Study population
In the present study, data from four waves (2010, 2012, 2014 and 2016) of the Swedish Longitudinal Occupational Survey of Health (SLOSH) were used. Starting in 2006, Statistics Sweden has collected data every second year by means of questionnaires, following the same respondents. In 2008 and 2014 more respondents were included. SLOSH is a largely representative cohort of the Swedish working population with the aim of examining work environment and health. For a more thorough description, see [35].
Inclusion criteria were that the respondent worked as a health or social care worker during at least two data collections. More waves were included if, on these occasions, the individual was also working in these occupations. Occupation was self-reported and then by Statistics Sweden coded into register data [36]. Thereafter, we listed occupational codes that were relevant with respect to the intended study population, that is, employees who worked as health or social care workers (e.g., a nurse, physician, care assistant, social worker, psychologist, therapist).
The final sample consisted of 2,333 individuals, of which 87.6% were women. The mean age was 49.9 (9.5) and 54.1% had a university degree. Of these, 78.4% were married/cohabitant and 51.8% had children living at home. Only associations were targeted where the individual kept his/her manager between two subsequent waves.
Variables
All variables were measured at all four time points.
Managerial leadership
Managerial leadership was assessed with nine questions [22] where the participants were asked to evaluate the behaviour of the closest manager, such as whether he/she gave enough information, communicated clear goals and expectations, gave sufficient power and was good at pushing through and carrying out changes, but also whether the manager was supportive, encouraged participation, gave positive feedback and cared about the employees’ professional development. The managerial leadership scale is the leadership climate dimension of the validated Stress profile instrument [23] that was originally developed for assessing psychosocial stressors in the workplace and in private life. The instrument is based on stress research and established theories, and has been validated and tested in several workplaces in Sweden [23]. The Stress profile has been influential on, for example, the development of the well-used Copenhagen Psychosocial Questionnaire (COPSOQ) [37]. Since one question was missing in one SLOSH wave, this particular item (concerning receiving criticism from the leader if something that was not good was done) was excluded from the analyses. The response alternatives ranged from 1: “often” to 4: “never”. The Cronbach’s alpha for the remaining nine questions ranged from 0.90 to 0.91 over the SLOSH waves. Table S1 in the supplementary material presents descriptives for the nine items at the first wave. For the purpose of the moderator analyses the participants were classified as having a poor manager or having a good manager (using the median value as the cut-off).
Psychosocial work stressors
The measures of psychological demands and decision authority were both obtained from the Swedish version of the Demand-Control-Support-Questionnaire (DCSQ) [38, 39]. Psychological demands were assessed with five items (working fast, working intensively, too much effort, (not) enough time and conflicting demands), and decision authority was measured with two items (what to do at work and how to do the work). The response alternatives ranged from 1: “often” to 4: “never/almost never”. For psychological demands Cronbach’s alpha ranged from 0.74 to 0.78 and for decision authority from 0.74 to 0.77. Items were reversed so that higher values represented more work stressors.
Workplace violence
Workplace violence was measured with a single question: “Were you exposed to violence or the threat of violence in your work during the last six (twelve) months?” In 2010 the time period that was referred to in the question was the last twelve months, whereas thereafter (2012, 2014 and 2016) the last six months were referred to. The answer alternatives were dichotomised to “no” or “yes”, where yes indicated at least once during the period in question. An increase in the number of workers exposed to violence between the first and the second measurement points was observed (20.9% in 2010, 25.9% in 2012, 24.3% in 2014, 24.2% in 2016).
Register-based sickness absence
In Sweden, if not returning to work after seven days of sickness absence, the employee needs a doctor’s certificate. After seven additional days, the employee can receive sickness benefit (preventive sickness benefit, rehabilitation allowance and occupational injury allowance) from the Swedish Social Insurance Agency. In the present study, register-based sickness absence (net days) refers to the number of days per year of receiving sickness benefit from the agency. One register-based sickness absence day could correspond to either a full day (100%) of sickness benefit, two days of 50% or four days of 25%. We dichotomised the variable into (1) no register-based sickness absence, and (2) one or more register-based sickness absence days.
Covariates
Information on gender, age and educational level were obtained from register data. Information on civil status (married/cohabiting or not) and parental status (children living at home or not) was self-reported.
Analytical strategy
Hypotheses were tested using autoregressive cross-lagged models within a multilevel structural equation modelling (MSEM) framework with both observed and latent variables. These models make it possible to address the reciprocal temporal relationships among exposure variable, mediators/moderators and outcome, and also account for the multiple levels of the data i.e. multiple measurement points (level 1-within-person level) nested within individuals (level 2-between level). More precisely we used a two-level structural equation model (SEM) that allows partitioning of between- and within-person effects to account for two inherent types of heterogeneity, within-person across time and between-person [40,41,42].
Before testing the hypotheses, we tested the measurement invariance in the latent variables of psychological demands, decision authority and leadership, which all showed good fit statistics and no indication of measurement variance over time.
In a first step, we examined the bivariate multilevel structural cross-lagged relationships between managerial leadership (exposure variable) and sickness absence (outcome). Managerial leadership was measured at the first time point (t-1, years 2010, 2012, and 2014) and sickness absence at the subsequent time point (t, years 2012, 2014 or 2016). The cross-lagged paths estimated the effect of one variable on the other with a two-year time lag. Each path in the models was adjusted for age, gender, education, civil status, and children living at home. Indicators of the latent variable leadership were allowed to correlate between waves. The same bivariate models were used in order to test associations between managerial leadership and psychological demands, between decision authority and workplace violence (putative mediators) as well as between putative mediators, as described above and sickness absence (outcome). The models were adjusted for the same set of covariates as above. If there were significant paths between the predictor and the mediator and between the mediator and the outcome, a mediation model under the MSEM framework could be fitted. The second step was to apply such a model to our data. A longitudinal mediation model within an MSEM framework, in which leadership was measured at t-2 (in the years 2010 or 2012), psychological demands, decision authority and violence at t-1 (in the years 2012 or 2014) and sickness absence at t (in the years 2014 or 2016), was fitted. The model was adjusted for the same set of covariates as in the bivariate models. Such a model makes it possible to estimate the direct effect (the part of the exposure effect which was not mediated through psychological demands, decision authority, or violence) as well as the indirect effect (the part of the exposure effect which was mediated through psychological demands, decision authority or violence) between leadership and sickness absence.
In a third step, in order to examine if managerial leadership moderates the association between psychosocial work stressors (psychological demands, decision authority, or workplace violence) and sickness absence, we utilised bivariate models described in step 1 in a stratified model by managerial leadership. To determine whether the cross-lagged paths differed between good and bad leadership we conducted multiple-group analyses testing differences in each hypothesised and reverse association separately. We created two groups based on leadership (using the median value as a cut-off), then compared a model in which the paths were allowed to vary freely with a model in which the paths were constrained to be equal between good and bad leadership. The likelihood ratio test was used for comparing restricted and non-restricted models. A significant change in chi-square (df) between the non-restricted model and the restricted one indicates a poorer fit for the restricted model. The multilevel SEM models were built in MPLUS 7. All variables were treated as time-varying. Standardised estimates were reported for the final models. The fit statistics chi-square (df), the root mean square error of approximation (RMSEA), and the standardised root mean square residual (SRMR) were considered. Model fit is assumed to be acceptable when RMSEA ≤ 0.08 and SRMR ≤ 0.08 [43].