Are Reports of Psychological Stress Higher in Occupational Studies? A Systematic Review across Occupational and Population Based Studies

Objectives The general health questionnaire (GHQ) is commonly used to assess symptoms of common mental disorder (CMD). Prevalence estimates for CMD caseness from UK population studies are thought to be in the range of 14–17%, and the UK occupational studies of which we are aware indicate a higher prevalence. This review will synthesise the existing research using the GHQ from both population and occupational studies and will compare the weighted prevalence estimates between them. Methods We conducted a systematic review and meta-analysis to examine the prevalence of CMD, as assessed by the GHQ, in all UK occupational and population studies conducted from 1990 onwards. Results The search revealed 65 occupational papers which met the search criteria and 15 relevant papers for UK population studies. The weighted prevalence estimate for CMD across all occupational studies which used the same version and cut-off for the GHQ was 29.6% (95% confidence intervals (CIs) 27.3–31.9%) and for comparable population studies was significantly lower at 19.1% (95% CIs 17.3–20.8%). This difference was reduced after restricting the studies by response rate and sampling method (23.9% (95% CIs 20.5%–27.4%) vs. 19.2% (95 CIs 17.1%–21.3%)). Conclusions Counter intuitively, the prevalence of CMD is higher in occupational studies, compared to population studies (which include individuals not in employment), although this difference narrowed after accounting for measures of study quality, including response rate and sampling method. This finding is inconsistent with the healthy worker effect, which would presume lower levels of psychological symptoms in individuals in employment. One explanation is that the GHQ is sensitive to contextual factors, and it seems possible that symptoms of CMD are over reported when participants know that they have been recruited to a study on the basis that they belong to a specific occupational group, as in nearly all “stress” surveys.


Introduction
The general health questionnaire (GHQ) is one of the most commonly used measures to assess symptoms of common mental disorder (CMD) in the UK [1]. It has been administered in population studies (e.g. British Household Panel Survey) and more commonly in smaller studies of particular groups, such as occupational studies (e.g. a study of UK doctors [2]) to estimate the level of CMD in a specific population. It is generally believed that there are particular occupational groups who are exposed to a higher level of stress than other occupations, e.g. police officers and military personnel; however, there are few studies comparing rates of CMD across these occupations or to the general population.
UK prevalence estimates for common mental disorder from population studies are in the range of 14-17% [3,4], with the prevalence of CMD in occupational studies, such as military personnel and London civil servants higher at 20% and 27% respectively [5,6]. This difference is inconsistent with the 'healthy worker effect' and the assumption that healthier individuals are more likely to be selected into work, which is well established [7]. Furthermore, it is conflicting with the fact that there are many aspects of working which have a positive impact on mental health [8]. Occupational samples are also less likely to include the disabled, those with long term physical or mental health disorders, and by definition exclude the unemployed which in itself is a strong risk factor for poor mental health [9].
We have therefore undertaken a systematic review to identify UK studies that use the GHQ in either a population or an occupational setting and to examine whether a comparison between these types of surveys reveals a difference in the level of CMD caseness reported. The specific aim is to examine the prevalence of CMD caseness in all UK occupational and population studies, conducted from 1990 onwards, which have administered the GHQ, and to compare the weighted prevalence estimates between them.

Search strategy
The literature search was conducted in November 2011 using Medline, EMBASE and PsycInfo electronic databases to identify UK studies which had used the GHQ, covering the period from January 1990 to the date of the search. The search was restricted to UK studies due to the high quantity of international studies which have used the GHQ and also due to between country differences in occupational issues. The search terms were used as free text terms and were combined with Boolean operators. The initial search using the terms (general health questionnaire OR GHQ) was combined with (United Kingdom OR UK OR Britain OR England OR Wales OR Scotland OR Ireland OR Northern Ireland OR British Isles) using the AND operator.

Inclusion criteria
Inclusion criteria for occupational and population studies 1. The study should have administered the General Health Questionnaire 2. The prevalence of CMD in the sample should be reported or available from the authors 3. Conducted in the UK 4. Sample size should be at least 100 participants 5. Sample should not include children, adolescents (,18years) or student groups 6. Should not use data from participants recruited in primary/ secondary care, from GP records or who were recruited from a health care register Inclusion criteria specific to occupational studies 1. The sample should be a particular occupational group(s) 2. Studies on individuals in a trainee position in a particular occupation (e.g. clinical psychology trainees) were included, but other types of student groups (e.g. medical undergraduates) were excluded Inclusion criteria specific to population studies 1. Should include at least 1000 participants (higher than for occupational studies because the sampling frame for population studies would be expected to be larger) 2. Should not be a study of older adults (.60 y)

Data extraction and analysis
Data were independently extracted by two researchers (LG & IBZ). Data from 20 occupational and 10 population studies were extracted by both researchers and agreement between the researchers was high. The data extracted from the articles were: author, title and date of publication; information about the population and the location of the study; study design and type of sampling; number of participants and response rate; sample characteristics (e.g. gender split and age); which version of the GHQ questionnaire was administered and the cut-off used; and data on the prevalence of CMD. The numerator (the number of CMD cases) and the denominator (the sample size or number of participants who had data on CMD) were entered into the review database so that the prevalence (%) of CMD, standard errors and 95% confidence intervals (CI) could be calculated. Stata v11.0 was used for all data analyses [10] and meta analyses were conducted to produce weighted estimates and to examine the between study heterogeneity.
1. The metan command was used to produce the forest plots, displaying the prevalence of CMD and 95% CIs for each sample, grouped by occupation for the occupational studies. Forest plots were produced across all of the occupational studies, displaying the weighted estimates by occupational group and overall across all of the studies. A further forest plot displayed the weighted estimates across the occupational studies and for the population studies, so that the estimates could be compared between these categories. 2. To further investigate the issue of heterogeneity, and the difference between the two types of study, analyses were conducted which restricted the studies included to those which used the GHQ-12 version of the questionnaire and with a cutoff of 3/4 (41 occupational and 19 population studies) so that the prevalence estimates between studies should be more comparable. Random effects meta-analysis models were conducted within these restricted studies. 3. Meta regressions were conducted using the metareg command to examine whether response rate (categorised as less than 50% vs. equal to or more than 50%) and sampling method (split into studies which used random sampling or aimed to recruit all participants from the sampling frame vs. studies which used non-random sampling) were associated with the prevalence of CMD. These variables were entered individually into meta regressions for the occupational and population studies separately. 4. A further meta-analysis was conducted which restricted the studies to those with used the same version and cut-off of the GHQ, which was also restricted to studies which had a response rate of at least 50% and which used random sampling and a forest plot was produced displaying the weighted estimates for the occupational and the population studies.

Ethical approval
Ethical approval was not required for this systematic review.

Study selection
The initial search terms general health questionnaire OR GHQ (restricted to abstracts in the English language and to the period from January 1990 -current) identified 10106 abstracts. A further search using the terms: United Kingdom OR UK OR Britain OR England OR Wales OR Scotland OR Ireland OR Northern Ireland OR British Isles was combined with the initial search using the AND operator, to restrict the search only to UK based studies. This reduced the number of abstracts to 1458 that met the initial search criteria: 494 of these abstracts were in Medline, 510 in EMBASE and 454 in PsycINFO. Examination of the titles revealed that 502 abstracts from the three databases were duplicates, so there were 956 articles that met the search criteria for this review.
Initial screening of the 956 abstracts showed that 262 of these were not relevant to the review, or were literature reviews. Further screening of the 694 remaining abstracts excluded a further 393 abstracts on the basis of the overall inclusion criteria with 301 abstracts remaining (see figure 1).  Selection of occupational studies 172 of these 301 abstracts were found to be occupational studies (see figure 1). Of these, 22 articles used data from the Whitehall II data and only the paper including data from the first wave was included, so a further 21 articles were excluded. 38 abstracts used data from the King's Centre for Military Health (KCMHR) military studies, of which 3 of the main studies were included in the review as the majority of the remaining papers used data from these same studies (or from a different phase of the same study). Therefore 56 papers were removed, which used these datasets, before the full articles were collated.
The 116 remaining articles were sent to full review, however, 27 did not provide data on the prevalence of CMD and this information was not available from the authors. Further reasons that papers were not included in the review included: 12 papers used the same data as another paper included in the review, 9 included samples ,100 participants (this information was not apparent from the abstract), 3 papers did not collect data on CMD. 65 occupational papers were included in the final review for which data on CMD were available.

Selection of population studies
129 of the 301 abstracts were found to be population or household studies and the full articles were collated which were sent to review (see figure 1). Out of these, 16 studies did not include at least 1000 participants, 7 studies were of older adults and 7 were excluded which either did not administer the GHQ or the prevalence of CMD was not available.
Within the 99 remaining studies, there were 13 population based studies which were eligible for inclusion in the review and some of the population studies were reported in multiple abstracts (e.g. 38 used data from the British Household Panel Survey). There was also 5 out of the 99 studies that reported data from more than one population study.
There were 11 papers identified within the original search that included the relevant data on the prevalence of CMD for either one, or in some cases two, of the population studies (due to some papers reporting data from multiple population studies). The decision was made to report data from only one time-point/ assessment for all of the longitudinal or repeated cross-sectional studies. Data was not available in the 11 identified papers for the Health Survey for England (data was extracted from an official published report from the NHS Information Centre), for the National Child Development Study (this data was extracted from the dataset by the researcher), for the Scottish Health Survey (extracted from an official report), nor for the West of Scotland Twenty-07 study (data extracted for the 3 age cohorts by the Twenty-07 data manager).
Data was therefore reported on 13 population studies: Aberdeen Children of the 1950s study; British Household Panel Survey wave 1; English and Welsh Civil and Social Justice Survey; Health and Lifestyle Survey; Health Survey for England 1995; MRC National Health and Development Study (GHQ administered at 53 y); National Child Occupational studies Study characteristics. Out of the 65 studies identified within this review (table 1), the majority had been conducted in health professionals and NHS staff (n = 41). Five had been carried out in military personnel, four in white collar workers, five in academics and teachers and the remaining 10 studies in social services staff, police, manual workers and chaplains. 47 studies recruited sample sizes ranging from 100 to 1000 participants; however, there were 18 studies which recruited a large sample, including a number of studies of NHS staff (e.g. a study of 1308 UK hospital consultants [11]), the Whitehall II study of civil servants (.10 000 staff from 20 London based civil service departments [6]) and studies of military personnel (e.g. a cross sectional survey of .10 000 UK armed forces personnel [5]). Twenty one studies used random sampling (or invited all participants in the sampling frame to participate in the study) and 44 used non-random sampling. Forty four studies had a response rate of at least 50%, 17 studies reported a response rate of less than 50% and 4 did not report this information.
How was common mental disorder measured? The version of the GHQ that was administered varied between studies, in addition to the cut-off that was utilised. 49 assessed CMD using the GHQ-12, 14 used the GHQ-28, and 2 used the GHQ-30. 39 out of the 49 studies that administered the GHQ-12 used a cut-off of 3/4 as the criteria for being a case of CMD and 11 out of the 14 studies which used the GHQ-28 used a cut-off of 4/5, suggesting that these were the most common criteria for occupational studies which assess CMD.
Prevalence of common mental disorder. Figure 2 displays the weighted estimates for CMD caseness across the different occupational groups, with an overall estimate of 31.6% (95% confidence intervals (CIs) 29.9-33.4%). There was evidence for high heterogeneity [12] from random effect models between the studies in all of the occupational groups (academics and teachers -I 2 = 96.4%; health professionals and NHS staff -I 2 = 96.3%; manual workers -I 2 = 97.0%; military personnel -I 2 = 99.0%; police -I 2 = 97.7%; social services staff -I 2 = 82.6%; white collar workers -I 2 = 91.4%). The prevalence of CMD appears to be highest within the studies of academics/teachers (4 out of the 5 studies included academics working at a university) and social services staff, but the differences between the occupational groups were not statistically significant as the CIs overlapped. The weighted estimate for the prevalence of CMD in studies of academics and teachers was 37.2% (95% CIs 27.8-46.7%), in health professionals and NHS staff was 32.4% (95% CIs 30.0-34.7%), in manual workers was 28.4% (95% CIs 17.0-39.7%), in    [13], with the highest in a study of substance misuse professionals of 82.4% [14].
Occupational studies which used the GHQ-12 (cut-off 3/ 4), random sampling and had a response rate $50%. In univariate meta regression analyses, neither the variable identifying studies as having conducted random sampling (b = 0.00, se[b] = 0.03, p = 0.93), nor that representing having a response rate $50% (b = 20.11, se[b] = 0.10, p = 0.28) were associated with the prevalence of CMD in the restricted occupational studies.
The meta-analysis restricted to the 10 studies [2,5,27,33,36,37,38,39,47,48] which used the same version of the GHQ and the same cut-off, in addition to those which conducted random sampling, with a response rate of at least 50%, resulted in a lower overall prevalence estimate for CMD of 23.9% (95% CIs 20.5-27.4%), with less difference between the occupational groups which were represented in the studies that met these criteria: 21

Population studies
Study characteristics. There were 13 population studies which had administered the GHQ, with the West of Scotland Twenty-07 study including 3 separate age cohorts, so there were 15 studies in total. The British Household Panel Survey, Health and Lifestyle Survey, MRC National Health and Development Study and the National Child Development Study all reported data from England, Scotland and Wales, with the English and Welsh Civil and Social Justice Survey only reporting data from two countries. The remaining studies included participants either from a single country (e.g. the Health Survey for England) or from a particular region in the UK (e.g. the Aberdeen Children of the 1950s Study) (table 2). All of the studies used random sampling or invited all participants in the sampling frame to participate in the study. Fourteen of the studies reported the response rate and 13 of these had a response rate of at least 50%.
How was common mental disorder measured? Eight of the 13 population studies administered the GHQ-12, with the most commonly used cut-off of 3/4. The follow-up to the Aberdeen Children of the 1950s study used a reduced 4-item version of the GHQ with a cut-off of 0/1. The MRC National Health and Development Study used the GHQ-28 (5/6) and the remaining three studies used the GHQ-30 (two with a 4/5 cut-off Prevalence of common mental disorder. Figure 4 displays the weighted estimates for CMD caseness for the population and occupational studies, showing that the estimate was 20.4% (95% CIs 18.2-22.6%) for population studies and there was high heterogeneity (I 2 = 98.4%). The prevalence across the studies ranged from 14.0% scoring above the cut-off for symptoms of CMD in the English and Welsh Civil and Social Justice Survey to 31.2% in the Health and Lifestyle Survey.
Population studies which used the GHQ-12 (cut-off 3/4), random sampling and had a response rate $50%. In a univariate meta regression analysis, having a response rate $50% (b = 20.06, se[b] = 0.16, p = 0.72) was not associated with the prevalence of CMD in the restricted population studies. The metaanalysis restricted to the 7 studies [49,50,51,52,54,55]which used the same version of the GHQ and the same cut-off, in addition to those which conducted random sampling, with a response rate of at least 50%, reported a prevalence estimate of CMD of 19.2% (95% CIs 17.1-21.3%), with evidence for high heterogeneity between these restricted studies (I 2 = 96.4%).
Comparing the prevalence of CMD in occupational and population studies Figure 4 highlights the trend for the weighted estimates of CMD to be higher overall in occupational studies, regardless of the occupational group, compared to the population studies and the confidence intervals do not overlap suggesting that the difference is statistically significant. When restricted to studies that used GHQ-12 (cut-off 3/4) the prevalence estimates (see figure 5) also suggest that the prevalence is significantly higher in occupational studies compared to population studies, but the differences is reduced and no longer significant after further restricting by response rate and sampling method (see figure 6; occupational studies 223.9%, 95% CIs 20.5%-27.4%; population studies 219.2%, 95 CIs 17.1%-21.3%).

Discussion
The main finding of this systematic review is that people appear less likely to report symptoms of CMD in the context of a population based study rather than in a study of the occupational group to which they belong. This difference was reduced but not fully accounted for by differences between these types of studies, in relation to the quality of the studies. This effect is unlikely to reflect true differences in the level of CMD symptoms experienced, because population studies aim to comprise all sections of the population, including those with chronic health problems, long term disabilities and the unemployed, whilst occupational studies will be subject to the health worker effect [7]; so one would anticipate the reverse. There are aspects of work which are associated with increased psychological distress (e.g. high demands and low decision latitude [56]), however, the benefits of employment, over not working, for mental health have consistently been outlined and it is established that the prevalence of CMD is lower in working populations compared to the general population [57]. There were also unexpected findings within the occupational groups: occupations for which one would expect a higher level of traumatic events, such as the military and police, were not found to have a higher prevalence of CMD than other occupational groups, including academics, teachers, white collar workers and social services staff.
We propose that studies directed at the mental health or ''stress levels'' of particular occupations may be subjected to a systematic bias, one that is not present in true population studies when participants are not selected purely because they belong to a specific occupation. Numerous occupational studies are actually labelled as studies of ''work stress'', which may give rise to a framing effect. There is evidence for strong contextual effects in previous experimental and observational studies [58,59,60,61], which can be defined as the effect of environmental factors on subjective outcomes, in addition to bias in self-report psychological measures [62]. The setting in which a questionnaire is completed is likely to influence responses. Responses in the occupational studies may have been biased by a framing effect [63], with the emphasis on job related questions potentially leading to individuals venting their work frustrations through questionnaires which provide an opportunity to report dissatisfaction. This framing effect may be heightened, or conversely lessened, depending on where the GHQ is embedded within the questionnaire and its positioning relative to other psychosocial measures.
Within the occupational groups certain other trends also became apparent, before any restrictions had been made to the meta-analyses. This study was constrained to published research studies (so we did not have data on all occupational groups in the UK), but the overall findings suggested that teachers and academics (with this category predominantly comprising studies of university academics) and social services staff had higher levels of CMD (although the differences were not statistically significant). There is further evidence for higher rates of CMD in teachers in other studies. Analyses of the 2000 UK Adult Psychiatric Morbidity Survey showed that the prevalence of psychological disorder was highest in managers & administrators, teaching professionals, other associate professionals, clerical and secretarial, 'other' sales and personal service occupations. However the prevalence of psychological disorder was higher in primary and secondary teachers, compared to higher education teachers [64], so is not completely in agreement with this review which predominantly included university academics. A further study examining sickness absence from 2001 to 2007 in the Netherlands showed that the percentage of sick days due to CMD was highest in the education sector, followed by financial services and health care sectors [65]. The Adult Psychiatric Morbidity Survey is a population study and the latter study measured actual behaviour in response to CMD, suggesting that the difference between occupational groups is unlikely to be a consequence of contextual effects which seem to be stable across occupations.
A surprising finding within this review was the variation in prevalence estimates for CMD caseness between the population studies, even when restricted to those which used the same questionnaire and threshold. The population studies identified within this review are often used as a population reference for measures of mental health [66], but this study shows that the comparison may differ depending on which population study is chosen. Whilst both the British Household Panel Survey and Health Survey for England are considered to recruit representative samples, there is still a difference between the prevalence of CMD caseness for these surveys, indicating that this either results from a sampling effect or from the sensitivity of the GHQ to other factors which may differ between surveys. These include the position of the GHQ within the overall questionnaire, the length of a questionnaire and finally whether it was self-completion or interview administered.

Strengths and weaknesses
We have conducted a comprehensive review of all papers published since 1990 that have administered the GHQ, within the two areas of research in which this questionnaire is most commonly used, both occupational and population research. This is the first study of which we are aware to compare occupational and population rates of CMD, that is not restricted to a particular occupation. Weaknesses of this review include the fact that the majority of occupational studies of mental health have been conducted within health professionals, most likely due to the salience of these issues to the authors, and there were fewer studies in occupations that are considered to be at a high risk of stress, such as the police. The version of the GHQ administered differed between studies, in addition to the cut-off that was applied. Whilst we tried to control for this by restricting some of the analyses to comparable studies, this is a general limitation of comparisons between GHQ studies which is not always taken into account by study authors. However, the findings do not suggest that the studies with the highest prevalence were those using the lowest cutoff. Further limitations include the sampling differences between the population studies and the difference in location both within the population studies and between the population and the occupational studies. The majority of the occupational studies were conducted in a specific area as opposed to being national studies of a particular occupation. There was also evidence that in general the population based studies were of a higher quality than occupational studies, and the difference between them was reduced after making restrictions to the meta-analysis based on study quality. Finally, there was very high heterogeneity in the meta analyses of both the occupational and population data, which remained when the studies included in the meta-analysis were restricted to be more comparable [67,68].

Implications
The primary implication of this research relates to the sensitivity of the GHQ, which asks about 'recent' symptoms of mental health, to factors other than objective mental health and to the potential framing effect resulting from the overall narrative of a questionnaire or interview. However, we suggest that other self-reported measures of mental health may be subject to contextual effects, and that interpreting the results of any single study without considering the context in which it was given, and the possible bias that introduces, may lead to flawed conclusions. Hence for example, if an individual reports higher levels of psychological symptoms within the context of an occupational study, this may be a reflection of dissatisfaction with their job as opposed to reflecting depression or unhappiness with their life outside of work. These types of effects have previously been shown in individual studies, but we have systematically reviewed evidence across a range of studies. The elevated levels of CMD evidenced in occupational studies may be reduced if mental health is assessed separately to job satisfaction and other occupational constructs and thought should be given as to where the mental health measures are incorporated in a questionnaire.

Conclusions
This review has shown that more individuals scored above the threshold for CMD caseness in occupational studies, compared to population studies, even when accounting for the version of the questionnaire that was administered and the threshold used to classify CMD. We propose that this difference may have resulted from the context of the occupational studies resulting in higher reports of psychological symptoms, in addition to differences between these types of studies in study quality.