Regression to the Mean

At a Glance

1. Quick Summary

Any score that mixes skill with luck drifts back toward its long-run average after an extreme reading, because the luck that produced the extreme is unlikely to repeat. A student who bombs one test tends to do better on the next; a batter hitting .400 in April cools off; a patient who sees the doctor on the worst day of a flare-up feels better a week later, treatment or no treatment. Francis Galton measured the pattern in family heights in 1886, and Kahneman and Tversky showed in 1973 that human intuition ignores it. The trouble is that we act on extremes: we praise after the best performance and punish after the worst, we medicate at the peak of symptoms, then we watch the numbers settle and conclude that the praise or the pill did it. That conclusion is the regression fallacy, and it fools trained statisticians too.

Simply About the Bias

When something has a really good day or a really bad day, the next day is usually more normal. That happens on its own. Luck does not stay extreme for long. The mistake is thinking that whatever you did in between caused the change.

Say you play a video game and get your worst score ever. Your big brother says you are holding the controller wrong and makes you change your grip. The next game, your score goes up. Did the new grip do it? Maybe. But a score that bad was mostly bad luck, and your next score was going to be better anyway. Your brother now believes his grip tip works, and so do you.

Here is a trick. When something improves after you changed it, ask one question: was it unusually bad right before I changed it? If yes, the change might have done nothing at all. Try the old way again for a few rounds and compare. Only then can you tell what really helped.

2. The Science Behind It

2.1. Discovery and History

Francis Galton noticed the pattern first in sweet peas in the 1870s: the offspring of very large seeds were large, but less large than their parents, and the offspring of very small seeds were less small. He called it "reversion" and took it for a fact about heredity.

The human data came in the 1886 paper "Regression towards mediocrity in hereditary stature." Galton collected the heights of 928 adult children from 205 sets of parents and plotted them against a combined "mid-parent" height. Tall parents had tall children, but on average the children sat only about two-thirds of the way from the population mean to their parents' height. "Mediocrity" meant the middle, not low quality. Galton first took this for a force in heredity, then saw that plotting parents against children gave the same result: regression appears wherever two measures are correlated but not perfectly, in both directions, with no biology required.

The fallacy got its clearest early demonstration in 1933, when Horace Secrist, a statistics professor at Northwestern, published The Triumph of Mediocrity in Business. He had tracked the profit rates of hundreds of firms over a decade and found that the most profitable firms in year one had slid toward average by year ten, while the least profitable had climbed. He read this as competition grinding excellence down. Harold Hotelling, reviewing the book in the Journal of the American Statistical Association, pointed out that Secrist had rediscovered regression, and that the spread of profit rates in his own data had not shrunk.

Donald Campbell and Julian Stanley carried the lesson into research design in 1963, listing "statistical regression" among the threats to internal validity: any group selected for an extreme score will look improved after any treatment. Psychology joined with Kahneman and Tversky's 1973 paper "On the psychology of prediction," which showed that people predict as if the link between evidence and outcome were perfect. Kahneman later told the story behind it: an Israeli Air Force flight instructor who had concluded from experience that praise makes cadets worse and shouting makes them better.

Key milestones:

• 1877: Galton reports "reversion" toward the average in sweet-pea seed sizes
• 1886: Galton's stature paper finds a regression ratio of about two-thirds in 928 adult children
• 1933: Secrist's The Triumph of Mediocrity in Business mistakes regression for a law of competition; Hotelling's review identifies the error
• 1963: Campbell and Stanley name statistical regression as a threat to internal validity
• 1973: Kahneman and Tversky show that intuitive predictions are "nonregressive"
• 1999: Campbell and Kenny publish A Primer on Regression Artifacts
• 2005: Barnett, van der Pols and Dobson publish their tutorial in the International Journal of Epidemiology

2.2. Key Researchers

2.3. Landmark Studies

"Regression towards mediocrity in hereditary stature" (Galton, 1886)

Galton grouped 928 adult children by mid-parent height and tabulated the medians. Children of mid-parents 3 inches above the population average were, on average, about 2 inches above it; children of mid-parents 3 inches below were about 2 inches below. The structure is the point: whenever a prediction is made from an imperfect predictor, the best prediction is less extreme than the predictor.

"On the psychology of prediction" (Kahneman and Tversky, 1973)

Participants read descriptions of people and predicted outcomes such as grade point average. The descriptions carried little diagnostic value, yet predictions were almost exactly as extreme as the impressions they created: a glowing paragraph produced a prediction of a glowing GPA. A weak predictor should produce a prediction near the average; participants behaved as if it were perfect. The authors called this "nonregressive" prediction and tied it to the representativeness heuristic. The paper also discusses the flight-training case, in which instructors saw praised cadets fly worse next time and reprimanded cadets fly better, and drew the wrong causal lesson.

The Triumph of Mediocrity in Business and Hotelling's review (Secrist, 1933; Hotelling, 1933)

Secrist sorted firms into deciles by a base year and followed each decile forward through the 1920s. Every top decile fell and every bottom decile rose. Hotelling observed that the same pattern appears if you sort by the final year and read backward: the "best" firms in 1930 were closer to average in 1920. Nothing was converging. Secrist had selected extremes and watched the luck component wash out. In the exchange that followed, Hotelling compared the method to proving the multiplication table by arranging elephants in rows and columns.

2.4. Neurological Basis

There is no body of brain-imaging work on the regression fallacy specifically, and this section should not pretend otherwise. What exists is a well-established picture of the machinery the fallacy runs on.

The brain's learning systems detect causes from temporal sequence: do something, see a change, strengthen the link. Dopamine neurons in the midbrain fire in proportion to how much an outcome differs from expectation (the reward prediction error documented by Wolfram Schultz and colleagues in the 1990s), so surprising events, including extreme performances, carry an outsized teaching signal. There is no comparable signal for "this would have happened anyway," because a counterfactual baseline is not an event and generates no surprise. Recognizing regression means overriding the automatic link with a deliberate statistical thought, which is slow and easily skipped.

3. Evolutionary Origins

Our ancestors did not need the concept of noise. They needed to learn fast which actions were followed by which outcomes, and a rule that treats every "action, then change" sequence as evidence of causation is a good rule when most changes have causes and a false belief costs little.

Regression only becomes a problem at a certain scale of measurement. It needs a quantity measured repeatedly, enough chance variation to matter, and selection on an extreme. Ancestral humans rarely tracked numbers across many trials and never had a control group. In small samples the chance component is enormous, so almost every observed change could be credited to the last thing someone did, and no feedback could teach the opposite lesson.

There is a social pressure behind the fallacy too. A leader who acts decisively when things are at their worst, and whose intervention is followed by improvement, gains authority. Since things at their worst usually do improve, interventions at that moment are rewarded by circumstance regardless of merit. The same mechanism rewards punishment: discipline applied after the worst behavior is followed by better behavior, so cultures that punish feel confirmed in punishing.

4. How This Bias Manifests

4.1. In Everyday Life

• Crediting a cold remedy for a recovery that began at the worst point of the cold, which is when most people reach for a remedy
• Believing a child's "talk" after a terrible report card produced the better grades that followed
• Concluding that a lucky charm or a pre-game ritual fixed a slump
• Expecting a child who scored brilliantly once to keep scoring brilliantly, and reading the normal next result as a decline

4.2. In the Workplace

• Performance reviews that reward last quarter's stars and reprimand the laggards, then read the next quarter's convergence as proof the reviews worked
• Training programs assessed only on employees sent because of poor scores, who would have improved anyway
• Managers who learn that criticism works and praise backfires, because the criticized had just done badly and the praised had just done well

4.3. In Business and Marketing

• Advertising campaigns evaluated on sales in the month after launch, when the launch was timed to a dip
• "Turnaround" case studies about firms selected because they were at the bottom when the study began
• Testimonials from customers who tried the product when their problem was at its worst
• "Most improved" store rankings, which are largely rankings of which stores had the worst luck in the base period

4.4. In Politics and Media

Governments act on extremes almost by definition. Crime waves prompt crackdowns and the worst-performing schools get targeted programs, so each policy is set up to look successful whether or not it works.

Manifestations include:

• Crime initiatives announced during a spike and credited for the fall that follows
• School accountability schemes in which the lowest-scoring schools post the largest "gains" and the highest-scoring post "declines"
• Speed cameras sited at junctions with the worst recent crash year, whose subsequent improvement overstates the camera's effect
• Polling narratives that treat an outlier poll as a trend and the return to average as a "collapse"

4.5. In Healthcare

Medicine is where the fallacy does the most damage, because patients seek care when symptoms are at their worst and illnesses fluctuate. Barnett, van der Pols and Dobson (2005) use blood pressure as their example: enroll people on one high reading and the group's average falls at the next measurement before any treatment, because some readings were high partly by chance. Bland and Altman made the same point in the BMJ in 1994. Much of what gets attributed to the placebo effect in uncontrolled studies is regression plus the natural course of illness, one reason the 2001 meta-analysis by Hróbjartsson and Gøtzsche found so little difference between placebo and no-treatment arms.

Implications:

• Conditions with flares (back pain, migraine, eczema, depression) make almost any treatment started during a flare look effective
• Alternative therapies thrive on this pattern, since people try them after conventional care "failed" at a peak of symptoms
• Doctors can learn from experience that a treatment works when they have only observed it at the extremes
• Patients who stop a drug because symptoms improved may be stopping the drug that helped, or one that did nothing

4.6. In Finance and Investing

• Chasing last year's top-performing funds, which on average deliver closer-to-average returns next year
• Firing a money manager after a bad year and hiring one after a good year, buying high and selling low
• Reading a company's record quarter as a new level rather than a high draw from a noisy distribution
• Crediting a central bank with a rebound that followed a panic low

5. Real-World Case Studies

Case Study 1: The Flight Instructors

• Context: In the 1960s Daniel Kahneman, then a young psychologist in Israel, was lecturing to Israeli Air Force flight instructors on the finding that reward improves skill learning more than punishment does.
• What happened: A senior instructor objected. Whenever he praised a cadet for a clean maneuver, the next attempt was usually worse; whenever he shouted at a cadet for a bad one, the next attempt was usually better. Experience showed, he said, that praise hurts and punishment helps.
• The bias at work: The instructor was right about what he had seen and wrong about why. A maneuver good enough to earn praise was partly luck, so the next was likely to be more ordinary; a maneuver bad enough to earn a shout was partly bad luck, so the next was likely to be better. Kahneman demonstrated this on the spot: each instructor tossed two coins at a target behind his back, and those who did best on the first toss mostly did worse on the second, with no feedback at all.
• Consequences: Kahneman called this one of the most satisfying insights of his career. Because we reward after good outcomes and punish after bad ones, we are, in his words, statistically punished for rewarding and rewarded for punishing.
• Lessons learned: Experience alone cannot distinguish an effective intervention from one applied at an extreme. Only a comparison against doing nothing can.

Case Study 2: The Sports Illustrated Cover Jinx

• Context: Since the 1950s, fans have believed that a Sports Illustrated cover brings misfortune. Eddie Mathews of the Milwaukee Braves was on the first issue in 1954; the team's winning streak ended and he broke a hand soon after.
• What happened: In January 2002 the magazine investigated its own curse. Writer Alexander Wolff went through every cover and found that a bit over a third of cover subjects had suffered a measurable drop in performance, an injury or some other misfortune soon afterward. The issue carried a black cat on the cover because no athlete would pose for it.
• The bias at work: Athletes get on the cover at a peak, and a peak is the point from which the next measurement is most likely to be lower. Add that a normal season after a record one looks like a slump, and a "jinx rate" of a third requires no jinx. The "Madden curse" and the "sophomore slump" are the same artifact with different names.
• Consequences: Teams extend contracts at peaks and sign players coming off career years, then pay for the regression. Analysts such as Bill James, who called the pattern the "Plexiglas principle," built forecasting systems that shrink extreme recent seasons toward each player's established level.
• Lessons learned: When the condition for noticing something is an extreme, the follow-up will almost always be less extreme. Any explanation for that drop has to beat the explanation that needs no cause at all.

Historical Example: The Triumph of Mediocrity in Business (1933)

Horace Secrist was a respected statistician and textbook author when he published The Triumph of Mediocrity in Business after roughly a decade of work. He had sorted firms in dozens of industries by profit ratios in a base year, then tracked each group forward through several hundred pages of tables. The charts were monotonous: the best groups drifted down toward the industry average, the worst drifted up. Secrist concluded that competition itself erodes excellence and lifts failure.

Harold Hotelling saw at once that this was regression to the mean and nothing else. The spread of profit rates across firms, which should have shrunk if mediocrity were triumphing, had not. Secrist defended himself in the correspondence pages, and Hotelling replied with the elephant comparison that outlived the book.

The episode remains the standard warning in statistics teaching because the person who fell for the fallacy was a professional, working carefully, with abundant data. Selection on an extreme guarantees the artifact however much data sits behind it.

6. The Cost of This Bias

6.1. Personal Costs

• Money spent on remedies, supplements and courses that "worked" only because they were tried at a low point
• Relationships in which criticism is believed to be effective and warmth believed to backfire
• Children raised under harsher discipline than their parents would otherwise choose

6.2. Professional Costs

• Promotions and firings based on one extreme quarter, followed by disappointment with the "star" and surprise at the recovery of the "problem" employee
• Training budgets spent on programs never tested against a comparison group
• Loss of credibility when a confident "this works" fails under proper evaluation

6.3. Societal Costs

• Public money committed to policies that only looked effective because they were introduced at a peak of the problem
• School systems that rank and punish on year-to-year changes, which are mostly noise
• Treatments and alternative therapies sustained by testimonials from patients treated at their worst
• Erosion of trust in evaluation when programs that "obviously worked" fail in randomized trials

6.4. Statistical Impact

• Galton's regression ratio for human height was about two-thirds: children of parents 3 inches from the mean averaged about 2 inches from it
• Expected regression is proportional to one minus the correlation between the two measurements; at a correlation of 0.5, an extreme first reading is expected to lose half its distance from the mean on the second
• Hróbjartsson and Gøtzsche's 2001 meta-analysis found little difference between placebo and no-treatment arms for most outcomes, implying that much of the "improvement on placebo" in uncontrolled studies is regression and natural history

7. The Hidden Benefits

The phenomenon of regression is pure gain for anyone who understands it. It is a free forecasting tool: when you must predict the next reading from a noisy last one, move your guess toward the average and you will beat most intuitions. The fallacy itself has side effects that are not all bad:

• It keeps people trying: A belief that interventions work, even when the evidence is regression, keeps patients taking medicine, managers coaching and parents engaging. Learned helplessness is the alternative.

• It supplies harmless comfort: Lucky charms and comeback stories give people a sense of control at low points. When the cost is low, a false sense of agency during a trough can be worth more than the truth.

• It often lands on a real effect: Many interventions applied at extremes do work: antibiotics for a bad infection, a serious talk with a struggling teenager. The fallacy overstates the effect, but the direction of the action is often right.

These benefits last only until the stakes rise. In medicine, policy and finance, the cost of treating noise as signal dwarfs the comfort of feeling in control.

8. Self-Assessment: Do You Have This Bias?

8.1. Warning Signs Checklist

• When something improves right after you change it, you assume the change caused the improvement
• You have a remedy, supplement or ritual you believe in because of one dramatic recovery
• In your experience, criticism works better on people than praise
• You expect someone who just had their best-ever result to keep performing at that level
• You expect someone who just had their worst-ever result to keep performing at that level
• You rarely ask "was this unusually bad just before I acted?" when judging whether an action worked
• You would judge a new program by comparing numbers before and after, without a comparison group
• You switch funds, suppliers, doctors or strategies after a single bad period
• You are most likely to try something new when things are at their worst, and to credit it afterward
• You can name several things that "turned around" a situation but few that were tested against doing nothing

Scoring:

• 0-2 checked: Low susceptibility
• 3-5 checked: Moderate susceptibility
• 6-8 checked: High susceptibility
• 9-10 checked: Very high susceptibility

8.2. Self-Reflection Questions

1. Think of the last treatment or technique you recommended because it "worked for you." How bad were things when you started it?
2. In your work, who gets singled out for praise or correction, and when? Is the timing tied to an extreme result?
3. When you last judged a change by before-and-after numbers, what would the "after" number have looked like with no change at all?
4. Can you recall a slump that ended on its own? What were you tempted to credit for its ending?
5. When someone you manage or teach has a breakthrough performance, do you treat the next, more ordinary one as a problem to fix?

8.3. Quick Diagnostic Scenario

Scenario: Your team's customer satisfaction score, normally around 7.5 out of 10, dropped to 6.1 last month. You ran a two-hour workshop on service standards. This month the score is 7.4.

How would you respond?

• A) Roll the workshop out to every team; it lifted the score by more than a point → High susceptibility
• B) Call it a probable success but keep an eye on next month's figure → Moderate susceptibility
• C) Note that 6.1 was an unusually low reading, so a return to 7.4 was the most likely outcome regardless; test the workshop on half the teams before concluding anything → Low susceptibility

9. Identifying This Bias in Others

9.1. Behavioral Indicators

• Acts most decisively at the extremes and cites the subsequent improvement as validation
• Keeps a list of things that "worked" without a single comparison against doing nothing
• Reads every "most improved" ranking as a list of successful interventions
• Reacts to one bad result with major changes and to one good result with major commitments
• Prefers punishment to praise on the basis of "what I've seen"

9.2. Conversational Red Flags

Phrases people say when under this bias:

• "I tried it when I was at my lowest, and within a week I was better."
• "Every time I praise him he slacks off; every time I come down hard he shapes up."
• "The numbers were terrible, we brought in the consultants, and look at them now."
• "Crime fell by 20 percent after the crackdown. What more proof do you need?"

Types of arguments they make:

• Before-and-after comparisons with no control group and no mention of what prompted the action
• Testimonials gathered from people who sought help at a peak of their problem

Questions they avoid asking:

• "What would have happened if we had done nothing?"
• "How unusual was the result that made us act in the first place?"

9.3. Situational Triggers

• Any decision taken because a number was unusually high or low
• Programs that enroll participants on an extreme score (remedial classes, high-risk patient lists, "problem" employees)
• Conditions that fluctuate, such as pain, mood, allergies and skin problems
• Crises, which arrive at peaks and are followed by relief

10. Cognitive Debiasing Strategies

10.1. Immediate Techniques

• The "Why Now?" Question: Before crediting any change, ask why the action was taken when it was. If the answer is "because things were unusually bad (or good)," assume some of the subsequent movement is regression.
• Shrink the Forecast: When predicting from a single extreme result, start from the long-run average and move only part of the way toward the result. The weaker the link between past and future, the less you move.
• Reverse the Sort: Do what Hotelling did to Secrist. Pick the extremes at the end of the period and look backward. If they also appear to have come from the middle, you are looking at regression.

10.2. Long-Term Strategies

• Learn the Mechanism, Not the Slogan: Work through one example with numbers (Galton's two-thirds, or a correlation of 0.5 and an extreme reading) until the idea that regression needs no cause feels obvious.
• Keep a Prediction Log for Extremes: Each time a result stands out, write down your prediction for the next one. Review quarterly. Most people find their predictions too extreme in the direction of the last result.
• Separate Selection from Evaluation: If you must select on a measurement, evaluate on a different one, or on a later one taken after the selection effect has passed.

10.3. Environmental Design

• Control Groups by Default: Any program that can be rolled out to half the eligible units first should be. The untreated half is the only honest baseline.
• Multiple Baseline Measurements: Require two or three pre-intervention readings before anyone is enrolled on the basis of an extreme. Averaging them removes most of the chance from the selection.
• Confidence Bands on Rankings: Any league table or "most improved" list should show the range of normal variation, so that movement inside the band is not read as change.

10.4. When to Seek External Input

• When a program or treatment was applied because a number was extreme and you are about to scale it up
• When you are drawing a lesson about people (praise, discipline, incentives) from your own experience of their next performance
• When a before-after comparison is the only evidence and the stakes are high
• When you notice that everything you try seems to work

Whom to ask: Someone trained in experimental design or epidemiology, who will ask about selection and controls before asking about the size of the effect. Frame the request as "tell me how this result could have appeared with no real effect at all."

11. Practical Exercises

Exercise 1: The Coin-Toss Instructor

• Objective: Experience regression with your own hands, where no causal story is possible
• Time required: 20 minutes
• Materials needed: Two coins, a target (a cup on the floor), a score sheet, at least four people
• Difficulty level: Beginner
• Instructions:
  1. Each person stands with their back to the target and tosses two coins over their shoulder toward it. Score each toss by distance from the target.
  2. Record every person's score for round one.
  3. Without any feedback or coaching, repeat for round two and record again.
  4. Sort people by their round-one score and compare the round-two scores of the top half with the bottom half. The best round-one performers mostly got worse and the worst mostly got better, with nothing done in between.
• Reflection questions:
  • If a coach had praised the top half and scolded the bottom half between rounds, what would the coach have concluded?
  • Where in your own life does feedback arrive shaped like this?
• Frequency: Once, then repeat as a demonstration whenever you need to explain the idea to a team

Exercise 2: The Shrinkage Forecast

• Objective: Practice making regressive predictions from extreme results
• Time required: 15 minutes per session
• Materials needed: A notebook and any source of repeated measurements (sports scores, sales figures, step counts)
• Difficulty level: Intermediate
• Instructions:
  1. Find a recent reading that stands out as unusually high or low.
  2. Write down your gut prediction for the next reading.
  3. Compute the long-run average from the previous ten or more readings.
  4. Estimate roughly how strongly one reading predicts the next (near zero for coin-like outcomes, near one for stable traits). Move your prediction from the average toward the extreme reading by about that fraction.
  5. Record both predictions. When the next reading arrives, note which was closer.
• Reflection questions:
  • How far apart were your gut prediction and the shrunken one?
  • Over ten trials, which prediction won more often?
• Frequency: Weekly for two months

Exercise 3: The Selection Audit

• Objective: Find regression hiding inside your existing beliefs about what works
• Time required: 45 minutes
• Materials needed: Paper and pen
• Difficulty level: Intermediate
• Instructions:
  1. List ten things you believe work because you have seen them work: remedies, management techniques, parenting tactics, investment rules.
  2. For each, write down what prompted you to try it. Was the situation unusually bad, or unusually good, at that moment?
  3. For each, write down whether you have ever seen the outcome without the intervention, under similar conditions.
  4. Mark every item where step 2 says "extreme" and step 3 says "no." Those are the beliefs regression could fully explain.
  5. Pick one and design the cheapest fair test you can: a comparison period, or a search for a proper trial.
• Reflection questions:
  • How many of your ten beliefs survived the audit?
  • Which belief were you most reluctant to mark?
• Frequency: Twice a year

Daily Practice

The "Was It Extreme?" Check

Once a day, when you hear or make a claim that something caused an improvement or decline, ask whether the starting point was unusually high or low.

• Suggested duration: 1 minute whenever a before-and-after claim comes up
• Best time of day: Whenever you read news, review results or hear "it worked"
• How to track progress: Keep a tally of claims checked and how many began at an extreme

Weekly Challenge

Predict the Comedown

Each week, pick one person, team, stock or statistic that just had a standout result, and write down a prediction for its next result that is closer to its average. Check the result the following week.

• Expected outcomes after 4 weeks: Less surprise at "slumps" and "recoveries"; a working feel for how much to shrink
• Journaling prompts for reflection:
  • Which comedowns did I predict correctly, and which extremes turned out to be real changes in level?
  • What explanation did the news offer for each comedown, and did it need one?

12. For Specific Audiences

For Leaders and Managers

• Performance Management: Reward and correct on multi-period averages. A single quarter at the top or bottom is mostly noise, and reacting to it teaches you that correction works and praise fails.
• Personal Practice: When your intervention is followed by improvement, say to yourself "and some of that was coming anyway" before you say anything else.

For Parents and Educators

• Age-Appropriate Explanation: "A really bad day is usually followed by a better one, all by itself. So if you change something on a bad day and the next day is better, you can't be sure the change did it. Try it a few times to find out."
• Discipline: Punishment after the worst behavior is always followed by better behavior, which is why harsh discipline feels effective. Judge your approach on weeks and months, not the next hour.
• Remedial Programs: Children selected for extra help on one low score will improve on the next test regardless. Measure the program on a later assessment, or against similar children who did not receive it.

For Healthcare Professionals

• Clinical Experience: You see most patients at the worst of fluctuating conditions, so any treatment given then will seem to work. Trust trials over the memory of patients who got better.
• Screening and Enrollment: Enrolling on a single abnormal reading guarantees improvement at the next. Use repeated baseline measurements before diagnosis or trial entry.
• Patient Communication: When a patient credits a supplement or alternative therapy, ask when they started it. Explaining regression kindly, with the "worst day" example, lands better than dismissing the therapy.

For Financial Professionals

• Fund Selection: Last year's top decile of funds is, on average, closer to the middle next year. Resist clients who want to chase it, and the same impulse in your own allocations.
• Client Narratives: Clients will credit the advisor hired at the bottom with the recovery and blame the advisor hired at the top for the slump. Tell the regression story before it happens.

13. Interactions with Other Biases

Biases That Amplify This One

Biases That Counteract This One

Common Bias Chains

Extreme Result → Representativeness Heuristic → Intervention → Regression to the Mean → Illusory Correlation → Confirmation Bias

A standout result arrives and the representativeness heuristic reads it as a real change in level, which prompts action. The next reading regresses, as it was always going to. The improvement is paired with the action, the pairing is remembered and the non-events forgotten, and a belief in the action is born that further extremes keep reinforcing.

Interrupting the chain: The cheapest point is the first link. If an extreme result is recognized as partly chance before any action is taken, the improvement that follows carries no surprise to be explained. Failing that, a control group at the point of action breaks the chain at the third link.

14. Cultural Perspectives

There is little direct cross-cultural research on the regression fallacy. The statistical phenomenon is the same everywhere, and the machinery that misreads it, causal learning from sequence, is not known to differ much between populations. What differs is the story each culture tells about the comedown after a peak, and those stories look like folk encodings of regression that attach a cause where none is needed.

Many societies warn against praising a child, a harvest or a success too loudly. The evil-eye traditions of the Mediterranean, the Middle East and South Asia hold that admiration invites misfortune, and the folk evidence is easy to see: admiration comes at peaks, and peaks are followed by declines. Greek thought paired hubris with nemesis, and Chinese proverbs about the tallest tree catching the wind carry the same lesson. In the English-speaking world the pattern surfaces as the jinx and the fear of "tempting fate."

Attitudes to punishment and praise interact with the fallacy too. Where strict discipline is the norm, the feedback Kahneman described confirms the norm daily. Where praise-based parenting is the norm, the same regression runs in reverse, and practitioners may privately feel the approach "doesn't work" without being able to say why.

15. Myths and Misconceptions

16. Expert Insights

"Because we tend to reward others when they do well and punish them when they do badly, and because there is regression to the mean, it is part of the human condition that we are statistically punished for rewarding others and rewarded for punishing them." — Daniel Kahneman, Thinking, Fast and Slow (2011), Chapter 17

"Our mind is strongly biased toward causal explanations and does not deal well with 'mere statistics.'" — Daniel Kahneman, Thinking, Fast and Slow (2011)

The average deviation of the children from the population mean was only a fraction, about two-thirds, of the deviation of their parents; the extremes of one generation are moderated in the next, whichever direction you look. — Francis Galton (1886), paraphrasing the central finding of "Regression towards mediocrity in hereditary stature"

Selecting people for a study because of one extreme measurement guarantees that their next measurement will, on average, be less extreme, whether or not anything is done to them; the reliable protections are a randomized control group and repeated baseline measurements. — Adrian Barnett, Jolieke van der Pols and Annette Dobson (2005), paraphrasing the conclusions of their International Journal of Epidemiology tutorial

17. Key Takeaways

1. The Phenomenon Is Not the Bias: Regression to the mean is a statistical fact about any quantity with a chance component. The bias is the regression fallacy: inventing a cause for a return to average that needed none.

2. Extremes Are Where We Act: We treat at the peak of symptoms, discipline after the worst behavior and reorganize after the worst quarter. Each of those moments is primed to make whatever follows look effective.

3. Praise and Punishment Are Both Confounded: Kahneman's flight instructors learned that shouting works because they only shouted after bad performances. The same logic makes praise look harmful. Neither conclusion follows.

4. Volume of Data Is No Protection: Secrist had a decade of figures on hundreds of firms and still read regression as a law of business. Selection on an extreme guarantees the artifact.

5. Shrink Your Predictions: The best prediction from an extreme reading lies between that reading and the long-run average, and closer to the average the weaker the correlation.

6. Control Groups Are the Cure for Evaluation: Before-and-after comparisons on units selected for extreme scores cannot distinguish an effective program from an inert one. Randomization, multiple baselines or a comparison group can.

7. Ask "Why Now?": Before crediting any change, ask whether the action was taken because things were unusually bad or good. If it was, assume part of the improvement was coming anyway.

18. Further Resources

Academic Papers

• Galton, F. (1886). Regression towards mediocrity in hereditary stature. Journal of the Anthropological Institute of Great Britain and Ireland, 15, 246-263.
• Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. Psychological Review, 80(4), 237-251.
• Hotelling, H. (1933). Review of The Triumph of Mediocrity in Business by Horace Secrist. Journal of the American Statistical Association, 28(184), 463-465.
• Barnett, A. G., van der Pols, J. C., & Dobson, A. J. (2005). Regression to the mean: What it is and how to deal with it. International Journal of Epidemiology, 34(1), 215-220.
• Bland, J. M., & Altman, D. G. (1994). Statistics notes: Regression towards the mean. BMJ, 308(6942), 1499.
• Stigler, S. M. (1997). Regression towards the mean, historically considered. Statistical Methods in Medical Research, 6(2), 103-114.
• Hróbjartsson, A., & Gøtzsche, P. C. (2001). Is the placebo powerless? An analysis of clinical trials comparing placebo with no treatment. New England Journal of Medicine, 344(21), 1594-1602.

Books

• Campbell, D. T., & Kenny, D. A. (1999). A Primer on Regression Artifacts. Guilford Press.
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
• Stigler, S. M. (1999). Statistics on the Table: The History of Statistical Concepts and Methods. Harvard University Press.

Book Chapters

• Kahneman, D. (2011). Regression to the mean. In Thinking, Fast and Slow (Chapter 17). Farrar, Straus and Giroux.
• Campbell, D. T., & Stanley, J. C. (1963). Experimental and quasi-experimental designs for research on teaching. In N. L. Gage (Ed.), Handbook of Research on Teaching (pp. 171-246). Rand McNally.

19. Summary Card

20. Glossary of Terms Used

21. Discussion Questions

For book clubs, classrooms, or self-reflection:

1. Kahneman's flight instructors were experienced, attentive and honest, and still drew the wrong lesson. What experience in your own life has the same structure, where feedback arrives only after extremes?

2. Horace Secrist worked for a decade with careful data and published a confident law of business. What institutional habits would have caught his error earlier?

3. Many cultures warn against praising success too loudly. Is the evil-eye tradition a folk theory of regression? What is lost and gained when a statistical pattern is encoded as a superstition?

4. Much of what is attributed to the placebo effect in uncontrolled settings is regression and natural history. How should that change the way patients and doctors talk about treatments that "worked for me"?

5. Regression can be misused as well as ignored: not every recovery is a return to average. How would you decide, for a specific policy or treatment, whether the no-cause explanation has been ruled out?

Viimati uuendatud 15 oktoober 2026