Constructing and Validating an Occupational Mechanical Job Exposure Index (MJEI) Based on Five Norwegian Nationwide Survey of Living Conditions on Work Environment


 Background: The aim of this study is to (i) construct and validate a job exposure index for mechanical job exposures and (ii) test its predictive validity for individual reported lower back pain and long-term sick absence, as well as disability benefits, long-term sick leave and mortality using register data. Methods: We utilize data from the Norwegian nationwide Survey of Living Conditions on work environment in 2006, 2009, 2013, 2016 and 2019. Occupations are classified on a 4-digit level based on the Norwegian version of the International Standard Classification of Occupations (ISCO-88). We constructed a 4-digit correspondence table between the occupational codes used in the 2006 and 2009 surveys (STYRK-98) and the codes used in 2013, 2016 and 2019 (STYRK-08). The mechanical exposures were collected by Statistics Norway using telephone interviews. We validate the agreement between the individual- and occupational-based mechanical exposure index (MJEI) estimates using Spearman`s Rho, sensitivity and specificity measures. The predictive validity of the mechanical job exposure matrix was tested investigating the association between individual reported lower back pain and long-term sick absence and the individual- and MJEI estimates. Furthermore, we investigate the association between disability benefits, mortality and the number of long-term sick absence periods, retrieved from register data, and the occupational-based mechanical exposure (MJEI) estimates. Results: The analysis shows a fair-to-moderate overlap between occupational-based mechanical exposure index (MJEI) and the index based on individual reported exposures. When assessing the predictive value of the occupational-based mechanical exposure index to the index based on the individual reported exposures, the MJEI showed lower and reproducible associations with lower back pain for both men and women. For long-term sick leave, the MJEI showed higher and reproducible associations for both genders. As for the register-based outcomes, the MJEI predicts disability and a significant higher number of long-term sickness benefits periods for both men and women. For men the MJEI also predicts higher mortality. Conclusion: The predictive value of the mechanical job exposure index (MJEI) is overall acceptable and will be useful in register-based studies lacking information on these types of exposures.


Introduction
The Nordic countries have a longstanding tradition of using high quality register data for research purposes. The fact that these data often include the entire population and consists of long time series makes them "a goldmine" for research (1). However, these data are not collected for research purposes and therefore often lack information that is vital for research into social inequalities in health and epidemiological register research in general. Varying dimensions of working conditions are usually such a "missing piece of the jigsaw puzzle" in register-based research. Knowledge about work environment is crucial in itself in a democratic society that cares about the population's living conditions and well-being. In addition, knowledge about the pathways into and out of employment and different types of jobs is important to assess the interrelationship between work, health and wellbeing over the life course (2; 3). One way to overcome the problem of missing information on work environment in register data is to use a job exposure matrix (JEM). Based on job titles JEM's have been developed for a number of different and speci c exposures and stressors (4; 5; 6). Despite the great advantage of creating this kind of information for use in register data, a notorious problem with JEM is that it entails the risk of misclassi cation which limits its applicability. This relates to the exact de nition of exposures as well as the de nition of exposed/non-exposed. These possibilities and challenges related to JEM are the motivation for this undertaking.
In Norway, Hanvold et. al. (7) have constructed a JEM for mechanical and psychosocial job exposures based on the Norwegian nationwide Survey of Living Conditions on work environment in 2006 and 2009. This paper is inspired by the innovative work done by these researchers. We have, however, moved beyond Hanvold et al. in two ways. First, we have added three Norwegian nationwide Survey of Living Conditions on work environment, i.e. for the years 2013, 2016 and 2019. Hence we have achieved a much larger number of observations. Second, we have applied JEM on Norwegian register data and investigated the associations between JEM and three health outcomes, disability bene ts, mortality, and the number of long-term sick absence periods.

Previous Research
Use of nation-wide and longitudinal register data for research purposes has many advantages (1). One challenge though is the lack of information on exposures to job hazards of different types (4). One solution to these problems is to use a Job Exposure Matrix (JEM). A JEM is used to assign exposures on the basis of occupational titles. Hence, a JEM is potentially convenient when information on individual occupation is available, but there is no information on job exposures or job hazards, as is the case in Norwegian register data. The JEM method is quite extensively used and has proved useful in several contexts (6; 8; 9). As indicated JEMs have been produced to capture several types of exposures and stressors, such as biological, mechanical, chemical and psychosocial. Since our JEM addresses mechanical exposures only, we will primarily review empirical studies that have assessed the reliability and validity of mechanical JEMs.
In the Netherlands Rijs et al. (10) found that use of force and work in uncomfortable positions were signi cantly associated with functional limitations and self-perceived health. A moderate probability of repetitive movements was associated with functional limitations in former workers. A high probability of repetitive movements was associated with functional limitations in current and former workers as well as with SPH and hip and knee. The authors conclude that the results suggest that the JEM accurately classi es jobs according to physical demands.
In Finland Solovieva et al. (5) reported that the speci city of the mechanical JEM was, in particular among women. The degrees of agreement, measured by kappa, were fair for most exposures. For men, all JEM exposures were signi cantly associated with one month prevalence of low back pain. For women, this applied to four out of six JEM exposures. The researchers conclude that the JEM can be «considered as a valid instrument for exposure assessment in large-scale epidemiological studies, when more precise but more labour-intensive methods are not feasible» (5: p 1).
In Norway Hanvold et al. (7) constructed and validated JEMs capturing mechanical as well as psychosocial work exposures. They found a general fair to moderate agreement between the JEM and individual work exposures. The JEM performed better for mechanical work exposures than for psychosocial stressors. The predictive validity of the mechanical JEM showed an acceptable relationship with the risk of low-back pain. The authors conclude that JEM «may be useful in large epidemiological register studies» (7: 239).
Against this background, the aim of this article is to propose a mechanical JEM for use in Norwegian register data. This implies to assess its statistical properties in various ways, as described below. The idea is to use this JEM for different purposes in our «research program» on work, health and welfare trajectories among vulnerable groups. Hitherto, available information in Norwegian register data has been limited to occupation (job titles), social class and employment status. Our ambition is to add a reliable and validated index variable describing mechanical exposures to this list. Speci cally, we will construct a composite mechanical job exposure index (MJEI), compromising eight different mechanical job exposures, and validate it by the assessment of a con rmative factor analysis, by investigating the correspondence between the individual reported exposures and the occupational exposures, by judging sensitivity and speci city measures, and lastly by examining whether the MJEI predicts self-reported lower-back pain and long-term sick leave using survey data, and disability and long-term sick leave using register data.

Study population
The populations included in the analysis are described according to age, educational level and major occupational groups in Table  1 -the survey data and Table 2 -the register data. As shown in Table 1 Table 3 the mean number of respondents per JEM group more than dobled in both men and woman.   (13), the eight mechanical exposures were dichotomized into exposed and not exposed at the individual level. The questions and cut-off values used are shown in Table   4 below. Table 4 Exposures, Questions and Non-exposed/Exposed Composite Mechanical Job Exposure Index Exposures Questions Not exposed/Exposed Heavy lifting (>20 kg) Do you have to lift something that weighs more than 20 kg daily, and in the case of how many times per. day?
"Yes, at least 20 times per. day", "Yes, 5-19 times per. day", "Yes, 1-4 times per day", "No" 0 = Not exposed (No), 1 = Exposed (≥1-4 times) Hands above shoulder height Do you work with your hands raised at shoulder height or higher? -"yes" or "no" All the exposure variables are constructed as the proportion of individuals within each JEM-group that are exposed to the speci c exposure. Thus, we have constructed variables that, in principle, goes from 0 to 100 percent based variables that are dichotomous (exposed = 1, not exposed = 0). This means that occupational codes with a value of 0 on one of the variables implies that none with these occupational codes, belonging to the same JEM-group, has provided an answer that involves exposure. In contrast, the value 100 means that all respondents with that occupational code, belonging to the same JEM-group, have provided an answer that involves exposure. In total, we have 323 unique occupational codes that are used when the index is merged to register data.
Constructing the composite mechanical job exposure index (MJEI) In order to investigate the factorial validity of the occupational-based mechanical exposure index (MJEI), con rmatory factor analysis was performed. The CFA model was tted in Stata v16 and for model estimation maximum likelihood was applied.
Model evaluation was based on chi-square tests for model t and further model t indices, including the root mean square error of approximation (RMSEA), the comparative t index (CFI), the Tucker-Lewis index (TLI) and the standardised root mean square residual (SRMR). For model t to be interpreted as 'acceptable', a RMSEA of < 0.05 was considered a close t, while a RMSEA and a SRMR of up to 0.08 were considered acceptable. Comparing the t of a target model to the t of an independent or null model, the CFI has a cut-off for good t CFI of ⩾0.90. A TLI of 0.95 indicates the model of interest improves the t by 95% relative to the null model, and the cut-off for good t was sat at TLI ⩾0.95. Furthermore, the correlations of residuals to improve model t when tting the nine one-factor models were considered. Correlated residuals < 0.2 were considered acceptable when tting the model (14; 15). Potential model adjustments were based on modi cation indices as provided in the Stata output using the 'estat gof, stats (all)' command. To obtain a clearer idea of the data and potential problematic items, a one-factor model was tted to the data. To test whether modi cations, in terms of correlated within factor residuals, led to signi cant model improvement, modi cation indices were obtained using the 'estat mindices' command in Stata. .072 Exposures *Standarised factor loading Standard error Share exposed -Heavy lifting (>20 kg) .744 .030 Share exposed -Hands above shoulder height .816 .023 Share exposed -Heavy physical work .758 .029 Share exposed -Squatting/kneeling .838 .021 Share exposed -Forward bending .728 .031 Share exposed -Awkward lifting .862 .019 Share exposed -Standing/walking .762 .028

*no cross-loadings and no correlated residuals
The results from tting a one-factor model is shown in Table 5. The "Original" row shows the results when tting the MJEI with no cross-loadings and no correlated residuals. All factor loadings were high (i.e. >0.7; see column "Standardised factor loading" in Table 4).
When tting the one-factor model, correlated residuals were sequentially added to respective models, which improved each model t signi cantly. As shown in Table 4, a model t with ten modi cations gave a satisfying model t. All the correlated residuals were <0.2. The MJEI showed good internal consistency with a Cronbach's alpha of 0.89 (see Table 5).

The MJEI performance
In order to assess the MJEI performance we used four different performance measures: Cohen sKappa, sensitivity, spec if icity and Spearmans Rho. Cohen`s Kappa mesures agreement between the group-based exporsure estimates and the individual expsosure estimats, taking into account that agreement may occur by chance. According to Cohen (16) the kappa values can be classi ed as poor (<0.20), fair (0.21-0.40), moderate (0.41-0.60), good (0.61-0.80) and excellent (0.81-1) agreement. Sensitivity measures the proportion of individuals who are identi ed as exposed based on individual estimates, that are also identi ed as exposed using the group-based estimates. Speci city measures the proportion of individuals who are identi ed as unexposed based on individual estimates, that are also identi ed as unexposed using the group-based estimates. Spearman rRhomeasuresthemo ¬onicrelationships, whetherl ∈ ear or ¬, betweentwovariab ≤ s. InthispaperweuseSpearman r Rho to investigate the correspondence, i.e. the rank order, between the exposures reported by the individual employee and the exposures linked to the individual using their occupational code.
As shown in Table 6, CohensKappaisfare → g∞d and mod eratef or allexposuresexceptf or " s gave a sensitivity of >50 percent for six out of eight exposures. The speci city was ⩾ 75 percent for six out of eight exposures for men and ve for women. Hanvold et. al. (7) used cut-off values when constructing their nal Job Exposure Matrix, being as theyre goal was to investigate each exposures association with lower-back pain, our goal however is to construct a Mechanical Job Exposure Index for the use in register data analysis. Thus, we choose not to reduce the information in the exposures using cut-offs values but have instead use the exposure variables measuring the percentage within each occupational code that is exposed. The sensitivity and speci city measures provide a valuable insight into the different exposures performance in identifying exposed and non-exposed individuals. However, since our goal is to measure the overall mechanical exposure in each occupation, it seems more fruitful to consider occupations as more or less exposed based on the percentage reporting to be exposed in each occupation. Thus, we have chosen to keep the measures, measuring the percentage exposed and used in the factor analysis, as is when constructing the Mechanical Job Exposure Index (MJEI). To test the correspondence between the exposures measured as percentage exposed within each occupation (group-based exposure) and the individual reported exposures we use Spearman`s Rho, the results from a rank correlation analysis is presented in Table 7. .637 (.000) The rank correlation between the Composite Mechanical Job Exposure Index based on the individual reported exposures and the group-based exposures is .642 for men and .626 for women (see Table 7). Thus, the correlation between the index based individual reported exposures and the group-based exposures is strong for both genders. For each of the eight exposures the correlation between the individual reported exposure and the group-based exposure is weak for "neck exion" and "forward bending". Whereas the correlation is moderate for "heavy lifting", "hands above head", "Heavy physical work", "Squatting/kneeling", "Awkward lifting" and strong for "Standing/walking". When comparing the sensitivity measures with the correlations it shows that those exposures with a low sensitivity, "work with neck exion" and "forward bending" for both genders and "hands above shoulder" for women, also have a weaker correlation. Nevertheless, an overall correlation of .642 for men and .626 for women demonstrates that the Composite Occupational Mechanical Job Exposure Index (MJEI), based on ve Norwegian nationwide Survey of Living Conditions on work environment, is strongly correlated with the overall mechanical job exposures experienced at the individual level.
Low-back pain, long-term sick leave, disability bene ts and mortality To test the predictive validity of the Composite Mechanical Job Exposure Index individual reported low-back pain and long-term sick leave is used as outcome variables in the analysis based on the ve surveys. Individual reported low-back pain is measured as a dummy-variable: "Have you during the last month been bothered by lower back pain?" "Very or quite bothered" = 1, "a little or not at all bothered" = 0. Individual reported sick leave is also measured as a dummy-variable: "Have you during the last 12 months had continuous sick leave for more than 14 days?" "Yes"=1, "No"= 0.
Furthermore, the predictive validity of the Composite Mechanical Job Exposure Index is tested merging the index to register data using receipt of disability bene t in the period 2008 to 2017, the number of long-term sick leave periods between 2008 and 2015 and mortality between 2008 and 2017 as outcome variables. "Disability" and "mortality" are both measured as dummy variables: "disabled during 2008 to 2017" = 1, "not disabled during 2008 to 2017" = 0 and "dead during 2008 to 2015" = 1, "not dead during 2008 to 2017" = 0. "Long-term sick leave periods" is measured as a continues variable and sums up the number of sick leave periods exceeding 16 days between 2008 and 2015.
Predictive validity of the Composite Mechanical Job Exposure Index As shown in gure 1, for both men and women, the unadjusted occupational MJEI estimate is not signi cantly lower than the individual MJEI estimates (unadjusted and adjusted), thus the occupational MJEI shows a reproduceable likelihood for lower-back pain for men. When adjusting for level of education and age, the reproduceable likelihood for lower-back pain is signi cantly lower for men, but still signi cant. Figure 2 shows the likelihood of reporting a long-term sick leave among men and women, according to the occupational MJEI and the individual MJEI. The occupational MJEI shows a reproduceable likelihood for long-term sick leave for both men and women, and the adjusted occupational MJEI estimate does not signi cantly differ from the individual estimates.
When investigating the association between the occupational MJEI and disability 2008-2017, the occupational MJEI does not predict a higher likelihood for disability among men when adjusting for age and level of education ( Figure 3). For women the occupational MJEI predicts a higher likelihood for disability during 2008 to 2017 both before and after adjusting for age and level of education.
As shown in gure 4, the occupational MJEI predicts higher mortality among men both before and after adjusting for age and level of education. For women the occupational MJEI predicts higher mortality after adjusting for age and level of education.
The occupational MJEI predicts a signi cantly higher probability of having ten or more long-term sick leave periods during 2008 to 2015 for both men and women, before and after adjusting for age and level of education. As shown in gure 5, the predicted likelihood is almost twice as high for women compared to men.

Summary, Discussion And Conclusion
In this paper we have tested key aspects of reliability and validity for a composite measure of a mechanical job exposure matrix, MJEI. Our main ndings may be summarized in the following three points. 1) A con rmatory factor analysis shows that the 8 items measuring different aspects of mechanical exposures re ects one underlying dimension, and thus indicates that the index MJEI is reliable.
2) The overall rank order correlation (Spearman's Rho) between individual and occupation based mechanical exposure is moderate-high and suggests that MJEI can be used when individual exposures are not available. 3) An analysis of the predictive validity of the MJEI suggests that overall the constructed MJEI index is related to several health outcomes in the expected way. This analysis was carried out on survey data and register data and involved health outcomes like self-reported back pain, long term sick leave, receipt of disability bene t and mortality. The validity of MJEI thus seems to be acceptable.
As alluded to above, the JEM approach has some weaknesses related to imprecision as compared with individually reported exposures. The reliability and validity tests carried out in this article suggest, however, that its statistical properties altogether are fair. One should also be aware of one advantage that JEM has as compared with a heavily debated problem pertaining to selfreported job exposures, namely systematic reporting bias (17). Using JEM, one can to a large extent ignore this possible problem.
A strength of this study is the high number of observations that is achieved by merging ve waves of The Survey of Living Conditions. This has resulted in 43 977 valid respondents. The mean number of respondents in each JEM group is 412 and the median is 261. The largest JEM group include 1503 respondents and the smallest 19 respondents. Only two JEM groups have the minimum number of 19 respondents. Overall, these high numbers increase the precision of the estimated exposures of mechanical al hazards. As pointed out by Choi (18) there might be a trade off between number of subjects in each JEM group and the precision of the estimated exposures: To collapse several occupational groups in order to obtain a higher number of observations, will to some extent result in lower precision. In our material, this is less of a problem since few occupations are collapsed and the remaining are "clean" occupational groups.
The validation literature referred to above seems to conclude that when individual information on job exposures is lacking, JEM is a useful proxy. Our results seem to con rm this body of research. The statistical properties of our mechanical job exposure index (MJEI) are overall acceptable. Since it was our purpose to construct a JEM index that could be used in analyses of register data, we are inclined to conclude that MJEI is a valid measure of mechanical exposures that can be informative in register-based studies in Norway.  Linear probability model using survey data only and individual reported long-term sick leave as dependent variable. Results when not adjusting and adjusting for level of education and age.  Linear probability model using mortality 2008-2017 as dependent variable. Results when not adjusting and adjusting for level of education and age. Register data. Men N=819 232. Women N=770 303. Figure 5