The Healthy User Problem: Why Vitamins Look Miraculous in Cohorts and Inert in Trials

For years, hormone therapy looked as though it protected women's hearts. The observational data were consistent and unembarrassed: women taking it had fewer cardiac events than women who weren't. Then the question was put to a randomised trial — assign it deliberately, follow what happens — and the heart-protection story did not survive the asking. The signal that had looked so solid in the cohorts largely dissolved.

Nothing about the chemistry had changed between the two kinds of study. What changed was who chose. The women taking hormone therapy in those cohorts were not a random slice of the population; they were, on average, wealthier, better educated, more engaged with their own health, more likely to be in front of a doctor in the first place. The drug was being credited with the benefits of being that woman. This is the healthy user effect, and once you have seen it you cannot unsee it — because most health journalism is built directly on top of it.

What the healthy user effect actually is

Start with the uncomfortable fact underneath all of it: health behaviours cluster. The person who takes a daily vitamin is more likely to wear a seatbelt, more likely to keep dental appointments, more likely to have a job with sick pay and a gym they occasionally use. None of these things causes the others, but they travel together, because they are all downstream of something larger — disposable income, education, a sense of agency over one's own life, the simple bandwidth to think about next year's health while this year's bills are paid.

An observational study of any one of those behaviours inherits the whole bundle. When researchers compare vitamin-takers with non-takers and find the takers live longer, they have compared two groups that differ in the vitamin and in a hundred other things at once. The vitamin is the variable they wrote down. The hundred other things are the variables doing the work.

There is a second mechanism, subtler and more telling: the adherence effect. People who reliably take their medication tend to do better than people who don't — and this holds even when the thing they are reliably taking is a placebo. The pill is sugar. The reliability is not. Sticking to a regimen, any regimen, marks you out as the kind of person whose life has enough order in it to stick to things, and that orderliness predicts better outcomes all by itself. When even faithful placebo-taking looks protective, you are no longer measuring pharmacology. You are measuring conscientiousness, wearing a lab coat.

Underneath both sits the engine: socioeconomic confounding. Wealth and education shape what you eat, where you live, how much you sleep, whether your work is grinding you down, how soon a worrying symptom gets looked at. They shape, in other words, almost everything that determines how long and how well a person lives — and they also shape who adopts the latest health behaviour first. The behaviour and the outcome share a common cause, and a common cause is the most efficient way in the world to manufacture a correlation that means nothing.

Where it shows up

Once you are looking for it, the pattern is everywhere, and it has a signature: a behaviour that looks wonderful in cohorts and unremarkable when finally randomised.

Supplement epidemiology is the cleanest example. Across observational data, a long list of vitamins and supplements has at various times appeared to lengthen life, protect the heart, or ward off cancer. Take many of the more prominent claims into a randomised trial, where a coin decides who gets the supplement rather than the person's own habits and means, and the effect characteristically shrinks toward nothing. The supplement didn't lose its powers in the lab. It never had them. What it had was a user base that was healthier before the first capsule was swallowed.

Moderate drinking is the case study they should teach in schools. For a generation, a glass of wine a night was reported as positively good for you — the famous curve where teetotallers and heavy drinkers both fared worse than the moderate middle. It was repeated until it felt like common sense. But the abstainers' group was quietly contaminated with people who had stopped drinking because they were already ill, and the moderate drinkers were disproportionately the comfortable, the employed, the socially embedded. As studies got better at separating lifelong non-drinkers from the sick quitters, the protective benefit thinned dramatically. The wine had been taking credit for the dinner party — for having the kind of life in which a relaxed glass with friends is a normal Tuesday.

And then there is the entire wellness economy, much of which is the healthy user effect monetised on purpose. The supplement, the tracker, the subscription, the retreat — their customers are people with money, health-consciousness and time, who would have done comparatively well without buying anything at all. The testimonials are real. The customers genuinely thrive. The product is, in a great many cases, simply along for the ride, billing for an outcome it had little hand in producing.

Why adjustment doesn't rescue it

At this point a reasonable person objects: surely the statisticians know all this. They adjust for it. The papers say so — "adjusted for age, sex, BMI, smoking and income." Problem solved.

It is not solved, for a reason that sounds like a technicality and is actually the whole game: you can only adjust for what you measured. Statistical adjustment is a method for holding constant the variables you wrote down in your dataset. It can do absolutely nothing about the variables you didn't — or couldn't — capture. And the variables that drive the healthy user effect are precisely the slippery, unmeasured ones: conscientiousness, health literacy, the texture of someone's daily stress, how supported they are, whether they are the sort of person who finishes what they start. You cannot put "has their life together" in a regression model, because no one knows how to measure it cleanly. So it sits there, unadjusted, doing its quiet work.

This residual — the confounding left over after you have controlled for everything you managed to record — is not a rounding error. In studies of self-selected health behaviours it is frequently the main event. The unmeasured stuff isn't noise around the signal; much of the time it is the signal, which is exactly why the trial and the cohort disagree.

So when you read "adjusted for age, sex and BMI," read it for what it honestly is: a statement about three things the authors happened to have, not a magic spell that purifies the comparison. Adjustment narrows the gap between the groups. It does not close it, and it cannot close the part that was never on the page. Treating a long list of covariates as proof of causation is one of the most common upgrades of confidence in all of health reporting, and one of the least earned.

None of this means observational research is worthless — that overcorrection is its own mistake. For a great many questions, randomising is impossible or flatly unethical: you cannot assign people to a lifetime of smoking, or to poverty, to see what happens. Observational work is the only honest tool we have for those, and done well — with prospective design, with sober handling of confounders, with conclusions sized to match — it carries real weight. The smoking-and-cancer case was won on exactly this kind of evidence. The discipline is not to dismiss cohorts. It is to know what a cohort can and cannot prove on its own.

The reader's filter

So you do not need an epidemiology degree to read health news better. You need one question, asked reflexively: who chooses this behaviour, and what else is true of them?

That is the entire filter. The instant a study reports that some habit goes with better health, picture the person who actually adopts it — buys the supplement, joins the gym, takes the daily pill without missing — and ask what else tends to come bundled with being that person. If the honest answer is "money, time, education and a generally well-ordered life," then those are your suspects, and the headline's chosen variable is more likely a passenger than a driver.

There is a second tell, just as useful, for when cohorts and trials openly disagree. The instinct is to split the difference, or to assume the trial was simply too small to detect a real effect. Resist it. When a well-run randomised trial flatly contradicts a pile of observational findings, the smart money says the cohort was confounded, not that the trial was underpowered. Randomisation exists precisely to break the link between the behaviour and the kind of person who adopts it — to let the coin choose instead of the income bracket. When the coin disagrees with the cohort, the coin is usually right. Believe the trial.

And the simplest signal of all lives in the verbs. Watch for "linked to," "associated with," "tied to" sitting in headline position while the body sells you a conclusion in plain causal language. Those phrases are not casual synonyms for "causes" — they are the technically correct, deliberately weaker terms for a correlation, and their job in a headline is to imply causation while retaining deniability. When a piece leans on "linked to" up top and then quietly graduates to cause-and-effect by the third paragraph, it is telling you, in its own grammar, what it could not actually demonstrate.

What this means

The healthy user effect is the closest thing health journalism has to a master key. Strip away the supplements and the tracking apps and the wine and the wellness retreats, and a startling share of "X linked to longer life" reduces to the same finding restated: people who do X were already the kind of people who live longer. The behaviour didn't earn the outcome. It was riding alongside it, and the study photographed them together and captioned it as cause.

This is not a counsel of nihilism, and it is not an excuse to believe nothing. It is a request to keep the conditional welded to the claim — to read "associated with" as the careful, deniable thing it is, to remember that adjustment cleans only what was measured, and to trust the coin over the cohort when they part ways. Once that filter is installed, it does not switch off. You will catch yourself running it on every confident health headline you meet for the rest of your life, asking the one question the headline was built to make you forget: not "what does this behaviour do?" but "who already does it, and what else is true of them?"