TwoStepCisMR: A Novel Method and R Package for Attenuating Bias in cis-Mendelian Randomization Analyses

Abstract

Mendelian randomisation (MR) is an increasingly popular method for strengthening causal inference in epidemiological studies. cis-MR in particular uses genetic variants in the gene region of a drug target protein as an instrumental variable to provide quasi-experimental evidence for on-target drug effects. A limitation of this framework is when the genetic variant is correlated to another variant that also effects the outcome of interest (confounding through linkage disequilibrium). Methods for correcting this bias, such as multivariable MR, struggle in a cis setting because of the high correlation among genetic variants. Here, through simulation experiments and an applied example considering the effect of interleukin 6 receptor signaling on coronary artery disease risk, we present an alternative method for attenuating bias that does not suffer from this problem. As our method uses both MR and the product and difference method for mediation analysis, our proposal inherits all assumptions of these methods. We have additionally developed an R package, TwoStepCisMR, to facilitate the implementation of the method.

Proteome-wide Mendelian randomization implicates nephronectin as an actionable mediator of the effect of obesity on COVID-19 severity

“…7 ) to avoid weak instrument bias ( Methods ). This approach was also used in previous studies 48 – 50 . In supplementary mediation analyses, we found that adjusting for the exposure when estimating β NPNT-to-severity (Sobel test) also support the role of NPNT as a mediator for the effect of obesity-related exposures (BMI, body fat percentage and body fat mass) on COVID-19 severity outcomes; however, the estimated proportion mediated was modest, likely due to weak instrument bias (Supplementary Table 20 ).…”

“…We used the product of coefficients method without adjusting for the exposure (BMI or body fat percentage) when estimating the effect of the mediator on the outcome ( β NPNT-to-severity ) to avoid weak instrument bias. This approach was also used in previous studies 48 – 50 . We did so because the above-mentioned exposure adjustment requires multivariable MR using NPNT and BMI (or body fat percentage; fat mass) as exposures; however, there are only two instrumental variables for NPNT ( cis -pQTL), but hundreds of instrumental variables for BMI.…”

Proteome-wide Mendelian randomization implicates nephronectin as an actionable mediator of the effect of obesity on COVID-19 severity

“…Throughout the above analyses, we did not adjust the exposure while estimating the effect of the mediator on the outcome (β NPNT-to-severity in Figure 7 ) to avoid weak instrument bias (see Methods ). This approach was also used in previous studies 48,49,50 . In supplementary mediation analyses, we found that adjusting for the exposure when estimating β NPNT-to-severity (i.e., Sobel test) also support the role of NPNT as a mediator for the effect of obesity-related exposures (BMI, body fat percentage, and body fat mass) on COVID-19 severity outcomes; however, the estimated proportion mediated was modest, likely due to weak instrument bias ( Supplementary Table 20 ).…”

“…We used the product of coefficients method without adjusting for the exposure (BMI or body fat percentage) when estimating the effect of the mediator on the outcome ( β NPNT-to-severity ) to avoid weak instrument bias. This approach was also used in previous studies 48,49,50 . We did so because the above-mentioned exposure adjustment requires multivariable MR using NPNT and BMI (or body fat percentage; fat mass) as exposures; however, there are only two instrumental variables for NPNT (cis-pQTL), but hundreds of instrumental variables for BMI.…”

Appraisal of the causal effect of plasma caffeine on adiposity, type 2 diabetes, and cardiovascular disease: two sample mendelian randomisation study

“…30 The standard error of the mediated effect was then estimated by use of bootstrap simulations with 100 000 iterations. 31 Most widely employedstatistical sensitivity analyses that use mendelian randomisation methods to detect and adjust for pleiotropy cannot be conducted with only two instrumental variables. To assess potential pleiotropy of the SNPs associated with caffeine, we instead performed a phenome-wide association analysis in the MR-Base platform (lookup on 1 August 2022).…”