Regression toward the mean is a statistical phenomenon in which observations selected because they are unusually high or low tend, on average, to be less extreme on another related measurement. A student selected for an exceptionally high examination score, for example, may score less exceptionally on a comparable second examination even if the student’s underlying ability has not changed. The phenomenon arises because an extreme observation can reflect both persistent characteristics and temporary influences that are unlikely to recur in exactly the same combination.
Regression toward the mean is important in statistics, scientific experimentation, performance assessment, and forecasting because it can make ordinary variation look like an effect of treatment, punishment, reward, or policy. It concerns conditional averages, not a rule that every extreme observation must be followed by a more ordinary one. Its magnitude depends on the relationship between measurements, the population being considered, and the way observations were selected.
Originally identified in nineteenth-century studies of heredity, regression toward the mean helped give regression analysis its name. Its modern explanation draws on conditional expectation, imperfect correlation, measurement error, and selection. These connections also clarify its relationship to statistical shrinkage and distinguish it from a physical or economic process that actively returns a variable to equilibrium.
Meaning, scope, and common misconceptions
The central feature of regression toward the mean is the comparison between an extreme selected observation and a second, imperfectly related observation. Suppose two examinations have the same average score and the same spread of scores. If the relationship between them is positive but imperfect, the average second-examination score of students selected for very high first-examination scores will commonly lie closer to the overall average. Students selected for very low first-examination scores will commonly show the opposite movement. Neither pattern requires any change in instruction or ability.
The relevant “mean” is the expected value for an appropriate reference population, sometimes after accounting for known characteristics. A professional athlete’s unusually strong performance need not predict a return to the average of all people; the relevant comparison may be the athlete’s own expected performance or that of similarly skilled competitors. Likewise, a pupil’s result may need to be evaluated relative to pupils of the same age or preparation. Using the wrong reference group can turn a valid statistical principle into a misleading prediction.
Regression toward the mean does not imply that everyone becomes average, that differences between individuals disappear, or that a high observation must be followed by a low one. With a strong positive relationship between measurements, an exceptionally high first score usually still predicts an above-average second score. The prediction is simply less extreme. Nor does the phenomenon require a sequence in time: one variable can show regression toward its mean when conditioned on another variable measured at the same time.
It is also distinct from the law of large numbers. That law concerns the stabilization of averages as the number of observations increases under suitable assumptions. Regression toward the mean concerns the conditional relationship between measurements after selection. A second individual result may regress toward the mean even though no larger average has been calculated. Conversely, averaging many measurements can reduce temporary variation and therefore reduce the amount of regression expected on a later measurement.
Imperfect correlation alone does not establish that every conditional average moves toward the mean at every point. The familiar statement is exact under important models, especially a bivariate normal model, and provides a useful linear-prediction description more generally. For arbitrary nonlinear relationships, conditional means can behave differently in different regions. Careful use therefore specifies whether the claim concerns an exact conditional expectation, a best linear prediction, or an empirical tendency within a particular dataset.
Historical origins in heredity research
Francis Galton developed the idea through investigations of biological inheritance. In experiments with sweet peas during the 1870s, he compared seed sizes across generations and found that descendants of unusually large seeds were, on average, less unusually large, while descendants of unusually small seeds were less unusually small. He initially described a related pattern using the term “reversion.” These investigations helped direct attention toward the relationship between parental and offspring measurements rather than simply toward population averages.
His best-known treatment appeared in the 1886 paper “Regression Towards Mediocrity in Hereditary Stature.” Galton compared adult offspring heights with a combined measure of parental height, making an adjustment to put male and female heights on a comparable scale. Tall parents tended to have tall offspring, but the offspring’s average departure from the population mean was smaller than the parents’ departure. Short parents displayed the corresponding pattern. In this historical usage, “mediocrity” referred to the population’s central or average stature, rather than to a judgment about personal worth.
The finding raised an apparent puzzle: if children of tall parents were less tall on average, and children of short parents were less short, why did variation in height not steadily disappear? The resolution is that a conditional average is not a complete description of the next generation. Offspring vary around the average predicted from parental height. In a suitable stable population model, this residual variation can preserve the overall distribution even while the conditional mean is closer to the population mean. A compressed set of predictions therefore does not imply a correspondingly compressed distribution of actual outcomes.
Galton’s work contributed to the development of correlation and regression, subsequently formalized by Karl Pearson and other statisticians. Galton’s Natural Inheritance, published in 1889, developed these statistical approaches further. The modern term “regression” survives in models whose subject matter has nothing to do with heredity and whose fitted relationships need not represent movement toward any average.
The historical example also exposes the limits of interpreting regression as a directional biological force. A statistical relationship can be examined in either direction: parental height can be predicted from offspring height, just as offspring height can be predicted from parental height. Both predictions may be less extreme on standardized scales. This symmetry is possible because the questions condition on different groups. Modern explanations consequently separate the mathematical phenomenon from the particular hereditary mechanisms and social assumptions associated with Galton’s research.
Conditional expectation and the mathematical model
Let and be two random variables with means and , positive finite standard deviations and , and correlation . If their joint probability distribution is bivariate normal, the conditional expectation of given is
On standardized scales, with and , this becomes
For , the expected standardized departure on the second measurement is smaller than the observed standardized departure on the first. If , a first measurement two standard deviations above its mean predicts a second measurement 1.2 standard deviations above its mean. For tests with mean 100 and standard deviation 15, a first score of 130 therefore predicts a second score of 118. These are model-based predictions, not guaranteed individual outcomes; the bivariate normal model also gives conditional variance .
When the two measurements have the same mean and the same standard deviation, the expected change has a particularly clear form:
An above-average starting value predicts a negative change, while a below-average starting value predicts a positive change. The expected proportional reduction in departure from the mean is . If the measurements are independent and identically distributed, the expected second value is simply , although the actual second result may still be extreme.
The corresponding linear expression remains the best affine predictor under mean squared error whenever the required second moments exist. Outside the bivariate normal model, however, it need not equal the true conditional expectation. A nonlinear conditional mean can depart substantially from its best-fitting straight line. Moreover, a raw regression slope below one is not by itself evidence of regression toward the mean: changing measurement units changes the slope. Standardization removes this arbitrary dependence on units.
The endpoint cases clarify the interpretation. With perfect positive correlation, standardized values are exactly preserved, so there is no regression on that scale. With negative correlation, the predicted standardized value reverses sign; when , its magnitude is still smaller under the linear model. Unequal means or variances also matter. A population-wide upward trend can coexist with regression toward the mean in standardized departures, so “less extreme” need not mean numerically lower.
Stable differences, temporary variation, and measurement reliability
A useful explanation separates a stable individual component from temporary variation. Suppose repeated measurements satisfy
where is individual ’s persistent departure from the population mean and the two error terms have mean zero. Assume that the errors are mutually independent, independent of , and have common variance . If the variance of is , then
The resulting repeated-measure correlation is
This ratio is also a classical reliability coefficient under the stated model. A high observed value can result from a large persistent component, a large positive temporary component, or both. Among observations selected for being high, temporary components tend to be positive. Because the second temporary component is independent, it does not inherit that selected positive contribution. Regression follows without any change in the stable component.
“Error” here is broader than a defective instrument. It can include temporary physiological fluctuations, favorable examination questions, variable playing conditions, or other influences not intended to represent the persistent attribute. A perfectly functioning measuring device can therefore record data that regress toward the mean. Conversely, systematic influences shared across occasions can make the errors correlated, altering the relationship between test–retest correlation and reliability. The simple variance ratio is not valid without its independence assumptions.
Consider a hypothetical assessment with stable between-person variance 144 and temporary variance 81. Total variance is 225, giving a standard deviation of 15 and repeated-measure correlation . Under the corresponding Gaussian model, a score 30 points above the population mean predicts a repeat score 19.2 points above it. The expected decline of 10.8 points reflects the statistical decomposition of the unusually high initial observation, not a predicted loss of the underlying attribute.
Averaging independent repeated measurements reduces the temporary variance. For an average of measurements, the temporary component has variance , while the stable between-person variance remains unchanged. Such an average generally predicts subsequent performance better than a single noisy measurement. The improvement is smaller when the errors are correlated or the underlying attribute changes over time. This distinction explains why instrument precision, temporal stability, and repeatability are related but not interchangeable concepts.
Selection of extremes and apparent improvement
Regression toward the mean becomes especially consequential when selection depends on the same measurement later used as a baseline. Examples include selecting schools after exceptionally low examination results, road junctions after unusually many collisions, or employees after unusually high output. The selected baseline is not a typical observation for those units. Even if their underlying characteristics remain unchanged, the average temporary component at selection can be unusually favorable or unfavorable.
Under the equal-mean, equal-variance linear conditional model, selection above a threshold gives
The expected difference between the selected baseline mean and the follow-up mean is therefore proportional to . For a normal distribution, write , and let and denote the standard normal density and distribution functions. Then
Selection farther into the tail increases the selected group’s initial departure and thus generally increases the expected absolute regression under this model.
Selection can occur indirectly. A prize awarded for the best annual result, a news story about the most successful investment fund, or a ranking of the fastest-growing businesses all emphasize extreme realized outcomes. When persistent differences and temporary influences both contribute, the observed leaders tend to include units that benefited from unusually favorable temporary conditions. Their subsequent average performance can fall without implying that the award, publicity, or ranking caused deterioration.
This connection is related to selection bias, but the concepts are not identical. Regression toward the mean describes a conditional relationship; selection bias describes a distortion relative to a specified inferential target. Conditioning on an extreme baseline does not automatically invalidate every comparison. It becomes problematic when the resulting expected change is treated as evidence of a causal effect or as a typical trajectory for the wider population.
There is also an algebraic reason for caution when relating baseline values to change scores. If , then
For equal variances , this equals , which is negative whenever . A negative association between baseline and change can therefore appear simply because baseline is subtracted in constructing the change score. Such a pattern is not, by itself, evidence that initially high-scoring individuals experienced a stronger adverse influence or that initially low-scoring individuals benefited more.
Consequences for experiments and causal evaluation
Regression toward the mean presents a basic problem for causal inference: change after an intervention is not necessarily change caused by it. If an intervention begins because an outcome has become unusually poor, some subsequent improvement may have occurred without the intervention. A simple before-and-after comparison combines any treatment effect with regression toward the mean, general time trends, and other changes affecting the outcome.
The distinction can be expressed using potential outcomes. Let denote a follow-up outcome under treatment and the follow-up outcome without treatment. The causal effect concerns , whereas a treated individual’s observed before-and-after difference is . These quantities differ because the untreated follow-up outcome need not equal the selected baseline . Regression toward the mean is one reason that equality can fail even in the absence of a genuine underlying trend.
A randomized controlled trial addresses this problem by comparing treatment groups formed through random assignment. If participants are selected for extreme baseline measurements and then randomized, both groups share the selection mechanism and, in expectation, the same regression tendency. A between-group follow-up contrast can identify an average treatment effect under the trial’s assumptions. The improvement within the treatment group alone still does not measure efficacy, even when that within-group change is statistically significant.
Baseline adjustment can improve precision. A common model is
where indicates treatment assignment. In an appropriately specified randomized-trial analysis, represents the treatment contrast conditional on baseline. This approach, commonly called analysis of covariance, does not work by mechanically deleting regression toward the mean. Its role is to use predictive baseline information while preserving the randomized comparison. Heterogeneous effects, missing observations, and nonlinear baseline relationships require additional care.
Nonrandomized evaluations need stronger design-specific justification. Repeated pre-intervention observations can reveal whether an extreme baseline was an isolated fluctuation or part of a persistent pattern. Comparison units must have a credible relationship to the treated units’ untreated outcomes, especially when treatment was triggered by an extreme event. Even difference-in-differences can be misleading if a selected temporary shock makes the groups’ subsequent untreated trajectories diverge. Statistical significance, a large sample, or a sophisticated regression equation does not by itself supply the missing counterfactual.
Bayesian prediction and statistical shrinkage
Regression toward the mean has a close mathematical relationship to Bayesian inference. Suppose an individual’s latent value has a prior distribution
and an observed measurement follows
After observing , the posterior distribution has mean
The estimate lies between the observation and the population mean. The weight expresses the relative strength of persistent variation and observational noise.
This operation is called shrinkage because departures from a common center are reduced. If measurement noise is small relative to between-individual variation, the estimate remains close to the observed value. If noise is large, the estimate moves farther toward the population mean. Under squared-error loss, the posterior mean minimizes posterior expected loss. Shrinkage therefore has a decision-theoretic justification rather than being an arbitrary preference for moderate results.
Suppose a future measurement is conditionally independent of given , with conditional mean . Then
The shrunken estimate of the latent value is also the predicted future measurement. This connects the Bayesian model directly to regression toward the mean. The corresponding posterior predictive distribution nevertheless includes both uncertainty about and the variability of the future observation. A moderate point prediction does not imply that the next observed result will be close to that prediction.
Hierarchical models extend this reasoning to many groups, such as schools, hospitals, athletes, or geographical areas. Groups with little data or highly variable outcomes generally receive more shrinkage than groups with precise measurements. In empirical Bayes methods, population-level quantities are estimated from the collection of groups. The chosen hierarchy matters: shrinking an institution toward a poorly matched population can obscure relevant differences rather than improve estimation.
The relationship to regularization is close but not an identity. Methods such as ridge regression shrink estimated coefficients to improve prediction or stabilize estimation. Regression toward the mean describes a relationship among observations or their conditional expectations, whereas shrinkage is an estimation procedure. A deliberately biased estimator can have lower overall prediction error, and an unadjusted estimator can be unbiased before selection yet exaggerate the latent values of units selected for extreme observed estimates. These are distinct questions about prediction, estimation, and conditioning.
Applications in performance, public services, and research
In sport, exceptional performances often combine persistent skill with temporary influences such as opponents, weather, health, and chance. A player selected for an outstanding season may remain highly skilled while producing a less exceptional following season. Differences in playing time also matter: a very high success rate based on a few attempts is generally less informative than the same rate based on many attempts. Appropriate prediction therefore considers both the observed rate and its precision, rather than treating every position in a ranking as equally stable.
Educational assessment presents a similar structure. A school selected after a single year of unusually low results can improve in the following year without a change in educational effectiveness. Cohort composition, examination content, attendance, and sampling variation may all contribute to annual fluctuations. Comparisons based on several cohorts, relevant pupil characteristics, and uncertainty estimates address different components of the problem. They do not make regression disappear, but they can distinguish a persistent school-level signal from an unusually favorable or unfavorable year.
Road-safety evaluation provides a count-data example. A junction may receive an intervention because its recent collision count is exceptionally high. Even without engineering changes, a high count can be followed by a lower count because realized events fluctuate around an underlying rate. Models based on a Poisson distribution or more flexible count distributions can distinguish observed counts from estimated long-run risk. Empirical Bayes evaluations commonly combine a site’s history with predictions from comparable sites, while accounting for factors such as traffic exposure and broader trends.
In health research, participants are sometimes recruited because a measured symptom level, physiological value, or other outcome exceeds a threshold. Repeated measurements can then be less extreme because of ordinary biological variability and measurement variation. This statistical issue is separate from any particular disease mechanism or treatment. A change in an untreated comparison group can reflect regression toward the mean, natural history, and other influences; it should not automatically be attributed to a single explanatory category.
Scientific replication introduces a related selection process. Studies with unusually large estimated effects are more likely to pass a significance threshold or attract attention. Their estimates may combine a genuine effect with favorable sampling error, so an independent replication can produce a smaller estimate even when the underlying effect is unchanged. Publication bias and selective reporting can intensify this pattern. Smaller replication estimates do not by themselves prove that an original effect was absent, but they show why selected discovery estimates and their uncertainty require careful interpretation. The relevant comparison is between noisy estimates and plausible underlying effects, not merely between two headline results.
Time-series mean reversion and limits of the analogy
The expression “mean reversion” is also used for dynamic behavior in a time series. A common example is the first-order autoregressive model
where the innovations have conditional mean zero. When , repeated substitution gives
For , the expected departure from decays progressively with the forecast horizon. With appropriate finite-variance innovations and initialization, the process has a stable long-run distribution.
This model resembles the standard regression-toward-the-mean equation, but the interpretations need not be identical. A dynamic model specifies how the process evolves through time. Regression toward the mean can arise merely from conditioning on a noisy observation of an unchanged latent value. For example, repeated measurements may all share the same fixed individual attribute. Conditional on one extreme observation, other measurements are predicted to be less extreme, but there need be no gradual temporal return or feedback mechanism. Under a simple independent-error model, the conditional prediction can be the same for every other measurement occasion.
A random walk provides an important contrast:
With zero-mean innovations, its conditional forecast is the current value, not a fixed population mean. An extreme level does not therefore imply an expected reversal. In financial or economic applications, the existence of a historical average is insufficient to establish mean reversion; trends, changing parameters, and nonstationarity can make a fixed reference mean inappropriate. Even a valid mean-reverting model need not yield a reliable prediction of the next individual movement.
Distributional assumptions also limit broader claims. A variable may have no finite mean, as in the standard Cauchy distribution, or no finite variance, making the usual correlation-based formulas unavailable. The phenomenon may still have analogues expressed through conditional medians or other summaries, but those require separate definitions. For multimodal populations, a single overall mean may lie between substantively different groups and provide an unsuitable predictive target.
Finally, apparent regression can coexist with real change. Learning, deterioration, treatment, environmental shifts, or institutional reform may move the underlying distribution while temporary variation still produces regression within it. Determining their relative contributions requires a specified reference population, a model for repeated measurements, and evidence about the untreated or otherwise relevant trajectory. The statistical pattern alone supplies neither a causal explanation nor a universal law of return.