The simple correlation matrix in Table revealed highly significant intercorrelations between all relevant study variables without exception, providing evidence for an unadjusted association between WLC and MSD, substantiating the suspicion of confounding in this association, supporting the assumption of WLC as a strong stressor and indicating a mediational relationship. There was a highly significant bivariate correlation between the independent or exposure variable (WLC) and the dependent or outcome variable (MSD) as expected (r = .20). In addition, the work-related variables of physical strain at work, workload, time pressure at work and job autonomy correlated consistently with both consolidated multiple-item measures of WLC and MSD and therefore had to be considered as potential confounders in a further multivariate analysis of association, or rather regression. Also, the assumed intervening variable of general stress was strongly correlated with both the dependent (r = .23) and independent (r = .44) variables.
Intercorrelations between all relevant study variables (Pearson's r)
With these preconditions satisfied, only one last condition remained to meet all sufficient statistical criteria for mediation according to Baron and Kenny's three-step approach for testing mediation with cross-sectional data [41
]: the effect of the independent variable on the dependent variable must be less when the intervening variable is controlled. To study this and the other hypotheses (see Figure ), four multiple linear regression models were computed using either the consolidated WLC measure or the two differentiated WLC subscales as the independent or exposure variable(s), and the MSD summary score as the dependent or outcome variable (see Table ). The effect of WLC on MSD was initially adjusted for the control variables (Model 1), then additionally adjusted for the confounding variables (Model 2), subsequently also controlled for the intervening variable (Model 3) and finally calculated separately using the two differentiated WLC subscales instead of the consolidated measure (Model 4). Overall, we found clear evidence for an association and confounding as well as support and indication for mediation if possible with cross-sectional data.
Work-life conflict and stress as the main explanatory factors and other covariates of musculoskeletal disorders
These multiple linear regression analyses (see Table ), stepwise including and adjusting for additional covariates, showed that the association initially found (β = .21) and controlled for the usual control variables (Model 1) was considerably reduced when firstly an adjustment was made for the considered work-related confounders (Model 2), and secondly, the intervening stress variable as the assumed mediator was additionally added to the regression model (Model 3). The reduction of effect size due to these adjustments was substantial. Overall, the beta coefficient as the measure of association halved, and the stress variable finally turned out to be the stronger risk factor for MSD (β = .17) than WLC. However, there still remained a significant and relevant association between WLC and MSD (β = .10), as shown in the fully adjusted regression model (Model 3). The completely specified and differentiated regression model (Model 4) came out as the best fit of all four computed linear regression models, and explained the highest percentage of the total variance (adjusted R squared = 11.8%). This model showed the multiple adjusted effects of the two assessed and considered directions of WLC separately and identified work-to-life conflict as the only predictive and MSD-relevant direction of WLC. The other measured direction of WLC, namely life-to-work conflict, turned out to be not at all predictive with regard to MSD. In other words, the association found between WLC and MSD can be ascribed exclusively to the effect of the work-to-life conflict. The two WLC directions or subscales were not that strongly correlated with each other (r = .24), so that their combined use in a regression analysis was not problematic with regard to likely collinearity.
Stratified analyses now focusing solely on work-to-life conflict as the only remaining explanatory variable basically confirmed the association between WLC and MSD for both measured types of MSD and for both sexes separately.
Table shows the number of cases (absolute frequencies) and the percentages (relative frequencies) for both measured types of MSD and both sexes separately, each differentiated by degree of exposure (work-to-life conflict). The results revealed that percentages of backache or low back pain as well as of neck or shoulder pain increase greatly and gradually with an increasing degree of work-to-life conflict, especially in women. Percentages in the most exposed group (very high work-to-life conflict) are five times higher throughout than in the least exposed group (very low work-to-life conflict), independent of sex or type of MSD.
Number of cases and percentages of strong (low) back pain and neck/shoulder pain in men and women by degree of work-to-life conflict
Multivariate logistic regression analyses with multiple adjusted OR as measures of association also revealed that the relative risk of suffering seriously from backache or low back pain and neck or shoulder pain was highest in those who were most exposed to work-to-life conflict compared to the reference group comprising those least exposed, irrespective of type of MSD or sex (see Table ). Multiple adjusted OR for this most exposed group ranged from 3.8 to 6.3 in the total effect model and from 1.9 to 4.1 in the direct effect model, depending on type of MSD and sex. A strong association and a clear gradient was found for both types of MSD and both sexes (total effect model). Even when perceived general stress was considered, i.e. when the corresponding intervening variable was included in the logistic regression model, adjusted OR were clearly reduced but still increased steadily with cumulative degree of work-to-life conflict (direct effect model). This applies basically to both types of MSD, although for back pain the increased OR mostly were no longer significant on the 95%-confidence interval level for the more exposed group (very high work-to-life conflict) compared to the reference group (very low work-to-life conflict). Again, as expected and already implied by the linear regression models (see Table ), stress was found to reduce the strength of association and therefore suspected to mediate the relationship between WLC and MSD more (backache or low back pain) or less (neck or shoulder pain). Stress turned out to have quite a strong and independent effect on back and neck or shoulder pain and therefore to be a strong risk factor for MSD itself. Although the results of these stratified and differentiated logistic regression analyses were largely consistent with the findings of the previous and consolidated linear regression analyses, slightly differing results were found with regard to sex and type of MSD. Associations were stronger and gradients were more pronounced in women and for neck/shoulder pain, whereas stress-adjusted direct effects were smaller for (low) back pain.
Work-to-life conflict as the principal risk factor for strong (low) back pain and neck/shoulder pain, stratified by sex and additionally adjusted for general stress