PLOS Genetics study identifies substantial biases in two-sample Mendelian randomisation from instrument selection and sample overlap

A study published in PLOS Genetics (8 May 2026) by Amanda Forde, Gibran Hemani, and John Ferguson examines methodological biases affecting two-sample Mendelian randomisation (MR) — a popular approach for using genetic variants as instrumental variables to estimate causal effects of exposures on health outcomes.

The authors use simulations and empirical analyses with UK Biobank data to demonstrate that three sources of bias — Winner's Curse (overestimation of variant effects at genome-wide significant thresholds), weak instrument bias (from instruments with modest F-statistics), and participant overlap between discovery and target samples — can together produce substantial distortions in estimated causal effects. The paper proposes a simulated sample-splitting approach to address these issues and provides practical guidance on its implementation.

Mendelian randomisation has become a core tool in genetic epidemiology, with results from MR studies increasingly cited in evidence reviews and policy discussions. Biases of the kind described in this paper are relevant to anyone interpreting MR results, including researchers, systematic reviewers, and methodologists. The findings complement a broader literature on MR assumptions and limitations and are of direct relevance to those designing or critically appraising MR studies.