In 1886, Francis Galton measured 930 adult children and their parents. He noticed something that bothered him. The children of unusually tall parents were usually tall, but not as tall as their parents. The children of unusually short parents were usually short, but closer to the middle. No one had intervened. The numbers had simply pulled back toward the average.
That pull is regression to the mean. It is a statistical effect, not a flaw in anyone's eyesight. The bias comes afterward, when we see the drift and invent a cause for it. A person feels awful, tries a remedy, feels better, and credits the remedy. Sometimes the remedy deserves the credit. Often part of the improvement was coming anyway.
Why Extreme Readings Slide Back
Any single measurement mixes a stable part with a random part. Take two blood pressure readings from the same person and they will differ. If you select people because the first reading was high, you have selected some who are truly high and some who were nudged up by chance. On the second reading the chance part is unlikely to repeat. The group's average falls, and nobody did anything.
Two things control the size of the effect. The weaker the correlation between the two measurements, the more regression you get. And the farther the starting value is from the population mean, the more room there is to fall back.
Two Numbers You Can Picture
Consider a blood sugar example from the UVA Library. A person's true mean is 99. One day the reading lands two standard deviations above that. It is a rare spike, and the follow-up reading comes in much closer to 99. The person did not get healthier in between. The spike was noise.
The same source describes a simulation of diastolic blood pressure for 1,000 people. The true mean was 90 mmHg, and the two readings correlated at about 0.80. A reading of 95 or higher counted as hypertensive. Even with a fairly strong correlation, the people flagged at the first reading averaged lower at the second. No drug was involved. Only selection and noise.
When the Drift Looks Like a Win
Picture a clinic that screens for high distress scores, enrolls the top scorers in a new therapy, and finds lower scores later. One story says the therapy worked. Another says many of those first scores were inflated by a bad week, and the group would have improved somewhat regardless. Both can be partly true.
Regression does not mean interventions never work. It means a simple before-and-after comparison cannot say how much of the change belongs to the treatment. The effect can run the other way too. A real benefit can look smaller than it is, or a failing program can look successful.
A recent opinion piece shows the same logic in public debate. Criminologist James Alan Fox, quoted in a Now Georgia commentary, described a drop of roughly 24 percent in U.S. mass killings in 2025 compared with 2024. The count was 17, the lowest since 2006. Fox attributes much of the drop to a return to average levels after spikes in 2018 and 2019. That is one expert's reading in an opinion column, but it illustrates the point: a decline after a spike is not proof that a policy succeeded.
The Fix Is a Comparison Group
The standard defense is a control group. If one group gets the program and a similar group does not, both experience the same regression. Whatever extra change the treated group shows is a better estimate of the program's effect. J-PAL describes its work as randomized impact evaluations, which are designed to separate program effects from other changes. The sources here do not offer a step-by-step correction for data that lacks a comparison group.
Three Questions Before You Credit the Fix
- Was the starting measurement unusually high or low compared with the usual level?
- Is there a comparison group that did not receive the treatment?
- Would the result have been expected to drift back toward average anyway?
A fourth is worth adding. How closely do the two measurements track each other? If the link is weak, expect more drift. When a change follows a spike, ask these questions first, and hold off on the celebration until you have answers.
Please seek a professional. This article is general information, not financial, tax or insurance advice. Please seek a licensed financial, tax or insurance professional about your own situation.
Sources
- Regression to the Mean: Statistical Bias Can Mislead Interpretation in Cartilage and Osteoarthritis Clinics and Research - PMC
- Effect of regression to the mean on decision making in health care - PMC
- Regression to the Mean — CASRAI
- Regression to the mean: What it is and why it matters for impact evaluations | The Abdul Latif Jameel Poverty Action Lab
- Regression to the Mean and Change Score Analysis | UVA Library
- https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGiBoZ7FPNbCVhBgW3W3hsNASk-1K3Y6ODn0BCp8TWvX17kZk030gGNpKD5vLJpGjUmugAYVcHkWE9H16uovXd9qer6lDHHlC_0ztnjGW4-3_oBVlJTdM6yAMNoNk3BrRbekn_g1BmFbMBCcPupnXAFsbxt2i7sMi_CWx649PHoxywevCf2lGwozbEh3TZ8w4Unm4LP0mpc6fa8i9au0LgrNA==
- The ‘regression’ Illusion: Why we shouldn't celebrate 2025's safety stats yet - Now Georgia
Siobhan O'Reilly investigates the friction between urban decay and digital expansion for Deeperpectives. She spent a decade reporting from transit hubs and temporary settlements across the rust belt. Her work prioritizes the raw data of lived experience over polished corporate narratives.
