Mendelian randomisation for mediation analysis: current methods and challenges for implementation

Key results

1. 01

   SuggestedUnder simulated unmeasured confounding, phenotypic difference- and product-of-coefficients estimators were equally biased, whereas multivariable MR (MVMR) and two-step MR recovered direct and indirect effects and the proportion mediated without bias.

   Confounding was simulated as residual covariance (Figure 3: simulated N = 5000; true total effects 0.2, 0.5 and 1.0); MR estimates were unbiased with or without it, and phenotypic estimates were unbiased when no confounding was simulated under a null total effect. Authors argue mediator adjustment in MR therefore avoids collider bias. · Unmeasured confounding; Figure legends

2. 02

   SuggestedNon-differential measurement error pushed phenotypic mediation estimates in opposite directions depending on which variable was mismeasured, yet left MVMR and two-step MR estimates of the mediated effect close to the truth.

   With continuous exposure and mediator, simulated error in the mediator made phenotypic methods underestimate the mediated effect, while error in the exposure made them overestimate it; MVMR and two-step MR showed little bias for error in either variable, matching prior evidence that MR is less error-sensitive than phenotypic analysis. · Measurement error in the exposure or mediator

3. 03

   SuggestedFor binary outcomes, MR mediation depended on scale: MVMR and two-step MR agreed with little to no bias on a linear scale, showed small bias on the log odds ratio scale and were biased on the odds ratio scale; phenotypic methods were biased for both rare and common outcomes.

   Simulated binary outcomes had 5% (rare) or 25% (common) prevalence; log odds ratio bias was small, tended to grow with the true proportion mediated and reflects non-collapsibility of odds ratios. Authors recommend a linear scale for individual-level data and the log odds ratio scale, transformed afterwards, for summary data. · Simulation study; Analysis of binary outcomes; Conclusions

4. 04

   SuggestedWith several mediators, MVMR adjusting for all gave an unbiased direct effect even when mediators were causally linked, whereas summing two-step MR estimates double-counted chained paths; MVMR is advised for joint effects and two-step MR for per-mediator effects.

   In simulations with three mediators where M2 causes M3, the summed two-step MR proportion mediated was over-estimated by 6%, matching the M3-via-M2 contribution, while M2's two-step indirect effect stayed unbiased; without mediator-mediator links both methods coincide. Separately, with all-continuous variables MVMR had smaller standard deviations than two-step MR. · Multiple Mediators; Analysis of multiple mediators; Which method and when

5. 05

   SuggestedMR mediation is fragile with weak instruments or small total effects: weak instruments for either exposure or mediator biased both methods' indirect effect and proportion mediated, and a null total effect gave implausible proportions mediated.

   Strong instruments are conventionally defined by an F-statistic or conditional F-statistic above 10; bias from weak exposure instruments was largest with a common binary outcome. With weak or imprecise total effects, MR indirect effects and proportions mediated had large standard deviations, and authors warn of inflated type 1 error. · Weak instrument bias; Small total effects