Mind · Dynamics
RegressionToMean
A result picked out because it was extreme is usually followed by one closer to typical, because part of what made it extreme was chance
Definition
Mechanism
A measurement that varies partly by chance comes out extreme: a spike in errors, a record quarter, an unusually bad score. Because part of the extremity was chance, a later measurement of the same thing is usually closer to typical, and the expected return grows with the share of chance in the measure. Extremes are often what prompt a response, so the return tends to arrive just after it, and any real effect of the response adds to the return or offsets it rather than replacing it. Estimating what the response did therefore needs a comparison that shares the regression, such as similar extremes left alone, or the return expected from how noisy the measure is. A measure with no chance component does not regress.
Invariants
- A measurement with a chance component was picked out or noticed because it was extreme.
- Later measurements of the same thing are expected to lie closer to typical, by an amount that grows with the share of chance in the measure.
Risks
Failure modes
Return read as effect
the move back toward typical after an extreme is credited wholly to the response.
Decline read as harm
praise or reward after an unusually good result is blamed for the ordinary result that follows.
Effect dismissed as regression
a real effect of a response is written off, although regression accounts only for the part of the change that the measure's noise predicts.
Relationships
Root pattern: builds on nothing in this library.
Nothing in this library builds on this pattern yet.