Regression to the Mean in Psychology Explained

In psychology, regression to the mean usually refers to a mistake people make, not to the statistics. After an extreme result, the next one tends to land closer to average. People then credit whatever they did in between. Psychologists call this the regression fallacy: treating a statistical rebound as proof that a reward or a punishment caused the change.

You run into it whenever someone acts on an extreme result and then judges the action by the next result. A manager criticizes a rep after a terrible month. A team ships a fix after a dashboard spikes. A coach benches a player after a bad game. Each of them sees the numbers move toward normal and concludes that the action worked. Often it did nothing at all.

Any single performance mixes skill with luck. Skill carries over from one attempt to the next. Luck doesn't. If you pick someone because their last result was unusually good, you have probably picked someone who was skilled and also lucky that day. Next time the skill is still there, but the luck is redrawn, so the result usually falls back toward that person's typical level. An unusually bad result works the same way in reverse.

That is all regression is, and it happens whenever two measurements of the same thing are imperfectly correlated, meaning they don't line up exactly. Farnam Street's explainer offers a useful pair of cases. Celsius and Fahrenheit readings correlate perfectly, so nothing regresses. Height and weight correlate only moderately, so extremes on one tend to come with less extreme values on the other. Kahneman put it bluntly: "The observed regression to the mean cannot be more interesting or more explainable than the imperfect correlation."

The psychology comes in after the statistics. In their 1973 paper "On the Psychology of Prediction" (Psychological Review, 80(4), 237 to 251), Daniel Kahneman and Amos Tversky argued that people make predictions by representativeness. That means they expect the next outcome to resemble the last one, instead of pulling their forecast back toward the average. Because we don't expect the rebound, we explain it with a cause when it arrives.

The best-known case comes from Kahneman's own work. In the mid-1960s he was at Hebrew University, lecturing Israeli Air Force flight instructors in Jerusalem on the finding that reward improves performance more reliably than punishment. A senior instructor objected. When he praised a cadet for an exceptionally smooth landing, the next landing was usually worse. When he shouted at a cadet after a bad one, the next was usually better.

The instructors concluded, in Oxford Reference's summary, "that praise was not merely ineffective but counter-productive and that punishment was highly effective." Tversky and Kahneman wrote the episode up in their 1974 Science paper, quoted by Decision Science News:

Kahneman then ran a demonstration with no feedback at all. Officers threw chalk at a target with their backs turned, twice. The best throwers on the first round mostly did worse on the second, and the worst mostly did better. Nobody had been praised or criticized. He retold the story in Thinking, Fast and Slow (2011) as one of the moments that started his work on judgment.

The same thing shows up beyond single employees. The BMJ's 2003 paper on regression to the mean warns that if top hospitals are rewarded and struggling ones get extra support, the leaders will seem to slip and the laggards will seem to improve, whether or not the reward or the support did anything. In 1933 the statistician H. Secrist published The Triumph of Mediocrity in Business, which showed top firms declining and weak firms improving the following year. He read it as a real trend. It was regression. A/B testing has the same problem: a variant chosen because it posted a +25% lift probably owes part of that lift to favorable noise, so a rerun should be expected to come in smaller.

It isn't a claim that feedback does nothing. Praise or criticism may well have real effects. The fallacy is using a before-and-after rebound as the evidence for them.

It isn't a force that pulls results back. As Peter Bevelin put it, "Regression to the mean is not a natural law. Merely a statistical tendency." A truly skilled rep's high results don't get dragged down. Only the luck in a single extreme result fails to repeat.

It isn't the statistical concept itself. Regression to the mean is the pattern in the data, covered in our definition of regression to the mean. The regression fallacy is the human habit of misreading that pattern.

Whenever someone says an action fixed a problem, ask how the people or units were picked. If they were picked because their numbers were extreme, some of the improvement would have happened anyway.

The clean test is a comparison group: similar people with equally extreme results who got no intervention. Control groups improve too when they are chosen for being extreme, so the real effect is the gap between the two groups, not the before-and-after change. Without a control group, gather several measurements before acting, so you are reacting to someone's typical level and not one lucky or unlucky stretch. To see how big the rebound can get, the same idea with real numbers walks through the arithmetic.

One more check from the A/B testing world: would you accept the same explanation if the result had gone the other way? If the manager credits yelling when a rep recovers but blames bad luck when a rep doesn't, that's a story built after the fact, not evidence.