Regression to the Mean: Why Extreme Measurements Drift Back, and How It Fools Us

When you measure something at its most extreme moment and measure it again later, the second reading tends to sit closer to the average, even if nothing was done in between. That drift is called regression to the mean, and it is one of the quietest reasons a treatment or a tweak gets credit it never earned. The mechanism is not mysterious. Any single measurement mixes a stable underlying value with a dose of luck, and when you select cases for being extreme, you have partly selected for unlucky moments that will not repeat. A second look catches the value on a more typical day, and it looks like improvement. This is an article about reading evidence, not medical advice; for anything about your own care, talk with a clinician who knows your history.

I meet this trap from both chairs. As a reviewer I have read confident before-and-after results that owed more to arithmetic than to any intervention. In my own work on the genetics of type 2 diabetes at the Lund University Diabetes Centre, where the same person's blood sugar wanders day to day, the gap between a true change and a lucky reading decides what you are allowed to believe.

What is regression to the mean, in one sentence?

Here is the quotable version. Regression to the mean is the tendency for an extreme measurement, chosen precisely because it was extreme, to be followed by a less extreme one, because part of what made it extreme was random and random parts do not persist. It is not a force pulling values toward the center. It is a consequence of how selection and chance interact, and it appears whenever a measurement is even slightly noisy.

The key word is selection. If you pick people, days, or hospitals because they posted an unusual number, you have picked the ones whose true value is unusual and also scooped up the ones who happened to catch a bad break that day. The true value sticks around at the next reading. The bad break does not.

A measurement is a signal plus a wobble

Think of any reading as two parts added together: a stable part, the person's real average level, and a wobble, the day to day noise from sleep, stress, the instrument, the meal before. A blood pressure of 165 today might be a true average of 150 plus a 15 point bad day, or a true 165 on a calm one. You cannot tell from one number which it was.

Now select everyone whose reading came in high. You have gathered people with genuinely high averages, but also people with ordinary averages who were caught on a wobble. Bring the whole group back next week and the wobbles reshuffle. The genuinely high stay roughly high, while the ordinary ones, having no reason to repeat last week's bad luck, drift down toward their real level. The group average falls, nothing was treated, and the arithmetic did the work.

Why this fools careful people

The illusion is convincing because it arrives wearing the costume of cause and effect. You did something between the two measurements, so the change feels like your doing.

Picture a clinic that enrolls patients with the highest readings into a new program, then measures them at three months and reports a drop. Some of that drop may be the program. Some of it is the inevitable settling of a group chosen for being at its worst. Without a comparison group chosen the same way and left untreated, you cannot separate the two, and the headline number quietly claims all the credit. The same shape appears whenever someone tries a remedy on their worst day, or praises and scolds staff after an unusual result: the thing that came next would have looked better anyway.

Where it shows up in health and research

The clearest case is any single screening value used to select people. Cholesterol, blood pressure, fasting glucose, a symptom score, a pain rating: all of them wobble, and all are routinely used to pick who gets an intervention. Select on a high reading and the follow-up will tend to look better whether or not the intervention did anything.

This is why the uncontrolled before-and-after study is so easy to oversell. It measures a group at a selected high point, applies something, measures again, and shows improvement, with regression baked into the design. I am not dismissing such studies; they are useful for generating ideas. The mistake is treating their improvement as proof of effect when a fair share was always going to happen. The same caution applies to my own field, where studying people at the extreme of a biomarker means their later readings often look more average, a softening easily mistaken for biology when part of it is statistical.

How good studies neutralize it

The fix is not clever statistics applied after the fact. It is design, decided before any data arrive.

A control group selected the same way

The single most powerful defense is a comparison group chosen by the same extreme criterion and then left untreated. Both groups regress toward the mean by roughly the same amount, so that drift cancels out, and any difference left between them is a fairer estimate of what the intervention actually did. This is one of the deepest reasons the randomized controlled trial earns its standing. It does not only balance confounders; it lets regression hit both arms equally so the comparison stays honest.

I have lived inside this logic. With EASY Diabetes, an AI clinical decision-support tool for type 2 diabetes that I co-developed, we ran a multi-clinic randomized controlled trial, EASY-1 (NCT03258268), evaluated against standard of care. Comparing program clinics against a control arm, rather than measuring enrolled patients before and after, is how you keep regression from masquerading as benefit.

More than one baseline measurement

When you must select on a value, averaging two or three readings before you start shrinks the wobble in your starting point, so there is less room left to regress. And the more a measurement varies on its own, the more regression you should expect, so a noisy reading taken once warrants more skepticism than a stable measure averaged over time.

What you can do as a reader

You do not need the formulas to read defensively. When you meet a claim that something improved after an intervention, ask the quiet question first: were these cases chosen for being extreme to begin with? If so, expect some improvement for free, and look for a control group selected the same way before you credit the intervention at all.

The point is not to distrust every improvement, because real effects are real and worth finding. It is to remember that the most extreme moment is, almost by definition, the one most likely to be followed by a calmer one. Calmer is not the same as cured. Tell the difference, and you will be fooled far less often, and frightened far less often too.