We linked baseline and 2 year follow-up data obtained in the Gradients of Occupational Health in Hospital Workers (GROW) study. The longitudinal component of this investigation followed injured workers (cases) with incident WRMSDs of the trunk, neck and upper and lower extremities from two separate hospital sites comparing them to non-injured referents from the same sites. The protocol was approved by the University of California San Francisco committee on research involving human subjects.
We collected baseline data in 2002–04 through structured telephone interviews employing computer-assisted telephone interviewing software. In addition, we conducted onsite ergonomics observations of work practices in a subset of subjects (75%). Details of the study design, recruitment and validity and reliability of the baseline interview and ergonomics instruments used in the subset analysis as well as the baseline study findings have been previously published [14
]. In brief, we recruited participants from a base of ~6000 hospital workers at two sites, representing all occupational groups (with the exception of physicians who were excluded from the study). Cases were defined by an incident WRMSD determined to be work related by physicians or nurse practitioners employed at each site’s employee health clinic. Referents were equally matched on the basis of (i) job group, (ii) shift work type (e.g. working a routine daytime schedule compared to various shift arrangements) or (iii) at random sequentially in time (incidence density matching), yielding an overall 3:1 ratio of referents to cases. Approximately 2 years after baseline participation, we attempted to recontact subjects for follow-up structured telephone interviews.
Data for age, sex, race–ethnicity, education, income, smoking status and medical co-morbidities were obtained at baseline. Occupational categories were grouped as administrators and managers, nursing, other clinical, clerical, technical or support staff. We assessed work organization factors using two measures: (i) job strain, derived from the Job Content Questionnaire (JCQ) [19
] and (ii) the effort–reward ratio, derived from the Effort-Reward Imbalance (ERI) Questionnaire [20
]. Three ergonomics measures, based on direct worksite observations, were assessed at baseline [16
]. The first ergonomic measure assessed upper body and neck strain [Upper Body Assessment—University of California (UBA-UC)]; the second, back and lower extremity strain (LBA-UC) and the third, the observed proportion of time spent using a computer. These measures were summarized for all individuals within each of 13 job categories within the study population and then applied to all individuals within that category [16
General health status was measured at both baseline and follow-up using the physical component summary of the Short-Form 12 (PCS) [21
], which theoretically ranges from 0 to 100 (higher scores reflecting better functional status). Body region-specific health status instruments assessing disability and pain were administered to cases and matched referents for each of four injury types [back, upper extremity (UE), lower extremity (LE) and neck]. Low back symptoms and pain were assessed using the Roland–Morris Scale [22
], which ranges from 0 to 24 (higher scores denote worse functional status). UE symptoms were measured using the 11-item Quick Disability of the Arm, Shoulder and Hand (DASH) instrument, which ranges from 0 to 100 (higher scores represent more severe and disabling symptoms) [23
]. The severity of LE symptoms was assessed using a shortened version of the Western Ontario and McMaster Universities (WOMAC) osteoarthritis index [24
], which ranges from 0 to 100 (higher scores indicate better functional status).
Two work status measures (work effectiveness and lost workdays) were evaluated. Work effectiveness was assessed using a self-reported work effectiveness score, ranging from 0 to 100% (0% corresponding to inability to work at all and 100% indicating greatest effectiveness). Lost workdays for any cause in the 4 weeks preceding the interview were also elicited. We also ascertained whether subjects were no longer working at their original job site or at any job.
To assess whether injured cases had regained functioning by 2 years relative to all referents, we compared scores for cases versus referents at baseline and at follow-up for all measures assessed at both time points (PCS, work effectiveness, lost workdays and injury-specific disability). Due to the non-normal distribution of these results, we used the Wilcoxon rank-sum test to test differences by injury status ( and ). Because statistical differences between continuous scales may be of marginal clinical relevance, we created a binary measure for each of these outcomes dichotomizing between poor and relatively better health or work status. For most outcomes, the threshold for poor functioning was determined based on the quartile distribution for each score among referents. For measures in which higher scores reflect better status (PCS and WOMAC), poor status was based on the lowest quartile; for measures in which higher scores reflected worse functioning (Roland–Morris and DASH), the highest quartile defined poor status. The threshold for self-rated work effectiveness was set to 90%, consistent with our previous dichotomization of this measure [15
]. For lost workdays, we defined poor status as two or more lost workdays in the past 4 weeks, a cut point approximating the 37th percentile.
Figure 1. Comparison of case and referent general health and work status distributions at baseline and follow-up. *Wilcoxon P < 0.05; **Wilcoxon P < 0.01; ***Wilcoxon P < 0.001. DASH, Disability of the Arm, Shoulder and Hand; WOMAC, Western (more ...)
Figure 2. Comparison of case and referent injury-specific health and work status distributions at baseline and follow-up. *Wilcoxon P < 0.05; **Wilcoxon P < 0.01; ***Wilcoxon P < 0.001. DASH, Disability of the Arm, Shoulder and Hand; WOMAC, (more ...)
Univariate logistic regression tested whether injury status was associated with poor functional status at baseline or at follow-up as well as the association between injury and changing to a job at another location (whether or not this involved job duty changes) or complete work cessation. After a screening step based on univariate analyses employing a statistical significance cut-off of P <0.05 for other baseline factors of interest (demographics, occupational category, clinical characteristics, job strain defined by the JCQ, ERI and ergonomics measures), we used multivariable logistic regression to ascertain (i) the degree of association between injury status and poor functional status, including adjustment for baseline cofactors of interest, and (ii) whether other baseline factors of interest were predictive of poor functional status at follow-up taking injury status into account. Age, sex and race–ethnicity were retained in the models regardless of statistical significance at the univariate screening step due to their known influence on health and functioning.
Because of collinearity between the education and income variables, we combined them for the regression analysis by adding one point each for higher levels of each and grouping that sum into quintiles, a measure of SES we had employed in the baseline analysis [15
]. Similarly, we combined data from the collinear JCQ and ERI measures by reducing each into a binary above-median or below-median dichotomous variable, creating four mutually exclusive indicator variables: above the median for both (worst quadrant), above the median for one but not the other (middle quadrants) and below the median for both (the default referent category).
We also investigated whether lowest quintile of SES or high combined job strain/effort–reward imbalance, which we hypothesized a priori could be effect modifiers for injury status, should be included in our final models. This was accomplished by rerunning the key multivariate models including each interaction term separately. Interactions that were statistically significant at P <0.10 and did not include sparse cells (sample n < 10) were considered for inclusion.
We imputed missing data for individuals without any follow-up information (n = 70) and for subjects missing one or more key dependent or independent variables on either the baseline or follow-up assessments (n = 95). This imputation was conducted using multiple imputation procedures (SAS version 9.1.3 PROC MI and PROC MIANALYZE). Specific variables imputed were family income at baseline, education at baseline and all outcome measures.