Study design and sample
Our study sample was derived from the Lifelines Cohort Study . Lifelines is a multi-disciplinary prospective population-based cohort study, examining in a unique three-generation design the health and health-related behaviors of 167,729 persons living in the north of the Netherlands. Lifelines employs a broad range of investigative procedures to assess the biomedical, socio-demographic, behavioral, physical and psychological factors that contribute to the health and disease of the general population, with a special focus on multi-morbidity and complex genetics. The study profile of Lifelines, the recruitment, and the data collection are described elsewhere . Baseline assessment (T1), consisting of a physical examination, collecting blood and urine samples, interviews and self-report questionnaires, was conducted between 2006 and 2013. Data collection was performed at baseline (T1), and on average 1.5 years (T2), 2.5 years (T3) and 3.9 years (T4) after baseline (Supplementary Fig. 1) .
The current study used a subsample of 120,177 participants of 18 years and older, who did not have MetS at T1, and whose data were complete for ≥70% of the variables at T1. Participants who were lost to follow-up at T4 (n = 29,889), for whom no MetS status could be determined based on data at T4 (n = 5877), or who had > 30% missing values at T1 (n = 31,195), were excluded from analysis (Supplementary Figure 2). Also excluded were participants with three or more MetS indicators missing, or who had provided information on only three or four indicators, so that we were unable to determine whether they had MetS . Finally, 53,216 participants were included in the analyses.
Measures and procedures
SEP was defined by years of education, equivalized household income, and occupational prestige, as measured with self-reported questionnaires at T1 . Educational level was recoded into years of education, using the number of years it would take to complete each level by the fastest route possible (see Supplementary Table 1 for measurements of the relevant variables in the Lifelines Cohort Study) . Income was recoded as equivalized household income, determined by dividing the midpoint of each participant’s net household income category by the square root of his or her household size . The amounts were divided by 100; the model estimates thus show the difference in odds ratio (OR) of MetS for a 100-euro difference in equivalized household income. Occupational prestige was recoded from the International Standard Classification of Occupations 2008 (ISCO08)  to the continuous Standard International Occupational Prestige Scale 2008 (SIOPS08)  and divided by 10; the model estimates thus show the difference in OR of MetS for a 10-point difference in occupational prestige score. SIOPS08 is a continuous scale, ranging from 0 to 100, and indicating low to high occupational prestige .
MetS indicators were measured during the physical examination and blood sample collection at T1 and T4 . MetS was considered present when, according to the National Cholesterol Education Program’s Adult Treatment Panel III (NCEP-ATPIII), at least three of the five indicators were present . At T2 and T3 MetS was not assessed. MetS criteria are: 1) Waist circumference ≥ 102 cm in male or ≥ 88 cm in female; 2) Systolic blood pressure ≥ 130 mmHg, diastolic blood pressure ≥ 85 mmHg, or use of blood pressure-lowering medication; 3) Triglycerides ≥150 mg/dL (1.7 mmol/l), or use of medication for elevated triglycerides; 4) HDL cholesterol < 40 mg/dL (1.0 mmol/L) in male, or < 50 mg/dL (1.3 mmol/L) in female, or use of lipid-lowering medication; 5) Fasting blood glucose level ≥ 100 mg/dL (≥ 5.6 mmol/l), diagnosis of type 2 diabetes, or use of blood glucose-lowering medication. Medication use at T1 was classified according to the Anatomical Therapeutic Chemical coding scheme , and at T4 with a general question about current medication use (yes/no). For every participant MetS status (yes/no) was dichotomized.
Long-term difficulties perceived as stressful
Long-term difficulties perceived as stressful (‘chronic stress’) were assessed using the Long-term Difficulties Inventory (LDI), a self-reported questionnaire [5, 22]. The LDI consists of 12 items evaluating to what extent various domains of life including housing, work, social relationships (relationships with friends or acquaintances, partner, children, parents or relatives), free time, finances, health, school/study, and religion had been perceived as stressful during the last year. A three-point Likert-scale was used for each item, ranging from 0 (not stressful) to 2 (very stressful). In this study, chronic stress was measured on average 1.5, 2.5 and 3.9 years after T1 (at T2, T3 and T4), after which a continuous variable indicating the total chronic stress between T1 and T4 was calculated (‘sum score’ range 0-72). In addition, chronic stress per domain of life was calculated (range 0-6) and grouped under three categories: not stressful (0), slightly stressful (1-3), and very stressful (4-6). Three domains, chronic stress: 1) at or with work, 2) with partner, and 3) with finances, are considered the most important chronic stress domains, and were displayed separately.
Age and sex at T1, and time between T1 and T4, were used as control variables in all models. Covariates that may influence specific models (e.g., partner status, for chronic stress from difficulties related to partners; or work status, for chronic stress from difficulties related to work) were added to the specific models.
Multivariable logistic- and linear regression analyses, controlling for age, sex, covariates that could influence the specific models, and other SEP measures at T1 and time between T1 and T4, were used to estimate the direct associations between SEP, chronic stress and MetS development (Fig. 1, paths 1, 2 and 3). The association between the chronic stress domains and MetS (path 3) was tested for moderation by sex by adding interaction terms between sex and chronic stress to the models. The total, direct, and indirect associations between SEP and MetS, via chronic stress and the mediating percentages of chronic stress, were estimated using the Karlson-Holm-Breen (KHB) method . The KHB method was used to decompose the total effects of SEP measures on MetS development in the non-linear models into the sum of direct and indirect effects. We used the KHB method, since parameter estimates across nested non-linear models cannot be directly compared because regression coefficients and their error variance are not separately identified; this results in different error variances across models . This problem of ‘rescaling’ of the error variance across nested models makes it impossible to simply examine the change in the effect of SEP on MetS development after inclusion of the chronic stress variables. The KHB method adjusts for this rescaling, and provides unbiased estimates of how much each domain-specific chronic stress variable mediates the association between the SEP measures and MetS development, depending on the presence of the other domain-specific chronic stress variables in the model [29, 30]. The results of all steps are presented as OR with 99% Confidence Intervals (CI), using ‘not stressful’ as reference category. Missing values on SEP measures and chronic stress were imputed using the Multiple Imputation by Chained Equation (MICE) method (10 imputed samples drawn every 100 iterations) . To improve the quality of the imputed values, in addition to the variables used in the substantive models we added length and weight as auxiliary variables to the imputation model . The imputation model included the independent variables, the mediating variables, the dependent variables, the auxiliary variables, age and sex.
By means of sensitivity analyses the robustness of the results was evaluated. To assess the potential role of misclassification of medication use at T4, analyses were repeated for a study sample only of participants who did not use medication at T4 (n = 31,358). To assess the potential role of selection bias from excluding participants with more than 30% missing variables, analyses were repeated with a study sample that included such participants (n = 85,957). Finally, a complete case analysis was performed to investigate differences in associations between the study population with imputed data and the complete cases (n = 41,455). In an additional analysis, the SEP-MetS relationship models were tested for moderation by sex by adding to the models interaction terms with SEP. All analyses were performed using StataMP 13 (64-bit). To allow for multiple testing, p-values< 0.01 were considered to be statistically significant.