The misconception
Select the ten highest scores from a large first test and measure those people again. If their average falls, an intervention introduced between tests may appear to have harmed performance. Select the ten lowest and their average may rise, making the same intervention appear helpful. Neither change requires an intervention.
An observed extreme often combines a genuinely unusual underlying value with unusually favorable or unfavorable measurement noise. On a second measurement, the stable part remains, but the new noise is unlikely to be equally extreme in the same direction. The group average therefore tends to move closer to the population average. This is regression to the mean.
A synthetic experiment
The linked experiment generates 1,000 synthetic cases with two normally distributed measurements. Both measurements are centered at 50 with standard deviation 10, and the measurement-correlation control sets their programmed correlation from 0.10 to 0.95. No treatment, event, learning, or population shift occurs between them.
At the default setting, rank the first measurements and retain the highest 10%, or 100 cases. That selected group begins well above 50. Its retest mean usually remains above 50 but lies below its first mean. The result is a group average from one reproducible simulation; individual points can still move upward, and New draw changes the particular simulated cases.
The scatterplot shows only selected cases, placing first measurement on the horizontal axis and retest on the vertical axis. A dashed diagonal marks equal values on both occasions. Points below it decreased and points above it increased. The three metrics compare the selected first mean, selected retest mean, and their difference; the table also reports selected count, threshold, and observed correlation across the full generated population.
Raise measurement correlation toward 0.95. More of the first ranking persists, so the selected retest mean generally remains closer to the selected first mean. Then move Select highest from 10% toward 30%. This less extreme selected group begins closer to the population center, so there is less distance available for its average to move inward. The current lab intentionally selects only the highest fraction; low-tail selection would be the reflected theoretical case, not another interface mode.
Derivation and reasoning
The simulation can be written in standardized units as:
Z2 = rho * Z1 + sqrt(1 - rho^2) * error
Here Z1 and error are independent standard-normal draws, and rho is the selected correlation. This construction gives both measurements the same center and spread while controlling their association. In an unselected population both standardized means remain near zero. Selection changes the conditional average because very high values of Z1 are multiplied by a number below one in the persistent part of Z2.
For a standardized bivariate normal model, let Z1 and Z2 be the first and second scores with correlation rho. The conditional expectation is:
E(Z2 | Z1 = z) = rho * z
When 0 < rho < 1 and z is positive, rho * z is positive but less extreme than z. A first score two standard deviations above the mean predicts a second conditional mean only one standard deviation above the mean when rho = 0.5. For a negative first score, multiplication by the same positive correlation moves the expectation upward toward zero.
The phrase toward the mean refers to a conditional average, not a force acting on every individual. Some selected high scorers will score even higher next time. The expected group movement emerges across repeated realizations.
Assumptions
The clean demonstration assumes a stable population distribution and repeat measurements with less than perfect correlation. If the overall process mean changes between sessions, observed movement can combine that trend with regression. A concurrent comparison group and a model for time are then important.
The temporary errors are assumed to be centered and measured on a consistent scale. A calibration shift, different task difficulty, ceiling effect, or changed definition can produce movement that is not regression to the mean.
The linear formula is exact for the stated standardized bivariate normal model. Regression effects occur more broadly, but their form need not be a straight line with slope rho. The simulation should label its distribution and avoid presenting one convenient model as universal.
Failure cases
Do not “correct” every before-and-after change by simply subtracting an estimated regression amount. Real changes, interventions, time trends, and measurement artifacts can coexist. Design the comparison before outcomes are observed and use repeated baseline measurements when feasible.
Do not conclude that an extreme individual must become ordinary. The statement concerns conditional expectation. It also does not imply that population variance necessarily shrinks over time; a stable population can produce the same number of new extreme observations on each occasion even though previously selected extremes regress on average.
Do not confuse regression to the mean with the regression line’s ordinary purpose of prediction. The concepts are related historically and mathematically, but the inferential error arises specifically when selection on an extreme noisy measure is followed by an uncontrolled comparison.
Source notes
The UC Berkeley statistics text derives the regression effect using standardized units and correlation, emphasizing that conditional averages are less extreme when correlation is imperfect. Its companion regression chapter relates the regression line to means, standard deviations, and correlation. This guide translates that reasoning into an original two-measurement simulation. It does not use observed human performance or claim an effect for any intervention.
Keep these distinctions
- Extreme observed scores usually combine an extreme signal with favorable or unfavorable noise.
- When repeated measurements are imperfectly correlated, conditional averages move toward the population average.
- A concurrent comparison group helps separate regression to the mean from an intervention effect.
Sources checked
These sources support the statistical definitions and boundaries. The scenario, prose, calculations, and figure on this page are original to Count & Chance.
- Errors in Regression: The Regression Effect, University of California, Berkeley Department of Statistics. Checked 2026-08-19.
- Regression, University of California, Berkeley Department of Statistics. Checked 2026-08-19.