Study population
Data from the Finnish Retirement and Aging study (FIREA) were used. FIREA is an ongoing longitudinal cohort study of aging public sector employees in Finland established in 2013. Detailed description of the study design has been provided elsewhere (Leskinen et al. 2018). The eligible population for the FIREA study cohort included all public sector employees whose individual pensionable date was between 2014 and 2019 and who were working in one of the 27 municipalities in Southwest Finland or the nine selected cities or five hospital districts around Finland in 2012 (n = 10,629).
Information on the participants’ official individual pensionable dates was obtained from the pension insurance institute for the public sector (Keva Public Sector Pensions). The participants reported their actual retirement date in questionnaires. Participants were first contacted by sending them a questionnaire 18 months prior to their individual pensionable date (i.e. the lowest age of eligibility for old-age pension). The questionnaire was thereafter sent to the participants annually, at least four times in total. By the end of December 2019, 6,783 cohort members (64% of the eligible sample) had responded to at least one questionnaire. In this study, we included participants who responded to the questionnaire at least once before the pensionable date between 2013 and 2019 and who either reported their actual retirement date or continued working beyond their individual pensionable date (for a minimum of six months) (n = 4,263). The FIREA study was conducted in line with the Declaration of Helsinki and it was approved by the Ethics Committee of Hospital District of Southwest Finland (84/1801/2014). All FIREA participants have given written informed consent to participate in the study.
Assessment of extended employment
In Finland, the retirement ages in the public sector are regulated by the Public Sector Pensions Act. Until the end of 2016, each employee had an individual pensionable date. In general, public sector employees could retire at age 63, but they also could work beyond this age until they reached the age of 68 years, and working beyond the pensionable date accrued a higher pension. After a pension reform in January 2017, each age group has its own retirement age range and the lowest age of eligibility for old-age pension will gradually increase; the increase being 3 months/year until year 2030 after which the increase will be tied to the average life expectancy (Keva 2018). Thus, the retirement age is flexible within a certain age range, being for example from 63 years and 3 months until turning 68 years for those born in 1955 and from 65 years until turning 70 years for those born in 1962. In addition, for some employees, such as primary school teachers, it was possible to keep their earlier retirement age from the previous pension acts in which it was below 63 years. When an employee reached the statutory retirement age window and decided to retire, his/her pension depended on the age at retirement, as the amount of the pension depended on the duration of the working career and the salary with working longer accumulating a higher pension. Information on participant’s pension scheme (i.e. whether the individual pensionable date was before or after the pension reform in 2017) was treated as a covariate in the analyses.
First, the time between the register-based individual pensionable date and the actual retirement date was first calculated. Participants were then divided into two groups: (1) those who did not extend their employment or extended it six months at the most beyond the individual pensionable date (i.e. no extension) and (2) those who extended their employment for over six months beyond the pensionable date (i.e. extended employment). This cut-off point for the extension of employment was chosen as the same cut-off point has been used in previous studies examining extended employment among Finnish employees (Virtanen et al. 2014, 2017). If the participant had not retired at the time of the last available survey, but the follow-up time between the pensionable date and the date of completing the last survey was over six months, they were categorized into the extended employment group.
Assessment of sex and potential explanatory factors
Information on sex and occupational titles of the last known occupation prior to pensionable date was obtained from Keva Public Sector Pensions. We use the term “sex” throughout the text, as the information was register-based, although psychosocial and cultural factors are also considered. Occupational titles were coded based on the Standard Classification of Occupations (ISCO) by Statistics Finland (2019) and occupational status was categorized into two groups: non-manual occupations (ISCO classes 1–4, e.g. teacher, physicians, registered nurses) and service and manual occupations (ISCO classes 5–9, e.g. cleaners, maintenance workers). A validated gender-specific job exposure matric was used to estimate physical workload at each occupation (low physical workload vs. no) (Solovieva et al. 2012).
Information on the other potential explanatory factors was derived from the last FIREA questionnaire preceding the individual pensionable date. We examined the following sociodemographic and work factors: marital status (married vs. no; the latter including single, divorced/separated, and widowed), working status of a spouse (spouse working full-time vs. no; the latter including also those with no spouse), caregiving status (not providing care vs. providing care for a family member), part-time retirement (yes vs. no; based on self-reported employment status), shift work status (regular working hours vs. shift work; the latter including shift work with or without night shifts, regular night work, and other irregular work), and good working capacity (yes vs. no; measured on a scale from 0 [i.e. incapable of working] to 10 [i.e. the best possible working capacity] with scores 8–10 presenting good working capacity (Ilmarinen et al. 1997)).
Information on job strain and work time control was also included. Job strain was measured using nine items assessing job control and five items assessing job demands from the short version of the Job Content Questionnaire (Fransson et al. 2012; Karasek et al. 1998). Job strain was defined as the difference between means scores of the job demand items and job control items (both evaluated on a 5-point Likert-type scale ranging from 1 = totally agree to 5 = totally disagree) and dichotomized into a measure of low job strain (yes, lowest tertile of the scores; no, the other two upper tertiles). Work time control was assessed by asking participants to evaluate how much they could influence several aspects of their working time (evaluated on a scale from 1 = very little to 5 = very much), such as the length of a workday, the starting and ending times of a workday, the handling of private matters during the workday, the scheduling of vacations and paid days off, and the taking of unpaid leave (Ala-Mursula et al. 2005). The highest tertile of the scores was set to indicate high work time control and the other two tertiles as not (yes vs. no). As the questions related to job strain and work time control were not included in FIREA study until the 2016 survey, we used additional information from another cohort study, the Finnish Public Sector (FPS) study (Kivimäki et al. 2007), in which most of the participants had also participated, for those with missing information on these factors who gave their permission to link their information from the FPS surveys (35% of the respondents).
In addition, the following factors associated with health and lifestyle were examined: self-rated health (good vs. suboptimal; assessed with a 5-point scale from 1 = good to 5 = poor with options 1 and 2 indicating good health), psychological distress (yes vs. no; based on a 12-item version of the General Health Questionnaire with a cut-off point of four or more symptoms indicating psychological distress (Goldberg 1972)), pain (no pain vs. mild or severe pain), chronic diseases (no chronic diseases vs. 1 or more chronic diseases, including coronary artery disease, myocardial infarction, stroke, intermitted claudication, hypertension, diabetes, osteoarthritis, osteoporosis, sciatica, fibromyalgia, rheumatoid arthritis, depression, other mental disorder, asthma, and cancer), self-reported sleep duration (≥ 6.5 h vs. less, based on usual sleep duration per 24 h), sleep difficulties (yes vs. no, based on the Jenkins Sleep Problem Scale (Jenkins et al. 1988) with a cut-off point of at least 4 nights/week indicating sleep difficulties), smoking status (non-smoker vs. currently smoking), alcohol use (no risk-use of alcohol vs. risk-use of alcohol; based on weekly alcohol use with the limit for risk-use set as > 288 g/week for men and > 192 g/week for women), physical activity (recommended physical activity vs. no, measured as the metabolic equivalent task [MET] hours with the cutoff point for recommended physical activity set as > 14 MET hours/week), and body mass index (normal weight vs. no; based on self-reported weight and height with the limit for normal weight set as < 25 kg/m2).
Statistical analyses
Differences between men and women in sociodemographic, work- and health-related factors at baseline were examined using χ2 test. In addition, we examined the associations between these potential explanatory factors and working beyond the pensionable date separately for men and women using log-binomial regression models. Log-binomial regression analysis is recommended when the outcome is common, as in this case (29% with extended employment) (Valeri and VanderWeele 2013). The results are expressed as unadjusted risk ratios (RR) and their 95% confidence intervals (CI).
To examine the contribution of explanatory factors on the association between sex and extended employment we used mediation analysis. Since mediation analysis requires that the exposure (here sex) and the explanatory factor (i.e. the mediator) are correlated and that both are associated with the outcome (extended employment), we included only those baseline factors in the mediation analyses that were associated with sex in the total sample and extended employment in men, women, or both. To be included in the mediation analyses, all participants were required to have information on each of the factors that were used in the mediation analyses. This resulted in an analytic sample of 2,513 participants. In the mediation analysis, we first examined the association between sex and extended employment and then serially adjusted for each of the explanatory factors. We report the results as RRs and their 95% CIs. The models have been adjusted for pension scheme (i.e. whether participant’s individual pensionable date was before or after the pension reform in 2017). We calculated the percentage of excess risk mediated (PERM) by each factor as follows: PERM = [RR(pension scheme adjusted)—RR(explanatory factor adjusted)]/[RR(pension scheme adjusted)—1] × 100.
However, the traditional mediation analysis described above, based on the method by Baron and Kenny (1986), does not take into account that the exposure and mediator may interact. This possibility has been addressed by the new causal interference methods that are based on the counterfactual framework. Using the SAS macro presented by Valeri and VanderWeele (2013), we used the counterfactual mediation analysis that allows the presence of exposure-mediator interaction and decomposes the effects into direct and indirect effects. Natural direct effect (NDE) provides RR for the association between sex and extended employment in a hypothetical scenario where the level of exposure to a potential explanatory factor (the mediator) is similar among both men and women. Natural indirect effect (NIE) refers to the excess risk of extended employment among men compared to women that is due to their exposure to potential explanatory factors. In total effect (TE), both natural direct and indirect effects are considered to estimate the RRs for the association between sex and extended employment. The SAS macro produces also the proportion (%) of the TE that the mediator in question explains. We performed the counterfactual mediation analyses using dichotomous exposure, mediator, and outcome variables. The outcome was modeled using a log-binomial regression and the mediators by using logistic regression. All statistical analyses were performed with SAS 9.4 Statistical Package (SAS Institute, Cary, North Carolina, USA).
Finally, we conducted an additional analysis to assess the robustness of our findings by using a cut-off point of working over 12 months beyond the pensionable date. The association between the potential explanatory factors and the extension of employment of over 12 months was examined separately for men and women and those factors that were associated with both sex and this longer extension of employment were included in the mediation analyses. The analytic sample for the extension of employment of over 12 months was 2,819 participants with no missing data on the factors that were used in the mediation analyses.