Is MR genetic prediction?

Its main purpose is causal inference about an exposure, rather than predicting an individual's outcome.

Can multiple genetic variants eliminate all bias?

No; they can add information and support sensitivity analyses, but may also introduce invalid pathways.

Does agreement across methods prove the assumptions?

No; methods can share vulnerabilities, and agreement must be interpreted in context.

AI Glossary / Statistics and probability / Intermediate

Mendelian randomization

Definition

Mendelian randomization, or MR, uses genetic variants as instrumental variables to investigate whether an exposure has a causal effect on an outcome. The approach draws on the allocation of inherited variants to help separate exposure differences from some sources of confounding and reverse causation. It is an observational causal-inference method with important assumptions, not a randomized trial conducted by assigning people a treatment.

What it means

An ordinary association between an exposure and an outcome may arise because the exposure changes the outcome, the outcome changes the exposure, or other factors influence both. MR seeks information from genetic variants associated with the exposure to help distinguish these possibilities.
The variants are instruments, not necessarily the main biological subject of interest. The intended question usually concerns the exposure's effect. Genetic association alone is insufficient: the variant must provide a defensible source of variation for the specific causal question.

How it works

Define the exposure, outcome, population, and effect scale. Select genetic instruments using transparent criteria and examine their biological and statistical relevance. A variant weakly associated with the exposure can produce unstable estimates and bias.
Three core instrumental-variable conditions organize the analysis. Relevance requires an association with the exposure. Independence requires the instrument not to share uncontrolled causes with the outcome. Exclusion requires its effect on the outcome to operate through the exposure rather than an alternative pathway.
Estimating a particular effect size requires additional assumptions about the causal relationship, such as an appropriate linear or homogeneous-effect model, or other conditions defining a more local effect. The three core conditions do not by themselves guarantee that every ratio estimates the same universal causal effect.
In a simple single-variant analysis, divide the variant-outcome association by the variant-exposure association, using consistent allele orientation. This is the Wald ratio. Methods using several variants combine information while accounting for uncertainty and relevant dependence.
One-sample MR measures the associations in the same sample. Two-sample MR can use associations from separate datasets, which must be sufficiently compatible for the interpretation. Population differences, overlapping participants, mismatched alleles, and different measurement definitions can matter. Sensitivity analyses investigate vulnerabilities but cannot certify all assumptions as true.

Three practical examples

Research prioritization: scientists use MR evidence alongside other study designs to evaluate whether a proposed biological exposure merits further investigation.
Evidence assessment: an analyst distinguishes an AI-discovered observational association from a separate MR analysis intended to support a causal hypothesis.
Research literacy: a reader checks whether a claim about causation depends on plausible genetic instruments rather than interpreting every genetic association as a causal finding.

Think of it this way

Think of an instrument as a proposed source of variation that moves one input while avoiding unrelated routes to the output. If it also moves another influential input, the output change cannot be attributed solely to the intended pathway. Genetic inheritance can provide useful variation, but the relevant pathways still need examination.

Worked example

Use entirely invented, unitless exposure and outcome scores. Suppose one additional copy of a particular allele is associated with a 0.20-unit increase in exposure X and a 0.12-unit increase in outcome Y. Under a valid instrument and an appropriate linear causal model, the Wald ratio is 0.12/0.20 = 0.60 outcome units per exposure unit.
This is a conditional interpretation, not a causal conclusion produced by arithmetic alone. Suppose instead that the allele has an alternative direct pathway contributing 0.04 outcome units independently of X. The outcome association attributable to the X pathway would then be 0.12 - 0.04 = 0.08, and its ratio would be 0.08/0.20 = 0.40. Ignoring that alternative pathway would report 0.60 rather than 0.40 in this constructed system.
The denominator also matters. A change of 0.01 in the outcome association changes the first ratio by 0.01/0.20 = 0.05. If the exposure association were much smaller, the same numerator change would be amplified more strongly. This calculation illustrates sensitivity to the denominator; it is not a substitute for instrument-strength diagnostics or an uncertainty interval.

Illustrative case study

Fictional case: Rowan Research uses an AI system to identify candidate exposure-outcome associations in published datasets. The team commissions an MR analysis for one candidate instead of treating prediction accuracy as causal evidence.
Researchers review instrument selection, allele alignment, population compatibility, and possible alternative biological pathways. They compare results under several justified analyses and report limitations when estimates disagree. This fictional case does not identify an actual treatment, establish a health benefit, or claim that adding MR automatically validates an AI-generated hypothesis.

Common misconceptions and mistakes

Do not interpret the word randomization as proof that an MR study is free of confounding. Population structure, assortative mating, family effects, and selection can undermine the intended comparison.
Horizontal pleiotropy occurs when a variant affects the outcome through a route outside the exposure pathway, threatening exclusion. Effects downstream of the exposure are not automatically the same violation. Correlation between nearby variants also needs appropriate handling.
Do not equate lifelong genetically influenced differences with the effect of a short-term intervention at a particular age. Exposure timing, magnitude, and biological mechanism may differ. A null estimate may reflect weak information or an unsuitable instrument rather than proof that an exposure has no effect.

Business use

MR can contribute to research portfolios and evidence reviews by testing causal hypotheses with assumptions different from those of conventional observational studies. It should be interpreted alongside experimental, mechanistic, and other observational evidence.
For an AI-enabled research workflow, retain data provenance, instrument choices, allele harmonization, populations, effect units, uncertainty, and sensitivity results. Avoid converting an association-screening pipeline into an automatic causal-claim generator. A reporting checklist improves transparency but does not itself establish validity or determine whether a proposed intervention will work.

Check your understanding

with compatible associations of 0.10 for the outcome and 0.25 for exposure, what is the Wald ratio?

Answer: 0.4 outcome units per exposure unit.

what assumption is threatened by an alternative variant-to-outcome pathway?

Answer: exclusion restriction.

is an MR estimate automatically the effect of starting a treatment tomorrow?

Answer: no; the exposure contrast and timing require interpretation.

Related terms

Instrumental variables
Confounding
Causal inference
Sensitivity analysis
Instrumental variable
Genetic variant
Genome-wide association study
Pleiotropy
Wald ratio

Learn next

Next, study instrumental-variable identification and weak-instrument problems. Then examine one-sample versus two-sample MR, allele harmonization, and how sensitivity methods rely on different assumptions.

Sources

Skrivankova and colleagues, STROBE-MR explanation and elaboration: https://pmc.ncbi.nlm.nih.gov/articles/PMC8546498/ Davies, Holmes, and Davey Smith, Reading Mendelian randomisation studies: https://www.bmj.com/content/362/bmj.k601

Previous: Memory map Next: Mention detection