Regression to the Mean
Why extreme scores tend to become less extreme
Regression to the mean describes a prediction pattern in which an unusually high or low measurement is followed, on average, by a measurement closer to the population mean. For best linear least-squares predictions, this shrinkage follows when repeated measurements have positive correlation below one.
Correlation alone does not determine the conditional mean E[Y∣X]; interpreting the fitted value as that conditional mean requires a linear conditional-mean model, such as an appropriate bivariate-normal model. Regression to the mean is not a force that changes people or samples.
Put both measurements on their own standardized scales. If zX is a first measurement and r is the correlation between first and second measurements, the best linear least-squares prediction of the standardized second score is
When 0<r<1, this linear prediction moves an extreme zX toward zero, the standardized mean. Convert back if needed with Y^=μY+zYsY.
With positive correlation below one, best linear least-squares predictions from extreme first measurements lie closer to the second-measurement mean.
Worked example
Suppose a first test is well above its mean and the retest correlation is 0.70. The best linear prediction for the retest is still above average, but it is only 70% as far from the retest mean on the standardized scale.
This does not mean every student must move closer to the mean, nor does correlation alone fix every conditional average. Individual retest scores also contain variation around the fitted line.
Best linear predicted standardized retest score versus first standardized score
The figure shows the least-squares prediction line crossing the means. The original first-score distance, marked A on the horizontal axis, projects vertically to B on the fitted line; B is nearer the horizontal retest-mean line.
When extreme scores are selected in a signal-plus-random-error setting, they often include a favourable noise component that is unlikely to repeat in the same direction.
Common mistake: claiming regression to the mean proves that an intervention failed or succeeded. A before-and-after change in an extreme group can happen even without an intervention. A proper comparison needs a suitable control group or design.
A student's first standardized score is zX=2.0. If the correlation between first and second measurements is r=0.60, what is the best linear least-squares predicted standardized second score?
A clinic selects patients whose first blood-pressure readings are unusually high, then measures them again one week later without treatment. The group's average is lower on the second reading. Which explanation is most justified?