Regression to the Mean

Regression to the Mean

Overview

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.

Named by Francis Galton (1886) studying parent-child height: tall parents have tall children, but slightly shorter; short parents have short children, but slightly taller. Kahneman's Israeli Air Force example (2011, Ch. 17) is the most-cited operational case — flight instructors concluded punishment works and praise doesn't, but were observing regression, not causation.

Composes with survivorship-bias (extreme survivors regress), probabilistic-thinking (regression is probabilistic), narrative-fallacy (regression drives post-hoc narratives), fundamental-attribution-error (attributing regression to character/intervention is FAE).

When to Use

• Evaluating the effect of an intervention on extreme performers (struggling teams, top sales reps, low-rated branches)
• Designing or interpreting A/B tests or pilot programs
• Reviewing year-over-year performance changes
• Hiring or promoting top performers
• Evaluating investment fund performance
• Analyzing acquisition outcomes
• Building or critiquing causal claims about training, coaching, or feedback
• Judging whether an AI startup's viral quarter, a fund's AI hot streak, or a model's benchmark spike is a durable trend or an outlier reverting toward average (AI hype extrapolation)
• Someone says "regression to the mean," "things will average out," "they always come back"

Not when: the measurement is noise-free (rare in business); the underlying signal is genuinely changing (e.g., the business model fundamentally improved); the intervention is so substantial that no plausible regression can explain the effect.

Coaching Novices (Adaptive Front Door)

• Engine mode: user brings a specific performance change → run The Process directly.
• Coach mode: user is new to the framework → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-line: when an extreme performance is followed by a less-extreme one, regression to the mean is the default explanation — not the intervention.
2. Check fit. Did we observe an extreme value followed by retest? If yes, regression applies.
3. Elicit the structure. What was the extreme observation? What was the follow-up? What intervention happened in between? Is there a control or baseline?

[WAIT — do not advance until user responds]

1. One question at a time: how much regression would we expect without the intervention? Does the observed change exceed it? What's the noise level in the measurement?

[WAIT — do not advance until user responds]

1. Close: regression vs. causal effect separated + control or baseline identified + claim calibrated.

[WAIT — do not advance until user responds]

The Process

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.

Output: Regression Analysis

# Regression Analysis: 
Extreme observation — Subject: | Value: | Period:
Intervention — What: | When: | By whom:
Retest — Follow-up value: | Toward mean (Y/N): | Magnitude:
Expected regression — Noise level: | Predicted magnitude:
Control — Group: | Change: | Intervention effect = treatment − control:
Causal calibration — % regression: | % intervention: | Confidence:
Adjusted action — Credit/blame verdict + future evaluation process:

→ Method in Action: Galton 1886 + Kahneman 2011 Chapter 17 Israeli Air Force → 2026 lens: The AI Hot Streak — viral quarters, fund streaks & benchmark spikes (2023–2026)

Pack: Regression Across Domains

Applying It Well

• Use longer time windows to reduce noise (yearly vs. quarterly)
• Always ask: "what would regression alone predict?" before crediting an intervention
• Control groups are not optional when evaluating interventions on extreme performers
• "Praise hurts, punishment helps" is the signature cognitive error of regression blindness

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags

• An intervention is being evaluated on extreme performers without a control group
• A single-period change is being treated as evidence of a stable trend
• Causal credit/blame is being assigned to interventions that coincided with regression
• "The intervention worked" claims lack baseline regression estimates
• Promotion / firing decisions are based on single-period extreme performance
• "Comeback" narratives credit decisive action, not regression

Verification

• Extreme observation identified; retest compared to it
• Noise level and regression magnitude estimated
• Control group (or baseline regression expectation) used
• Causal effect separated from regression; claim calibrated
• Long-window measurements used where possible; future evaluation process set

→ Primary sources: references/sources.md

---

Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/regression-to-the-mean · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/regression-to-the-mean.json