Data
All analyses were conducted on linkages drawn from two data sources: the Swedish Level of Living Surveys (LNU) and the Swedish Longitudinal Study of Living Conditions of the Oldest Old (SWEOLD). Each linkage entail data from one wave of LNU with follow-up in a wave of SWEOLD. LNU and SWEOLD are both longitudinal social surveys, based on random samples of the Swedish population. The first wave of LNU was carried out in 1968 using a nationally representative sample of individuals aged 18 to 75 years. The sample was then followed-up in 1974, 1981, 1991, 2000, and 2010 [41]. The response rates for the waves used in the present study varies between 78.3% and 90.8% (n ≈ 6000–8000). In LNU, respondents were asked about a wide variety of topics, including their present health status, educational achievements, and occupation. SWEOLD is a continuation of LNU: it includes those who have ‘aged out’ of the LNU sample (i.e. those who are 75 years or older), and who were previously included in the LNU sample. SWEOLD has been conducted in 1992, 2002, 2004, and 2011; and the response rates varied between 84.4% and 95.4% (n ≈ 600–1000). In the 2004 wave, the age-span included was wider, and included those aged 69 years and older [42].
The independent variables were assessed at baseline and dependent variables in the designated follow-up. Respondents from LNU 1968 were followed-up in SWEOLD 1992 (linkage 1), respondents from LNU 1981 were followed-up in SWEOLD 2002 (linkage 2) and SWEOLD 2004 (linkage 3), and respondents from LNU 1991 were followed-up in SWEOLD 2011 (linkage 4).
The four linkages were analysed separately and as pooled data. The magnitude of the associations between the independent variables and the outcomes varied by linkage, but the differences were small and not systematic. Thus, we present the results from the full sample, with all the linkages pooled.
The study was restricted to people who were 46 to 64 years at baseline, and who were alive and participated in the designated follow-up (n = 2342). The age restriction was based on the possibility to be included in a follow-up in old age (SWEOLD data). Social class and occupational complexity level is based on occupation, therefore we included only those that were in paid employment (or self-employed) at baseline in the main analyses (n = 2027; 87% of the total sample). However, we performed separate analyses for those who were not in paid employment (or self-employed) at baseline (Table 3).
In the analyses of mobility limitations, the study sample was restricted to people without any mobility limitations at baseline (n = 1772), and in the analyses of psychological distress, to people without psychological distress at baseline (n = 1616). Sensitivity analyses, comparing the estimates to estimates based on the full sample showed similar results. Respondents who did not answer the questions about the outcome variables were excluded, which resulted in a different number of observations in the analyses of mobility limitations (n = 1763) and psychological distress (n = 1596). Since ADL limitations was not assessed in the LNU (baseline) all respondents were analysed studying ADL limitations (n = 2027).
Measurements
Mobility limitations were assessed by the question: ‘Can you walk 100 meters at a fairly brisk pace without problems?’ and ‘Can you climb stairs (up and down) without problems?’ The response alternatives were ‘yes’ and ‘no’. Responses were summarised in an index that ranged from 0 (no problems) to 2 (problems with both tasks).
Limitations in activity of daily living (ADL) were assessed by five questions: ‘Can you eat by yourself?’, ‘Can you go to the toilet by yourself?’, ‘Can you dress and undress yourself?’, ‘Can you get in and out of bed by yourself?’, and ‘Can you wash your hair by yourself?’. The response alternatives were ‘yes, manage completely by myself’, ‘yes, with help’, and ‘no, not at all’. The responses were summarised in an index from 0 (managed all five tasks without help) to 10 (not able to perform any of the tasks).
Psychological distress was assessed by a general question: ‘Have you had any of the following diseases or disorders during the last 12 months?’, followed by a multi-item list of symptoms and disorders. We calculated an index based on the responses regarding anxiety and depressive symptoms. The response alternatives were ‘no’, ‘yes, slight’, and ‘yes, severe’. The responses were summarised in an index ranging from 0 (no problems) to 4 (severe psychological distress).
Education was measured as educational attainment. Self-reported educational attainment was divided into three groups: high, medium, or low. Upper secondary school or above was considered a high level of educational attainment. Compulsory school complemented with vocational training was considered a medium level of education. Attending compulsory school only or no schooling was considered as low level of education.
Social class was based on self-reported occupation at baseline, classified in accordance with the Swedish socioeconomic index (SEI), which is very similar to the Erikson, Goldthorpe, and Portocarero’s (EGP) schema [43]. The social classes were collapsed into three groups: low (unskilled and skilled blue-collar workers, small farmers and entrepreneurs without employees); medium (lower-level white-collar workers, farmers and entrepreneurs with 1 to 19 employees); and high (intermediate and upper-level white-collar workers, farmers and entrepreneurs with 20 or more employees, and academic professionals).
Occupational complexity was assessed by assigning a ‘substantive complexity’ score to each occupational category. The substantive complexity scores indicate the level of intellectual flexibility, engagement, and skills needed to perform working tasks of greater or lesser complexity. The measure of substantive complexity used in this study was developed by Roos and Treiman [44], and is based on the U.S. Dictionary of Occupational Titles (DOT) and the U.S. Census 1970. The DOT included 46 worker characteristics, assessed by job analysts. Roos and Treiman performed a principal component factor analysis and found a factor, including eight of the characteristics (general educational development, specific vocational preparation, complexity of work with data, intelligence aptitude, verbal aptitude, numerical aptitude, abstract interest in the job, and temperament for repetitive and continuous processes), that they called ‘substantive complexity’. This measure forms the basis for our measure of occupational complexity [44]. These scores have later been matched to Swedish occupational categories. See Andel et al. [45] for a description of the matching procedure and Darin-Mattsson et al. [29] for a more thorough description of occupational complexity. Occupational complexity ranges 0–10, the scale was divided into three categories on the complexity scale: 0–3.3 = low complexity, 3.4–6.4 = medium complexity, and 6.5–10 = high complexity.
Individual income was assessed from Swedish tax registers the year before baseline. Income was standardised to the purchasing power of 1991, log transformed and divided into quintiles. To increase the comparability of the models and to increase model fit, we divided income into three categories. We collapsed the two lowest quintiles into category 1 and quintiles three and four into category 2. Category 3 consisted of the fifth quintile, which included those with the highest incomes. This categorization was data-driven and based on tests of spline lines, which showed that the used categorization gave the best fit of the data.
We also used all of the SES indicators described above to construct a composite measure of SES (the SES-index). We used education, social class, occupational complexity, and income as classified above, and summarized them. The index was then divided into tertiles. This was done to investigate whether a composite measure could, statistically, capture as much, or more, of the variance in late-life health as the individual indicators.
Covariates in all analyses were age, sex, and linkage. Age and sex was assessed by self-reports during the interview. Age was measured by birth year and given continuous representation in all analyses. Sex was categorized as either woman or man.
Statistical methods
The results are presented in terms of Average Marginal Effects (AME) multiplied by 100. AMEs are estimated as the average difference in probability of the given outcome across all observations with covariates at their observed values. Thus, the estimates can be interpreted as the differences in the probability of the outcome in percentage points. We use AMEs, rather than odds ratios or β-coefficients, as they are comparable across models, intuitively interpretable, and provide absolute measures of inequality [46]. The AMEs from ordered logistic regression should be interpreted as the difference in the probability of reporting an outcome one-step higher on the scale than the outcome in the reference group. The proportional odds assumption was tested with the gologit2 command in Stata [47], and the assumption was accepted.
In addition, we used McKelvey & Zavoina’s pseudo-R2 to compare estimates of explained variance from different models using the same dataset [48]. To study how each of the main independent variables contributed to model fit, we calculated the change in pseudo-R2 obtained by adding each independent variables (education, social class, income, and occupational complexity) one at a time to a model including only the outcomes and adjustments for sex, age, and linkage. For each indicator of socioeconomic status, we also calculated the change in pseudo-R2 associated to the exclusion of that variable from the full model (model 2).
We also ran all analyses stratified by sex. We found only small, statistically non-significant, differences between women and men except in the association between social class and mobility limitations (see results).
Our study design allow individuals to be included in several linkages. Thus, 282 people were included in both linkage 2 and 3 (<15% of the sample) and the indicators of health in old age could be clustered on individual level. As this could lead to artificially low standard errors, we used cluster-correlated robust estimates of variance in the analyses [49].
We also used multiple imputation to impute missing data on mobility limitations at baseline (113 imputations were included). Sensitivity analyses showed none or small differences between the imputed and non-imputed data. There were no internal non-responses of psychological distress at baseline in the analytical sample. No imputations were made on the outcome variables.
Spearman’s correlations between the independent variables were moderate. The correlation between education and social class was 0.51; between education and income, 0.38; between education and occupational complexity, 0.29; between class and income, 0.39; between class and occupational complexity, 0.38; and between income and occupational complexity, 0.22. All correlations were statistically significant (p < 0.001). Correlations between the indicators differed between the linkages, and there was a general trend towards minor, statistically non-significant, increases in the correlations over time. Model 2 was tested for multicollinearity with the VIF command in Stata, and the result indicated no multicollinearity.