Adjusting for Medication Use in GWAS and Its Impact on Mendelian Randomisation Analyses: An Example of Systolic Blood Pressure in UK Biobank

Figures

• 

  Figure 1. Scatter plot of p values of independent genetic variants from unadjusted and medication‐adjusted GWAS. GWAS, genome‐wide association study; P, p value. The x axis and y axis shows the p values in the unadjusted and adjusted GWAS respectively. Each dot represents an independent genetic variant from the unadjusted or adjusted GWAS. Purple dots indicate variants which are genome‐wide significant in both unadjusted and adjusted GWAS; Dark green dots show variants which are genome‐wide significant in the unadjusted GWAS but not the adjusted GWAS; Light green dots show variants which are genome‐wide significant in the adjusted GWAS but not the unadjusted GWAS. The dotted diagonal green line has a gradient of 1; if p values from both GWAS are similar, we would expect to observe dots along the dotted green line. A steeper slope greater than 1 indicates that the adjusted GWAS has a more significant p value than the unadjusted GWAS, and vice‐versa.

• 

  Figure 2. Scatter plot of effect estimates of independent genetic variants from unadjusted GWAS (reference) vs. adjusted GWAS. GWAS, genome‐wide association study; P: p value. The x axis shows the absolute effect estimates of independent genetic variants from unadjusted GWAS. The y‐axis shows the corresponding effect estimates of these variants in the adjusted GWAS. Each dot represents a single genetic variant with its corresponding standard error intervals. Green dots indicate that there is no evidence for heterogeneity of effect estimates for genetic variants. Purple dots indicate evidence (p < 0.05) of heterogeneity between effect estimates for the same genetic variant in both GWAS after correction for multiple testing. The grey shaded region shows the 95% confidence intervals around the fitted regression line. The dotted green line has a gradient of 1; if effect estimates from both GWAS are similar, we will observe dots close to the green line. A positive slope > 1 indicates stronger absolute effect estimates in the adjusted GWAS, a positive slope < 1 indicates stronger absolute effect estimated in the unadjusted GWAS. The reported p values are from a two‐tailed T‐test evaluating if…

• 

  Figure 3. Scatter plot of independent genetic variants effect estimates of untreated vs. treated GWAS. GWAS, genome‐wide association study; P, p value. The x axis shows the absolute effect estimates of independent genetic variants from GWAS restricted to untreated individuals. The y axis shows the corresponding effect estimates of these variants in the GWAS restricted to treated individuals. Each dot represents a single genetic variant with its corresponding standard error intervals. Green dots indicate that there is no evidence for heterogeneity of effect estimates for genetic variants. Purple dots indicate evidence (p < 0.05) of heterogeneity between effect estimates for the same genetic variant in both GWAS after correction for multiple testing. The grey shaded region shows the 95% confidence intervals around the fitted regression line. The dotted green line has a gradient of 1; if effect estimates from both GWAS are similar, we will observe dots close to the green line. A positive slope < 1 indicates stronger absolute effect estimated in the GWAS restricted to untreated individuals. The reported p values are from a two‐tailed T‐test evaluating if the slope term differs from 1.

• 

  Figure 4. Inverse variance weighted MR results. GWAS, genome‐wide association study; MR, Mendelian randomisation; MVMR, multivariable Mendelian randomisation; SBP, systolic blood pressure; SD, standard deviation; SNPs, single nucleotide polymorphisms. Summary‐level data MR inverse variance weighted estimates are shown in the figure, with corresponding 95% confidence intervals and p values for each applied adjustment method. The following effects were estimated: (a) odds ratio for each standard deviation increase in SBP on coronary artery disease, (b) effect estimates of body mass index on SBP, (c) effect estimates of SBP on natural hair colour as a negative control outcome. There was less variation in effects when SBP was the exposure, but more variation in effects when SBP was the outcome (and when the exposure influences the probability of receiving treatment and hence altering the outcome phenotype level directly). MVMR analysis was subjected to weak instrument bias.

• 

  Figure 5. Weak instrument bias in multivariable Mendelian randomisation. SNP, single nucleotide polymorphism, used as an instrumental variable in a Mendelian randomisation analysis. X1(t1) represents the first exposure at time 1, X1(t2) represents the first exposure at time 2, and X2 represents the second exposure. In our study, X1(t1) is untreated systolic blood pressure, X2 is antihypertensive use, X1(t2) is measured systolic blood pressure. We postulate this relationship between the variables which could result in the observed weak instrument bias. Although both X1(t2) and X2 could be both robustly associated with identified SNPs, the association of X1(t2) conditional on X2 and X2 conditional on X1(t2) can be weak.