The present study is cross-sectional and based on register data from AFA Insurance in Stockholm, Sweden. The register consists of questionnaire data on health and the psychosocial work environment collected among employees from the public and private sector in Sweden, including employees of private businesses, municipalities, and county councils in different parts of Sweden. It was administered in connection with the employees participating in a project (from 2002 to 2007) aimed at improving workers’ health. This initiative was not primarily a research project but it was based on valid measurement instruments and used methods developed and evaluated in former research projects (Bergstrom et al.
2008; Vingard et al.
2005). Because different organizations were involved in the study, the response rate varied. The response rates were between 65 and 94% (mean 78%) and are described in an earlier report on this occupational health initiative (Järvholm et al.
2008). Furthermore, in some organizations the questionnaire was modified and tailored to different occupational groups by removing or adding items. The survey was administered by regular mail to the employees with two reminders.
Participants
To be included in the analyses the following criteria had to be met: (1) having been at the current workplace for at least 1 year, (2) between 18 and 65 years of age, and (3) have a complete set of data on presenteeism and sickness absenteeism available. The database originally consisted of 193,640 employees but after database clearing for research purposes and applying the mentioned criteria, 46,069 subjects were excluded (72% women, 28% men), leaving a population of 147,571 individuals. There was no information on the actual number of employees that were invited to respond to the survey, the only information available was the response rates given above. Because the main analyses used listwise deletion, 17,410 subjects with missing data for any of the variables included in the analyses were also excluded, giving a final study population of 130,161 employees, of which 73% were women and 27% men. The study population is described in Table
1. The majority was employed for municipalities, followed by county councils. The most commonly reported workplaces were schools (municipalities), hospitals (county councils), and post offices (state-owned). This information should be treated tentatively, however, since information about subject’s workplaces was missing for 45% of the population. As can be seen in the table, 12% of the study group reported being sickness present on more than 5 occasions over the previous year, and 8% reported more than 25 days on sick leave during this period. Among the 17,410 excluded employees, 63% were woman and 37% men; the mean age of this group was 47.0 years (sd 10.8) and the extent of presenteeism and absenteeism was almost identical to the study group.
Table 1
Descriptive data for the study population
Gender, n (%) |
Women | 95,640 (73) |
Men | 34,521 (27) |
Age, m (SD) | 46.8 (10.6) |
Education, n (%) |
Compulsory school | 13,078 (18) |
High school | 31,781 (44) |
University | 27,402 (38) |
Employer, n% |
Private or state-owned company | 22,616 (17) |
Municipality | 73,602 (57) |
County council | 33,943 (26) |
Sickness presenteeism during the previous year, n% |
None | 43,140 (33) |
1 time | 25,241 (19) |
2–5 times | 46,546 (36) |
> 5 times | 15,234 (12) |
Sickness absenteeism during the previous year, n% |
None | 40,833 (31) |
1–7 days | 58,022 (45) |
8–24 days | 20,735 (16) |
25–365 | 10,571 (8) |
Work-to-family conflict1, n% |
Very seldom or never | 45,116 (35) |
Rather seldom | 31,729 (24) |
Sometimes | 38,111 (29) |
Quite often | 12,398 (10) |
Very often or always | 2807 (2) |
Job demands 1–51, mean (sd) | 2.96 (.70) |
Job control 1–51, mean (sd) | 2.88 (.80) |
Job support 1–51, mean (sd) | 3.80 (.80) |
Role conflict 1–51, mean (sd) | 2.37 (.84) |
WTFC 1–51, mean (sd) | 2.20 (1.08) |
General health, 0–1002, mean (sd) | 86.94 (14.01) |
Job motivation, 1–53, mean (sd) | 4.23 (.75) |
Measurement instruments
The psychosocial workplace factors were assessed using the General Nordic Questionnaire for Psychological and Social Factors at Work (QPSNordic) if no other means of assessment is given (Dallner et al.
2000). The variables included in the questionnaire were primarily chosen based on the theoretical assumptions described in the introduction and on the results from the meta-analysis of Miraglia and Johns (
2016). Furthermore, in line with this meta-analysis, global variables were preferred when possible, that is, different aspects of a construct were merged into one single index (see below). All QPS-scales have five scale steps from 1 to 5.
For the background factors of gender, age, education, and employer, we used single items (Bergstrom et al.
2008).
A global job control index (8 items, QPSNordic) was constructed based on all but one of the items in the control over decisions index (4 items; e.g., “Can you influence the amount of work assigned to you?”) and the control over work pace index (4 items; e.g., “Can you set your own work pace?”). The item “Can you decide when to be in contact with clients?” from the control over decisions index was omitted, because many of the respondents did not work with clients. Cronbach’s alpha was 0.82 for this index.
A global index of job demands (7 items, QPSNordic) was calculated based on 4 items measuring quantitative job demands (e.g., “Do you have to work overtime?”) and 3 items assessing decision demands (e.g., “Does your work require quick decisions?”). Cronbach’s alpha was 0.80.
Furthermore, a global index of job support (5 items, QPSNordic) at work was constructed based on indices on support from colleagues (2 items; e.g., “If needed, can you get support and help with your work from your coworkers?”) and on support from supervisor (3 items; e.g., “If needed, is your immediate superior willing to listen to your work-related problems?”). Cronbach’s alpha was low (0.61) and considered to be acceptable.
Role conflicts (QPSNordic) were assessed using 3 items (e.g., “Do you have to do things that you feel should be done differently?). Cronbach’s alpha was 0.72.
Work-to-family conflict (QPSNordic) was assessed with the item, “Do the demands of your work interfere with your home and family life?”.
Work motivation was measured with a modified work motivation scale developed by Björklund (Björklund et al.
2013). The questions included were the following: “Do you feel stimulated by your work tasks?”, “Are you motivated to work?”, “How often do you feel a strong will to work?” and “Would you spend less time at work if possible?” The response format ranged from “1 = never” to “5 = always”. Cronbach’s alpha was 0.78.
Sickness presenteeism was measured using the following item: “Has it happened over the previous 12 months that you have gone to work despite feeling that you should have taken sick leave due to your state of health?” The response options were “1 = no, never,” “2 = yes, once,” “3 = yes, 2–5 times,” and “4 = yes, more than five times” (Aronsson et al.
2000).
Sickness absenteeism was measured using a question from the Work Ability Index (Ilmarinen
2007). The question used was “How many days in total have you been away from work due to your own illness (sick leave, health care, treatment or examination)?” The response options were “no days,” “1–7 days,” “8–24 days,” “25–99 days,” and “100–365 days.” The categories “25–99 days” and “100–365 days” were merged for our analyses.
General health. Health-related quality of life (
HRQoL) was assessed using the Swedish version of the EuroQol EQ-5D-3L (Bjork and Norinder
1999; Brooks
1996). The EQ-5D is a generic HRQoL questionnaire based on five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression. These dimensions can be used to generate individual health profiles that represent health states. In this study, we applied utility weights of health states from a general population using the Danish time-trade-off values to convert the health profiles to values between 0 and 1 (Wittrup-Jensen et al.
2009). A value of 0.00 indicates the worst possible health state while a value of 1.00 indicates the best.
Statistical analyses
The variables were first checked for their statistical distribution and then correlational analyses were conducted using Pearson’s r. Secondly, we estimated the different parameters of the path model with manifest indicators, and their direct and indirect effects on sickness absence and sickness presence. Structural equation models (SEM) and the data program AMOS 25 was used. To assess model fit, we used different indicators. These were the root mean square residual (RMR), which is the square root of the discrepancy between the sample covariance matrix and the implied covariance matrix. Values close to 0 represent a good fit and, as a rule of thumb, values < 0.08 are deemed to be acceptable (Hooper et al.
2008). We also used the goodness of fit index (GFI). The GFI can be compared to R-squared, and ranges from 0 to 1 and is considered satisfactory when > 0.90. Furthermore, we used the adjusted goodness of fit (AGFI) acceptable value > 0.90, the comparative fit index (CFI) acceptable value > 0.95 and the Root mean square error of approximation (RMSEA), acceptable value < 0.10. We also calculated a Bayesian posterior predictive p-value, which should be near 0.5 for a correct model, with values toward the extremes of 0 or 1 indicating that a model is not plausible.
Since some of our variables were categorical and/or skewed, we decided to use the asymptotic distribution-free (ADF) method (Jones and Waller
2015). Moreover, to handle skewness resulting from the indirect effects (mediating effects) produced by the direct effects and their distributions, we used 5000 bootstrap replicates to estimate bias-corrected confidence intervals and p-values (Lockwood CM and DP.
1998; Valente et al.
2016). Finally, to check our estimates, we ran Bayesian estimation via the Markov chain Monte Carlo (MCMC) algorithm with a uniform prior (Byrne
2010; Kaplan and Depaoli
2012). The estimates of the two methods were close and in general identical when rounded to the second or third decimal place. In the sequel, we report the estimates from the ADF bootstrap procedure if not stated differently. Listwise deletion of data was used, since Amos cannot accept missing data when using any estimation criterion, such as ADF, except for when using maximum likelihood. In the results section, we will use the concept of “effect” in a statistical sense; that is, we do not imply causal effects between variables.