A method with a narrow promise
Rheumatologists should treat Mendelian randomisation as a way to examine difficult causal questions, not as a shortcut around the method's assumptions, a clinically focused review says. The approach may be useful when randomized trials are impractical and observational studies are vulnerable to confounding, reverse causation or selection bias. Its value depends on whether the genetic instruments are relevant and the assumptions are credible.
The article is designed to help rheumatologists distinguish robust MR studies from weak or overinterpreted analyses. It also uses illustrative analyses based on publicly available summary statistics and instruments from relevant genome-wide association studies, commonly known as GWAS.
The assumptions matter
At the heart of MR are three conditions. Genetic instruments must be relevant to the exposure under study, independent of factors that could distort the result, and subject to the exclusion restriction, meaning they should not influence the outcome through a separate route. The review identifies relevance, independence and exclusion restriction as the core assumptions that make an MR result interpretable.
Two-sample MR estimates the gene-exposure association in one dataset and the gene-outcome association in another, commonly using GWAS summary statistics. The review describes this design as scalable, low cost and often more statistically powerful because it can draw on large GWAS. Those practical advantages do not make the underlying assumptions optional.
The examples are not treatment instructions
The review's examples show why a neat-looking association still needs context. In the illustrative analyses, genetically proxied body mass index, or BMI, was associated with higher risk of both beta-blocker use and knee osteoarthritis. The odds ratio was 1.59 per standard deviation for beta-blocker use and 1.78 per standard deviation for knee osteoarthritis, with both p values below 0.001.
LDL cholesterol produced a contrasting pattern. Genetically proxied LDL was associated with higher risk of beta-blocker use, with an odds ratio of 1.16 per standard deviation, and lower risk of knee osteoarthritis, with an odds ratio of 0.92 per standard deviation. Both p values were below 0.001. Genetically proxied HbA1c was unrelated to beta-blocker use and knee osteoarthritis risk in the illustrative analyses.
These examples are best read as demonstrations of how MR results are reported and interpreted, not as standalone evidence that a clinical intervention would have the same effect. The review's broader message is that MR can strengthen causal inference only when its assumptions are credible and the analysis is built around a defensible causal question.
The proxy is not the prescription
The distinction is especially important in drug research. Drug-target MR is presented as a tool for validating a biological target, assessing safety and anticipating long-term effects of perturbing that target. It is not a direct estimate of the treatment response a patient with established disease would have after being prescribed the drug.
The review also warns against treating genetic predictors of medication use as direct stand-ins for pharmacological drug effects. Medication use can reflect indication, disease severity, access to healthcare, prescribing, adherence and comorbidity. Using those predictors as proxies for drug effects is therefore described as likely to be uninformative and biased.
For binary disease exposures, MR usually reflects genetic liability rather than the clinical state of being diagnosed, treated or followed for the disease. That distinction becomes critical when researchers ask what happens after diagnosis. The review says MR is better suited to disease onset or long-term liability than to progression, severity, treatment response or prognosis, because restricting an analysis to people with established disease can create a selection problem known as collider bias.
A quick screen for credible studies
To bring some discipline to the first read of a paper, the review proposes six first-pass questions for screening MR submissions. The framework is practical guidance for deciding which analyses merit closer methodological and clinical appraisal, not a validated scoring system.
Not every useful application is causal. Genetic correlation and MR-style concordance are presented as scalable ways to benchmark phenotype definitions, but they do not establish causality and cannot replace clinical adjudication, classification criteria or careful outcome definition.
The paper's status
Taken together, the review argues for a narrower, more useful reading of MR. It may strengthen causal inference when trials are impractical, but readers still need to ask whether the genetic instruments and core assumptions are credible and whether the design matches the question.
The supplied document is an accepted article following full peer review, but this version has not undergone final copyediting, proofreading or typesetting and will not be identical to the final published version.
The funding statement reports support for S.S. Zhao from Arthritis UK and the NIHR Manchester Biomedical Research Centre. S.L. Au Yeung reported unrelated honoraria from Standard BioTools, while the other author declared no conflict of interest.
Paper data and sources
Original title: Beyond the Hype of Mendelian Randomization for Rheumatology: What It Is Good for, What It Cannot Do, and How to Read It Critically.
Authors: Gabriel Chun Yin Sung, Shiu Lun Au Yeung, Sizheng Steven Zhao
Journal/Repository: The Journal of rheumatology
Status: Peer-reviewed
First online: 2026-10-01
DOI: 10.3899/jrheum.2026-0574
Original paper · Full text