The role of glycaemic and lipid risk factors in mediating the effect of BMI on coronary heart disease: a two-step, two-sample Mendelian randomisation study

As an indication of the strength of the association between genetic instruments and phenotypes, we report the proportion of variation in BMI and all mediators explained by their genetic variant instruments and also the F-statistic for the regression of BMI and all mediators on their genetic instruments. The proportion of the BMI–CHD effect that is explained by a group of mediators will be estimated with bias if the mediators are related to each other, and/or if the outcome has an effect on the mediator (i.e. there is reverse causality) and the instrument affects the mediators through the outcome. Therefore, we tested for potential bi-directional causal effects of BMI, potential mediators and CHD with each other using the inverse variance weighted (IVW) approach described below.

Horizontal pleiotropy, where the genetic variant influences the outcome through a pathway other than the exposure, violates an assumption of MR and can bias causal estimates. To guard against this we used three different analytical approaches for both step one (effect of BMI on CHD and potential mediators) and step two (effect of potential mediators on CHD) of the two-step MR mediation approach. Each of the three methods assumes different models of horizontal pleiotropy. The value of comparing results from all three is that we have more confidence in results that are consistent across the different methods. Full details of these approaches, including their different assumptions, are provided in Table 1 and ESM Methods 1.

To estimate the effect of BMI on CHD taking account of genetically determined potential mediators, we used the IVW MR method, adjusting for the SNP–potential mediator effect [23]. The proportion of the effect that is mediated by any of the potential mediators was estimated by the changes in the total effect of the genetically determined BMI on CHD risk (for more details see ESM Methods 2). This method assumes that mediators are continuously measured variables and as type 2 diabetes is dichotomised we did not assess the proportion of the BMI–CHD effect due to type 2 diabetes. An analysis diagram is shown in Fig. 2. All statistical analysis was performed using STATA 13.1 (Stata Corp LP, College Station, TX, USA) and R (version 3.2.5, the R Foundation for statistical Computing, Vienna, Austria) software.

As expected, we observed evidence for association between fasting plasma glucose and type 2 diabetes, and that both fasting plasma glucose and type 2 diabetes were associated with HbA_1c (Table 2). LDL-cholesterol, HDL-cholesterol and triacylglycerols were associated with each other. CHD appears to be causally positively related to type 2 diabetes, but was not related to other potential mediators (Table 2).

There was consistent support across all three MR methods for a causal effect of higher BMI on higher CHD and type 2 diabetes risk, and higher levels of fasting glucose, HbA_1c, fasting insulin and triacylglycerols, together with lower HDL-cholesterol (Table 3). None of the methods supported a causal effect of BMI on LDL-cholesterol (Table 3 and ESM Table 10).

There was broadly consistent support across all three MR methods for a positive effect of type 2 diabetes, HbA_1c, triacylglycerols and LDL-cholesterol on CHD risk (Table 4). For type 2 diabetes the MR-Egger 95% CI just included the null value, but this method has lower statistical power than the others and the point estimates were similar across all methods. For triacylglycerols the estimate of effect (slope) from MR-Egger was a little weaker than for all of the other methods (e.g. 1.24 vs 1.13 comparing the IVW and MR-Egger methods), suggesting that some but not all of the effect of triacylglycerols estimated by IVW and other methods might be due to horizontal pleiotropy. In IVW and the median method analyses, there is evidence for the causal association of lower HDL-cholesterol and higher fasting glucose and insulin with higher risk of CHD. However, for all of these MR-Egger suggested that effects were largely due to horizontal pleiotropy, with effect estimates markedly attenuated to the null and the intercepts all being non-zero.

We explored those potential mediators that had causal support from MR for both an effect of BMI on them (step one) and of the mediators on CHD (step two): type 2 diabetes, HbA_1c and triacylglycerols (Table 5). Our results suggested that triacylglycerols were an important mediator, with either type 2 diabetes or HbA_1c further contributing to mediation of BMI on CHD. The BMI–CHD effect reduced from 1.45 (95% CI 1.27, 1.66) to 1.16 (95% CI 0.99, 1.36) and 1.36 (95% CI 1.19, 1.57) with adjustment for the estimated effects of triacylglycerols and HbA_1c, respectively, and to 1.09 (95% CI 0.94, 1.27) with adjustment for both.

Two-sample MR exploits the fact that it is not necessary to obtain both gene–exposure association (ratio denominator) and gene–outcome association (ratio numerator) from the same sample of participants. There are some advantages to obtaining them from two different sets of participants. For example, ‘winners’ curse’ [16] can bias true causal effects towards observational results in one-sample MR but is less likely to generate false-positive findings in two-sample MR. In addition, the weak instrument bias, which biases effects towards the confounded multivariable regression result in one-sample MR, biases the effect towards the null in two-sample MR (with non-overlapping datasets). The main advantage of two-sample MR is the increased statistical power, particularly in relation to testing effects on binary disease outcomes (i.e. CHD or type 2 diabetes) because of the use of summary data from GWAS consortia [16].

Our study is extremely large and uses genetic variants to avoid some of the key limitations of traditional multivariable regression approaches to mediation. Horizontal pleiotropy is one of the major concerns in relation to limitations of MR studies. However, to explore the potential effects of this pleiotropy, we used different MR methods (IVW, median-based estimators and MR-Egger) that have different assumptions and we assessed the consistency across each of these estimators. The mediators that we took forward into MR-based mediation analyses (triacylglycerols, HbA_1c and type 2 diabetes) had consistent causal effects across these different methods for both steps (i.e. the effect of BMI on them and of them on CHD). In the mediation analyses, where we include both genetically predicted triacylglycerols and HbA_1c, we are assuming that these factors are not causally related to each other. We tested for causal relationships between potential mediators prior to our main two-step MR analyses and these do not suggest any causal effects between triacylglycerols and HbA_lc or other glycaemic traits. However, MR studies cannot completely rule out a causal relationship between the two. Previous large prospective studies showed triacylglycerols predicted the development of type 2 diabetes [43, 44], if this association is casual, the estimated mediation effect by dysglycaemia and triacylglycerols could be inflated. Our results would be biased if the mediators we have tested caused variation in BMI (i.e. there was reverse causality from mediators to BMI). If this were the case, we would expect a bi-directional MR effect between BMI and the mediating risk factors. However, we found no evidence that triacylglycerols or HbA_1c caused variation in BMI (though the causal effect of BMI to these mediators was present).

Whilst all three MR methods suggest a casual effect of triacylglycerols on CHD, the MR-Egger intercept suggests that directional horizontal pleiotropy may be exhibited by the instruments. It is plausible that the genetic variants we used as instruments for triacylglycerols also affect other remnant cholesterols or other lipids and those also contribute to mediating BMI effects on CHD. Another potential limitation to our study is that we have assumed no interaction between BMI and mediators, but we are not able to test for this because we have used aggregated genome-wide data. Previous observational studies suggest that the association between BMI and CHD may be modified by hypertension [45], but have not found effect modification by the glycaemic and lipid traits that we have examined here [46]. In two-sample MR, with independent samples, weak instrument bias can result in bias towards the null. In mediation analyses this could result in an underestimation of mediating effects. However, given our large sample size and the fact that our genetic instruments explained 2.1% and 2.4% of the variation in triacylglycerols and HbA_1c, respectively, and had very large first stage F-statistics, we think this is unlikely to have had a major effect on our results. In addition, there is a partial overlap in studies that contributed to both GWAS (i.e. some cohort studies have contributed both to GWAS of exposure and also of outcomes). Of the 38 studies included in the CARDIoGRAMplusC4D, 24 appear in GIANT (about 30% of participant overlap) [47]. In the case of weak instruments, the sample overlap between the exposure- and outcome-consortia could bias two-sample MR estimates towards the confounded association between the exposure and the outcome [16, 47]. Nevertheless, as we used genetic instruments strongly associated with our exposure, as suggested by large F-statistics, it is unlikely that our results were biased by weak instruments [47].

In conclusion, our results support a causal effect of higher BMI on CHD risk that is, at least partially, mediated through the effect of BMI on triacylglycerols, HbA_1c and type 2 diabetes. These findings support the need for interventional studies examining whether lowering triacylglycerols or providing glucose-lowering therapy for people who are overweight or obese is effective at reducing their increased risk (in comparison with people of healthy weight) of CHD.