Introduction

Epidemiological studies consistently demonstrate an association between childhood socioeconomic position (SEP), often measured by parental occupation or education, and mortality later in life, especially mortality from cardiovascular diseases (CVD).1–8 The association generally holds true for both men and women and across older and newer generations.1 Furthermore, the burden of CVD risk factors in adulthood such as tobacco smoking, alcohol consumption, physical inactivity, unhealthy diet and body mass index (BMI) has been found to be disproportionately distributed across different strata of the social hierarchies in childhood and adulthood, where people with low SEP show more health-damaging behaviours than those with high SEP.4 5 9 10

In life course epidemiology, two complementary models are proposed to explain the associations between social environment during childhood and risks of disease and mortality later in life: a pathway model and a critical period model. According to the former, early life circumstances affect health outcomes in adulthood by shaping later exposures operating at different stages across the life span.11 12 Thus, a large body of research have suggested that socioeconomic background in childhood affects adult CVDs and mortality by influencing social trajectories such as education and employment and acting through behavioural risk factors such as smoking and drinking.8 13–15 The critical period model, however, refers to a time period in life, particularly in early life, during which exposure to a risk factor may have an irreversible effect on subsequent health.16 In line with this hypothesis, several studies have shown that adverse social circumstances in childhood are associated with increased risks of adult CVDs, mortality and other health outcomes independent of educational attainment, adult social position and other risk factors, implying a latent biological path unexplained by circumstances in adulthood.8 17–21

Typically, previous studies have examined underlying pathways by controlling for risk factors thought to mediate the associations between exposures and health outcomes.1 5 8 15 The common statistical practice has been to fit and compare two regression models: one model without the mediators and another model adjusting for the mediators. The difference in estimates from the two models is interpreted as the mediated effect, that is, the effect operating through the mediators. Findings in those studies may suffer from severe biases as they relied on traditional regression models and often violated some of the fundamental assumptions underlying causal mediation analyses. The recent literature on causal inference lists a set of strong assumptions important for the identification of direct and indirect effects: no unmeasured confounding of the exposure–outcome relationship, no unmeasured confounding of the mediator–outcome relationship, no unmeasured confounding of the exposure–mediator relationship and no intermediate confounding, that is, confounding of the mediator–outcome relationship by a descendent of the exposure.22 These no-confounding assumptions must hold in order for the direct and indirect effects to be causally interpreted. Unfortunately, the mediator–outcome confounding has often been overlooked in the mediation literature based on the traditional regression approach.22–24

Another major limitation is exposure–mediator interaction in the presence of which the traditional regression method fails to decompose the exposure effect.25–27 An important advance in mediation analysis came with Pearl’s28 29 mediation formula that effectively decomposes the total effect into the sum of the natural direct and indirect effects even when an exposure–mediator interaction is at play. Drawing on the Robins and Greenland’s counterfactual framework,25 Pearl’s mediation formula makes a causal contrast between two hypothetical worlds: every individual is exposed in one world, while no individual is exposed in the other. In both worlds, the mediator is set to a value that each individual would naturally take in the absence of the exposure. The difference in the two hypothetical worlds is interpreted as the natural direct effect. Similarly, the natural indirect effect is defined by fixing the exposure in both worlds while letting the mediator for each individual vary between the two worlds.28 Although theoretically appealing, the existing counterfactual mediation approach has limited utility since it is not generally suited to multicategorical, multiple mediator and survival settings.22 30–32

The current study has the ambition to circumvent some of the aforementioned methodological limitations by applying a recently developed weighting approach to mediation analysis.30 31 The aim is to quantify the joint mediation effect of educational attainment, adult social class and behavioural risk factors in the association between social class in childhood and CVD mortality in adulthood in a population-based cohort in Stockholm, Sweden. To compare with CVD mortality, we additionally analysed mortality from all causes except CVDs (henceforth non-CVD mortality) with a view to replicating the current evidence that the causes of CVD mortality, such as coronary heart diseases and stroke, are more strongly related to adverse childhood experiences than other causes of death.17 18

Methods

Measures

Outcomes

The two outcome measures were CVD mortality and non-CVD mortality. Data on mortality were derived from the Cause of Death Register. The study subjects were followed from 1 July 2002 until deaths or the end of the study on 31 December 2011, whichever occurred first. The WHO’s 10th Revision of the International Classification of Diseases (ICD) was used to define CVD mortality (ICD codes I00-I99). All other causes of death were classified as non-CVD mortality.

Exposure

The exposure was father’s SEP measured by father’s occupational social class. Data on father’s occupation was retrospectively collected in the baseline survey in 2002. Based on the Swedish socioeconomic classification,34 Statistics Sweden coded the occupational information into the following eight categories: unskilled manual workers; skilled manual workers; non-manual workers at low level; non-manual workers at mid-level; non-manual workers at high level; self-employed; farmers; and unclassifiable. We treated the non-classified as missing observations and dichotomised the remaining categories into non-manual SEP (low-level, mid-level and high-level non-manual workers, self-employed and farmers) and manual SEP (unskilled and skilled manual workers). Since the self-employed and the farmers are in general considered to be advantaged in the Swedish socioeconomic context,6 we chose to merge them into the non-manual group.

Mediators

We have used two distinct sets of mediators: (1) the social mediators comprising participants’ own education and adult SEP measured by own occupation; and (2) the behavioural mediators, that is, smoking, risky drinking, physical inactivity, diet as well as BMI. All mediators were assessed at baseline. We derived information on participants’ level of education from Statistics Sweden and classified it into three groups: low (primary schooling); medium (secondary schooling); and high (postsecondary/university education). Adult SEP was measured through the survey questionnaire where the participants were asked to report their current/previous occupation and tasks in as much detail as possible. These responses were later used by Statistics Sweden for the Swedish socioeconomic classification: unskilled manual workers; skilled manual workers; non-manual workers at low level; non-manual workers at mid-level; non-manual workers at high level; and unclassifiable. We categorised adult SEP into three groups: non-manual SEP (low-level, mid-level and high-level non-manual workers); manual SEP (unskilled and skilled manual workers); and others (unclassifiable).

The measure of smoking was derived from two questions assessing current and former smoking respectively. Current smoking was defined as smoking tobacco daily during the survey, and former smoking was defined as smoking tobacco daily for at least 6 months in the past. Participants were also asked to report the average amount of alcohol consumption per week and the frequency of binge drinking. As in a previous study,9 we defined risky drinking as consumption of >168 g of pure alcohol per week for men and >108 g of pure alcohol per week for women (high consumption); or consumption of alcohol equivalent to half bottle of spirits/two bottles of wine on a single occasion at least one time per month (binge drinking). Physical activity was measured by using the question ‘How much have you moved/exercised yourself physically in your leisure time during the past 12 months?’ and was coded into four levels: active (at least 30 min of physical exercise >2 times per week with sweating); moderately active (ie, at least 30 min of physical exercise 1–2 times per week with sweating, eg, running and swimming); slightly active (more than 2 hours of physical activity per week without sweating); and inactive (less than 2 hours per week). Diet was assessed by a question ‘How often do you consume fruits or berries?’ and was coded into three categories: more than once a day; almost daily/a few times a week; and once a week or less. BMI was calculated from self-reported height and weight and was conventionally defined as a ratio of weight in kilograms divided by height in metres squared. The BMI score was split into four groups: underweight (<18.5); normal weight (18.5–<25); overweight (25–<30); and obesity (≥30).

Covariates

The covariates used in the study were age (continuous), gender (men and women), country of birth (Sweden, Nordic and others) and marital status (married and single/divorced/widowed) of the study subjects. Whereas age, gender and marital status were register-based data and were considered as mediator–outcome confounders, country of birth was measured through the survey questionnaire and was considered as a confounder potentially affecting the exposure–outcome, exposure–mediator and mediator–outcome relationships (figure 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

A simple causal diagram of the association between father’s socioeconomic position (SEP) and mortality. Social mediators=own education and adult SEP; behavioural mediators=smoking, alcohol drinking, physical inactivity, poor diet and body mass index; confounders=country of birth, age, gender and marital status. 

Results

The study results were based on 19 720 individuals (54% women) born during 1918–1962 and followed for mortality during 2002–2011. The mean age at baseline was 58.2 years (range 40–84 years), and the mean attained age at the end of follow-up was 63.1 years (range 41–94 years). Eighty-two per cent of the study members were born in Sweden, 8% were born in other Nordic countries (Finland, Norway, Denmark and Iceland) and 10% were born outside the Nordic region. During a mean follow-up of 9 years (range 0.37–9.50 years), a total of 2036 deaths occurred of which 751 were due to CVDs. The proportions of deaths from CVDs and non-CVDs in the sample were 3.8% and 6.5%, respectively. Less than half of the sample (46%) had fathers with manual occupations. Table 1 shows the baseline characteristics of both the imputed sample and the sample with complete cases.

Table 2 shows the distribution of the risk factors of CVD mortality by father’s SEP. Results indicate that compared with offspring of non-manual fathers, offspring of manual fathers are themselves more likely to attain low education (23% vs 35%, p<0.001) and manual occupations (23% vs 39%, p<0.001) as adults. The degree of correlation of father’s SEP with participants’ own SEP in adulthood was 0.24 (p<0.001). Similarly, the study subjects whose fathers had a manual occupation showed a more unhealthy behavioural risk profile in terms of adult smoking, risky alcohol drinking, physical inactivity, poor diet as well as overweight and obesity.

In table 3, we show the associations of each social and behavioural risk factor with CVD mortality and non-CVD mortality estimated on the IRR scale, adjusting for father’s SEP and baseline covariates. Overall, all risk factors were found to be associated with both outcomes. However, overweight and obesity did not exhibit any significant association with non-CVD mortality.

The estimated total ‘causal effect’ as well as the direct and indirect effects of father’s SEP on CVD and non-CVD mortality are shown in table 4. Compared with father’s non-manual SEP, manual SEP increased the risk of CVD mortality by 24% (IRR^{Total Effect} (IRR^TE) 1.24; 95% CI 1.09 to 1.41). Formal tests did not yield any effect modification by age (p value for interaction=0.391) or gender (p value for interaction=0.419). Own education and SEP jointly mediated 29% (IRR^{Natural Indirect Effect} (IRR^NIE) 1.06; 95% CI 1.01 to 1.11) of the total effect, while the whole set of mediators including behavioural risk factors jointly mediated 44% (IRR^NIE 1.09; 95% CI 1.04 to 1.14). Thus, the magnitude of the mediated effect by the behavioural factors independent of education and adult SEP was (44%–29%)=15%. Moreover, father’s SEP was associated with CVD mortality independent of the adult social mediators (IRR^NDE 1.17; 95% CI 1.00 to 1.35). The gender-stratified results (online supplementary file 2) further indicate that the total mediation effect was larger for women than for men (64% vs 27%).

With regard to non-CVD mortality, the effect of father’s manual SEP was 1.15 times higher (95% CI 1.04 to 1.27) compared with non-manual SEP. The effect mediated by all social and behavioural intermediates was equivalent to 42% (IRR^NIE 1.06: 95% CI 1.01 to 1.10), whereas an effect equivalent to 38% (IRR^NIE 1.05: 95% CI 1.02 to 1.09) was mediated by the two social intermediates, that is, education and adult SEP. The magnitude of the mediation was generally overestimated by the traditional mediation models when compared with the results from IOW-based models, as evident from table 4 as well as the online tables in supplementary files 2 and 3.

Discussion

The results suggest that a difference by family social class does exist in the risks of both CVD mortality and non-CVD mortality, although the risk of non-CVD mortality tends to be less strong than that of CVD mortality. Using the IOW method, our study further demonstrates that education and social class position in adulthood together with the behavioural risk factors and BMI account for 44% of the increased risk of CVD mortality among the participants. Almost the same magnitude of mediation was observed in the association between childhood social class and non-CVD mortality. The previous literature investigating the magnitude of mediation has generated inconsistent evidence,8 14 15 19 which partially reflects the difference in methodological approaches and the measurement of the mediators. The observed difference in mediation between men and women in the gender-specific analysis (supplementary file 2) needs to be verified in subsequent studies with larger numbers of CVD death.

The natural indirect effects accounting for over forty percent of the total effects of childhood SEP in our study represent the joint mediation effects carried forward by the social and behavioural risk factors. We did not, however, estimate the indirect effects of individual mediators separately as it may not be an appropriate analytic strategy when the mediators affect one another.22 32 We instead chose to estimate the path-specific indirect effects using the sequential mediation approach,32 which required us to make an additional assumption that the social structural pathway comprising education and adult SEP precedes and impacts the behavioural mediators, although one may argue that the health behaviours are already shaped by family background and personality traits during childhood and adolescence. The findings reveal that the social pathway explained large proportions of the studied associations, whereas relatively small proportions were explained by the addition of behavioural mediators, that is, 15% and 4% for CVD and non-CVD mortality, respectively. The findings from the sequential mediation analysis thus point to education and adult SEP as constituting a more powerful set of mediators than smoking, alcohol consumption, unhealthy diet, physical inactivity and BMI taken together. A qualitatively similar conclusion has been drawn in recent studies examining the mediating roles of material and behavioural pathways.19 40 Compared with non-CVD mortality, however, the behavioural factors turn out to be more important for CVD mortality.

We also observed a direct effect of childhood SEP on CVD mortality, that is, an effect that remains after accounting for the socioeconomic indicators and health damaging behaviours measured in adulthood. This finding is in agreement with several earlier studies that documented an increased risk of CVD mortality associated with parental social background even when adulthood circumstances were held constant.8 41 However, the estimated natural direct effect in this study as well as in prior studies requires a cautious interpretation. A majority of the prior literature interpreted the direct effect as a ‘critical period’ effect, thereby defining it as a latent biological pathway unaffected by adult circumstances regardless of the number of adult risk factors considered. Given that we have considered a limited set of social and behavioural mediators, there is room for additional unmeasured mediators or other potentially interlinked mediating pathways (eg, health conditions in childhood) which, if taken into account, could possibly explain some of the ‘direct’ effect.

Similar to other mediation approaches, the mediation parameters obtained through the IOW approach rely on the assumptions that there are no unmeasured confounders affecting the exposure–outcome, exposure–mediator and mediator–outcome relationships and that there are no unmeasured mediator–outcome confounders affected by the exposure. If the models were correctly specified and the no-confounding assumptions held, the IOW-based mediation parameters in our study deserve causal interpretations. Although the bias due to unmeasured confounding cannot be ruled out, the use of the IOW method has offered the current study an advantage over prior research in estimating causally interpretable parameters in the context of multimediators and exposure–mediator interactions in the presence of which the traditional mediation framework is often likely to generate biased results.22 42 43

The traditional regression models presuppose that there are no exposure–mediator or mediator–mediator interactions, although such a presupposition sounds unrealistic given the complexity of the contexts within which diseases and health inequalities emerge. Ignoring interactions, even when the interaction terms are not statistically significant, potentially leads to biased conclusions.22 The main analytic challenge arises due to an exposure–mediator interaction that does not allow decomposing the total effect into direct and indirect effects. We tried to overcome this analytic challenge with the IOW method, which is robust to the inherent interaction structure in the data.31 Similar to a previous study using the same method,35 the current study finds that the mediation parameters derived from the traditional regression models are not entirely compatible with those from the IOW models. A general inflation of the mediated effect has been found in traditional models. Since the traditional models require the assumption of no interaction, an inflation or diminution of the extent of mediation may depend on the directions and magnitudes of the underlying exposure–mediator and mediator–mediator interactions. In the absence of such interactions, however, the traditional regression approach to mediation can yield valid estimates.