors entered into the final regression model. Table 5 reports manual stepped logistic regression modelling. Explanatory variables were recoded to reduce the number of levels per variable. LR tests were used to justify addition and retention of each of the explanatory variables when added to the basic model. Using this manual stepped approach, we were able to investigate which specific risk factors, when added to the model, led to a reduction in risk in specific ethnic groups or all ethnic groups. The basic model included CFS as an outcome, as well as age (in years), socioeconomic position (measured by education, employment type and whether employed) and ethnicity. We entered social strain, perceived discrimination and social support first; only negative aspects of social support contributed to the model and were retained. Then, in a separate model, we entered physical activity intensity and frequency. The frequency of physical activity did not contribute significantly to the model and was excluded. Starting with the basic model again, we then added anxiety and depression scores (see Table 5). A full model showing independent effects of all of these variables is shown in Table 5.

The weighted prevalence of CFS for the whole sample was 2.3% (108 of 4,273). The weighted prevalence of CF across the whole sample was 19.7% (881 of 4,276). CFS prevalence varied by ethnic group, age and educational level and was most prevalent among Pakistani people (Table 1). In contrast, the prevalence of CF varied only by age, sex and marital and employment status, but not by ethnic group (Table 1). The demographic factors associated with CF and CFS are similar (Table 1), but apart from the ethnic group differences, women and retired people have a higher risk of CF but not of CFS. Manual labour workers have a higher risk of CFS but not of CF. People over the age of 35 have a higher risk of CF and CFS in the univariate analyses (reported in Table 1). In the final model (Table 5), a 1-year increase in age is associated with a 2% (95% confidence interval (95% CI), 1% to 4%) higher risk of CFS. As CF did not vary by ethnic group, we did not undertake any further analyses of CF.

When considering univariate associations, chronic social strain, perceived discrimination, the negative aspects of social support, physical inactivity and anxiety and depression were each associated with a higher risk of CFS (Table 2).

Table 3 sets out ethnic variations in measures of social adversity, physical inactivity and anxiety and depression. The Bangladeshi group was more likely to score in the high range on social support (positive and negative) and on social strain. Perceived discrimination scores were highest among the Black Caribbean, Pakistani and Indian groups. Bangladeshi and Pakistani groups were least active as measured on the frequency and intensity of physical activity. Bangladeshi and Pakistani groups had the highest mean depression scores.

Before reporting regression models, Table 4 sets out the nonparametric correlations between demographic and risk factors entered into the final regression model. This shows highly significant correlations, which are addressed by including these variables in the regression models to show their independent effects.

In regression models, adding chronic strain, perceived discrimination and negative aspects of social support to the basic model led to some reduction in the risk among specific ethnic groups, thus partially explaining the higher prevalence of CFS in some ethnic groups (see Table 5). However, there was little reduction in risk for the Pakistani group (from an odds ratio (OR) of 4.1 (95% CI, 1.6 to 10.4) to an OR of 3.5 (95% CI, 1.4 to 9.1)). Adding physical inactivity reduced the risk of CFS to a nonsignificant level in all but the Pakistani group. Anxiety and depression alone did not account for ethnic variations in CFS prevalence.

In the full model, age, social strain, negative aspects of social support, physical inactivity and anxiety and depression were independent risk factors for CFS and together explained ethnic variations in CFS prevalence. The point estimates for the Pakistani group especially, and for the other ethnic groups, remained elevated (>1) but were not statistically significant. The full model explained 24% of the variance.