Mediators of the association between educational attainment and type 2 diabetes mellitus: a two-step multivariable Mendelian randomisation study

Type 2 diabetes is a multifactorial group of disorders in which impaired insulin secretion and/or insulin resistance results in dysregulated carbohydrate, lipid and protein metabolism [1]. It confers an increase in risk of cardiovascular disease and all-cause mortality [2, 3]. Global prevalence has been continuously rising over the past few decades, with type 2 diabetes projected to affect 9.9% of the world population by the year 2045 [4‐8], thus posing an increasingly unsustainable global health burden [9].

A large and consistent body of epidemiological data suggests that those with lower educational attainment (EA) are disproportionately affected by type 2 diabetes [10]. This association is likely mediated by modifiable risk factors, such as obesity, sedentary behaviour, physical activity (PA), smoking and blood pressure [11‐15]. Knowledge of mediation in the EA–type 2 diabetes association will inform public health policies, e.g. by prioritising targets for intervention to reduce the excess risk of type 2 diabetes due to low EA. Our current knowledge of mediating pathways is predominantly based on traditional observational studies that are sensitive to confounding and reverse causation. Therefore, it is uncertain to what extent the associations between EA and type 2 diabetes, and their intermediates, are confounded or affected by reverse causation.

A well-acknowledged method to support causal inference in observational data is Mendelian randomisation (MR). This method uses SNPs, identified in genome-wide association studies (GWAS) to be strongly associated with an exposure, as instrumental variables [16]. Under a number of assumptions, MR yields estimates of an exposure–outcome relation that are less likely to be biased due to unobserved confounding. Recent advances in MR methodology include multivariable MR (MVMR), which can be applied to investigate mediation [17].

Previous MR studies have provided support for a potential causal effect of EA (measured by years of schooling) on coronary artery disease [18], with evidence of mediation through risk factors such as BMI, smoking and blood pressure [19]. For type 2 diabetes risk, recent MR studies have also provided evidence of such a causal effect of EA [20‐22]. However, these studies did not assess mediation by modifiable factors [21, 22], while one recent study only examined mediation by BMI and smoking [23]. Furthermore, some MR studies relied on genetic associations leveraged from less recent, less precise GWAS data [20, 21]. Recent GWAS on EA [24] and type 2 diabetes [25] have yielded more precise estimates of SNP effects due to their larger sample size compared with less recent GWAS. Updating the results from previous MR studies on EA and type 2 diabetes, as well as assessing potential mediation, using the most recent GWAS data would result in more precise insights into the causal structure underlying the EA–type 2 diabetes association.

We therefore aimed to obtain causal estimates of the association between EA and type 2 diabetes and to characterise the causal structure by assessing mediation effects of BMI, sedentary behaviour, PA, smoking and blood pressure in an MVMR framework. In addition, we aimed to obtain observational mediation estimates from the 2013–2014 National Health and Nutrition Examination Survey (NHANES).

In this two-step MVMR study, we found evidence suggestive of a causal, protective effect of EA on type 2 diabetes, with up to 84% mediation by a combination of the modifiable factors BMI, television watching, blood pressure and smoking. Observational mediation estimates in the NHANES 2013–2014 were consistent with the MR mediation estimates with regard to directionality and priority ranking of mediators, but overall suggested less pronounced mediation by the risk factors of interest.

In the present study, the MR-estimated causal effect of a 1 SD (4.2 years of schooling) increase in EA was a 47% reduction in odds of type 2 diabetes (OR 0.53, similar to previous MR studies, ORs ranging from 0.39 to 0.61) [20‐23].

Previous observational studies of the EA–type 2 diabetes association reported 31–53% mediation by a range of risk factors [12, 13]. A previous MR study found that 64% of the association between EA and type 2 diabetes was mediated by BMI and smoking [23], similar to the combined mediation estimate of BMI and smoking in the present study (58%). In general, observational estimates of mediation are lower than those derived from MR. This could be due to underestimation of associations in observational studies due to confounding or measurement error. Although MR is less sensitive to confounding or measurement error, it has been suggested to yield higher associations given that SNP effects represent an estimate of lifetime exposure [19].

Up to ~84% of the EA–type 2 diabetes association was mediated by traditional (i.e. clinical) risk factors, while ~16% remains unexplained. Potential factors that may explain the remainder of the association include factors such as area deprivation, income, diet, health literacy, healthcare access and psychosocial factors. Many of these factors may be not heritable and therefore not suitable for GWAS and consequently unsuitable for 2SMR. However, these factors are expected to show high overlap with factors investigated in the present study, i.e. BMI, television watching, smoking and blood pressure; we therefore expect that these omitted factors would not have contributed substantially to explaining the EA–type 2 diabetes relation. They might however play a role in intervention strategies to reduce type 2 diabetes risk, e.g. reducing BMI through improving diet and health literacy.

Estimates generated from MR are generally insensitive to reverse causation due to the random assignment of alleles at conception. However, our results suggested a bidirectional negative relationship of television watching and smoking with EA, which may imply that EA could also be a mediator of these two traits, complicating the hypothesised model. Additionally, given the low instrument strength for these two traits, results for these two factors should be cautiously interpreted.

Directly intervening on EA by raising the school-leaving age has been shown to be effective in improving adult health (including type 2 diabetes) and reducing mortality in the UK [40]. Other interventions may involve improving access to education, and improving quality of (health) education. However, such interventions are impractical short-term solutions to reducing the burden of type 2 diabetes. In the present study, we provide evidence of substantial mediation of the EA–type 2 diabetes relation through several risk factors that are more easily modifiable than EA. Although population-wide intervention strategies on these modifiable mediators are expected to increase public health, such an approach may widen the inequality gap of type 2 diabetes risk [41]; a high-risk prevention approach (i.e. interventions that target mediators in those with low EA) may therefore be necessary to reduce socioeconomic disparities in type 2 diabetes risk. Our results ranked BMI and television watching to be the strongest contributing factors in the EA–type 2 diabetes association, interestingly, with relatively little overlap, thus suggesting partly independent effects. Interventions on BMI may involve addressing the obesogenic environment associated with low-socioeconomic status neighbourhoods [42, 43], e.g. by limiting fast food outlets in these neighbourhoods. Screen time interventions (television or otherwise) have previously been successfully implemented to improve diet, weight and PA in children [44]. Future studies may further investigate the feasibility and potential impact of such interventions on adult type 2 diabetes risk.

The present MR study yields population-averaged causal estimates of association and mediation. Given the sex differences in both EA [45] and type 2 diabetes [46], it is likely that associations and mediators are also different between sexes, as shown previously [11]. Future MR studies may investigate this using sex-stratified GWAS data.

The predominance of GWAS, including those used in the present study, were performed in white European ancestry populations from high-income countries; generalisation to other ethnicities and low- and medium-income countries is therefore uncertain. Furthermore, a strong relation exists among ethnicity/race, socioeconomic status and health [47] in multi-ethnic communities. Future studies (including GWAS and MR) should therefore be more inclusive with regard to non-white community dwellers.

Strengths of the present study include that it uses SNPs as genetic instruments to minimise bias due to confounding and reverse causation. We used the most recent large-scale GWAS data to generate highly precise SNP effect estimates, facilitating precise MR analysis. The mediated effects estimated were consistent across the two MR mediation approaches and in the statistical sensitivity analyses. Furthermore, MR estimates were corroborated by observational mediation analyses in NHANES 2013–2014, allowing for triangulation [48] and thus improving the robustness of our findings.

Several limitations must be addressed. First, MR may be biased by pleiotropic effects of SNPs, i.e. genetic variants directly influencing both exposure and outcome: a violation of the exclusion restriction criterion. While sensitivity analyses (i.e. MR-Egger, weighted median) robust to pleiotropy [49, 50] showed consistent results, we did not adjust for cognitive ability, which is highly related to EA and a potential confounder in any EA–outcome relation. However, a recent study showed that MVMR adjustment of EA for cognitive ability did not meaningfully affect MR estimates [22]. Second, a potential limitation of using genetic data on social traits such as EA is that ‘population phenomena’ play a role. These phenomena include population stratification, dynastic effects (i.e. transgenerational effects of non-inherited parental SNP alleles) and assortative mating (e.g. non-random mating based on educational level). Whereas population stratification is usually accounted for in GWAS, dynastic effects and assortative mating are not; SNP–EA associations might thus be confounded and therefore may bias MR estimates [51‐54]. Future MR studies might exploit within-family genetic data (e.g. parent–offspring trios, siblings) that have the potential of accounting for such phenomena [55]. Third, the present study assumes absence of exposure × mediator interaction, which currently cannot be modelled in the present 2SMR setting. Fourth, type 2 diabetes and smoking were binary traits, requiring the use of log-odds (as per the original GWAS) in MR analysis for estimating direct and indirect effects. This is non-ideal as ORs are non-collapsible, i.e. marginal ORs are not directly comparable with conditional ORs [56]. Fifth, SNP effects on blood pressure traits were adjusted for BMI in the original GWAS, which subjects MR estimates involving SBP or DBP to potential bias with unpredictable direction [57]. Sixth, sample overlap between GWAS studies may have biased MR estimates towards observational association estimates [58].

To conclude, these results support a potentially causal protective effect of higher EA against type 2 diabetes, with substantial mediation by the modifiable risk factors BMI, television watching and, to a lesser extent, smoking, SBP and DBP. Interventions on these factors thus have the potential of substantially reducing the burden of type 2 diabetes attributable to low EA.