Mendelian Randomization: Using Genetic Lottery Tickets to Argue About Causality

Assumptions · Two-sample MR · IVW, MR-Egger, weighted median · Pleiotropy · Colocalization · Reporting

Published 2026 · Last updated 2026-09-26 · Hailu Zhou

Mendelian randomization (MR) exploits a simple idea: if genetic variants are assigned at conception like randomized labels, and a variant reliably changes an exposure, then comparing outcomes across that variant can reveal whether the exposure causally affects the outcome. Observational epidemiology asks "do people with high X get more disease?" and is haunted by confounding and reverse causation. MR asks "do people randomly dealt a variant that raises X also get more disease?" It is not magic, but under its assumptions it is far harder to confuse cause and correlation.

The core assumptions

The core assumptions are three. The instrument must associate with the exposure. It must not associate with confounders of the exposure–outcome relationship. And it must influence the outcome only through the exposure, not through some other pathway—the dreaded horizontal pleiotropy. Vertical pleiotropy (variant → LDL → heart disease) is the mechanism you want; horizontal pleiotropy (variant → LDL and inflammation independently) breaks the clean story.

Two-sample MR

Two-sample MR is the dominant modern form. You take SNP–exposure effects from one GWAS and SNP–outcome effects from another, harmonize alleles so both reflect the same effect allele, and combine them. Platforms like MR-Base aggregate thousands of GWAS summaries so analysts can test "does trait A cause outcome B" without collecting cohorts. This democratized causal inference—and also flooded the literature with underpowered, pleiotropy-blind papers.

Estimation methods

Estimation methods matter. Inverse-variance weighted (IVW) is the default but assumes no horizontal pleiotropy; if that fails, IVW lies confidently. MR-Egger allows balanced pleiotropy and gives an intercept test. Weighted median tolerates up to 50% invalid instruments if enough are valid. MR-PRESSO detects and removes outliers. Multivariable MR handles correlated exposures (e.g., BMI and waist-to-hip ratio) so you can ask which trait matters after adjusting for the other. Each method trades assumptions for robustness; reporting one number without sensitivity analyses is malpractice.

Instrument selection

Instrument selection is where rigor starts. Genome-wide significance (p < 5×10⁻⁸), clumping for linkage disequilibrium, excluding HLA or structurally complex regions when needed, and checking that instruments are not also associated with smoking, socioeconomic status, or population structure. Population mismatch between exposure and outcome GWAS introduces stratified ancestry effects that mimic causation. Allele harmonization errors—flipping strands incorrectly—can reverse a conclusion. MR is statistically mechanical but biologically unforgiving.

Pleiotropy diagnostics

Pleiotropy diagnostics are not decorations. Cochran's Q tests heterogeneity. MR-Egger intercept tests directional pleiotropy. Leave-one-out analysis checks whether one SNP drives the result. Steiger filtering asks whether the instrument explains more variance in exposure than outcome, as it should. Colocalization asks whether the same causal variant underlies both traits or whether two nearby associations are coincidentally overlapping. MR says "genetic evidence supports a causal link"; colocalization says "the same locus likely mediates both." Different claims, different weight.

What a mature MR study reads like

A mature MR study reads like this: preregister or at least prestate the exposure/outcome, justify instrument thresholds, show F-statistics for weakness, harmonize alleles transparently, run IVW plus Egger plus median plus MR-PRESSO, report heterogeneity and pleiotropy, perform leave-one-out, discuss power, and interpret with biological mechanism. If eQTL/pQTL instruments are used, the story becomes more mechanistic: not just "higher X causes disease," but "genetically predicted higher circulating protein Y increases risk," which can prioritize drug targets.

The honest limitation

The field's honest limitation is that MR estimates a lifetime, population-average effect of exposure differences encoded by genetics. It does not prove a drug will work, it does not capture developmental compensation, and it can be overturned by developmental plasticity. MR-Base and UK Biobank made MR a button; they did not make causal inference automatic. The analyst's job is to resist the button.

References

  1. Hemani G, Zheng J, Elsworth B, et al. The MR-Base platform supports systematic causal inference across the human phenome. eLife. 2018;7:e34408.
  2. Bowden J, Davey Smith G, Haycock PC, Burgess S. Consistent estimation in Mendelian randomization with some invalid instruments using the MR-Egger method. Genet Epidemiol. 2016;40(4):304–314.
  3. Verbanck M, Chen CY, Neale B, Do R. Detection of widespread horizontal pleiotropy in causal relationships inferred from Mendelian randomization. Nat Genet. 2018;50(5):693–698.
  4. Burgess S, Thompson SG. Mendelian Randomization: Methods for Using Genetic Variants in Causal Estimation. Chapman & Hall/CRC; 2015.
  5. Davies NM, Holmes MV, Davey Smith G. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians. BMJ. 2018;362:k601.
  6. Skrivankova VW, Richmond RC, Woolf BAR, et al. Strengthening the reporting of observational studies in epidemiology using Mendelian randomisation (STROBE-MR): explanation and elaboration. BMJ. 2021;375:n2233.
  7. Giambartolomei C, Vukcevic D, Schadt EE, et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 2014;10(5):e1004383.
  8. Sanderson E, Glymour MM, Holmes MV, et al. Mendelian randomization. Nat Rev Methods Primers. 2022;2:6.