AI Glossary / Statistics and probability / Intermediate

Regression to the mean

Definition

Regression to the mean is the tendency for observations selected because they are unusually high or low to be followed, on average, by less extreme observations when repeated measurements are imperfectly correlated. It can occur without any intervention or underlying improvement. The tendency reflects the relationship between persistent differences and temporary variation. It is an average conditional pattern, not a rule that every extreme case must move toward the mean on its next measurement.

What it means

An extreme observation can combine a genuinely unusual underlying level with a temporary fluctuation. When cases are selected using that observation, the fluctuation is part of why they were selected. On repetition, the same favorable or unfavorable fluctuation need not recur. The next measurement can therefore be less extreme even if nothing about the underlying process changed. The relevant mean and magnitude of the tendency depend on the population, measurement process and relationship between repeated observations.

How it works

Examine whether cases were chosen because of extreme baseline values, recent peaks or poor performance. Assess how stable the measurement is across repetitions and whether temporary variation is substantial. Compare changes with an appropriate control or design rather than attributing every before-and-after difference to an intervention. Repeated baseline measurements can help distinguish persistent level from transient noise. Statistical models can describe expected conditional changes, but they must reflect selection, dependence and the intended target population.

Three practical examples

Example 1: A team selected for unusually poor monthly performance performs less poorly the next month even without a process change. Example 2: Customers chosen for an exceptional spending spike spend less in the following period. Example 3: Items selected because of unusually high noisy measurements show lower averages when remeasured. These examples do not imply that interventions never work. They show why improvement after selecting extreme cases is not sufficient evidence that an intervention caused it.

Think of it this way

Think of a strong underlying signal with a temporary gust added to it. Selecting the highest reading tends to select both strong signals and favorable gusts. On the next reading, the gust may be ordinary, so the value is less extreme even though the underlying signal has not weakened. The analogy explains selection on noisy outcomes, not a force physically pulling every observation back to a central value.

Worked example

Suppose two measurements X and Y follow a bivariate normal model with means 100, standard deviations 15 and correlation 0.60. The conditional mean of Y given X = x is 100 + 0.60(x - 100). For a first measurement of 130, the expected repeat is 118, twelve points lower. For a first measurement of 70, the expected repeat is 82, twelve points higher. No treatment effect is included in this model. Conditional standard deviation is 15 × square root of (1 - 0.60 squared) = 12, so individual repeats still vary substantially around those expectations. The formula's shrinkage toward 100 is a property of this specified model, not a universal correction that can be applied to every dataset.

Illustrative case study

Fictional case study: Cedar Operations introduces coaching for teams selected after their worst month. Their next month's average improves, and an initial report credits coaching entirely. Analysts explain the selection effect and design a comparison that can separate coaching from expected fluctuation and broader time trends. This fictional case does not claim coaching is ineffective. It illustrates why a simple before-and-after improvement among extreme performers is insufficient to estimate its effect.

Common misconceptions and mistakes

Common mistakes include interpreting regression to the mean as inevitable decline after success, confusing it with a genuine time trend, and assuming the mean itself must remain fixed in a changing process. It is not the gambler's fallacy: independent future coin flips do not become more likely to compensate for earlier outcomes. The phenomenon also does not say that persistent high performers become average. Their later values can remain above average while being less extreme than the selected measurement.

Business use

Business use: The concept is important in performance management, customer targeting, quality investigations and program evaluation. Identify selection based on unusual recent outcomes before interpreting change. Use suitable comparison groups, randomization where appropriate or models that account for baseline selection and measurement variation. Report what change would be expected without the intervention. Avoid rewarding or penalizing individuals solely because a noisy selected extreme was followed by a more ordinary value.

Check your understanding

Why does the model predict 118 after observing 130 even without treatment?

Answer: Repeated measurements are imperfectly correlated, so the conditional expectation is less extreme.

Does 118 mean that every repeat will equal 118?

Answer: No. The conditional standard deviation is twelve in this example, representing substantial remaining variation.

Learn next

Simulate paired measurements with different correlations and select the highest baseline observations. Compare their average repeats without adding any intervention. Then add a real intervention effect and examine why both components can be present. Study evaluation designs that distinguish selection-driven change, secular trends and treatment effects before interpreting performance improvements.

Sources

Sources: Bland and Altman, Statistics Notes: Some examples of regression towards the mean, BMJ: https://www.bmj.com/content/309/6957/780 ; Harrell, Biostatistics Glossary: https://hbiostat.org/glossary/ . These author-written methodological sources explain the phenomenon and interpretation. The bivariate-normal calculation and fictional coaching evaluation are original illustrations with explicit modelling assumptions.

Previous: Regression surface Next: Regression tree