Why Does Regression to the Mean Happen?

Why Does Regression to the Mean Happen?

Regression to the mean describes a common pattern: when an observation is unusually high or unusually low, a later observation tends to be closer to the average. The phrase can sound as though something is actively pulling results toward the center. It is not. Regression to the mean is a consequence of variation, repeated measurement, and the fact that extreme observations often contain an unusually large random component.

The pattern appears in sports, education, health measurements, business metrics, and everyday decisions. A person may perform exceptionally well on one test and less exceptionally on the next. A patient may visit a clinic when symptoms are unusually severe and feel better later. A company may report an unusually strong month followed by a more ordinary month. These changes may be real, but regression to the mean alone does not tell us why they happened.

Signal Plus Random Variation

A useful way to think about an observed result is as a combination of a relatively stable signal and a temporary random component. The signal might represent underlying skill, long-term health, or a typical level of performance. The random component might include difficulty differences, fatigue, weather, chance events, sampling noise, or measurement error.

Suppose an observation is unusually high. It may be high partly because the underlying signal is high, but it may also be high because random influences happened to add to it. On a later measurement, those random influences may disappear, reverse, or simply be smaller. The later result therefore tends to be less extreme. The same reasoning applies in the other direction: an unusually low result may be followed by a higher one because an unusually negative random component is unlikely to repeat exactly.

This explanation does not require the underlying signal to change. In fact, regression to the mean can occur even when the person, process, or system remains exactly as capable as before. The key condition is that the measurements are imperfectly related: the first result gives useful information about the second, but it does not determine it completely. Penn State’s discussion of regression equations describes this relationship mathematically through predicted values that are generally less extreme than the original predictor when the correlation is below a perfect value.

A Synthetic Worked Example

Consider this entirely synthetic example. Imagine a fictional class taking a test twice, with each score centered around a synthetic class average of 70 points. Assume that every student’s underlying ability contributes to the score, while temporary test conditions contribute random variation. These numbers are illustrative and do not describe a real class or real students.

A fictional student earns a first score of 94 points. That result is 24 points above the synthetic average. For illustration, suppose the score consists of a stable ability component of 78 points plus a temporary positive component of 16 points, giving 94 points. On the second test, suppose the stable component remains 78 points, but the temporary component is only 4 points. The second score is then 82 points. The student has not necessarily lost ability; the unusually favorable first-test conditions simply did not repeat.

Now consider another fictional student with a first score of 46 points. That result is 24 points below the same synthetic average. Suppose the stable component is 62 points and the temporary negative component is negative 16 points, producing 46 points. On the second test, suppose the stable component remains 62 points and the temporary component is negative 4 points. The second score becomes 58 points. This is an improvement toward the average, but it does not by itself show that the student received effective help or suddenly worked harder.

The example is deliberately simple. Real measurements do not usually separate neatly into visible stable and random pieces. The decomposition is a teaching model, and every number in it is synthetic. Its purpose is to show why selecting people because they had extreme results can make later results look improved or worsened even without a causal intervention.

Why Selection Creates the Pattern

Regression to the mean becomes especially noticeable when people or cases are selected because of an extreme first result. If a school identifies students with the lowest synthetic test scores, a later average for that group may rise even without a new program. If a clinic studies patients on the day their symptoms are unusually intense, many may report less intense symptoms later. If a manager praises only the employee with an unusually high sales week, the next week may look disappointing by comparison.

The selection rule matters. Before selection, random variation is spread across high, middle, and low observations. After selecting only the extremes, the selected group is enriched for unusually favorable or unfavorable random components. Those components are not expected to persist at the same size. A later average therefore tends to move toward the overall average.

This is one reason before-and-after comparisons can be misleading. The change may be a mixture of genuine change, ordinary fluctuation, and regression to the mean. The larger the role of random variation in the initial measurement, the more room there is for regression. The stronger the relationship between the first and second measurements, the less movement toward the mean is generally expected.

Regression Is Not Causation

A common mistake is to interpret a return toward average as evidence that a treatment, punishment, reward, or policy caused the change. For example, if an athlete is given special coaching after an unusually poor performance and then performs better, the coaching may have helped. But the improvement could also reflect regression to the mean, normal variation, or both.

The reverse mistake is possible as well. A person selected after an unusually strong result may perform less strongly later, leading observers to conclude that praise, a bonus, or a change in routine harmed performance. That conclusion is also unjustified without a suitable comparison. Extreme results naturally create a difficult starting point for evaluating cause and effect.

A randomized comparison group helps. If selected participants are randomly assigned to an intervention or control condition, both groups can experience regression to the mean. The difference between their changes provides more useful evidence than the change in the treated group alone. Repeated measurements can also help estimate a typical level and reveal whether the apparent change is larger than ordinary fluctuation.

How Statistics Represents the Effect

Regression to the mean is closely related to correlation and prediction. A correlation describes how strongly paired measurements tend to move together; it does not mean that one measurement equals the other. When the relationship is less than perfect, the best prediction for a very high first result is usually high but closer to the average. The best prediction for a very low first result is usually low but also closer to the average.

NIST’s explanation of least-squares regression describes fitting a line that summarizes the relationship between variables. The fitted line can be used to predict one measurement from another, but predictions are not guarantees for individual cases. OpenStax likewise presents correlation and linear regression as tools for describing relationships and making predictions, with uncertainty remaining around those predictions.

This statistical idea is not a law that every extreme value must be followed by a less extreme value. Some high results are followed by high results, and some low results are followed by low results. Regression to the mean describes a tendency across many comparable cases, especially when the first measurement includes substantial random variation. It is a statement about expected patterns, not a promise about the next observation.

Practical Checklist

Before interpreting a change after an unusually high or low result, ask:

• Was the case selected because the first result was extreme?
• How much random variation or measurement error might the first result contain?
• Is there a comparison group that was not selected or treated in the same way?
• Were the same outcome and measurement process used again?
• Could ordinary regression to the mean explain some of the apparent change?
• Is the evidence strong enough to claim causation, or does it show only that the result changed?

The central lesson is simple: an extreme result is often a mixture of lasting signal and temporary chance. When the temporary part does not repeat, the next result tends to be closer to average. Understanding that mechanism helps prevent ordinary statistical variation from being mistaken for treatment success, failure, or a meaningful change in ability.