Regression to the Mean: Why Extreme Results Tend to Moderate

A student aces a test, then scores closer to average on the retest. A sports rookie has a phenomenal debut season, then a more ordinary "sophomore slump." A patient's blood pressure reading is unusually high at one visit, then more normal at the next, even with no treatment in between. In every case, a tempting explanation reaches for a real cause — the student got complacent, the athlete was overhyped, the treatment worked. Often, a large part of what's actually happening is regression to the mean: a purely statistical phenomenon, not a real change in the underlying thing being measured.

The Core Mechanism

Regression to the mean shows up whenever two things are true: a measurement involves some element of chance or noise on top of a real underlying value, and that measurement is repeated. Any individual reading is a mix of the true underlying value plus some random noise. An unusually extreme reading is more likely than an average reading to have been pushed to that extreme partly by favorable (or unfavorable) noise, not entirely by an equally extreme true value. On a second measurement, that particular noise doesn't repeat — so the second reading tends to land closer to the true underlying value than the first extreme reading did, which looks like "moving toward average" even though nothing about the underlying reality necessarily changed.

Worked Example: Predicting a Retest Score

Problem: A class's test scores have a mean of 70 and a standard deviation of 10. The correlation between students' scores on this test and a similar retest, across the whole class, is r = 0.6 (imperfect, since each test also reflects some random luck — which questions came up, how a student felt that day). A student scores 90 on the first test — 2 standard deviations above the mean. What score does the statistical relationship predict for that same student's retest?

When two variables share the same standard deviation, the predicted deviation on the second variable, in the same units as the first, is simply r times the original deviation — this follows directly from the regression slope formula, which reduces to exactly r when both variables have equal spread:

The student's original deviation from the mean was 90 − 70 = 20. Predicted retest deviation = 0.6 × 20 = 12. Predicted retest score = 70 + 12 = 82.

The model predicts this student's retest score will be 82 — still well above average, reflecting genuine underlying ability, but noticeably closer to the mean than the original 90. Nothing in this calculation assumes the student got lazier, was overconfident, or changed in any way — it's a direct, purely statistical consequence of the correlation between the two tests being less than a perfect 1.0.

Why It Specifically Affects Extreme Values

It's worth being precise about why this doesn't equally affect every score, only extreme ones. A student who scored exactly 70 (right at the mean) is predicted to score 70 again — 0.6 × 0 = 0 predicted deviation. Regression to the mean only pulls scores toward the average when they start away from it, and it pulls harder the further out they start (a perfect 100 would regress further, in absolute terms, than the 90 in the worked example). This is exactly why the phenomenon is most visible, and most likely to fool someone, at the extremes — a "best of the best" group selected specifically for an outstanding first performance is, as a group, guaranteed to look somewhat less outstanding on average the second time around, purely from this effect. The same logic runs in the opposite direction too: a "worst of the worst" group, selected the same way, tends to look somewhat less poor the second time around, for exactly the same underlying reason.

The Classic Trap: Mistaking It for a Real Effect

Regression to the mean becomes a genuine analytical trap whenever an intervention is applied specifically because a measurement was extreme, and then credited (or blamed) for the change that follows. A common real-world version: a company gives extra coaching specifically to its worst-performing salespeople this quarter, and next quarter their numbers improve — tempting to credit the coaching, but at least part of that improvement would likely have happened anyway, purely because an unusually bad quarter is disproportionately likely to be followed by a more typical one, coaching or not. Properly isolating the coaching's real effect requires comparing against a similarly extreme group that didn't receive the coaching — exactly the logic behind a genuine controlled comparison, rather than a simple before-and-after look at the group that was selected for being extreme in the first place.

A closely related version shows up constantly in medicine: a patient with an unusually high reading (blood pressure, cholesterol, a lab value) at one visit is often retested or given treatment, and the follow-up reading tends to be lower — a real phenomenon that a clinician might reasonably want to attribute partly to treatment, but that also reflects, at least in part, the same regression-to-the-mean pattern any two imperfectly correlated measurements of the same person would show even without any treatment at all. This is exactly why well-designed clinical studies use control groups rather than relying on simple before-and-after comparisons within a single group of patients selected for having an extreme initial reading.

Common Mistakes

Selecting a group specifically because it was extreme, then attributing its later moderation to an intervention. This is the single most consequential version of the mistake — always ask whether a comparison group that didn't receive the intervention would have regressed toward the mean by a similar amount on its own.

Assuming regression to the mean means "no real underlying difference exists at all." The worked example's predicted 82 is still meaningfully above the class average of 70 — regression to the mean moderates an extreme score, it doesn't erase the real underlying signal entirely, especially when the correlation r isn't extremely low.