What Is Regression to the Mean?
Regression to the mean is the tendency for an unusually high or unusually low score to move closer to the average when it is measured again. If a student scores far above their usual level on one test, their next score will probably be lower — not because they got worse, but because the first score was partly luck. If a patient records an extremely high blood pressure reading on one visit, the next reading will probably be lower.
This is not a theory about people changing. It is a statistical fact about measurement. Whenever a score is a mix of a person’s real level and some random variation — a lucky guess, a bad night’s sleep, a noisy instrument — the extreme scores are the ones where the random part happened to push hardest in one direction. On a second measurement, that random push is unlikely to repeat in the same direction, so the score drifts back toward the person’s typical level. In other words, extreme results are rarely as extreme the second time around.
Regression to the mean shows up everywhere in research: test scores, sports performance, blood pressure readings, and before-and-after studies of all kinds. If you do not understand it, you can easily mistake a purely statistical drift for a real effect — and conclude that an intervention worked when nothing actually changed.
Where the Idea Comes From
The idea was first described in the 19th century by Francis Galton, who noticed that very tall parents tended to have children shorter than they were, while very short parents tended to have children taller than they were. Children’s heights, he observed, “regressed” toward the population average. Galton called it “regression towards mediocrity” — an unfortunate name that gave statistics the term “regression”, and eventually the whole technique of regression analysis.
The same logic applies to any measurement taken twice. Think of a student’s score on a test as two parts added together: their true ability, which is fairly stable, plus a random component — how they felt that day, which questions happened to appear, whether they guessed well on a few items. A student who scores far above their usual level probably had a good day as well as real ability. Next time, the good day is unlikely to repeat exactly, so the score falls back toward their true level. Nothing about the student has changed. Only the luck has evened out.
A Worked Example: The “Miracle” Tutoring Programme
Imagine a school identifies the 20 students with the lowest scores on a maths test and enrols them in a special tutoring programme. At the end of term, the students are tested again, and their average score has risen by 15 marks. The headteacher declares the programme a triumph. But was it?
Part of the improvement is probably real — tutoring does help. But part of it is regression to the mean. The students were chosen precisely because their first scores were extremely low. Some of those low scores reflected genuine difficulty, but some reflected a bad day: illness, anxiety, a misread question, a run of unlucky guesses. On the retest, those unlucky students score closer to their true level, so the group average rises even if the tutoring did nothing at all.
This is why a comparison group matters. If the school had also tracked a similar group of low scorers who received no tutoring, their scores would probably have risen too — by the amount explained by regression to the mean alone. The real effect of the tutoring is the difference between the two groups’ improvement, not the improvement of the tutored group on its own. Without a control group, regression to the mean masquerades as a treatment effect.
The same trap catches the other end of the scale. A company rewards its best-performing salespeople with a bonus, then watches their sales fall the next quarter and concludes the bonus made them lazy. More likely, their record quarter was partly a lucky one, and their sales simply regressed toward their usual level.
Why Extreme Scores Are the Unreliable Ones
Regression to the mean happens whenever two conditions hold:
- Measurement is imperfect. Every score contains some random error — mood, fatigue, guessing, instrument noise. The more error in the measurement, the stronger the regression.
- The same people are measured twice, and selected because of an extreme first score. Selection on an extreme value is what sets the trap. If you pick people at random and measure them twice, their group average barely moves. If you pick the top or bottom 10 percent, their average will move toward the middle on the second measurement.
The further a score is from the average, the more of it is likely to be luck. A score exactly at the average is probably close to the person’s true level; a score far above it very likely contains a large lucky component. That is why the most extreme cases always regress the most.
A famous illustration is the so-called sophomore slump: an athlete has a spectacular season, then a merely good one, and commentators search for explanations — complacency, pressure, opponents adapting. Often the simplest explanation is regression to the mean: the spectacular season was partly a lucky one, and a return to the athlete’s usual (still excellent) level was the most likely outcome all along.
Regression to the Mean vs Other Threats to Validity
Regression to the mean is often confused with other reasons scores change between two measurements. The table below separates them.
| Explanation | What actually happens | How to tell it apart |
|---|---|---|
| Regression to the mean | Extreme first scores drift toward the average on retesting because part of the first score was random error. | Only affects groups selected for extreme scores; a control group selected the same way shows the same drift. |
| Maturation | People genuinely change over time — children grow, patients recover, students learn. | Affects everyone, not just extreme scorers; the change follows a plausible developmental or recovery pattern. |
| Testing effect | Taking the test once changes the second score — people remember items or learn the format. | Scores improve even for average scorers; using a different form of the test reduces it. |
| History | Something outside the study happens between the two measurements and changes scores. | Affects the whole group at once, often suddenly; a control group exposed to the same event shows the same change. |
| A real treatment effect | The intervention genuinely changes the outcome. | The treated group improves more than an equivalent control group measured the same way. |
The key point: regression to the mean, maturation, testing, and history can all produce a change in scores with no treatment at all. A control group separates a real effect from these lookalikes, because the control group is exposed to all of them except the treatment.
How to Avoid Being Fooled
If you are designing a study or reading one, here are the practical defences:
1. Always use a control group. This is the single most important safeguard. If both groups were selected the same way, both regress by the same amount, and the difference between them is the real effect.
2. Measure the baseline more than once. Averaging two or three pre-treatment measurements gives a steadier estimate of each person’s true level, so the selection is less driven by one lucky or unlucky day. The less error in the baseline, the weaker the regression.
3. Do not select participants on a single extreme score. If you must study low scorers, confirm the low score with a second measurement before enrolling them. Better still, select on a variable related to the outcome rather than on the outcome itself.
4. Analyse with the baseline as a covariate. Instead of comparing simple change scores, methods such as analysis of covariance (ANCOVA) adjust each person’s follow-up score for where they started, which accounts for regression to the mean in the analysis itself.
5. Be suspicious of before-and-after claims without a comparison. “Patients’ pain fell by 40 percent after our programme” means little if the patients were enrolled at their worst — when pain was at its peak and most likely to ease on its own. Ask what happened to a similar group that received no treatment.
Common Mistakes
Mistake 1: Crediting an intervention for a natural drift. The classic error. A remedial class, a new drug, a coaching programme or a management intervention is introduced for the worst performers, scores improve, and the intervention takes the credit. Always ask for the control group.
Mistake 2: Punishing success. A team has its best month ever, then returns to normal, and the manager assumes motivation dropped. Before blaming people, check whether the exceptional month was partly luck.
Mistake 3: Thinking regression means scores become average. Regression to the mean is a drift toward the average, not a collapse into it. A genuinely gifted student who scores exceptionally will probably score a little lower next time — but still well above average. The true level does not disappear; only the lucky extra fades.
Mistake 4: Confusing it with the placebo effect. The two often occur together in medical studies, but they are different. The placebo effect is a genuine improvement caused by expectation and belief. Regression to the mean is a statistical artefact of selecting extreme cases. A good control group separates them.
Frequently Asked Questions
Is regression to the mean the same as the law of averages?
Not quite. The “law of averages” is a loose folk idea that things even out over time — after five heads in a row, a tail is “due”. That is a misunderstanding of randomness. Regression to the mean is a precise statistical effect: it applies to measurements that combine a stable true value with random error, and it only produces a drift when people are selected because of an extreme first measurement.
Does regression to the mean mean my intervention did nothing?
No — it means you cannot tell from a before-and-after change alone. Your intervention may have worked very well. The point is that part of the observed change is expected even with no intervention, so you need a control group to find out how much of the change is really yours.
Can regression to the mean happen with only one measurement?
No. It is a statement about what happens between two measurements. With a single measurement there is nothing to regress — although a single extreme score is still likely to contain a large random component, which is why retesting extreme cases is good practice.
How strong is the effect?
It depends on how unreliable the measurement is. With a highly reliable test, the drift is small; with a noisy measure — a single blood pressure reading, a one-off performance rating — it can be large. The weaker the correlation between the first and second measurement, the stronger the regression.
Key Takeaways
- Regression to the mean is the tendency for extreme scores to move closer to the average when measured again, because part of every extreme score is random error.
- It is strongest when people are selected because of a single extreme measurement, and when the measurement itself is unreliable.
- It explains the sophomore slump, the “miracle” improvement of the worst students, and many apparent treatment effects in before-and-after studies.
- It is not a real change in people — it is a statistical drift — and it is not the same as maturation, testing effects, or the placebo effect.
- The defences are simple: use a control group, measure the baseline more than once, and be sceptical of any before-and-after claim that has no comparison group.
Once you can spot regression to the mean, you will see it everywhere — in research papers, in performance reviews, in sports commentary, and in health claims. Recognising it does not make you cynical; it makes you a more careful reader of evidence, and a better designer of studies of your own.