Mendelian randomization (MR) has become something of a darling in modern epidemiology, and for good reason. When done well, it offers a clever workaround for one of the field’s oldest problems: bias. By using genetic variants as proxies for exposures, MR can approximate randomized experiments without actually randomizing anyone. A couple of weeks ago, we released a premium article covering the technique and walked through a compelling example of how it works—using multiple genes associated with low LDL cholesterol to test whether low LDL-C causally drives dementia risk. (If you want a deeper understanding of what good MR looks like in practice, that one’s worth a read.) That’s MR at its best: a clearly defined biological trait with robust genetic architecture, strong instruments, and a plausible causal pathway.
A recent study tried to apply this same logic to a very different question: Do popular diets—specifically low-calorie, vegetarian, or gluten-free—cause inflammatory skin diseases like psoriasis and eczema?1 The main finding was that low-calorie diets may increase psoriatic arthritis risk. But before you start stress-eating in self-defense, there’s a more fundamental problem here—one that starts well before the statistics—with whether the question itself is one that genetics can meaningfully answer.
Building the case
The premise behind this study is an interesting one. Low-calorie, vegetarian, and gluten-free diets have surged in popularity over the past two decades—driven in no small part by social media—and each carries a distinct biological rationale. Low-calorie diets are claimed to extend lifespan, induce autophagy, and improve outcomes in cardiovascular disease and type 2 diabetes. Vegetarian diets can be rich in fiber and antioxidants, low in saturated fat, and contain no meat (sometimes on the belief that meat is harmful). Gluten-free diets, originally developed for celiac disease, are now widely adopted on the belief that gluten promotes inflammation even in people without a formal diagnosis. But when poorly balanced, all three can lead to deficiencies in iron, zinc, calcium, vitamin B12, vitamin D, and omega-3 fatty acids—nutrients that play important roles in immune regulation, skin barrier function, and wound healing. The relationship between these diets and inflammation, in other words, is far from one-directional. The existing observational literature reflects that ambiguity—some studies suggest a vegetarian diet improves atopic dermatitis symptoms, others find no association at all—and traditional observational studies struggle with confounding and selection bias that make definitive conclusions elusive.
To cut through that noise, this study employed MR, drawing dietary exposure data from the UK Biobank—a large-scale biomedical database containing genetic and health information from roughly half a million UK participants—and pairing it with data from the GWAS Catalog, a repository of genome-wide association study results spanning hundreds of traits and diseases. Genetic variants associated with the three dietary patterns were examined alongside four inflammatory skin diseases: psoriasis, psoriatic arthritis, atopic dermatitis, and acne. Sample sizes were large—around 65,000 participants for the dietary exposure data and up to 463,000 for the disease outcome data—and the authors deployed five different MR methods alongside sensitivity analyses, which test whether the results hold up when you probe for potential weaknesses.
On the surface, this looks like a well-designed study.
Our research team spends hundreds of hours each month vetting studies and distilling dense literature to deliver evidence-informed insights on health and longevity. If you find value in our work, consider becoming a premium member and supporting our mission.
Using a hammer to drive a screw
Even the best tool must be used for the right job. You wouldn’t use a hammer to screw two pieces of wood together—and if you did, you’d expect some critique of your craftsmanship.
For MR to work, three core assumptions must hold. The genetic variants need to strongly predict the exposure, they can’t be associated with confounders, and they can only affect the outcome through the exposure itself. When the exposure is something like LDL cholesterol or blood pressure—biological traits with clear, well-characterized genetic determinants—these assumptions are reasonable. You can point to specific genes involved in lipid metabolism that reliably predict higher or lower LDL-C levels across populations.
But what gene meaningfully predicts whether someone chooses to eat vegetarian? Unlike LDL cholesterol or blood pressure—biological traits governed by specific metabolic pathways—dietary patterns are complex human behaviors shaped by culture, ethics, economics, geography, and personal preference. Any genetic signal here is unlikely to represent the diet itself. At best it may capture indirect correlates: taste preferences, personality traits, educational attainment, or other socioeconomic factors. In other words, the genetic instrument may not be measuring diet at all—it may be measuring the kinds of people who tend to choose certain diets.
The authors’ methods reveal this problem plainly. They were forced to relax their significance threshold by several orders of magnitude—from the standard genome-wide level of p<5×10⁻⁸ to p<5×10⁻⁵—because too few variants met the usual criteria. In practical terms, this is the genetic equivalent of admitting that strong predictors of the exposure simply do not exist. The analysis therefore proceeds using weaker signals that are far more likely to reflect noise or indirect correlations than true causal instruments. Returning to the analogy: Having hammered the screw through and noticed the join isn’t holding well, you decide to simply lower your standards for what counts as a solid join—and proceed to advertise the results anyway. It looks fine in the workshop, but put any real weight on it and the problems become apparent quickly.
The instrument validity problem runs deeper than that, though. MR implicitly assumes the exposure is relatively stable across a lifetime, because the genetic variants are present from conception. This works well for genetically predicted LDL cholesterol, which tends to track consistently over time. But vanishingly few people are vegetarian, low-calorie, or gluten-free for their entire lives. Someone may follow the diet at the time of the survey, have adopted it only recently, and abandon it next year. The genetic instrument is supposed to proxy for a lifelong exposure—but the actual exposure is likely transient and inconsistent.
In other words, LDL cholesterol is a biological variable governed by enzymes and transport proteins encoded by specific genes. Dietary patterns are human behaviors influenced by culture, economics, identity, and environment. The former has a genetic architecture. The latter does not.
Applying MR to self-reported dietary patterns isn’t just a methodological limitation to note in passing—it’s a fundamental mismatch between tool and task that colors every result in the paper.
The underlying problem remains
Since genetics can’t reliably predict what diet someone actually ate, the study is left exposed to the usual biases that plague dietary research. Self-reported dietary data are notoriously unreliable—people misreport what they eat, and the categories themselves are poorly defined. “Low-calorie diet” means different things to different people. Someone who tried Meatless Mondays for a month might report as vegetarian.
It’s not surprising, then, that the study found very little. Just one statistically significant result emerged: Low-calorie diets were associated with a modestly increased risk of psoriatic arthritis, with an odds ratio of 1.05 (95% CI: 1.01–1.10)—a 5% relative increase that is clinically tiny even if you take it at face value. No significant associations were found for low-calorie diets and psoriasis, atopic dermatitis, or acne. Nothing for vegetarian diets. Nothing for gluten-free diets.
And even that solitary positive finding is deeply suspect. An extensive body of literature shows that caloric restriction reduces inflammation—so a small positive association running in the opposite direction, derived from weak instruments and self-reported data at a relaxed significance threshold, looks far more like noise than signal.
There’s also the question of reverse causation. What if the association between low-calorie diets and psoriatic arthritis is because people with the condition take up low-calorie diets as an alternative therapy for it? The authors rightly point out that MR is theoretically designed to eliminate this problem—since genetic variants are fixed at conception, they can’t be caused by a disease you develop later in life. In the classic MR setup, that logic is valid. But that protection only works if the genetic variants truly proxy the exposure of interest. If the variants associated with “choosing a low-calorie diet” are actually tagging something broader—health consciousness, weight concerns, or early symptoms that prompt dietary change—then the instrument was never measuring diet in the first place. Mendelian randomization cannot eliminate bias when the instrument itself is mis-specified. It simply relocates the confounding upstream into the genetics.
The bottom line
None of this is an indictment of MR as a technique, and it isn’t an indictment of the authors either. To their credit, they are transparent about the limitations—acknowledging insufficient statistical power, the problems with self-reported dietary data, and explicitly cautioning that their findings should not be interpreted as definitive causal evidence.
The problem is that many people don’t read the limitations section. They read the title, skim the abstract, and walk away with the takeaway that low-calorie diets might cause psoriatic arthritis. In an environment where that kind of finding can travel fast, the gap between what a paper actually claims and what people understand it to claim can cause real confusion.
The fundamental issue isn’t the statistics, the sample sizes, or even the one borderline finding on psoriatic arthritis. It’s that the genetic instruments underpinning this entire analysis were never up to the task of capturing something as behaviorally complex and inconsistently defined as a dietary pattern. When the foundation is shaky, no amount of methodological rigor built on top of it changes what you’re standing on.
When you see an MR study on a behavioral exposure, ask whether the genetic instruments actually make biological sense for what they’re supposed to predict. In this case, the honest answer was no—and no amount of analytical sophistication changes that.
For a list of all previous weekly emails, click here.
References
- Yu Y, Wang S, Lin J, et al. Causal relationships between popular diets (low-calorie, vegetarian, and gluten-free diets) and inflammatory skin diseases: A Mendelian randomization study. Clin Cosmet Investig Dermatol. 2025;18:2605-2615. doi:10.2147/CCID.S538761
