Perceived stress as a risk factor of unemployment: a register-based cohort study

We conducted a cohort study with 98 weeks of follow-up (from 22nd March 2010 to 5th February 2012). Survey data from the North Denmark Health Profile 2010 [16] was individually linked with information from national administrative registers using the unique personal identification number of all Danish residents [17]. The North Denmark Health Profile 2010 included a Danish version of the 10-item Cohen’s Perceived Stress Scale (PSS) to assess the level of perceived stress [11]. Information on employment status was provided by the Danish Register for Evaluation of Marginalization (DREAM) [18]. The register contains weekly information on public transfer payments to all Danish citizens including unemployment benefits, sickness benefits, retirement pensions, state educational grants and maternity leaves [18]. The Population’s Education Register provided information on current and highest completed education authorised by the Danish Ministry of Education [19]. Self-reported information on education level and employment status [16] was used if information was missing in the registers. The Income Statistics Register, which is provided by the Danish Tax authorities, contains information on taxable income of Danish residents [20]. The Danish Civil Registration System provided information on age, gender and emigration [17]. The Danish Register of Causes of Death [21] was used to identify participants who died during follow-up.

The study population was drawn from respondents of the North Denmark Health Profile 2010, which was sent to a random sample of 35,700 persons aged 16 years and above. The sample was stratified based on the 11 municipalities of the northern Region of Jutland, Denmark (570,000 inhabitants). 23,392 persons responded (response rate 65.5%) of which 21,842 persons answered all PSS items. Generally, the response rate was lowest among men. Furthermore, especially 16–34-year-old men and women did not respond [16]. We included respondents, who were 18–64 years old (N = 16,138), to represent a working population. At the time of the study, public retirement was possible at the age of 65. Only respondents, who were working at baseline and the preceding 3 months, were included to minimize the risk of including recently employed people, who might be on probation or very short-term temporary contracts. Employment status was assessed through DREAM (N = 16,024). Employed respondents were identified as citizens with no public transfer payments from 5th November 2009 to 22nd March 2010. Self-reported employment status [16] was used if people were not registered in DREAM (N = 114). Non-employed respondents (N = 6795) and respondents with missing information (N = 8) were excluded. This resulted in a study population of 9335 persons (Fig. 1).

The level of perceived stress was assessed through PSS, which measures the extent of unpredictability, lack of control and overload in the respondent’s life during the last month [11, 16]. The answers of each of the 10 items were scored from 0 to 4 ranging the five possible answers “never” to “very often”. The scores were added up ranging from 0 to 40 with a high score indicating a high level of perceived stress [11]. As the PSS do not have score cut-offs [22], the sum of the PSS-score was divided into quintiles for interpretation purpose.

Identification of the first episode of unemployment during follow-up was based on the weekly transfer payments of two types of unemployment benefits; voluntary unemployment insurance benefits (DREAM-codes 111–113, 151–152, 211–299) and social security benefits (DREAM-codes 130–149) registered in DREAM (see groupings of transfer payments in Additional file 1: Table S1). These codes represent unemployed who are assessed ready to undertake work by the social office. In Denmark, receiving unemployment insurance benefits requires membership of an unemployment fund for at least 1 year prior to unemployment and a specific amount of education or work experience [23]. Those who do not meet these requirements are entitled to receive social security benefits although specific financial requirements have to be met [24].

First episode of sickness absence during follow-up was identified as receipt of sickness benefit registered in DREAM [18]. In DREAM, registration of sickness absence is not uniform. Only sickness absence > 21 days was registered unless the employee suffered from a certified chronic disease or the company was insured by which the employee would receive sickness benefit from the first day of sickness absence [25].

Age was grouped into 18–25, 26–54 and 55–64 years of age. Household income was calculated as the average of 2007, 2008 and 2009. If the respondent lived with a partner, the household income was divided by 1.5 to allow for sharing of resources [26]. Household income (Euro) was divided into quartiles; 1: < 44,763, 2: 44,763–55,978, 3: 55,979–67,524 and 4: > 67,524. Highest achieved education in 2009 was divided into three groups based on the International Standard Classification of Education (ISCED) [27]; basic (ISCED level 0–3; early childhood education, primary education, lower secondary education and general upper secondary education), vocational (ISCED level 3; vocational upper secondary education) and higher education (ISCED level 5–8; bachelor’s, master’s and doctor or equivalent level).

Information on lifestyle factors and self-rated health was self-reported [16]. Smoking status was categorized into three groups; never smoker, former smoker and current smoker. Body Mass Index (BMI) was calculated from information on height and weight [16] and divided into four groups based on WHO standards; underweight (BMI < 18.5), normal weight (18.5 ≤ BMI < 25), overweight (25 ≤ BMI < 30) and obesity (BMI ≥ 30). Alcohol consumption was based on information on units per day during the week [16]. One unit equals 12 g of pure alcohol. Alcohol consumption was dichotomized into low (< 180 g/week for women and < 264 g/week for men) and high (> 168 g/week for women and > 252 g/week for men) level based on recommendations on alcohol consumption from the Danish Health Authorities [28]. Self-rated health was assessed through the question: “In general how do you assess your own health?” [16] and dichotomized into good (“Excellent”/“Very good”/“Good”) and poor (“Moderate”/“Poor”).

Descriptive statistics of baseline characteristics and outcome in the five stress quintiles were presented including outcome of non-respondents of the PSS. Cumulative incidence proportion of unemployment by perceived stress was estimated using the Aalen-Johansen estimator. Cox proportional hazards model was used to investigate the association between perceived stress quintiles at baseline and risk of unemployment during follow-up with lowest stress quintile as reference. To allow for the stratified sample design, the R-package ‘survey’ was used [29]. Unadjusted (model 1) and multivariate analyses including possible confounders (age, gender, SES) (model 2: fully adjusted) were performed. All analyses were further performed in subgroups defined by SES adjusting for gender and age.

To test the proportional hazards assumption, Martingale residuals were calculated and plotted. Stratified Cox models were used if the assumption was violated. Statistical interactions between perceived stress and covariates were tested and none were found. People who retired (N = 456), received benefits due to reduced ability to work (N = 21), emigrated (N = 25) or died (N = 15) during follow-up were censored. No loss to follow-up occurred due to the register-based design. Adjusting for lifestyle factors and self-rated health was performed as sensitivity analysis. Furthermore, the association between perceived stress and unemployment with and without preceding sickness absence as outcomes was performed as sensitivity analysis. A sample restricted to those who worked 6 months preceding baseline was also performed as a sensitivity analysis.

Results are presented as hazard ratios (HR) with corresponding 95% confidence intervals (CI). Data management was performed using SAS software, version 9.4 (SAS Institute Inc., Cary, North Carolina, USA). Data analysis was performed using R statistical software package, version 3.3.3 [30].

A high level of perceived stress increased the risk of unemployment with preceding sickness absence (adjusted HR: 3.26 95% CI: 1.35;7.84) and unemployment without preceding sickness absence (adjusted HR: 1.51, 95% CI: 1.16;1.97).

Adjusting for lifestyle factors and self-rated health, showed that the risk of unemployment by perceived stress was similar to the results not adjusted for this (HR: 1.65 95% CI: 1.25;2.19) (N = 8046) (Additional file 2: Table S2).

The association between perceived stress and risk of unemployment in the restricted study population was similar for all respondents (adjusted HR: 1.65 95% CI: 1.27;2.14) (N = 8877) and SES-groups (Additional file 3: Fig. S1).

The longitudinal setup separated perceived stress and unemployment in time. We only included 18–64-year-olds who were working at baseline and the preceding 3 months to represent a working population. The register-based design enabled complete follow-up data by which selection during follow-up was prevented. However, the incidence of unemployment was higher among non-respondents compared to respondents (11.8% vs. 7.6%) of the North Denmark Health Profile 2010. Non-respondents were likely to be in the high stress group and selection bias due to non-response might therefore occur. Particularly young men did not respond and had a lower level of perceived stress compared to women [16]. Our results were similar for men and women (Additional file 4: Table S3), but due to the higher risk of unemployment among non-respondents, the results of our analysis might be overestimated.

The comprehensive use of registers limited the risk of information bias compared with studies only using self-reported data [8]. The Income Statistics and Population’s Education Register are of high quality [19, 20]. The DREAM register was feasible for historical register-based follow-up studies of exit from the labour market [18]. Those who did not receive any social transfer payments at baseline and the preceding 3 months were included to represent a working population. However, an unknown number of unemployed individuals might be categorized as working. The two Danish types of unemployment benefits have certain economic and education requirements [23, 24] and some unemployed individuals will not meet the eligibility criteria for neither unemployment insurance benefits nor social insurance benefits. Selection bias only occurs if the possible included unemployed differ concerning both perceived stress and risk of unemployment. Generally, unemployed people experience higher levels of perceived stress compared to those working [1, 2] and a different risk of unemployment. The possible inclusion of unemployed in the employed population might therefore increase the level of perceived stress of the study population at baseline. This would lead to bias towards null with an underestimation of the results. However, the positive predictive value of employment in DREAM has been shown to be more than 90% [18], meaning that 90% of those registered as employed are in fact holding a job. This indicates that the risk of reduced representativeness of the included study population will be of minor importance. The predictive value is lower concerning unemployment benefit, social assistance and sickness benefit [18]. The risk of non-differential misclassification of unemployment might thus be increased. However, the predictive value of positive registration in DREAM was compared to self-reported data and self-reported data on temporary payments like sickness benefits might be biased [18].

Information on stress was assessed with the previously validated PSS, which has been used extensively [41]. We conceptualized the effect of income through household income, which was based on the average from 3 years to get stable values, as income often change over time. The household income was divided by 1.5 when living with a partner to partially allow for sharing of resources when living more people in a household, although this did not allow for children living at home [26].

This study was based on data from the North Denmark Region in 2010. Around this time many manufacturing companies were hit by large cutbacks and company closures due to the financial crisis. In Denmark, large layoffs were preceded by notice to employees [42]. We only included people employed at baseline and 3 months prior to baseline to exclude recently employed. Despite this, notifications of layoffs could be more than 3 months in cases of cutbacks and company closures and cause perceived job insecurity. Hereby, this unmeasured confounder might upwardly bias the results. However, similar results were found when the study population was restricted to employed people at baseline and preceding 6 months. Furthermore, no statistically significant interactions were found between perceived stress and income or education level. The influence of the layoffs might thus be limited as mainly low income and education groups were affected.

The risk of unemployment might depend on the unemployment rate in specific occupations, but data on occupation was unavailable. The unemployment rate in the society could also influence the risk of unemployment [1]. However, in this study the societal unemployment rate was irrelevant as the participants were followed in the same period. Due to the economic situation in the North Denmark Region, a higher unemployment rate was evident especially among low SES-groups. The higher unemployment rate could decrease the generalizability to the rest of Denmark, other comparable countries and time periods. However, adjustment for SES might partly account for the higher unemployment rate and this did not change the results considerably. The underrepresentation of especially young men might though decrease the generalizability to this group. In contrast to previous studies [8, 32], we included participants, independently of their occupation and gender and covered a wide age range, which improved the generalizability.