Mendelian Randomization (MR)

MR uses single nucleotide polymorphisms (SNPs) as proxies for modifiable exposures. These genetic variants are:

• Fixed at conception, preceding disease onset
• Randomly allocated during meiosis, analogous to treatment assignment in an RCT
• Immune to reverse causation, as disease status cannot alter germline genetics

A valid instrument must satisfy three core assumptions: relevance (robustly associated with the exposure), independence (not associated with confounders), and the exclusion restriction (affects the outcome only through the exposure).

A primary advantage of MR is its ability to minimize unobserved confounding that plagues traditional observational epidemiology. Because genetic variants are randomly assorted at meiosis, they are generally independent of the behavioral, environmental, and socioeconomic factors that typically confound exposure-outcome relationships. This property allows MR to estimate causal effects even when comprehensive confounder measurement is impossible.

A powerful study design where gene-exposure associations and gene-outcome associations are estimated from two independent, non-overlapping populations. This leverages publicly available GWAS summary statistics, allowing researchers to test causal hypotheses without access to individual-level data. The method dramatically increases statistical power by combining summary data from large consortia, even when the exposure and outcome have never been measured in the same cohort.

Horizontal pleiotropy—where a genetic variant affects the outcome through pathways independent of the exposure—violates the exclusion restriction and can bias MR estimates. Modern sensitivity analyses address this:

• MR-Egger regression: Allows for directional pleiotropy via an unconstrained intercept
• Weighted median estimator: Provides consistent estimates when up to 50% of instruments are invalid
• MR-PRESSO: Detects and removes outlier variants with disproportionate pleiotropic effects
• Cis-MR: Restricts instruments to variants within or near the gene encoding the exposure protein

MR can be applied bidirectionally to disentangle the direction of causality between two correlated traits. By testing whether trait A causes trait B and vice versa, researchers can resolve whether an observed association represents causation or reverse causation. Multivariable MR (MVMR) extends this by estimating the direct causal effect of multiple correlated exposures simultaneously, accounting for shared genetic architecture and mediating pathways.

Drug-target MR uses genetic variants within or near a gene encoding a druggable protein to predict the efficacy and safety of pharmacologically modulating that target. By mimicking the effect of a drug acting on a specific pathway, MR can:

• Predict on-target efficacy for proposed indications
• Identify potential adverse effects through phenome-wide scans
• Prioritize drug targets before costly clinical trials

This approach has successfully predicted trial outcomes for PCSK9 inhibitors, IL-6 receptor antagonists, and other therapeutics.

A critical violation of the instrumental variable exclusion restriction where a genetic variant influences the outcome through pathways independent of the exposure under study. This introduces systematic bias into causal estimates. For example, a variant used to instrument BMI might also affect the outcome through inflammatory pathways unrelated to adiposity. Robust MR methods like MR-Egger regression and MR-PRESSO are specifically designed to detect and correct for this phenomenon.

A statistical artifact that arises when genetic variants used as instruments are only weakly associated with the exposure of interest. This bias is amplified in two-sample MR settings and pulls causal estimates toward the confounded observational association in the direction of the null. The standard diagnostic metric is the F-statistic, with a value greater than 10 conventionally indicating sufficient instrument strength. Weak instruments can inflate Type I error rates even in large samples.

The primary input data for two-sample Mendelian randomization. These aggregated results from genome-wide association studies contain effect sizes (beta coefficients), standard errors, p-values, and allele frequencies for millions of single nucleotide polymorphisms (SNPs). MR analyses require harmonization of these summary statistics to ensure that effect alleles are aligned between the exposure and outcome datasets before causal estimation can proceed.

A formal graphical model used to encode causal assumptions before analysis. Nodes represent variables (genetic variants, exposures, outcomes, confounders), and directed edges represent causal effects. The acyclic constraint prohibits feedback loops. DAGs are essential for identifying sources of bias—such as collider bias and confounding—and for determining whether a proposed genetic instrument satisfies the necessary conditions for valid causal inference.

A statistical method that assesses whether two traits share the same causal genetic variant at a given genomic locus. In MR contexts, colocalization distinguishes a causal relationship from a situation where distinct variants in linkage disequilibrium independently affect exposure and outcome. A high posterior probability of a shared causal variant (typically PP.H4 > 0.8) strengthens the evidence that the exposure-outcome association is genuinely causal rather than an artifact of genomic confounding.