Aims and objective: The objective of this study, was to assess a) the amount of variation between work-groups in future sickness absence rates explained by work-group level psychosocial work characteristics in comparison to other health related information expressed as explained variance of future sickness absence rates by concurrently available information and b) to estimate the excess sickness absence rate observed in work-groups with impaired psychosocial work characteristics as compared to the most favorable quartile of work-groups, operationalized using the population attributable fraction concept. The population attributable risk (by some authors referred to as etiological fraction) offers an estimate for the proportion of a given risk (i.e. death, disease, sickness absence) that might theoretically be removed if all exposed subjects had the lowest level of exposure (i.e. work groups with favorable psychosocial work environment) [11,19-21].
Conceptual model and operationalization: The underlying conceptual model derived from the Whitehall II studies posits that multiple, albeit often correlated compositional factors (i.e. age, manual vs. non-manual work), health behaviors, medical outcomes and psychosocial risk factors or resources are prospectively related to sickness absence rates [14,15]. Rather than individual sickness absence data from a potentially biased sample, the most relevant work-group level outcome is the objectively recorded sickness absence rate including every employee. Likewise, the ideal exposure measure would be an objective external measurement of psychosocial work characteristics. Yet, constructs such as perceived appreciation are almost impossible to measure externally. The next best option to obtain a measure is to ask every employee and to average these ratings, like school grades from a class. Unfortunately, participation rates vary between work-groups, introducing unknown selection or recruitment biases. An option to estimate the size of this bias is to randomly re-sample from the existing participants and to observe the obtained variation. Then, appropriate multivariable regression models provide a useful estimation of the true underlying relationship .
Study setting and study population: The data for this prospective cohort study  were at seven work sites operated by a large German automotive manufacturer over a three-year period. As part of the program roll-out in 2014 and to maximize the potential generalizability of study results, the company´s human resource management team at each site selected work groups involved in production, engineering, development and administration and representing the range of work group sickness absence rates observed at the time of planning the program (see discussion section for consideration of possible selection bias).
All permanent employees aged 18 to 65 in the selected work-groups (size ranging from 25 to 1480) were eligible. For this study we included work-groups where at least 10 participants had specifically consented to scientific evaluation of their anonymized data (average consent rate 67%). A work-group was defined as the organizational unit in the organigram of the company that has one clearly identified superior and clearly identified organizational purpose. We excluded work-groups affected by a company-wide reorganization in 2015. These criteria resulted in 3992 participants from 29 work-groups from the original 5444 employees partaking in the comprehensive health evaluation in 2014. To address the issue of unequal work-group sizes in our analysis, we used a repeated bootstrap random sampling procedure as explained in detail below, which samples a larger proportion of participants from small work-groups (i.e. 70%) as compared to large work-groups (i.e. 5%)).
Baseline measurements: Comprehensive evaluations at baseline, conducted between June 23rd and December 15th 2014, consisted of self-completed health questionnaires, detailed medical examinations performed by members of the company’s occupational health services and an assessment of psychosocial work characteristics. Participants recorded age, gender and main type of work (manual vs. non-manual). All of these variables or measures explained in detail below were considered as candidate predictors potentially related to future sickness absence.
Perceived health and health-related behaviors were assessed using the SF-12 for health-related quality of life (mental and physical summary scale), the Copenhagen burnout inventory , daytime sleepiness , the IPAQ  for physical activity, the AUDIT-C  for alcohol consumption, and eight additional items assessing nutritional habits and current smoking status. The questionnaire further assessed self-reported regular intake of medications and physician approved medical conditions.
The detailed medical examination included body-mass-index, waist circumference, blood pressure, lipid levels, glycosylated hemoglobin (HbA1c), C-reactive protein, criteria for the metabolic syndrome based on the new IDF definition . As no fasting glucose was obtained, we used a HbA1c level exceeding 5.7% (prediabetes) instead of the glucose > 100 mg/dl criterion. Cardiovascular risk was estimated using the Framingham-Algorithm . To adjust for age effects, we calculated a “Framingham relative risk index” as the absolute predicted risk for an individual divided by the absolute predicted risk for a non-smoking, non-diabetic person of the same age and gender with other variables at the upper boundary of the most favorable quartile for all participants.
Psychosocial characteristics of the work environment were evaluated using a self-completed questionnaire based on the Copenhagen Psychosocial Questionnaire (COPSOQ V1, German version, that includes scales on quality of leadership, cognitive stress perception, quantitative and emotional demands, influence at work, predictability, job-satisfaction, possibilities for development, meaning of work, social support from colleagues and work life conflict . Psychosocial work characteristics can either act as resource or as adverse factor. For example, recognition or supportive leadership are consistently viewed as a resource. For the context of this investigation, we defined either stressors (e.g. high demands) or lack of resources (e.g. low resources such as lack of supportive leadership) as potentially adverse work characteristics.
Work ability was assessed using the 22-item Workability Index (range: 7-49)  which also contains one item on self-reported sickness absence during the past 12 months. The latter was used to assess possible selection bias in sampling as described below. Questionnaires were either completed online or by use of a paper-pencil version. For each candidate variable where a work-group environmental exposure was conceptually conceivable (i.e. quality of leadership) we calculated the work-group average as the mean of the randomly selected participants (see below) and the individual perception as the difference of the work-group mean and the individual value.
With the exception of the SF-12, most of the used scales use some arbitrary enumeration directly derived from converting Likert-scale coding (e.g. 0-5) to total scores. To enhance comparability of scales using such different metrics, we report data in a transformed fashion that was employed to facilitate communication with management resembling the grading experienced by most managers during the final years in German high schools. There the best grade is 15 points and the population averages around 10 points and giving rise to a standard deviation of 2.2 points. This has been the standard reporting system in the Mannheim Industrial Cohort Study, with expected German working population average scores of 10.
The work ability index is widely used self-administered questionnaire capturing seven dimensions of health including present and expected future work-ability. Dimension scores add to a total ranging from 7 (unable to work) to 49 (excellent work ability). The work ability index and its short form predict future long-term sickness absence (area under the receiver operating characteristic curve = 0.82 for manual workers and 0.79 for non-manual workers .
Outcomes: We obtained the work-group sickness absence from the official company records for 2014 and 2015. The averaged work-group sickness absence rate was determined from the proportion of total workdays missed during 2015 due to sickness for all members of each work-group (both participating individuals and non-participants). Total workdays in a year was assumed to be 220 after excluding holidays and the average number of days of vacation for employees throughout the company. For example, a sickness absence rate of 5% for a work-group of 100 employees implies that 1.100 of the 22.000 possible workdays in 2015 were lost due to sickness absence.
In Germany, any sickness absence longer than three days requires a physician´s medical certificate with an indication provided to the employer of the duration of the certified sickness leave. For up to six weeks of cumulative sickness absence per 12 months, the employer has to continue paying the salary. Thereafter, employees receive renumeration from the statutory health insurance. Thus, company records combining short-term non-medically certified absence and >3 day medically certified absence are the most accurate source available, above any social security registry data. Due to the European General Data Protection Regulation no further detail on individual sickness absences was available from company records. Due to these restraints, the company required a minimum number of employees per work-group. Thus, in highly fragmented engineering departments with small work-groups, the study work-group operationalization did not represent the lowest organizational level.
As explicated above, the outcome was measured sickness absence rates for all employees in 2015 at the work-group level, while candidate predictors were measured during the baseline assessment in 2014 at the individual level . To address the possibility of participation selection bias, we compared the self-reported sickness absences from the baseline with the company records for sickness absence averaged for all employees from the included work groups .
Aggregating all data to the work group level would have obscured important information such as within work group heterogeneity (i.e. work groups with a differing distributions of characteristics).Thus, to account for these multilevel analytical structure issues, we employed a bootstrap procedure as recommended for complex data  in which individuals from each work-group are randomly selected to form samples for analysis. We used a sampling rate of 0.7 for small work-groups and 0.05 for the largest work-group. Random sampling was repeated for 200 cycles to allow the generation of empirical confidence intervals. This method resulted in an average total of 870 participants for the analytical sample (95% range 821 – 918). In all analyses we employed generalized linear models with a binominal logit link function. This acknowledge that the observed outcome rate (i.e. 5%) arises from multiple binary events (an employee being absent or present). We also conducted least-square linear regression analyses that essentially yielded similar results (data not shown).
First, we employed univariate analyses to establish the strength of the independent association of each candidate factor from psychosocial work characteristics, medical findings, self-reported health, health behaviors, compositional and contextual factors identified at baseline in 2014 with future sickness absence. For each candidate variable we explored linear and quadratic terms to account for possible non-linearity of the relationship. To facilitate interpretation of the results, we show univariate associations as effect sizes, expressed as explained variance of the prediction vs. observed data (R-square) .
In the multivariable analyses we explored combinations of candidate predictor variables for psychosocial work characteristics, medical information and health behavior, subjective health and compositional characteristics (age, gender, main type of work). For each model, the predicted sickness absence rate was compared to the observed absence rate using least square regression. Predictive accuracy was expressed as variance explained (adjusted R-square).
We expected substantial correlation across the medical and behavioral variables and amongst the psychosocial work characteristics. Therefore, we used backward elimination strategies to arrive at parsimonious models. Variables were eliminated until further reduction lead to an average decline in adjusted R-square over 200 bootstrap sampling cycles by more than 0.03. All analyses were repeated adjusting for age, gender and main type of work.
As the last analytical step, we estimated the excess sickness absence over the expected “unavoidable sickness absence” attributed to the set of predictors by the respective parsimonious models. For each of the models we determined the sickness absence rate for the quartile of work-groups with the most favorable prediction scores. Following the method originally suggested by Miettinen we calculated the population attributable risk, also known as etiological fraction (PAF) [11,21]. The PAF calculates the excess work-group sickness absence that would theoretically be removed if all work-groups had favorable psychosocial work-group characteristics. Here, we present a slightly different conceptualization, the excess absence expressed as proportion of unavoidable baseline sickness.
All analyses were carried out in STATA (StataCorp. 2017. Stata Statistical Software: Release 15. College Station, TX: StataCorp LLC.).