Baseline characteristics for the 324,830 participants included in the analyses are provided in ESM Table 2. There were 6006 incident CHD events. The allele score explained 1.8% of the variance in HbA1c, corresponding to an F-statistic of 144.5 and a low risk of substantial weak instrument bias. For UK Biobank participants, associations of the variants incorporated in the allele score with HbA1c were generally of greater magnitude in men compared with women (ESM Fig. 1).
Linear Mendelian randomisation
Linear Mendelian randomisation analyses identified a positive association between higher genetically proxied average blood glucose levels and incident CHD risk when considering men and women together (ESM Fig. 2). For a one mmol/mol increase in HbA1c, the HR for incident CHD was 1.11 (95% CI 1.05, 1.18; p = 2 × 10−4). In sex-stratified analyses, the association was stronger in men (HR 1.12, 95% CI 1.05, 1.19; p = 4 × 10−4) than in women (HR 1.08, 95% CI 0.96, 1.20; p = 0.20) (ESM Table 3). Similar point estimates were obtained in sensitivity analyses using alternative Mendelian randomisation methods (ESM Table 4).
Non-linear Mendelian randomisation
In non-linear Mendelian randomisation, we observed no statistical evidence favouring a non-linear relationship between genetically proxied HbA1c and incident CHD over a linear one in any of the analyses (Fig. 1). Subgroup analyses presenting Mendelian randomisation estimates in quintiles of the population based on HbA1c levels (corrected for genetic predisposition) are presented in ESM Table 3.
We assessed genetic associations of the allele score and its individual variants with other glycaemic traits and lipid fractions (ESM Table 5 and ESM Fig. 3). The allele score was associated with 2-h glucose (p < 0.001) and fasting glucose (p < 0.001), but not fasting insulin (p = 0.22). The allele score was also associated with LDL-cholesterol (LDL-cholesterol, p = 0.03), but not HDL-cholesterol (p = 0.86) or triacylglycerols (p = 0.99). Although multivariable non-linear Mendelian randomisation adjusting for genetically proxied LDL-cholesterol showed some attenuation in the coefficient for genetically proxied HbA1c, an association persisted (HR 1.07, 95% CI 1.01, 1.14; p = 0.018), and the best-fitting fractional polynomial was the linear model (ESM Fig. 4). Similar results were also obtained after excluding the five variants that associated with LDL-cholesterol at p < 0.01 (rs1260326, rs10184004, rs11708067, rs505922 and rs174541): HR 1.10 (95% CI 1.03, 1.17; p = 0.003) (ESM Fig. 5).