Chapter 17: Regression to the Mean

Type: Chapter 17

Category: Psychology

Difficulty: Intermediate

Reading time: 6 min

Core idea

Kahneman discovered regression to the mean in an argument with Israeli Air Force flight instructors. The instructors had noticed a troubling pattern: when they praised pilots for excellent maneuvers, the pilots’ next maneuvers were typically worse; when they criticized pilots for poor maneuvers, the next maneuvers were typically better. The instructors concluded that praise hurts performance and criticism helps it.

They were wrong. They had discovered regression to the mean. An exceptionally good maneuver is partly skill and partly luck — a favorable convergence of conditions. A maneuver that follows it, on average, will be slightly less lucky — closer to average. The same logic applies to poor maneuvers. Praise and criticism were happening after extreme events; the regression toward average would have happened regardless of what the instructors said.

Regression to the mean is the mathematical inevitability that extreme measurements in any variable will be followed, on average, by measurements closer to the average — whenever the measurements are imperfect (which they always are). This is not a psychological phenomenon; it is a statistical fact. The psychological phenomenon is that people systematically fail to recognize it and instead generate causal explanations for statistical artifacts.

Why it matters

The causal attribution mistake

When an extreme outcome is followed by a more average one, System 1 attributes the change to whatever intervention occurred between the two observations. Flight instructors attribute regression to the effect of their criticism. Managers attribute a team’s improvement after a poor quarter to a new initiative. A coach attributes an athlete’s decline after a record performance to overconfidence. In each case, the intervention may be real but insufficient — some or all of the observed change would have happened anyway through regression.

This is one of the most consequential statistical errors in management, medicine, and policy. It produces:

• Reward for doing nothing: effective interventions are attributed to random fluctuations back toward average.
• Credit for punishing: punishment administered after extreme poor performance appears to “work” because regression brings performance back toward average — regardless of whether the punishment did anything.
• Persistence of ineffective treatments: a placebo administered to patients at the peak of their symptoms appears effective as symptoms regress naturally.

The correlation-regression relationship

Regression to the mean is a direct consequence of imperfect correlation. When two measurements of the same underlying variable are taken (e.g., exam score on two occasions, athletic performance in two seasons), they correlate imperfectly because both contain error. Whenever the correlation is less than perfect, extreme scores on the first measurement will be followed by less extreme scores on the second — by a magnitude proportional to the imperfection of the correlation.

Kahneman’s formula: the predicted score on the second measurement is the average plus a fraction of the first score’s deviation from average, where the fraction is the correlation coefficient. A correlation of 0.5 means half the deviation from average is expected to persist; the other half will regress.

Why System 1 misses it

Regression to the mean requires thinking in terms of distributions, correlations, and error — all System 2 operations. System 1 operates in terms of individual cases and causal mechanisms. When performance changes, System 1 seeks a cause. It finds interventions, changes in context, or personal qualities — never “imperfect correlation between two measurements.”

Key takeaways

Key takeaways

• Regression to the mean is a statistical inevitability: whenever measurements are imperfect (always), extreme scores on one measurement are followed, on average, by less extreme scores on the next.
• The Israeli Air Force example: instructors concluded praise hurts and criticism helps, but they were observing regression — extreme performances regress toward average regardless of feedback.
• The attribution error: people assign the change caused by regression to whatever intervention happened between the two measurements — producing false confidence in punishments and false skepticism about rewards.
• Correlation and regression are mathematically linked: the more imperfect the correlation, the stronger the regression toward the mean.
• Medical implication: patients seek treatment at the peak of their symptoms — when symptoms are most extreme, regression toward the average occurs regardless of treatment. This inflates apparent treatment effects.
• The fix: always ask, before attributing change to an intervention, what would have happened without it — especially after extreme observations.

Mental model

click to zoom

Regression to the mean — how imperfect measurement produces the appearance of intervention effects from statistical inevitability.

Read it as: An extreme observation reflects true performance plus luck. When luck returns to average on the next observation, performance regresses toward the mean. Any intervention that happened in between gets credited for the statistical inevitability. The intervention may have been real — but the regression would have happened regardless. Disentangling the two requires a control condition, which intuition never supplies.

Practical application

The peak-symptom problem in medicine

Patients most often seek treatment when their symptoms are at their worst — the peak of a naturally fluctuating condition. Treatment administered at the peak will frequently appear to work as symptoms regress toward the individual’s baseline. This is the main reason placebo effects are difficult to distinguish from regression effects in observational data. Controlled trials randomize the timing of treatment relative to symptom severity to correct for this.

To detect and correct for regression:

• Always ask “what would happen without this intervention?”: when performance improves after an intervention following extreme underperformance, consider how much improvement was expected from regression alone.
• Evaluate over multiple observations: a single before-after comparison cannot distinguish intervention effects from regression. Multiple observations at baseline (before the extreme) provide a better reference point.
• Be suspicious of training programs that follow performance crises: if a team gets coaching after their worst month, improvement will occur through regression regardless of coaching quality. Use a control group or measure performance relative to baseline, not relative to the worst event.

Example

A mutual fund that finished in the bottom quartile of its category last year is featured in a fund-of-funds manager’s “turnaround candidates” list. The analysis: management is aware of the problem, the fund’s strategy is sound, and recent underperformance was driven by sector rotation that has now reversed.

But statistically, bottom-quartile funds in any given year will, on average, move toward the median in the following year — not because management improved, but because extreme years contain more luck-against than the average year. Of the 25 funds in the bottom quartile, roughly half will move to the second quartile next year through regression alone, with no change in management quality. Attributing those improvements to turnaround analysis is pattern-matching on a statistical phenomenon.

Related lessons

Related

Chapter 10: The Law of Small Numbers Over-interpretation of extreme results from small samples — the same causal attribution bias that generates false lessons from regression. Chapter 18: Taming Intuitive Predictions The statistical method for correcting intuitive predictions by adjusting for regression — the prescriptive complement to this chapter's descriptive account. Heuristics The mental shortcuts System 1 uses — the causal attribution heuristic that drives regression blindness.