Study populations

The participating studies were recruited within the European Network for Genetic and Genomic Epidemiology (ENGAGE) consortium including 22 193 individuals with up to 3062 incident cardiovascular events from nine prospective studies (Tables S1, S2, available as Supplementary data at IJE online). Information on genotyping and quality control filters in each study is described in Supplementary data at IJE online. A non-weighted genetic risk score, as well as sensitivity analysis for a weighted score, was calculated from up to 32 independent BMI-associated single nucleotide polymorphisms (SNPs) reported by Speliotes et al.¹⁷(Tables S3, S4, available as Supplementary data at IJE online).

Outcomes

For each participant, the earliest available BMI measurement was used as baseline, and z-transformed for standardization, in each study. The cardiovascular outcomes were provided by the prospective follow-up studies and all were incident, i.e. occurring for the first time during follow-up (after baseline). The diagnoses were based on health registries and/or validated medical records (Table S5, available as Supplementary data at IJE online).

Association analyses

Cox proportional hazards models were used to study associations of BMI and the genetic score with time from BMI measurement to incident cardiovascular disease. Linear regression models were fitted for the association of the genetic score with BMI (Section 4 of Supplementary Data at IJE online). The software used for statistical analysis within each cohort is listed in Table S2. To allow for heterogeneity between studies, random-effects models were used in the meta-analysis (Section 5 of Supplementary Data at IJE online).

Instrumental variable analyses

The genetic risk score was used as the instrumental variable (IV) in the MR analysis, and the IV estimator was then calculated by dividing the corresponding untransformed beta from the meta-analysis of associations of genetic score with cardiovascular outcomes (separately for each outcome) by the beta from the meta-analysis of the association of the genetic score with BMI (Figure 1; Section 6 of Supplementary Data at IJE online).

Secondary analyses

Secondary analyses were performed to study age at event and sex effects (Section 7 of Supplementary Data at IJE online). Each stratum was meta-analyzed separately before MR analyses were undertaken. To test for sex effects, the difference between the effect size estimates for men and women were calculated (Section 8 of Supplementary Data at IJE online).

Additional cross-sectional analyses in ENGAGE (Sections 4.2, 7.2 and 9 of Supplementary Data at IJE online) and CARDIoGRAMplusC4D data (Section 10 of Supplementary Data at IJE online), including sensitivity analysis for pleiotropic effects (Figure S7, available as Supplementary data at IJE online), are described in the Supplementary material. Here, cardiovascular outcomes were binary, so the relationships between BMI and outcomes as well as between genetic score and outcome were modelled via logistic regression.¹⁹

Association analyses

The random-effects meta-analysis confirmed the association between the genetic score and BMI (β = 0.030 SD increase of BMI per allele, 95% CI, 0.028–0.033, P = 2.77·10⁻¹⁰⁷; Table S6, available as Supplementary data at IJE online). The sample size weighted mean BMI was 25.9 kg/m² (SD 4.5) and the sample size weighted mean age was 49.5 years (SD 12.2) in all cohorts. The observational meta-analyses showed that higher BMI was associated with higher risk of incident CHD (HR = 1.20 per SD increase of BMI, 95% CI, 1.12–1.28, P = 1.88·10⁻⁷), heart failure (HR = 1.47 per SD increase of BMI, 95% CI, 1.35–1.60, P = 9.27·10⁻¹⁹) and ischaemic stroke (HR = 1.15 per SD increase of BMI, 95% CI, 1.06–1.24, P = 0.00076; Table 1; Figure S2, available as Supplementary data at IJE online). The genetic risk score meta-analysis for associations with outcome were for incident CHD (HR = 1.00 SD increase of BMI per allele, 95% CI, 0.99–1.02, P = 0.62), heart failure (HR = 1.02 SD increase of BMI per allele, 95% CI, 1.00–1.04, P = 0.017) and ischaemic stroke (HR = 1.02 SD increase of BMI per allele, 95% CI, 1.00–1.04, P = 0.034; Table 1; Figure S3, available as Supplementary data at IJE online).

Instrumental variable analysis

The IV analyses suggested a causal effect of adiposity on incident heart failure (HR = 1.93, per SD increase of BMI, 95% CI, 1.12–1.30, P = 0.017) and ischaemic stroke (HR = 1.83 per SD increase of BMI, 95% CI, 1.05–3.20, P = 0.034; Table 1). There was no support for a comparable causal effect of BMI on incident CHD (HR = 1.13 per SD increase of BMI, 95% CI, 0.70–1.84, P = 0.62; Table 1), despite post hoc power calculations indicating that the current design had greater power for CHD than for heart failure or ischaemic stroke (84% power assuming HR = 2, similar to the estimated effects for the other outcomes; Section 11, Table S12, available as Supplementary data at IJE online). Despite the large differences in point estimates between observational and IV estimators, especially for ischaemic stroke (HR = 1.147 compared with 1.827), we could not provide statistical evidence because of overlapping confidence intervals (Table 1).

Secondary analyses

We performed analyses stratified by age at event (cut-off 55 years) and sex to investigate differences between these groups. The IV analysis found strong associations of BMI with incident heart failure in women only (HR = 3.33 per SD increase of BMI, 95% CI, 1.60–6.93, P = 0.001) and with ischaemic stroke in men only (HR = 2.01 per SD increase of BMI, 95% CI, 1.02–3.98, P = 0.04) (Table S9, Figure S4, available as Supplementary data at IJE online). However, z-tests provided little support for a true sex difference in heart failure (HR_men-women = 2.35, 95% CI, 0.87–6.36, P = 0.09) or in ischaemic stroke (HR_men-women = 1.95, 95% CI, 0.47–8.13, P = 0.36). Overall, the results were driven by the late-onset disease events, as estimates in the early-onset strata were either unavailable due to insufficient number of events, or came with wide confidence intervals (Table S9; Figure S4).

Cross-sectional analyses in the ENGAGE data revealed consistent observational estimates of the BMI–cardiovascular association with the main analyses (Table S7). The IV estimate supported a causal association between BMI and prevalent CHD [odds ratio (OR) = 1.47, 95% CI, 1.04–2.07, P = 0.03] (Table 2; Table S9). In addition, we included data on CHD from the CARDIoGRAMplusC4D consortium²⁰ and performed equivalent analyses (Figure 1; Supplementary Data, Section 10, available at IJE online). The associations were strong for the genetic BMI effect on prevalent CHD (OR = 1.010 per BMI-increasing allele, 95% CI, 1.007–1.014, P = 7.9·10⁻⁹; Table 2; Figure S5, available as Supplementary data at IJE online) as well as for the IV effect on prevalent CHD (OR = 1.40 per SD increase of BMI, 95% CI, 1.24–1.58, P = 2.4·10⁻⁸; Table 2; Figure S6, available as Supplementary data at IJE online) with similar effect sizes as in ENGAGE. Moreover, IV estimators were also calculated for each SNP separately to illustrate possible effect dissimilarities (Figure S6).¹⁸

Comparison with other MR studies

There are a few previous studies that have addressed the causality of adiposity on cardiovascular disease. Recently, we published a paper using the rs9939609 FTO variant as an IV in MR analyses of cardiometabolic traits.¹⁴ The study provided evidence for causality on heart failure, an observation that we now replicated using a genetic risk score providing more precise causal estimates and increased power. In the previous study, we did not find evidence for causal effects of BMI on ischaemic stroke, which we now could establish.

Another MR study by Nordestgaard and co-workers proposed a causal association of BMI on CHD using a genetic score derived from three SNPs.¹³ We used a genetic score derived from 32 SNPs, and thus have a stronger instrument than Nordestgaard and co-workers; nevertheless, the effect estimates are similar to our cross-sectional results. However, they included a larger study sample with more CHD events collected during a long follow-up time, which resulted in a higher overall statistical power than our present study. Otherwise, the studies were comparable in terms of age, BMI and sex distribution.

In 2014, Holmes and co-workers presented an MR analysis of BMI on cardiometabolic traits.¹⁵ They used a genetic score comprising 14 SNPs selected based on their genetic association study of BMI using the CardioChip.²¹ They were unable to provide any support for a causal association between BMI and stroke, or between BMI and CHD.

Another way to address causality for adiposity-related outcomes is to include offspring BMI as an instrument of an individual's own BMI to avoid reverse causation. This has been illustrated by Davey Smith and co-workers, who concluded that estimates for associations using offspring BMI and cardiovascular mortality rates are higher than traditional observational estimates.²²

Adiposity and ischaemic stroke

Adiposity increases the risk of hypertension and type 2 diabetes, which in turn are risk factors for ischaemic stroke⁶. The underlying pathological processes of adiposity on ischaemic stroke could be atherosclerosis, disturbed blood flow and atherogenesis. Ischaemic stroke has been positively associated with adiposity in large observational studies;^5,6,23 thus, our causal estimate is in line with previous findings. The fact that we only estimate a reliable causal effect between adiposity and ischaemic stroke in men and not in women may partly be due to fewer events in women and lower power. Although we did find a fairly large effect difference between sexes, in line with previous findings,²⁴ the lack of precision precludes us from offering firm evidence.

Adiposity and heart failure

Adiposity has been shown to be a risk factor for heart failure,^7,8 and the association is likely to be mediated through increased blood pressure, dyslipidaemia or insulin resistance. Our previous study, using a different transformation of BMI which resulted in a different OR, indicated a causal relationship,¹⁴ and the present study further strengthens the evidence for a causal association by using a genetic score as IV. Sex differences in the relation between adiposity and heart failure have been observed in previous epidemiological studies, which have described a somewhat higher risk in women using sex-specific baseline modelling.^25,26 This agrees with our estimates, even though the precision for the difference is again low. However, the fact that we detected a stronger effect in women, although our calculation showed more power in men, might lend some additional support for these earlier findings.

Adiposity and coronary heart disease

The association between adiposity and CHD has been thoroughly studied in the past decades.^2,27. The underlying cause of CHD is atherosclerosis, which provides a plausible mechanistic link for the relationship with adiposity. Randomized intervention trials of weight loss have been inconclusive.^9,10 However, it should be noted that atherosclerosis is driven by a long-term, low-grade inflammatory process, and short-term interventions on adiposity late in life might not reflect the same exposure as indicators of adiposity in MR studies. Further adding to the uncertainty, a meta-analysis of published MR studies could not provide evidence for a causal link between adiposity and CHD,¹⁵ although the likely reason was underpowered analyses. In the present study, we suggest a causal association using cross-sectional ENGAGE and CARDIoGRAMplusC4D data. The effect size was smaller for CHD than for ischaemic stroke and heart failure, which likely explains the lack of significant effects in previous smaller MR studies.

Strengths and limitations

Strengths of the present investigation include a large sample size with a large number of incident cardiovascular events, age- and sex-stratified analyses, high quality follow-up data and a strong IV based on multiple genetic variants.

However, there are also potential limitations to our investigation. Different disease definitions are used and some cohorts might have selection bias from genotyping at follow-up and not at baseline. Caution should always be taken regarding the assumptions of MR studies.^19,28 First, the genetic variants used as proxies for adiposity must have a reliable and independent association with BMI. Here, we report a strong association between genetic score and BMI, for variants robustly related to BMI.¹⁷ Worth noting, however, is that there is a partial overlap in studies that contributed to this effort and to the discovery of the BMI-associated loci in the Speliotes et al. paper.¹⁷ The SNPs are independent of confounders given the randomization during meiosis and conception, and analyses in TWINGENE showed no associations between the genetic score and smoking, education or exercise (Table S13, available as Supplementary data at IJE online). Population stratification is unlikely to be an issue because we include only individuals of European ancestry. Second, if the causal pathways from genotypes to cardiovascular outcomes do not go through adiposity, one of the assumptions would be violated. The well-known variants in the FTO, MC4R and TMEM18 loci have been reported to have biological functions important for adiposity.^17,29,30 It is possible that many of the other BMI loci that are not yet well annotated will also be found to be of importance for biological mechanisms underlying adiposity, although at this point this is unknown. Third, no other phenotypes should be related to variants outside the causal pathway; i.e. there should be no pleiotropy. By investigating effects of individual adiposity SNPs on CHD using CARDIoGRAMplusC4D data, we could conclude that large pleiotropic effects were unlikely. We also conducted sensitivity analysis in CARDIoGRAMplusC4D excluding SNPs from the genetic score with tendency of outlying effect size in the IV estimator, with similar results (Figure S7, available as Supplementary data at IJE online). Fourth, there should be a log-linear association between the exposure and outcome which is not true for BMI in observational studies.^5,31 However, for a BMI value ≥ 25 kg/m², the association has been reported to be linear and therefore our findings are primarily applicable to those individuals. In any case, if linearity would infer bias, it would most likely lead to underestimation of the associations (as estimates would be driven towards the null). Fifth, observational studies could suffer from confounding at baseline where time from and age at BMI measurement could infer regression dilution bias. This type of error is avoided in the MR method where genetic variants are used as proxy for life-course BMI changes.²⁸

Conclusions

The use of MR methods to draw conclusions on non-confounded causal inference in large population-based studies is rapidly gaining ground. By use of a multiple variant genetic score instrument as a proxy for the intermediate phenotype, it is possible to enhance power in the analyses and to suggest causal effects in disease aetiology.

In the current study, we used data from individuals of European descent to provide support for a causal relationship between adiposity and CHD, heart failure and, for the first time, ischaemic stroke. Although we present the largest study of adiposity as a causal risk factor for ischaemic stroke so far, the confidence intervals were wide, and future larger studies are called for to further establish this relation.