Upper body and lower limbs musculoskeletal symptoms and health inequalities in Europe: an analysis of cross-sectional data

The European Working Conditions Survey (EWCS) is a representative cross-sectional survey of workers (employees and self-employed) in the European Union conducted every five years [14]. The statistical population of the EWCS 2010 within each country represents all (non-institutionalised) persons aged 15 and over whose usual place of residence is in one of the countries included in the survey and who were employed the week that preceded the beginning of the interview. In general, the sampling scheme of the EWCS is a multistage stratified random sample. Strata are defined by NUTS regions level 2/3 or equivalent sampling units and, theoretically, all members of the population had a known non-zero inclusion probability [15]. In most countries the targeted sample was 1000 and the overall response rate for the survey was nearly 44%. The survey consists of a standardised questionnaire focusing on the working conditions and health status of workers across member states of the European Union and some other countries such as Norway, Switzerland or Turkey.

The descriptive and analytical analyses in this paper are restricted to the EU-15 countries for which data on physical working conditions is available (Austria, Belgium, Germany, Denmark, Spain, France, Finland, United Kingdom, Greece, Italy, Ireland, Luxembourg, The Netherlands, Portugal, and Sweden). Weighted descriptive statistics at the national and European level were estimated with EWCS waves 1995, 2000, 2005, and 2010 by using EU-15-specific sampling weights. Regression models were estimated with data from the EWCS 2010 as complete-case analyses. Only employed survey participants aged 18 to 65 years were included. The occupation of participants is operationalised by the International Standard Classification of Occupations 1988 (ISCO-88). Occupations are defined as a set of jobs whose main tasks and duties are characterised by a high degree of similarity. The ISCO-88 identifies the following 10 major occupational groups ISCO 1: managers, ISCO 2: professionals, ISCO 3: technicians and associate professionals, ISCO 4: clerical support workers, ISCO 5: service and sales workers, ISCO 6: skilled agricultural, forestry and fishery workers, ISCO 7: craft and related trade workers, ISCO 8: plant and machine operators, and assemblers, ISCO 9: elementary occupations, and ISCO 0: armed forces occupations [16]. In this paper ISCO 0 workers were excluded due to the fact that institutionalised persons are not covered appropriately in the EWCS waves.

Two generalised linear mixed models (GLMM) with logit link are estimated by using the Markov chain Monte Carlo Sampler in the framework of Bayesian inference [17‐19]. These models are also known as mixed logistic regression or multilevel logistic models for dichotomous variables and were chosen in order to take into account the nested structure of the sampling scheme (i.e. countries and regions). The dependent variable of the first GLMM regression is a dichotomous variable corresponding to the EWCS survey question: “Over the last 12 months, did you suffer from muscular pains in shoulders, neck and/or upper limbs?”. The dependent variable in the second GLMM regression corresponds to the question: “Over the last 12 months, did you suffer from muscular pains in lower limbs?”. The independent variables include several physical and psychosocial risk factors that have been identified in the literature [2, 3, 6, 7, 20]. A description of the variables and summary statistics are reported in Tables 1 and 2. The original survey questions related to risk factors associated with MSD are reproduced in Table 1. Ordinal variables in the regression models were not dichotomised.

For the GLMM regressions, the Metropolis-Hastings and Gibbs sampler are used to update the models. The Markov Chain was iterated 13000 times with burn-in set at 3000. In order to take into account the clustered structure of the data, a nested random-effects structure is defined with regions nested within countries. An inverse-Wishart with parameters V = 1,ν = 0.002 is the prior distribution of the covariance matrix of the random-effects. The inverse-Wishart seemed appropriate since iterations series for each estimand did not suggest strong autocorrelations after 500 lags (range of autocorrelations: (−0.07, 0.09) and (−0.09, 0.09) respectively), and relative large effective sample sizes were obtained (ranges: (229, 391) and (173, 331) respectively). In addition, graphical inspection of the estimates in each iteration confirmed a stable mixing property of the chain after the 13000 iterations specified (graphics are available from the author). Predicted estimates of prevalence rates were obtained by marginalising over the random effects and residuals (u, e, respectively) and using the approximation

where E(y) is the expected predicted value of the dependent variable (i.e. presence or absence of upper body and lower limbs complaints), X is the design matrix of covariates, β the vector of estimated parameters and σ ² the sum of the variance components [21]. In order to analyse the consistency of results and to test potential moderation effects of the occupational position of workers, the fully adjusted GLMM regressions are compared with the crude estimates obtained by simple logistic regression models ignoring random effects and excluding the occupational position of respondents. The GLMM regressions with logit link are estimated with the R-package MCMCglmm [17] and the crude estimates are calculated with the routines implemented in the function glm available for the programming language and statistical environment R. The analyses adhered to the STROBE guidelines for cross-sectional studies in epidemiology (see Additional file 1 for the detailed checklist of STROBE criteria) [22].

In order to evaluate the conditions of work across occupational categories and to compare the relative exposures of workers to adverse work environments, the so-called polar plots were used which allow a straightforward comparison of the relative weight of prevalences across groups. The polar plots are constructed for the EWCS waves by applying the following two-step procedure. First, the EU-15 weighted frequencies of the questions covering the physical conditions of work (see Table 1) are estimated for each wave and for each ISCO occupational category. Second, in order to consider only employees at high risk, the percentages of workers responding “All the time” and “Almost all the time” to the corresponding questions are added up for each wave i and each occupation o, say ρ _i,o .

Each polar plot corresponds to a circle or radius 1 divided in segments representing each variable. The radius of the variable with the highest value of ρ _i,o , say ρ _i,max, is assigned the value 1 in wave i. All other values of ρ _i,o across ISCO categories are expressed as fractions of the highest value ρ _i,max. In case that all occupations have the same score ρ _i,o in wave i (i.e. all workers being equally exposed), the circle would have radius 1 for all variables. For instance, the percentage of employees in the ISCO category 8 reporting in the EWCS 1995 being exposed “All the time” and “Almost all the time” to vibrations was highest (i.e. ρ _1995,max=ρ _{1995,ISCO 8}). Hence, for the ISCO category 8 the radius of the variable “Vibrations” is 1. Workers in other ISCO groups reported less frequently being exposed to vibrations so that all other radii have values lower than 1.

In Figure 1 the relative magnitude of the proportion of workers exposed to adverse physical working conditions across occupational categories is represented in the polar plots. A comparison of the estimates since EWCS 1995 reveals that the inequalities of exposure to physical risk factors remain not only consistent but also extremely concentrated in specific occupational groups. Agricultural workers (ISCO 6), craft workers (ISCO 7), operators (ISCO 8), and workers in elementary occupations (ISCO 9) report more frequently being exposed to adverse conditions in comparison with managers (ISCO 1), professionals (ISCO 2), technicians (ISCO 3), clerks (ISCO 4), and service workers (ISCO 5). Particularly, workers in ISCO categories 6, 7, and 8 report consistently very frequent exposure to tiring positions, extreme temperatures, carrying or moving heavy loads, working at very high speed and repetitive tasks. Service workers (ISCO 5) complain of being frequently exposed to tiring positions and carrying heavy loads. The percentage of managers, professionals, technicians and clerks reporting being frequently exposed to adverse physical working conditions across all waves is in comparison extremely small.

After row-wise deletion of observations with missing data the datasets underlying the regression analyses resulted in 17874 and 17868 complete cases, respectively, or about 84% of the original EWCS sample will all age groups (21201). The flow diagram of the complete-case datasets utilised in the regression models is reproduced in Figure 2. Descriptive statistics of the variables considered in the regression models are reported in Table 2. Prevalence of upper body and lower limbs complaints are large with upper body symptoms being more frequently reported than lower limbs complaints (45% vs. 30%). The median of the variables corresponding to exposure to tiring positions, carrying heavy loads, being exposed to very high or very low temperatures and vibrations is located at the right limit of the variable range (i.e. 7). These estimates suggest that most workers are not frequently exposed to those physical risk factors (see Table 2). About 27% of the EU-15 workforce is employed in occupations included in the ISCO groups 7, 8 and 9, whereas the majority of workers is employed in the professional, technical, clerical and services occupations (ISCO groups 2, 3, 4, and 5, respectively).

This paper suffers from several limitations. (1) Causal relationships between risk factors and self-reported symptoms cannot be identified since cross-sectional data were analysed. For instance, the extensive systematic review of Mayer and colleagues based on longitudinal studies found strong evidence for the association between neck and/or shoulder complaints and vibration (OR between 1.6 and 2.5) [26]. In this paper, on the contrary, exposure to vibrations was not associated with upper body symptoms. This may be due to the fact that workers change jobs in order to avoid or reduce exposure to risk factors such as vibrations. (2) Exposure levels to physical risk factors in the EWCS are not assessed by biomechanical measurements at the workplace but rely on self-reported frequency scales. The instruments used to assess exposure to psychosocial risks are not psychometrically valid. These methodological deficiencies might bias the odds ratios and prevalence rates estimates, assess different constructs (e.g. coping instead of stress levels), and inflate or dilute real associations. (3) Even though all EWCS questionnaires in the different languages were based on the English version, there might be additional variance related systematically to translation and interpretation differences across countries and regions. However, this was taken into account to some extent by the nested structure of the regression models. (4) As stated in the Background section, the EWCS suffers from a large proportion of non-response (overall response rate about 44%). This might limit the generalisability of the results to the targeted statistical population in the selected countries. (5) The outcomes were not based on clinical diagnoses but on self-reported presence or absence of musculoskeletal pain in the last 12 months prior to the survey. Hence, there might be substantial misclassification of outcome status biasing the odds ratios and prevalence estimates.