How to Read a Nutrition Study Without Getting Fooled

Nutrition reporting has a structural problem. A single study, often modest in size, gets compressed into a headline written to be clicked, and the compression usually destroys the very details that tell you whether to care. The good news is that the most common distortions follow a few patterns, and once you can name them you can spot them quickly. This is the vocabulary the rest of this site leans on when it weighs evidence, so it earns its own page.

Relative risk versus absolute risk

This is the distinction that ruins the most headlines, so it goes first. Relative risk tells you how much a risk changes in proportion. Absolute risk tells you how big the risk is to begin with, and therefore how much the change actually amounts to for you.

Suppose something raises your risk of a disease by 18%. That sounds like a lot. But 18% of what? If your baseline lifetime risk of that disease is about 5 in 100, then an 18% relative increase takes it to roughly 5.9 in 100, an extra case for every hundred-odd people. The same 18% applied to a common condition would mean far more cases; applied to a rare one, almost none. A relative-risk figure with no baseline attached is close to meaningless, and a headline that gives you the relative number while hiding the absolute one is doing that on purpose, because the relative number is the scarier of the two.

When you read that a food changes risk by some percentage, your first question should always be: a percentage change to what underlying number? If the article does not say, that is itself a finding about the article.

Hazard versus risk

A second distinction trips up even careful readers, and the processed-meat story is the textbook case. In 2015 the World Health Organization's cancer agency classified processed meat as a Group 1 carcinogen and red meat as Group 2A (IARC 2015). Headlines put processed meat in the same group as tobacco and asbestos, and many readers concluded that a slice of bacon was as dangerous as a cigarette.

That conclusion confuses hazard with risk. The classification grades how confident we are that something can cause cancer at all, the strength of the evidence, not how much it raises your risk in practice (IARC 2015). Group 1 means the evidence that the thing is carcinogenic is strong. It says nothing about potency. Tobacco and processed meat can share a category while differing enormously in how much each one actually moves your odds. Putting them in the same group is a statement about certainty, not about danger per serving.

The absolute numbers make the point. The risk increase for bowel cancer from regular processed-meat eating is real but modest in absolute terms, a small addition to a baseline risk, not the order of magnitude that tobacco represents. The classification was honest. The headlines that flattened "we are confident this is a carcinogen" into "this is as bad as smoking" were not. Hold hazard and risk apart and the whole story reads differently, without making the finding disappear.

Correlation versus cause, and the healthy-user effect

Most nutrition findings come from observational studies, which watch what people eat and track what happens to them. These studies can establish that two things move together. They cannot, by themselves, prove that one causes the other, because the people who eat differently also differ in other ways.

The cleanest illustration is a study that split plant-based eating into two kinds. It followed about 209,000 adults and found that plant-based diets built on whole foods were linked to lower coronary heart disease risk, while plant-based diets heavy in refined grains, sugary drinks, and processed snacks were not (Satija et al. 2017). The label "plant-based" did not predict the outcome on its own. The quality of the food did. That is exactly the kind of detail a cruder study, lumping all plant-based diets together, would have missed.

It also gestures at the trap underneath all observational nutrition research: the people who eat well tend to differ from those who do not in ways that have nothing to do with food. They smoke less, move more, drink less, and see doctors sooner. This is the healthy-user effect, and it means a food associated with good health might be riding along with those other habits rather than causing anything. Good studies measure these confounders and adjust for them statistically. No adjustment is ever complete, which is why a single observational study, however large, is a piece of evidence and not a verdict.

Knowing your study designs

The weight a finding deserves depends partly on how it was produced. A rough hierarchy helps.

An RCT can support causal claims that a cohort study cannot, because randomizing who eats what breaks the link between the diet and people's other habits. But diet RCTs are hard to run for long, since you cannot easily control what thousands of people eat for years. So nutrition leans heavily on cohorts and on meta-analyses that pool them. A meta-analysis sits at the top of the hierarchy only when the studies feeding it are sound; pool weak studies and you get a precise-looking average of weak evidence.

Statistical significance is not clinical significance

A result can be statistically significant, meaning it is unlikely to be a fluke, while being clinically trivial, meaning the effect is too small to matter to a real person. With a large enough sample, a tiny difference becomes "significant" in the statistical sense. The word does not mean important. When a headline reports a "significant" finding, check whether the effect is large enough to change anything in practice, rather than only large enough to clear a statistical threshold.

Spotting a press-release overclaim

Most distorted science stories are distorted before a journalist touches them, in the press release the institution issues. A few tells recur. The release describes a relative risk with no baseline. It treats a single new study as overturning a settled field. It states an observational association in causal language ("eating X causes Y") when the study could only show a link. It buries the sample size, the study design, and the confidence interval, the range the true effect probably falls within. And it leads with the most dramatic number available rather than the most representative one.

None of this means studies are worthless or that nutrition science cannot be trusted. It means the unit of trust is the body of evidence, not the headline, and that reading even one study carefully tells you more than reading ten headlines. The IARC case is the model: a careful, honest classification that became misleading only once it was squeezed into a sentence. Keep hazard apart from risk, demand the absolute number behind every relative one, remember that association is not cause, and you will be fooled far less often. For how these ideas play out on specific questions, see plant-based diets and cancer risk and the science library behind the rest of the site.