Trajectories of sickness absence and disability pension days among 189,321 white-collar workers in the trade and retail industry; a 7-year longitudinal Swedish cohort study

We used data from three nation-wide Swedish administrative registers linked at individual level by use of the Personal Identity Number [26] (PIN, a unique ten-digit number assigned to all Swedish residents): (1) Longitudinal Integration Database for Health Insurance and Labour Market Studies (LISA by Swedish acronym) held by Statistics Sweden: to identify the source population and for information on age, sex, country of birth, type of living area, family situation, educational level, income, size of workplace, occupational code according to the Swedish Standard for Occupational Classification (SSYK by Swedish acronym, the Swedish version of the International Classification of Occupations, ISCO), sector, and branch of industry according to the Swedish Standard for Industrial Classification (SNI by Swedish acronym) in 2012–2016; (2) MicroData for Analysis of the Social Insurance database (MiDAS) held by the Social Insurance Agency: for information on SA spells > 14 days and DP (dates, extent (25, 50, 75, or 100% of full-time) for 2010–2016); (3) the Cause of Death Register held by the National Board of Health and Welfare: for information on date of death.

The study population was all those who were aged 18–67 years and registered as living in Sweden on both 31 December 2011 and 31 December 2012, had an occupational code according to SSYK that indicated a white-collar occupation, were employed at a private sector company in the trade and retail industry according to SNI, and during 2012 had income from work, parental benefits, and/or SA/DP that amounted to at least 7920 SEK (that is, 75% of the necessary income level to qualify for SA benefits from the Social Insurance Agency). The limit of 75% of the minimum income to qualify for SA benefits was set since in many cases SA benefits is about 75% of the work income; without this adjustment, people with low incomes and long-term SA might have fallen below the minimum income level to be included in the study [27]. Those who were employed in the public sector, self-employed, or who had full-time DP the whole year 2012 were excluded, and those who died or emigrated before 1 January 2017 were excluded. The final study population was 189,321 individuals.

All people living in Sweden aged ≥ 16 years with an income from work or unemployment benefits, who due to disease or injury have a reduced work capacity, are covered by the national public SA insurance providing SA benefits. After a first qualifying day, the employer pays sick pay for the first 14 days of a SA spell, thereafter, SA benefits are paid by the Social Insurance Agency. Self-employed have more qualifying days. Unemployed get SA benefits from the Social Insurance Agency after the first qualifying day. A physician certificate is required after 7 days of self-certification. In this study, data on SA with benefits from the Social Insurance Agency were used. SA spells < 15 days were not included in the study, so as not to introduce bias regarding those who might have been unemployed part of the year. All residents in Sweden aged 19–64 years, whose work capacity is permanently or long-term reduced due to disease or injury, can be granted DP from the Social Insurance Agency.

Both SA and DP can be granted for part- or full-time (25, 50, 75, or 100% of ordinary work hours). SA benefits cover 80% and DP benefits 64% of lost income, both up to a certain level [28, 29]. We used net days of SA/DP so that partial days of SA/DP were combined, e.g., two days of part-time absence for 50% were summed to one net day. The first 14 days in SA spells were counted as being of the same extent as day 15 for the purpose of calculating net days. In all analyses SA and DP days were combined. Mean number of SA/DP net days/calendar year were calculated for each individual [28].

We used information on the following variables from 2012 unless otherwise noted: Sex: woman or man; Age: 18–24, 25–34, 35–44, 45–54, 55–64, or 65–67 years; Country of birth: Sweden, other Nordic country, other EU25, or rest of world including missing; Educational level: compulsory school (≤ 9 years or missing), high school (10–12 years), or college/university (≥ 13 years); Family situation: married/cohabiting with children < 18 at home, married/cohabiting without children at home, single with children at home, or single without children at home; Type of living area: large city (Stockholm, Gothenburg or Malmö), medium-sized town > 90,000 inhabitants within 30 km of city centre), or small town/rural (< 90,000 inhabitants within 30 km of city centre or rural); Job demands/control: based on a job exposure matrix derived from the Swedish Work Environment Survey from Statistics Sweden [30], categorised based on tertiles into the following nine groups: high demands/high control, high demands/medium control, high demands/low control, medium demands/high control, medium demands/medium control, medium demands/low control, low demands/high control, low demands/medium control, low demands/low control (for more information on the job exposure matrix and the categorisation, see Norberg et al. [31]); Having changed branch of industry in 2016, based on SNI, categorised into the following six groups: trade and retail (no change), manufacturing, services, transport, construction and installation, care and education, or hospitality; Having changed sector in 2016: state, region, municipal, private company (no change), or other; Change of occupation (based on SSYK code) 2012–2016: no change or change within sub-major group (according to SSYK classifications), change of sub-major group within major group, change to higher major group (e.g., from major group 2 Professionals to major group 1 Managers), and change to lower major group (e.g., from major group 1 Managers to major group 2 Professionals).

We used a procedure known as Group-Based Trajectory Modelling (Proc traj in SAS) to identify possible trajectories of mean SA/DP days/year in the seven years 2010–2016. Those who had 0 days during all of the seven studied years were categorized into a group of their own, since their trajectory was directly observable. We then fit a group-based trajectory model using only those who had any SA and/or DP during at least one of the seven years.

We determined the best-fitted model related to the number of trajectories using the Bayesian Information Criterion (BIC) [32], where values closer to 0 indicate better fit. The BIC value of a n trajectory model was compared to a n-1 trajectory model, and if the BIC value was smaller, the n group model was considered a better fit than the n-1 trajectory model. In order to keep the number of trajectories within reach for interpretation, a requirement of a minimum of 5% of the study population for the smallest trajectory was introduced (of the population used to build the model, i.e., excluding those who had 0 days each year, due to the size of this group relative to all others) [33]. We also used the method described by Coté et al. [34] to evaluate model fit, whereby the average probability for belonging to the trajectory should be above 0.70 in all trajectories. The final model was the one that fulfilled all the criteria of decreasing BIC value, at least 5% of the study population in each trajectory, and a ≥ 0.70 average probability of belonging to the trajectory in all trajectories.

In the next type of analyses, once the optimal number of trajectories was computed, the individual probabilities of belonging to a particular trajectory were estimated using a multinomial logit function. Individuals were assigned to the trajectory that they had the greatest probability of belonging to.

We used multinomial logistic regression to determine the association between each of the variables and belonging to each trajectory with crude and mutually adjusted odds ratios (OR) and 95% confidence intervals (CI). The reference groups for each respective variable were chosen as the largest groups, or the groups expected to have the lowest risk of belonging to trajectories other than “No SA/DP” (since ORs > 1 are easier to interpret).

Since those who are ≥ 65 years are not eligible for DP benefits (but can still be eligible for SA benefits), we ran sensitivity analyses where we excluded all those who were ≥ 61 years in 2012, and thus were ≥ 65 years during at least one of the follow-up years.

All methods were performed in accordance with the relevant guidelines and regulations. The project was approved of by the Regional Ethical Review Board of Stockholm. All analyses were done in SAS v 9.4.

The main strength of this study is the large and population-based cohort including all 189,321 individuals who lived in Sweden all of 2012, were 18–67 years old, and were employed in a white-collar occupation by a private company in the trade and retail industries, and who were still alive and living in Sweden in 2016. This means that the study is not based on a sample, and that the study population was large enough for subgroup analysis. Another important strength is that we could use microdata from three nationwide administrative registers of good quality [42], that there were no drop-outs (all could be followed up from inclusion to emigration, death, or end of follow-up), and that no self-reports, possibly affected by recall bias, were used. The long follow-up, stretching over seven years, is another strength, and that we can show developments and trends over time in different groups, rather than just measuring the outcome at one time point or the time to a specific event, as is traditionally done in this type of study.

Limitations are the exploratory and observational nature of the study, meaning that we are unable to draw any causal inferences from the results. That we only used SA spells > 14 days can be seen as both a strength and a limitation. Similarly, the use of a JEM to measure job demands and control can be seen as both a strength and a limitation. While the use of a JEM avoids reporting bias from people with poor work capacity potentially estimating the demands and control of their job as different from people with good work capacity, it may have introduced some misclassification of the exposure, since not all people in the same occupation have the same levels of demands and control. Many of the included factors had only a weak association with SA/DP trajectories. This indicates that there might be additional factors of importance that were not included in this study. Another limitation is that the occupational code is not updated each year for each person, meaning that we might have missed some changes of occupation.