Regression to the mean is the statistical tendency for an extreme measurement (very high or very low) to be followed by one closer to the average, simply because extreme values often owe part of their size to chance. It is not a force that “pulls” scores back, and it does not mean performance is bound to fall after a good result or rise after a bad one — it describes what random variation predictably does over repeated measurements. This guide defines regression to the mean, explains its causes, walks through a worked example, and sets out exactly how to avoid letting it distort your conclusions when you design and analyse research.
Regression to the mean is a concept that has circulated in scholarly source materials and statistical literature for over a century, ever since Sir Francis Galton observed in the 1880s that the children of very tall parents tended to be tall, but on average closer to the population mean than their parents. The term appears constantly in discussions of performance, investments, clinical trials and health outcomes, yet it is one of the most frequently misunderstood ideas in quantitative research. Below we draw on both primary source and secondary source evidence to explain what it really means, why it trips up researchers, and how to design studies that do not mistake a statistical artefact for a genuine effect.
What Is Regression to the Mean?
Regression to the mean is a statistical phenomenon in which a variable that records an extremely high or low value on one measurement tends to be closer to its average value on the next measurement. It arises whenever a measured outcome is not perfectly reliable — that is, whenever it combines a stable underlying value with some random fluctuation. When an observation lands at the extreme, it is disproportionately likely that the random component was unusually large at that moment; on the next occasion that random luck is unlikely to repeat, so the value drifts back toward the typical level.
Crucially, regression to the mean is a property of imperfect reliability and validity in measurement, not a description of how the world behaves. Imagine testing students with a series of quizzes. If a normally average student scores exceptionally high on one quiz, their next score is likely to be lower — closer to their personal average. If they score exceptionally low, their next score is likely to be higher. The students have not changed; the random element (a lucky guess, a good night’s sleep, a familiar question) simply did not recur. The phenomenon does not predict that any specific person will improve or decline. It predicts that, as a group, the extremes will be less extreme next time.
Worked Example: The Basketball Free Throws
What actually happened? The 90% night was an outlier — his true skill plus an unusually good run of luck. The 60% night is regression to the mean: the luck did not repeat, so the score moved back toward his real ability. It is a common mistake to reason, “he did worse because he got overconfident” or “he practised less because he thought he had peaked.” That reasoning is a form of confirmation bias — we invent a causal story to fit the change we expected to see. In reality the 60% is just natural variability. The same logic runs the other way: if Alex had a terrible night and sank only 10% — the kind of floor that can sit alongside a ceiling effect at the opposite extreme of a scale — we would expect his next game to climb back toward 50%, with no change in skill required.
The free-throw case captures the essential danger for researchers: whenever you select cases because they are extreme — the best performers, the sickest patients, the most accident-prone roads — and then measure them again, some movement toward the average is guaranteed before any intervention has had a chance to act. Confuse that movement with a treatment effect and your conclusions collapse.
What Causes Regression to the Mean?
Regression to the mean has a single root cause expressed through several mechanisms. The root cause is that observed values mix a stable “true” component with random noise. The strength of the effect depends on how much noise there is — formally, on the correlation between the two measurements. The weaker that correlation (the less reliable the measure), the stronger the regression.
| Driver | Why it produces regression | Where you see it |
|---|---|---|
| Measurement error | Random noise inflates or deflates a single reading; the noise rarely repeats on remeasurement. | Blood pressure, test scores, reaction times |
| Selection on an extreme | Choosing cases because they are top or bottom guarantees they were partly lucky/unlucky. | “Worst” schools, “best” funds, sickest patients |
| Low test–retest correlation | The less two measurements correlate, the further extremes fall back toward the mean. | Subjective ratings, noisy biomarkers |
| Natural fluctuation | Many quantities oscillate around a baseline; a peak is usually followed by a return. | Symptoms, sales, weather, economic output |
Note what is not on this list: motivation, effort, willpower, or any psychological rebound. Those are the stories observers tell to explain a change that statistics already predicted. The single inline figure below shows the mechanism in one picture.
Why Regression to the Mean Is a Problem in Research
Regression to the mean is not a flaw in nature; it is a flaw in inference. The problem appears when researchers, clinicians or policymakers fail to account for it and credit an intervention with a change that would have occurred anyway. The following contexts are where it does the most damage.
Misinterpreting treatment efficacy
Suppose patients with the most severe symptoms are recruited for a new drug trial and their symptoms ease over the following weeks. It is tempting to credit the drug. But because the patients were selected at their worst, some improvement was statistically inevitable — they were partly at an unlucky low. Without a comparison group, you cannot separate the drug’s effect from regression. This is one reason uncontrolled “before-and-after” studies overstate benefits.
Performance evaluation
An athlete or sales team has a stellar season and then a merely average one. Managers may blame complacency when the earlier peak was simply an outlier. The reverse — punishing a poor quarter that then “recovers” — wrongly rewards the punishment.
Educational assessment
If pupils who scored poorly on a test are given extra tuition and then score nearer their average, the tuition may be credited with the entire gain when regression accounts for part of it. Designing the evaluation without a control group bakes in an overestimate.
Policy and decision-making
When an intervention follows an extreme event — a year with a very high crime rate, say — and the next year returns to normal, an affinity bias may push officials toward measures they already favour and credit them with the fall. The intervention is praised when the extreme year was the anomaly and a return to the mean was always likely. The remedy is a bias for action grounded in evidence rather than in the comforting assumption that whatever we did must have worked.
Investment decisions
Investors often assume a top-performing fund will keep winning, or a laggard will keep losing, ignoring that extreme returns regress. “Reversion to the mean” is the financial sibling of this idea.
Research bias more broadly
In any empirical field, failing to account for regression is itself a recognised form of research bias. It is especially dangerous in medicine, where mistaking a statistical artefact for a real effect can shape clinical guidelines.
Regression to the Mean vs Related Effects
Researchers often confuse regression to the mean with other reasons a follow-up measurement might differ. Distinguishing them is essential because each demands a different control.
| Effect | What changes the second measurement | How to counter it |
|---|---|---|
| Regression to the mean | Random error that inflated the extreme does not recur. | Randomised control group, repeated baselines |
| Hawthorne effect | Participants change behaviour because they know they are observed. | Blinding, unobtrusive measures |
| Pygmalion effect | Researcher or teacher expectations alter participant performance. | Blinding researchers to group allocation |
| Placebo / natural recovery | Genuine improvement from belief or the passage of time. | Placebo control arm |
For the full picture of how the Hawthorne effect distorts observed behaviour, and how the Pygmalion effect works through expectations, see our dedicated guides — both can mimic regression in a poorly controlled design.
Common Examples of Regression to the Mean
These everyday cases show how widely the phenomenon reaches once you start looking for it.
Sports performance
An athlete has an extraordinary season followed by one nearer their career average. The dip is often a return to the mean, not a loss of skill — the basis of the famous “sophomore slump” and the “Sports Illustrated cover jinx”.
Medical treatment
A person with chronic pain tries a new remedy on a particularly bad day. The next few days feel better, and the remedy gets the credit — even though the bad day was the outlier and some easing was likely regardless.
Academic testing
A student scores exceptionally high on one standardised test, then closer to their average on the next. This need not signal declining ability; it is regression at work.
Traffic enforcement
A road safety team installs speed cameras where accidents peaked the previous year — sometimes influenced by explicit bias about which roads are “dangerous”. Accidents fall the next year and the cameras are praised, even though sites chosen at an accident peak would have improved anyway. This is one of the most documented real-world traps in safety evaluation.
Economic indicators
A year of exceptionally high growth is often followed by a more modest year. Many forces drive growth, but a simple return to typical levels can be one of them.
Business sales
A new product enjoys a launch spike from novelty, then settles at a lower but more sustainable level — a regression to a realistic mean rather than a sign of failure.
How to Identify Regression to the Mean
Before you can avoid it, you must spot it. Work through these checks whenever a follow-up measurement looks better or worse than the first.
- Examine the selection criterion. Were cases chosen because they were extreme (highest, lowest, sickest, riskiest)? If yes, expect regression.
- Plot the paired data. A scatter plot of first versus second measurement reveals the pattern: extreme first-measurement points sitting nearer the mean on the second axis.
- Compare against a control group. If an untreated group selected the same way improves by a similar amount, regression — not your intervention — is the driver.
- Use repeated measures. Several baselines before the intervention let you see whether values were already drifting toward the mean.
- Weigh alternative explanations. Confounders, seasonality or external events can all masquerade as treatment effects.
- Beware vivid memories. People recall extreme events strongly, creating a perception bias in which any less-extreme follow-up feels like meaningful change. A disciplined source-evaluation method helps separate genuine patterns from randomness.
How to Avoid Regression to the Mean
You cannot abolish regression to the mean — it is a mathematical consequence of imperfect measurement. What you can do is design and analyse studies so that it does not masquerade as a real effect. The following strategies, used together, neutralise it.
1. Use a randomised controlled design
A control group selected by the same extreme criterion will regress by the same amount. Randomly allocating participants to treatment and control means regression affects both arms equally, so any difference between them reflects the intervention, not the artefact. This is the single most powerful safeguard, which is why randomised controlled trials are the gold standard.
2. Take multiple baseline measurements
Averaging two or more pre-intervention readings gives a more reliable starting point. Because the random error partly cancels across readings, the averaged baseline is less extreme — and there is less room for spurious regression.
3. Avoid selecting on a single extreme score
Where possible, recruit on a stable characteristic rather than on one unusually high or low reading. If you must target extreme cases, select them using a measurement that is separate from the one you will later use to judge the outcome.
4. Improve measurement reliability
The stronger the test–retest correlation, the weaker the regression. Better instruments, clearer protocols and trained assessors all raise reliability — a direct link back to the principles of reliability and validity that underpin sound measurement.
5. Adjust statistically
Analysis of covariance (ANCOVA), which uses the baseline score as a covariate, is the standard way to correct change scores for regression. Galton’s own regression coefficient quantifies expected movement toward the mean and can be built into the model.
6. Pre-register and use a control for natural history
Pre-specifying the analysis stops you from rationalising a regression-driven change after the fact, and a control arm captures any natural recovery or trend that would have happened anyway.
“Regression to the mean is as inevitable as the law of gravity. If a population is selected to have an extreme value at one point, it will tend to be less extreme on a later measurement.” — J. Martin Bland and Douglas G. Altman, BMJ
A Practical Checklist for Researchers
Before you attribute a change to your intervention, run through this list. If you cannot tick the relevant boxes, regression to the mean may be inflating your result.
- I included a randomised control group selected by the same criterion.
- I used more than one baseline measurement and averaged them.
- I checked my instrument’s test–retest reliability.
- I adjusted for baseline scores using ANCOVA where appropriate.
- I did not select participants on a single extreme reading judged by the outcome measure.
- I considered and ruled out confounders, the Hawthorne and Pygmalion effects, and natural recovery.
Treating regression to the mean as a design problem rather than an afterthought is one of the clearest markers of a rigorous quantitative study — and a recurring theme across the wider family of pitfalls catalogued in our research bias hub.
Worried regression to the mean is skewing your results?
Our subject specialists can design controls, run the right statistical analysis and write up your findings with confidence.
If your concern is specifically the numbers — choosing the right model, running ANCOVA, or interpreting paired measurements — our Learn More statistical analysis service can take that weight off your shoulders.
