Regression to the Mean

Regression to the mean is a statistical phenomenon in which extreme values on a first measurement tend to be closer to the population average on subsequent measurements, regardless of any intervention applied between the measurements.^[4][5] It occurs whenever there is natural variability in the measured characteristic and the correlation between repeated measurements is less than perfect.

Regression to the mean is one of the most common sources of misinterpretation in clinical and social science research. When participants are selected for a study because they have extreme values on a measure (such as very high blood pressure or very low test scores), their subsequent measurements will tend to move toward the average even without any treatment.^[4][5] If this natural regression is not accounted for, it can be mistaken for a genuine treatment effect, leading to overestimation of the intervention's efficacy. Understanding regression to the mean is essential for designing studies that can distinguish true intervention effects from statistical artifacts.

Regression to the mean arises from the combination of a true underlying value and random measurement error (or natural biological variability). When an individual is measured on a variable that has both a stable component and a random component, an extreme observed value is likely to reflect a combination of a moderately extreme true value and a favorable (or unfavorable) random component. On remeasurement, the random component will, on average, be closer to zero, pulling the observed value toward the mean.

Consider blood pressure measurement. A patient whose systolic blood pressure reads 160 mmHg on one visit may have a true average blood pressure of 145 mmHg, with the remaining 15 mmHg attributable to situational factors (anxiety, caffeine, measurement error). On a second visit, the situational component is unlikely to be as extreme, so the reading will typically be closer to 145 mmHg. If a clinician started an antihypertensive between visits, the decline might be attributed to the medication when it was partly or entirely due to regression to the mean.

The magnitude of regression to the mean depends on the test retest correlation (reliability) of the measure and the degree to which extreme values were used as selection criteria. Lower reliability and more extreme selection thresholds produce greater regression to the mean.^[4] In a perfectly reliable measure (test retest correlation of 1.0), there is no regression to the mean because there is no random component.

Study designs that protect against misinterpreting regression to the mean include randomized controlled trials (where both treatment and control groups are expected to regress to the same extent, so the difference between them is attributable to the intervention rather than to regression to the mean), designs with multiple baseline measurements (to establish a stable baseline before intervention), and run in periods that reduce the influence of extreme values attributable to transient factors.^[4]

Consider regression to the mean whenever a study selects participants based on extreme values of the outcome measure and does not include a concurrent control group. It should be discussed as a threat to validity in any pre post design, any study that enrolls patients based on abnormal lab values, and any educational or performance evaluation that selects underperforming individuals for remediation.^[3] When interpreting published studies, assess whether the study design can distinguish true intervention effects from regression to the mean.

A school district identifies the 50 lowest scoring students on a standardized math test and enrolls them in an after school tutoring program. After six months, the students are retested, and their average score improves by 12 points. However, a control group of 50 similarly low scoring students who did not receive tutoring also improves by 8 points on retesting. The 8 point improvement in the control group reflects regression to the mean, together with any other changes common to both groups (such as practice with the test), and the tutoring effect is estimated at 4 points (12 minus 8), substantially less than the unadjusted improvement.