Robust inference for stochastic intermediate interventions in Mendelian randomization

Abstract

Mendelian randomization (MR) has established itself as a fundamental tool for inferring the causal effects of modifiable exposures on disease outcomes. Standard MR methods typically estimate the effect of intervening on the exposure itself. However, in many biomedical applications, direct intervention on the primary exposure is often ethically restricted or practically infeasible. Consequently, identifying modifiable intermediate variables (mediators) on the causal pathway offers a more accessible strategy for therapeutic intervention. In this work, we propose a novel statistical framework to assess the causal effect of stochastic intermediate interventions—interventions that shift the population distribution of the mediator rather than fixing it to a deterministic value. We develop a triply robust estimator that guarantees consistency in the union of three observed data models, offering rigorous protection against model misspecification. Furthermore, we extend our framework to robustly accommodate invalid instrumental variables. We evaluate the finite-sample performance of the proposed method through extensive simulations and demonstrate its utility by analyzing genomic data from the religious orders study and memory and aging project to prioritize protein targets for Alzheimer's disease.

Please click here for participating the talk online.