statistical mirages · metaphor 28 of 100
Scold the worst performer and they improve. Praise the best and they slip. Every coach, parent, and manager has run this experiment, and nearly all of them have drawn the same conclusion: harshness works, kindness backfires. The conclusion is false — and the reason it feels so true is arithmetic.
Daniel Kahneman was lecturing Israeli flight instructors on the psychology of reward when one of them objected. Praise a cadet for a beautiful landing, he said, and the next landing is worse. Scream at a cadet for a dangerous one, and the next is better. He had watched it happen for years. The room agreed: punishment teaches; praise spoils. Their observations were accurate. Their inference was one of the most reliably wrong conclusions the human mind produces.
Because a great landing is not pure skill. It is skill plus a good day — a favorable gust, sharp sleep, a lucky line through the turbulence — and luck is not a character trait. It does not repeat. The cadet at the top of today's board was disproportionately lent there by chance, and chance calls in its loan on the very next flight, sermon or no sermon.
—
—
Borrowed extremes
Split any performance into what persists and what the day contributed. The best score of the round is almost never the most skilled performer on an average day; it is a skilled performer on a very good day. Selecting the extreme selects for luck as surely as it selects for skill — you have skimmed the top of two distributions at once, and only one of them belongs to the person.
Luck has no memory. Next round it is re-dealt from scratch, and the champion's merely ordinary new luck subtracts exactly what yesterday's extraordinary luck had lent. The fall looks like complacency; the slump-breaker's rise looks like fear finally doing its job. Neither needs a cause. Nothing pulled anyone anywhere — the loan simply came due. And because the instructor's scream and the coach's benching were delivered precisely at the extreme, they stand at the scene collecting credit for a tide that was already turning.
What to try
Leave the true effect at 0.00 — feedback certifiably inert — and run 20 rounds. The ledger will report that scolding worked and praise backfired in the large majority of trials. Every tally is real. Every tally is luck being returned.
Drag the luck share down toward pure skill and run again. Extremes now persist — the best stay best — and the illusion starves. Then crank luck up and watch the mirage bloom. Then flip on reveal hidden skill and see who the extremes really were.
Set the effect to +0.50: praise now genuinely helps and scolding genuinely harms, by half a standard deviation. Run 20 rounds and read the ledger. It still votes for harshness. The illusion is bigger than the truth it buries.
In the wild
Here is the general recipe for a remedy with a perfect record: wait for an extreme, then act. Anything applied at an extreme gets credit for the return trip — because the return trip was leaving anyway, with or without passengers.
Coaches are sacked after historically bad streaks, and the team "responds to new leadership." Slumping stars are benched and "come back hungry." The turnaround was scheduled by the arithmetic before the decision was made.
You seek the doctor, the supplement, the healer on your worst day — that is what a worst day is for. Whatever you took as the pain crested gets the credit as it recedes. Try the pain levels skin: every remedy in the ledger "works."
Speed cameras go where last year's accidents spiked; accidents then fall there — as the freak cluster would have anyway. Programs aimed at the worst schools, wards, and quarters harvest the same free miracle.
The slope as honesty
Plot everyone's score this round against their score the next, and fit a line. Its slope — Var(skill) / (Var(skill)+Var(luck)), computed live under the scatter above — is the domain confessing how much of itself is the person and how much is the weather. A slope near 0.9 says extremes are mostly earned and will mostly persist; a slope near 0.2 says today's board is mostly a lottery reading.
The slope is also a calibration for your surprise: expect anyone to return 1 − slope of their distance from the mean, with no cause whatsoever. That is the baseline against which praise, punishment, remedies, and reorganizations must prove themselves — and it is a high bar precisely where luck is loud. Before crediting any intervention timed to an extreme, ask: did they beat the slope, or merely ride it home?
The mapping
| Mathematics | Life |
|---|---|
| skill | What persists about a person — the part of today's result you will meet again tomorrow. |
| noise | What the day contributed: sleep, gusts, mood, the bounce of the ball. Real, but not theirs. |
| the extreme | Mostly the day, misread as the person — a compound of merit and weather, credited entirely to merit. |
| return to the mean | Time reclaiming luck's loan: no force, no lesson learned, just a fresh deal of the cards. |
| scolding / praise | Interventions timed to extremes, standing on the beach collecting credit for the tide. |
| the ledger | How perfectly honest observation, aimed at extremes, compiles a false theory of human nature. |
Where the metaphor tears
Nothing pulls anyone toward the middle — no rubber band, no cosmic fairness. It is selection arithmetic: condition on an extreme and you have oversampled luck, which then merely fails to repeat. And it operates only where luck contributes variance. In near-pure-skill domains — chess ratings, the hundred-metre dash — the extremes largely persist, and "they'll come back to the pack" is just wishful thinking wearing a statistics costume.
The lesson is not that praise and punishment are inert — decades of evidence say encouragement and consequences both move people. The lesson is that evidence timed to extremes cannot tell you: as the instrument shows, a genuine half-standard-deviation effect barely dents the ledger. Distinguishing the tide from the intervention takes comparison groups and randomization — which is to say, it takes deliberately not doing what every instinct at the extreme demands.
Once you see regression everywhere, "it's just regression to the mean" becomes a universal solvent for dissolving inconvenient improvements — the rehab that held, the student who actually turned it around, the therapy that worked. Regression is the null hypothesis, not the verdict. Sometimes people change; the arithmetic only insists you not conclude it from a single trip back from the edge.