Data and analytic sample
Two linked nationally representative surveys conducted in Sweden were used: The Level of Living Survey (LNU) [24] and the Swedish Panel Study of Living Conditions of the Oldest Old (SWEOLD) [25]. The first LNU sample was selected in 1968. New waves were conducted in 1974, 1981, 1991, 2000 and 2010. The age range of the sample was 18–75 years. To represent the Swedish population, randomized quota sampling was conducted to include immigrants and young people. When crossing the upper age limit of 75 years, participants were re-interviewed in the SWEOLD study, which was conducted in five waves: 1992, 2002, 2004, 2011 and 2014. For LNU, structured face-to-face interviews were conducted with professional interviewers. SWEOLD primarily followed the same procedure. However, if a participant had a cognitive impairment or was unable to participate, proxy interviews were conducted with either the spouse, close relative, friend or relevant healthcare professional.
Data used in this study comprises three linked sets of data combined into one longitudinal dataset. Data from LNU 1968 were linked with re-interviews of the sample in SWEOLD 1992 (linkage 1), data from LNU 1981 were linked with SWEOLD 2002 (linkage 2) and data from LNU 1991 were linked to SWEOLD 2011 (linkage 3). The three linkages were compiled in one dataset and analyzed together. If respondents from the baseline in 1968 or 1981 did not respond to any of the questions, the same respondent’s answer from LNU 1974 was used instead. Non-response in baseline 1991 was replaced with answers from LNU 1981. However, this imputation did not exceed 10% in any variable. Response rates in LNU range from 78 to 91%, and in SWEOLD, from 86 to 95%.
In order to follow-up the participants from LNU in SWEOLD, they had to be 53 years or older at baseline. Long-term unemployed, students and housewives were excluded from the analyses because their working conditions could not be assessed. Baseline item nonresponse (n = 46, 4.9%) reduced the sample from 891 to 845 participants and item nonresponse at follow-up further reduced it to 676 or 814 depending on outcomes (Fig. 1). Age varied between 53 and 72 years (mean age 59) at baseline and 77 and 95 years (mean age 81) at follow-up.
Measures of physical function in older age
The main outcome variable was physical function in older age measured using two components available in the SWEOLD study: mobility limitations and physical impairment, used in earlier research [8, 15,16,17, 19]. Mobility limitations were assessed with the responses to three dichotomous “yes/no” questions: “Can you walk 100 meters briskly without difficulty?”, “Can you walk up and down stairs with no problem?”, and “Can you stand without support?”. The combined responses were coded into a three-value ordinal scale wherein 0 = “no mobility limitations” (completed all activities without difficulty), 1 = “mild” (unable to complete one of the activities without difficulty), and 2 = “severe” (difficulty in completing two or all three activities). Physical impairment was assessed by combining nine objective tests as follows: Picking up a pen from the floor, lifting 1 kilogram, touching the left ear with the right hand, touching the right ear with the left hand, touching the left toes with the right hand, touching the right toes with the left hand, placing both hands under the thighs/bottom, turning both palms up and down and getting up from a chair with arms crossed. The respondents were categorized on each task as “managed without difficulty,” “managed with difficulty,” and “did not manage.” To increase power, the resultant variable was dichotomized (“absent” i.e. participant managed all tasks without help, and “present” participant was unable to manage all tasks without help. The participants’ responses on Activities of Daily Living (ADL), recorded in the SWEOLD surveys, were used to impute missing values. Participants who were unable to carry out one or more ADL (e.g. dressing/undressing, bathroom visits, eating) were coded as “did not manage” based on the theoretical assumption that they were unable to participate in the physical performance tests (n = 34).
Measures of social class
Occupational social class was assessed at baseline. Occupation was categorized in accordance with the official Swedish socioeconomic classification (SEI) made by Statistics Sweden, and have been previously published [8, 10, 16]. This classification resembles the internationally well-known Erikson, Goldthorpe, and Portocarero’s (EGP) schema [26]. In order to include small-scale farmers and entrepreneurs, as well as allocate ranks and make social class an ordinal variable, the SEI categories for self-employed individuals were regrouped. The social class variable was coded as unskilled blue-collar workers, skilled blue-collar workers, including small-scale farmers and entrepreneurs with no employees, lower white-collar workers, including medium-scale farmers and entrepreneurs with one to nine employees, and upper white-collar workers, including large-scale farmers, academic professionals and entrepreneurs with at least 10 employees. The SEI coding incorporates educational achievement in the ordering of occupations as part of the classification criteria.
Measures of working conditions
Physical working conditions were assessed at baseline using a six-item index consisting of exposure to various components during the week before the interview. These were: “Do you have to be capable of lifting 60 kilos (heavy lifting) at your work?” (yes/no), “Are you exposed to heavy vibrations at your work?” (yes/no), “Is your job physically demanding?” (yes/no), “Do you sweat every day?” (yes/no), “Are you exposed to poison/acid/explosives?” (yes/no), “Are you exposed to gas, smoke or dust at your work?” (yes/no). The responses were then grouped into three categories. If participants responded negatively to all questions, they were coded as “no exposure”; if they responded “yes” to two or less questions, they were coded as “low exposure” and positive responses to more than two questions were coded as “high exposure.” This categorization was made to preserve power as few participants were exposed to all components. Psychosocial working conditions were assessed at baseline through responses to self-reported yes/no questions and categorized based on the job demand-control model [27]. Demand was derived from the questions “Is your job psychologically demanding/taxing?” and “Is your job hectic?” (yes/no). To answer “yes” to both questions was classified as psychologically demanding. Control consists of personal schedule freedom (decision-making authority) and intellectual discretion (skill discretion). In this study, control was measured through intellectual discretion exclusively, by the questions “What is the level of education required by your job?” and “Is your job monotonous?”. Based on the combined responses, jobs were categorized as: repetitious/monotonous, not repetitious/monotonous and minimum skill level, not repetitious/monotonous and 1–4 years of training (skill level required); and not repetitious/monotonous and more than 4 years of training (skill level required). This was dichotomized into low and high control, in which repetitious/monotonous work and non-repetitious/monotonous work with a minimum skill level was coded as “low control” and the other two were coded as high control.” The demand and control variables were then combined to form four job categories: “passive” = low control and low demand, “high-strain” = low control and high demand, “low-strain” = high control and low demand, and “active” = high control and high demand.
Covariates
Age (continuous) and sex (binary) were retrieved from the registers and confirmed during the interviews. Mobility at baseline was created in the same manner as described for mobility limitations in old age. Health problems at baseline were assessed using an index of self-reported diseases and symptoms experienced over the last 12 months (pain in shoulders, back, hips, joints and/or stomach, cardio-vascular diseases and symptoms, including hypertension, chest pain, swollen legs, myocardial infarction and heart failure, diabetes, leg ulcer, dizziness, breathlessness, fatigue and sleep problems). Education was assessed by highest level of education at baseline and divided into two groups: compulsory and beyond compulsory (i.e., vocational, upper secondary, and university).
Statistical analyses
Stata version 14 was used for all analyses. Binary and ordered logistic regression were used to investigate associations between independent and dependent variables. The proportional odds assumption was tested using generalized ordered logistic regression and approved. To test if the three linkages could be merged into one, we analyzed the linkages separately. Although estimate size and significance level slightly varied between the three linkages, the associations were in the same direction. Also, interaction between linkage and social class was tested. The interactions were statistically non-significant. Thus, the three linkages were merged into one data set to increase power. Supplementary analyses were conducted by including an interaction term between sex and social class. The interaction term was only statistically significant for one out of the 28 interactions, likely by chance. The only statistically significant interaction showed that women in lower white-collar occupations reported more mobility limitations in older age than men. Thus, we analyzed women and men together. All analyses were adjusted for baseline characteristics: age, sex, mobility and health problems (Model 1). Model 2 was additionally adjusted for physical working conditions. Model 3 was adjusted like Model 1 and additionally for psychosocial working conditions. Model 4 was adjusted like model 1 and additionally for physical and psychosocial working conditions.
KHB (Karlson, Holm, and Breen) is a user command in STATA that conducts mediation analyses of a variable in the relationship between an independent and dependent variable by comparing the beta coefficients of two nested non-linear probability models [28]. The output is in the form of total effects, direct effects and indirect effects. When there is an indirect effect, but no direct effect, it is called ‘full mediation’. When there are both direct and indirect effects, it is called ‘partial mediation’ [29].