Physical exposure data were obtained from the CONSTANCES project (“Cohorte des consultants des Centres d’examens de santé”), a large (expected n = 200,000) prospective French cohort study investigating occupational and social determinants of health in the general population [18]. CONSTANCES was designed to create a representative sample of French salaried workers. Detailed information on CONSTANCES is available at: www.constances.fr. CONSTANCES participants answered questions estimating 27 different physical risk factors in each participant’s current job. Exposure questions were patterned after the SALTSA criteria [19] and other sources [20]. Overall intensity of physical workload was assessed with the Borg rating of perceived exertion (RPE) scale, ranging from 6 (No Effort at All) to 20 (Exhausting). Questions pertaining to the duration or frequency of performing specific actions, including postures, repetitive motion, and the use of vibrating tools, were evaluated on a 4-point Likert scale (text of each question listed in Table 2). Generally, the Likert scale was formatted with the following anchor points: “Never or nearly never”, “Rarely (< 2 hours per day)”, “Often (2 to 4 hours per day)”, and “Always or nearly always”. Questions pertaining to regular handling, moving, or carrying loads asked participants to report whether they handle objects greater than 1 kg [yes/no], and if yes, asked the frequency of handling objects based on different ranges of weights, following the 4-point Likert scale above.

We used data from the first 81,425 CONSTANCES participants. Reported job titles were assigned a French 4-digit PCS (Profession et Catégorie Sociale) job code using the SiCore automated coding system [21]. The PCS classification system involves three nested levels of classification, from the 1-digit socio-professional job categories (Table 1) to the 4-digit PCS job code. This assignment resulted in 418 PCS job codes. Participants who were not currently working (n = 35,466), those who did not report a job title or who were not assigned a PCS job code through automatic coding (n = 10,396), and those who had missing exposure data (n = 30), were excluded.

To produce reliable estimates, we required a minimum of 10 valid responses for each risk factor within each PCS job code. PCS jobs with fewer than 10 responses were grouped with other similar PCS jobs to create adequately sized groups (a minimum of 10 valid responses for each exposure for each PCS code). This method has been previously applied in grouping American standard occupational classification (SOC) codes [22]. To create groups of similar jobs, we first used a PCS to ISCO-88 (International Standard Classification of Occupations) crosswalk (Codage Assisté des Professions et Secteurs d’activité) and an existing French auto-coding system tool [23]. Many PCS codes share a single ISCO-88 code, thus creating natural groupings. To group the remaining PCS job codes with few respondents, we used an ISCO-88 to ISCO-08 crosswalk, and an ISCO-08 to SOC crosswalk. All such groupings were reviewed, and PCS job codes that were not successfully grouped via crosswalks were grouped manually based on consensus opinions from three of the authors (BE, AD, AMD). PCS codes with a small sample size that could not be meaningfully merged with other jobs were excluded (n=7 participants). After all exclusions and job code grouping, the JEM was comprised of 27 physical exposures assigned to 407 PCS codes from 35,526 eligible participants.

We computed descriptive statistics to assess the demographics of the cohort, the overall distributions of each of the 27 risk factors, and proportion of symptomatic and asymptomatic participants. To better enable interpretation of JEM assigned exposure estimates and comparison with exposures based on other methods, the ordinal questionnaire responses were re-coded to time-based variables (i.e., minutes of activity per day). We selected the median value of the questionnaire time interval: 0 minutes (ordinal rating of 0 on the 5-pt ordinal scale), 5 minutes (ordinal rating of 1 = “Never or nearly never”), 60 minutes (rating of 2 = “Rarely (< 2 hours per day)”), 180 minutes (rating of 3 = “Often (2 to 4 hours per day”), and 360 minutes (rating of 4 = “Always or nearly always).

When reporting JEM-assigned exposure values, studies have used different exposure metrics [29,30]. MSD-focused JEMs have typically reported arithmetic means [1] and medians [31], therefore we reported both metrics. We also corrected the JEM mean value using empirical quantile mapping (EQM) methods [32] to adjust the group-level data to better reflect the distributions of individual-level exposure estimates. Using EQM, JEM mean values falling within every 1% quantile range were adjusted to reflect respective 1% quantiles of the individual-level self-reported values; this adjusted JEM mean is referred to as bias-corrected mean.

To compare exposure metrics, we calculated the within-job variance, between-job variance, and r-squared values for these three exposure metrics for all 27 physical exposures. Within-job variance was defined as the average of the squared deviation from group metric values (Eq. 1). Between-job variance was the average of the squared deviation of metric values from the global mean (Eq. 2). (Eq. 1)Within-job variance =1N−K∑j=1K∑i=1NXji−X˜j2 (Eq. 2)Between-job variance =1K−1∑j=1KnjX˜j−X¯2 where X˜ji is the estimated metric value for the j^th group.

Eligible participants represented 407 PCS job titles nested within six broad socio-professional categories. Twenty-three percent of the cohort reported musculoskeletal pain in one or more body regions (Table 1). A linear mixed model compared exposure values between symptomatic and asymptomatic participants; 23 of 26 risk factors demonstrated statistically significant differences (Table 2). Positive beta coefficients from these models indicated that symptomatic individuals reported higher exposure values than asymptomatic individuals within the same PCS job code. Of the 26 linear mixed models, 21 exposures variables had statistically significant positive beta estimates. Eleven exposure variables had beta estimates greater than 0.2. Negative beta estimates indicated that symptomatic workers reported lower exposures than asymptomatic workers. Significant negative beta estimates were observed with two variables: Change Task (β = −0.11) and Rest Eyes (β = −0.18).

The asymptomatic cohort (range: 0.15 to 6.13) demonstrated lower within-job variance than the full cohort (range: 0.16 to 6.65), resulting in more favorable homogenous exposure groups (Table 2). As a result, only exposure estimates from asymptomatic workers were included in the JEM.

As expected for the general population in an industrialized country, the risk factors with the highest mean and median duration of daily activity were related to computer or office work, with much lower daily durations of heavy lifting or hand exertion (Table 3). Examining individually reported exposures at the level of the job, NPMANOVA analysis showed significantly higher between-job variance than within-job variance among the 27 exposures (229 PCS Groupings; F[228,21989] = 67.18, p < 0.0001). PCS job codes explained 41.4% of the variance in individual self-reported exposures in the overall model. The univariate analysis (Table 3) for each risk factor variable revealed r-squared values ranging from 5% (Reaching for Items Behind Back) to 55% (Standing). This indicates that the amount of variance explained by PCS job codes was different between risk factor variables; of the 27 risk factors, 12 variables achieved r-squares greater than 30%, while three variables resulted in explained variance less than 10%. Despite the large range of explained variance, all univariate models were statistically significant (all p < 0.0001) indicating a relationship between exposures estimated by PCS code and self-reported exposure variables among asymptomatic workers.

Taking all reported risk factors into account, we observed non-overlapping relationships between individual PCS codes (shown by ellipses in Figure 1), indicating separation between different jobs. We also noted clustering of PCS codes within the same socio-professional categories (represented by color).

We observed minimal differences between the three exposure metrics (mean, median, bias-corrected mean) based on the within-job variance (Supplement, Table 1). Trends indicate a comparable within-job variance using the means (variance = 0.15 to 6.13), medians (variance = 0.18 to 6.73), and bias-corrected means (variance = 0.22 to 7.62). In contrast to the within-job variance, the bias-corrected mean (variance = 25.60 to 1193.55) showed markedly higher between-job variance than means (variance = 2.35 to 492.03) or medians (variance = 5.93 to 764.15). R-square values of the 27 physical risk factors ranged from 0.06 to 0.57 (JEM mean), 0.17 to 0.64 (JEM median), and 0.38 to 0.65 (JEM bias-corrected mean). Thus, compared to means or medians, use of bias-corrected means resulted in more homogenous exposure groups at the job level (greater contrast of within- and between-job variance), and explained more of the variance in individually-reported exposures.

Examination of residual plots shows increasing differences between individually reported versus group-level exposure estimates with increasing exposure level (Example: Figure 2; For all physical risk factors see Supplement, Figures 1–27). JEM-assigned exposure estimates were attenuated as the exposure level increased; this effect was most pronounced when assigning individual exposure values based on group-level mean values. Use of the bias-corrected mean led to smaller differences at all exposure levels compared to the JEM mean and median plots. A representative example of these box-plots for JEM mean, bias-corrected mean, and median exposure metrics is shown in Figure 2.

When using job-means, the mean differences were near-zero for all exposure variables [−0.002 (Repetition) to 0.003 (Drive Car or Truck)]; job-medians led to a mean difference ranging from −0.27 (Rest Eyes) to 0.40 (Repetition) [Supplement, Table 2]. JEM bias-corrected mean ranged between −0.05 (Handle Objects 1–4 kg) and 0.007 (Repetition and Drive Car or Truck). The bias-corrected mean also led to lower variance differences compared to JEM median values.