Galton's Regression: The Mean's Quiet Gravity

"Francis Galton noticed that the sons of very tall fathers tend to be tall but shorter than their fathers, and the sons of very short fathers tend to be short but taller than their fathers. He called it 'regression toward mediocrity.' We call it regression to the mean. It runs your life whether you see it or not."

The Observation

In the 1880s, Francis Galton plotted the heights of fathers against the heights of their adult sons. He found a clear pattern: extreme fathers had extreme sons, but the sons were less extreme than the fathers. Very tall fathers (say, two standard deviations above the mean) tended to have sons who were about one standard deviation above the mean — still tall, but closer to the population average. Very short fathers had similarly less-short sons. Galton called this 'regression toward mediocrity' (which sounds harsher than it is — he just meant 'movement back toward the typical').

What Galton was seeing was a fundamental statistical phenomenon, not a biological one. The same pattern appears in any noisy measurement: extreme observations tend to be followed by less-extreme observations, because the extreme was partly signal and partly noise, and the noise refreshes between observations.

Why It Happens

An extreme observation is, almost by definition, an above-average accumulation of luck (positive or negative). The signal (the underlying true value) is plus the noise (the random fluctuation) is what produced the extreme. The signal persists; the noise resets. On the next observation, the signal is roughly the same, but the noise has rolled fresh dice — and on average, the new dice are less extreme than the previous ones.

This is why extreme test scores tend to be followed by more average test scores. Why athletes who win an award one year often perform less impressively the next. Why a company that had a banner year tends to have a more normal year following. The 'regression' is not punishment for past success or reversion to mediocrity; it is the noise component of the previous extreme not repeating.

The Citizen Cost of Not Seeing It

Failing to see regression to the mean is the source of many wrong inferences:

The second album curse: a band's first album was a huge hit, partly luck. The second album reverts to the band's actual talent level, which is less extreme. The album is judged a disappointment. The 'curse' is the noise resetting.
Sports slumps: a player who had an exceptional season is partly skilled and partly lucky. Next season, the luck refreshes; the player is still skilled but less lucky; performance regresses. Commentators construct elaborate narratives about why the slump happened.
Educational interventions: a school in the bottom 10% receives an intervention. Next year, its scores improve. The intervention takes credit. But the bottom 10% were extreme on noise, and would have improved anyway by regression alone.
'The therapy worked': people seek therapy when they are at extremes (very anxious, very depressed). They improve. Therapy gets credit. Some of the improvement is regression to the mean.

The Operating Principle

When you see an extreme observation followed by a less-extreme one, your default explanation should include regression to the mean as a candidate cause — possibly the dominant cause. Reach for a specific causal explanation only after you've ruled out the noise-resetting hypothesis. Most stories in everyday life that have the shape 'X used to be extreme, then it changed' are partly regression to the mean. Naming the candidate is the citizen's defense against attribution to causes that may not exist.

Code

Galton-style regression to the mean, simulated·python

import numpy as np
rng = np.random.default_rng(240)

# Galton-style simulation: father's height = true component + noise.
# Son's height = same true component + fresh noise.
# Show that sons of extreme fathers regress toward the mean.

N = 100_000
true_signal = rng.normal(loc=170, scale=5, size=N)        # population mean 170, sd 5
noise_father = rng.normal(loc=0, scale=4, size=N)
noise_son = rng.normal(loc=0, scale=4, size=N)

father_height = true_signal + noise_father
son_height = true_signal + noise_son

# Group fathers by their height bins and compute average son height.
for lo, hi in [(155, 160), (165, 170), (170, 175), (180, 185), (190, 195)]:
    mask = (father_height >= lo) & (father_height < hi)
    if mask.sum() == 0:
        continue
    avg_father = father_height[mask].mean()
    avg_son = son_height[mask].mean()
    print(f"Fathers in [{lo}, {hi}):  avg father = {avg_father:.2f}cm   avg son = {avg_son:.2f}cm")

# Extreme-tall fathers have tall sons but the sons are CLOSER to the mean.
# Extreme-short fathers have short sons but the sons are CLOSER to the mean.
# That gap is regression to the mean: the noise component of the father's
# extremity does not transmit to the son, who gets fresh noise.

Exercise

Identify three 'X used to be great and then declined' stories in your life or in news (a band, an athlete, a friend's startup, a stock). For each, ask: how much of the original 'great' was the signal and how much was noise? Most of the time, you cannot decompose precisely, but the question itself prevents you from over-attributing the decline to causes ('he got complacent') that may be regression in disguise.

Hint

The shape 'first was extreme, then normal' is the regression signature. The signature alone is not evidence of decline — only of the previous extreme having a noise component that did not repeat.