── ── Mental model
Regression to the Mean
Regression to the mean is the statistical regularity that any noisy measurement producing an extreme value tends to be followed on retest by a less-extreme value — because the extreme portion was partly driven by non-repeating random noise. There is no "force pulling back to average"; it is a mathematical consequence of signal + noise structure.
How it works
1. Identify the extreme observation — value, subject, how extreme, any intervention. 2. Identify the retest — follow-up measurement, period, direction (toward mean?). 3. Estimate expected regression — noise level, historical variance; formula: regression ≈ (1 − reliability) × distance from mean. 4. Compare to control — untreated extreme performers; intervention effect = treatment change − control change. Without control, regression cannot be ruled out. 5. Calibrate causal claim — did the change exceed the regression baseline? By how much? Confidence? 6. Adjust action — credit/blame intervention only if it exceeds regression; recommend control structure for future tests.
When to use it
- someone says 'things always bounce back,' wonders why an intervention seemed to work on a struggling team, is surprised their star performer regressed, is evaluating whether a coaching/bonus/firing had an effect, or is interpreting before-vs-after performance changes without a control group
When not to use it
the measurement is genuinely noise-free (e.g., deterministic process outputs); the underlying signal has demonstrably changed due to a structural shift (new product launch, technology change).
Worked example
Galton 1886 + Kahneman 2011 Chapter 17 Israeli Air Force
Francis Galton (1822–1911) was a Victorian polymath: cousin of Charles Darwin, founder of the field of biometrics, and (less savorily) the founder of eugenics. His 1886 paper documenting the regression phenomenon was a methodological breakthrough. He had collected height data on 928 adult children and 205 sets of parents at the International Health Exhibition in London (1884) — a dataset that was, for its time, extraordinarily large and well-controlled.
Install this skill (free, MIT)
npx skills add deciqAI/knowledge-skills