Mendelian Randomization
A blood test tells you what a number is. It does not tell you whether that number is doing anything. Some markers cause the disease they predict. Others sit beside it, rising for the same reasons the disease does, carrying no traffic of their own.
Mendelian randomization A method that uses gene variants people are born with as a natural experiment, to test whether a factor causes an outcome. is the method that separates the two, using the fact that gene variants are dealt out at conception and cannot be rearranged afterwards by diet, income, illness or habit. It is the closest thing to a randomized trial that biology performs on its own.
This page explains how it works, shows the markers where it changes the picture, and states plainly where it fails. It also states what it is not allowed to do on this site.
Last reviewed: August 2026
Genetics describe the marker. The grade reflects what happened when people took something.
Not one letter on this site was assigned because a genetic analysis found a marker causal, and no letter will ever be raised on that basis. Grades come from published human randomized trials of an intervention, at the dose and form actually studied, with the PMID PubMed ID: the permanent number the US National Library of Medicine gives a published paper. Each one links to that paper's record on PubMed. attached. A genetic study is not a trial of anything, because nobody was assigned to take anything.
The rule runs in both directions, which matters more than it sounds. Lipoprotein(a) has genetic causality about as convincing as LDL Low-density lipoprotein cholesterol, the cholesterol fraction most closely tied to plaque in the arteries. cholesterol, and zero cardiovascular outcome trials of any drug that lowers it. That combination does not produce a grade. Homocysteine An amino acid in the blood that rises when vitamin B12, folate or B6 are low. runs the other way: the genetics are null, and that is equally not a reason to move a B-vitamin grade downward. Grades move on trials, in both directions, or they do not move.
This is enforced, not promised. A build gate fails the site if a genetic citation ever appears underneath a letter grade, and a second gate fails it if a causal claim is published without a named outcome, a stated population and a genetic citation to rest on.
A randomization that happens before you are born.
Observational Research that watches what people already do rather than assigning a treatment. It can show that two things travel together, not that one causes the other. studies have a permanent problem. People with higher levels of almost anything differ from people with lower levels in a hundred other ways, and some of those ways cause disease. Adjusting for the ones you measured leaves the ones you did not. And when a disease process itself raises the marker, the arrow can point backwards entirely.
Gene variants dodge both problems. Which variants you inherit is settled at conception, before diet, income, smoking, exercise or illness have had any chance to act, and nothing that happens later reaches back to change them. So if a variant that raises a marker also raises disease risk, in the right proportion, the marker is doing something. If it does not, the marker was reading the room rather than changing it.
An observational study compares people who happen to have different levels today, and hopes the groups are otherwise alike.
A genetic analysis compares people who inherited different genotypes, which is a comparison of groups assembled by nothing but chance.
A randomized trial compares people assigned to receive something or not, for a defined period, with a defined dose. It is the only one of the three that tests an intervention, which is why it is the only one that sets a grade here.
One consequence deserves stating early, because it explains most of the apparent contradictions further down this page. A genetic analysis measures a small difference sustained across an entire lifetime. A trial measures a large difference sustained for a few years. Those are different exposures, and they can honestly give different answers about the same substance.
Every marker below is a strong predictor.
They are not all causes.
These are the markers on the EBL panel that carry a genetic causal analysis. The interesting ones are not where the evidence agrees. They are where a marker predicts risk reliably and the genetics say it is not carrying the signal, because that is precisely the situation in which lowering the number achieves nothing.
Carriers of a variant that raises HDL cholesterol did not have lower myocardial infarction risk, and a 14-variant HDL score showed no association, while the LDL score used as a positive control in the same analysis behaved exactly as expected.
This is the clearest case on the panel of a marker that predicts risk without sitting on the causal path. The drug programs that raised HDL cholesterol went on to produce no reduction in cardiovascular events, which is what the genetics had already indicated. EBL grades those interventions on their trials, not on this analysis.
In 19 previously unpublished datasets covering 48,175 coronary cases, the odds ratio for the MTHFR genotype that raises homocysteine was 1.02, confidence interval The range the true effect probably falls in. A wide interval means the estimate is uncertain. 0.98 to 1.07. The published literature on the same genotype gave 1.15, a difference the authors attribute to publication bias.
This is the clearest published demonstration on the panel that the literature a marker generates can point one way while the unpublished data point another.
In 386,406 individuals including 33,546 with coronary heart disease and 18,166 with stroke, genetically predicted 25(OH)D was not associated with coronary heart disease (OR 0.98 per 10 nmol/L, 95% CI 0.95 to 1.01) or stroke (1.01, 0.97 to 1.05). Stratified analyses were null at every observed concentration.
The authors' own reading is that substantial reductions in cardiovascular morbidity from long-term low-dose supplementation are unlikely even if targeted at people with low vitamin D status.
Across 194,418 participants in 47 studies, the risk ratio for coronary heart disease per one standard deviation of genetically raised C-reactive protein was 1.00, with a confidence interval from 0.90 to 1.13.
hsCRP remains one of the strongest observational predictors on this panel. Predicting risk and causing it are separate questions, and this marker answers yes to the first and no to the second. The canakinumab trial cited on this row is not a counter-example. Blocking interleukin-1 beta reduced cardiovascular events and lowered CRP as a downstream readout, without touching lipids. Genetics test whether CRP itself carries the risk, and they say it does not. CRP marks the inflammatory pathway; it is not the pathway.
Conventional analysis gave an odds ratio of 1.18 per standard deviation of urate and multivariable analysis gave 1.10, but MR-Egger, which corrects for pleiotropy at the cost of statistical power, gave 1.05 with a confidence interval from 0.92 to 1.20.
The answer here depends on which method is applied, which is what pleiotropy does to a genetic instrument. Urate variants also associate with blood pressure and lipids, so a naive instrument carries those effects with it.
Two Mendelian randomization analyses of overlapping UK Biobank data reach different answers and are shown together. Sutherland 2022 (n=307,601; PMID 36279545) found an L-shaped relationship: risk fell steeply up to 50 nmol/L then flattened, with odds of death 25 percent higher at a measured 25 nmol/L than at 50. Sofianopoulou 2024 (PMID 38048800) (n=386,406 across UK Biobank, EPIC-CVD and two Copenhagen cohorts, 27,885 deaths) found no association of genetically predicted 25(OH)D with all-cause mortality (OR 0.99 per 10 nmol/L, 95% CI 0.95 to 1.02), and reported that null explicitly at all observed concentrations including the deficient stratum. Note also that Trummer 2013 (PMID 23319826), previously cited here as support, concluded that variants predicting 25(OH)D did not predict mortality and that low vitamin D is unlikely to be causal.
The disagreement is likely methodological rather than substantive. Nonlinear Mendelian randomization by the residual stratification method violates the constant-genetic-effect assumption; the doubly-ranked method (PMID 37390109) was developed as the fix, and reanalysis of a comparable nonlinear finding by the corrected method weakened it (PMID 38749693). Which method Sutherland used could not be confirmed from the abstract, so this is offered as the likely explanation and not as a verdict. The genetic analysis and the supplementation evidence are not in conflict. Genetics cover a lifetime across the whole range including genuine deficiency. The large supplementation trials enrolled populations mostly already above the deficient range and are graded on this site under vitamin D3. Both results can be correct because they test different exposures.
When LDL cholesterol, triglycerides and apolipoprotein B were separated in one multivariable analysis, apolipoprotein B was the trait that accounted for the relationship between lipoprotein lipids and coronary risk.
Variants that lower LDL cholesterol from birth are associated with proportionally lower coronary risk, and the size of the association tracks both how much lower the exposure is and how long it has been lower.
Across 35 cancers and 48 non-neoplastic diseases, genetically LONGER telomeres were associated with higher risk of several cancers. Odds ratios per one standard deviation longer were 5.27 for glioma, 3.19 for lung adenocarcinoma, 1.87 for melanoma and 1.55 for kidney cancer. Associations were stronger for rarer cancers and at tissue sites with lower rates of stem-cell division.
This is the reason a causal verdict on this site is attached to an outcome rather than to a marker. The same genetic evidence points in opposite directions for cancer and for cardiovascular disease, so there is no single answer to the question of whether longer telomeres are better.
In the same analysis, genetically longer telomeres were associated with LOWER risk of coronary heart disease, odds ratio 0.78 per standard deviation, and of abdominal aortic aneurysm at 0.63. There was little evidence of association with psychiatric, autoimmune, diabetic or most other non-neoplastic disease.
The KIV-2 size polymorphism, which sets lifelong Lp(a) Lipoprotein(a), a cholesterol-carrying particle whose level is set largely by genetics. level and is fixed at conception, tracks myocardial infarction risk in the same direction and roughly the same magnitude as measured Lp(a) does.
Causality here is not the same as a treatment answer. No cardiovascular outcome trial of any Lp(a)-lowering drug has reported. The largest, Lp(a)HORIZON, passed its stated mid-2026 readout window in June 2026 without a disclosed result.
Each entry is rendered from the same record that produces the corresponding row on the longevity lab panel, so the two cannot drift apart.
A method that corrects one bias can carry six others.
Genetic analysis arrived in longevity writing as a trump card, which is how every method arrives before anyone audits it. It is not a trump card. It has failure modes as specific as the ones it fixes, and a site that publishes it without them has simply swapped one unexamined authority for another.
A variant may affect the outcome by a route that has nothing to do with the marker. Then the analysis credits the marker for someone else's work. Uric acid on this panel is the live example: the conventional estimate says causal, and the estimate that corrects for pleiotropy says the confidence interval crosses one.
When the variants explain only a sliver of the variation in a marker, the estimate drifts back toward the confounded observational answer it was supposed to replace, while still looking like a genetic result.
Most of this literature runs in European-ancestry biobanks. Variant frequencies, linkage patterns and baseline risk all differ across populations, so an estimate does not automatically travel. Every causal entry on this site states the population it was measured in, for that reason.
A genetic estimate reflects a small difference held for seventy years. It is not the effect size How large a difference the treatment made, as distinct from whether the difference could be detected at all. a five-year trial will produce, and treating it as one systematically overstates what any intervention can deliver.
A body exposed to a genetic difference from conception may compensate for it. A short sharp change later in life gets no such adjustment period, so the genetic estimate can understate what an intervention would do.
Genetic studies are as publishable as any other kind, which means they are as selectively publishable. The homocysteine entry on this panel exists because someone assembled nineteen unpublished datasets and found the answer differed from the published literature on the same genotype.
Vitamin D, where both answers are correct.
Low vitamin D is associated with almost every disease anyone has looked for. Large supplementation trials have mostly come back neutral. The usual conclusion drawn from that pair is that the association was an artifact and the whole thing collapsed. The genetics say something more precise.
The relationship between genetically predicted vitamin D and mortality is L-shaped rather than straight. Risk falls steeply as levels rise out of true deficiency, and then flattens. Above the deficient range, more is not doing anything detectable.
Now look at who the trials enrolled. Largely people already above that range. A trial that supplements a population sitting on the flat part of the curve is testing the flat part of the curve, and a neutral result is the correct answer to the question it asked. The two findings are not in tension. They describe different sections of the same shape.
This is the general lesson and it applies well beyond vitamin D. When a genetic analysis and a trial disagree, the first question is not which one is wrong. It is which exposure each one measured, in whom, and over what span. Quite often they were never answering the same question.
What the genetics say about living longer.
A handful of Mendelian randomization studies have used lifespan itself as the outcome, usually measured as surviving past the 90th percentile of age or as how long a person's parents lived. Each one asks whether being born with a tendency toward a higher or lower level of something is linked to a longer or shorter life.
Read these as the method intends. Each one describes a difference people carried from birth, not what happens when someone changes a habit or takes a product later in life. None of them sets or changes a grade on this site; grades come only from human randomized trials of an intervention.
What this page is built on.
The method itself, the standard technique for detecting pleiotropy, and the reporting guideline that exists because the field filled with low-quality analyses once public summary data made them cheap to produce.