What Is Regression to the Mean, and Why Can It Mislead Before-and-After Studies?

Can Participants Improve Simply Because Their Initial Scores Were Unusually Low?

You recruit students who performed poorly on a diagnostic assessment, introduce a learning intervention, and administer another assessment several weeks later. Their average scores improve. The intervention appears successful.

But suppose some students would have obtained higher scores on the second assessment even without receiving the intervention. Their initial performance might have been affected by fatigue, temporary anxiety, unfamiliar questions, or ordinary measurement variation.

When participants are selected because of unusually extreme observations, subsequent measurements often become less extreme. This statistical phenomenon is called regression to the mean. It can make an ineffective intervention appear beneficial, particularly when researchers compare only one pretest with one posttest.

Extreme Observations Tend to Be Less Extreme When Measured Again

How Regression to the Mean Produces Misleading Changes

Why Do Extreme Scores Tend to Become Less Extreme?

Observed measurements often contain both relatively persistent characteristics and temporary influences. A student's examination score, for example, may reflect underlying knowledge alongside fatigue, distraction, test-specific familiarity, and random measurement variation.

When a student obtains an exceptionally low score, part of that extremity may reflect unfavorable temporary circumstances. Those circumstances are unlikely to recur with precisely the same magnitude during the next assessment.

Consequently, the student's expected subsequent score may be closer to the population average than the original observation.

The same reasoning applies to unusually high observations. Participants who perform exceptionally well on one occasion may obtain less exceptional scores when assessed again.

The phenomenon concerns expected patterns across repeated measurements, not a deterministic movement of every observation toward the mean.

Does Regression to the Mean Require Measurement Error?

No. Measurement error is one important source of imperfect correlation between repeated observations, but it is not the only source.

Genuine short-term fluctuations in the underlying outcome can also produce regression to the mean. Blood pressure, classroom engagement, test performance, and many behavioral measures may vary over time even when the measurement instrument functions correctly.

For regression to the mean to arise in the conventional repeated-measurement setting, the observations must be less than perfectly correlated, and attention is commonly directed toward observations selected for their extremity.

Researchers should therefore distinguish random measurement error from genuine temporary variation. Both can contribute to the observed pattern, but they have different substantive interpretations.

Why Selection Based on Extreme Scores Creates a Particular Problem

Many interventions deliberately target participants with unusually poor outcomes. Students with low achievement receive remedial instruction. Patients with high symptom scores enter treatment programs. Organizations with exceptionally poor performance receive additional support.

Such selection can be appropriate for practical and ethical reasons. However, it creates a methodological challenge when intervention effectiveness is evaluated using the same extreme observation that determined eligibility.

If participants are selected because their initial scores were unusually low, their subsequent scores may improve partly because the temporary influences contributing to the original extremity have diminished.

Without an appropriate comparison, researchers may attribute that improvement to the intervention.

This concern is distinct from the general possibility that participants naturally develop or encounter external events. Regression to the mean arises from the statistical relationship between selection, variability, and repeated measurement.

How Does Measurement Reliability Affect Regression to the Mean?

The relationship between repeated measurements matters. When scores are highly consistent across occasions, an extreme observation is more informative about a participant's expected subsequent performance.

When repeated measurements are less strongly correlated, an extreme score generally provides weaker evidence that the next observation will be equally extreme.

Under a classical measurement model, lower reliability can therefore increase the expected regression of extreme observed scores toward the population mean.

However, reliability is not the only determinant. Selection thresholds, population characteristics, the outcome distribution, and genuine changes between measurements also matter.

Researchers should not assume that a published reliability coefficient completely characterizes the amount of regression to the mean in their particular study.

What Does the Statistical Relationship Look Like?

A simplified bivariate normal model illustrates why an extreme first observation predicts a less extreme second observation when their correlation is below one.

The expected second score is 38, which is closer to the population mean of 50 than the initial score of 30.

This does not mean the individual must score 38. It is a conditional expectation under the specified model. Individual outcomes can move in either direction.

Nor does the calculation imply an intervention effect. No intervention is required for the expected movement from 30 toward 38.

When the two measurements have equal means and standard deviations, the expected standardized deviation from the mean at the second occasion is the first standardized deviation multiplied by their correlation.

Regression to the Mean Is Not the Same as Natural Improvement

Natural improvement refers to a genuine change in the underlying outcome that may occur without the intervention being evaluated.

Regression to the mean concerns the expected relationship between repeated observations when an extreme measurement is used to select or characterize participants.

Both phenomena can occur simultaneously. A student might genuinely learn through ordinary coursework while also obtaining a less extreme score because temporary influences on the original assessment have changed.

Why Can Regression to the Mean Make an Intervention Look Effective?

Suppose researchers recruit participants whose baseline scores fall below a particular threshold. They then provide an intervention and observe higher scores afterward.

The observed improvement contains several potentially contributing components, including any genuine intervention effect, other changes occurring over time, and regression to the mean.

A single-group comparison does not ordinarily separate these components.

Consequently, a statistically significant pretest-posttest difference may coexist with substantial uncertainty about whether the intervention produced the improvement.

This is one reason that before-and-after improvement does not automatically establish causation.

Can Regression to the Mean Also Make an Intervention Look Harmful?

Yes. If participants are selected because their initial outcomes are unusually favorable, subsequent measurements may be less favorable even without a harmful intervention.

For example, an institution might introduce a new administrative procedure immediately after recording an exceptionally high performance indicator. A later decline could partly reflect regression to the mean rather than deterioration caused by the procedure.

The direction of the apparent change depends on whether the selected observations were unusually low or unusually high.

Regression to the mean is therefore not inherently an improvement effect. It is a statistical consequence of imperfectly correlated repeated observations and selection or attention based on extremity.

Why Is a Control Group Useful?

A credible comparison group can help distinguish intervention-related changes from changes that would occur without the intervention.

For example, if both intervention and comparison participants are selected using the same low-score criterion, both groups may experience regression to the mean.

Comparing their subsequent outcomes may help estimate the additional effect associated with the intervention, provided the comparison groups are otherwise appropriate.

Random assignment after eligibility determination can be particularly valuable because it helps distribute selection-related influences between groups in expectation.

However, simply adding a comparison group does not guarantee validity. The appropriateness of the comparison group and its selection process remain consequential.

Can Multiple Baseline Measurements Reduce the Problem?

Repeated pre-intervention observations can help researchers determine whether an extreme score is unusual relative to a participant's broader pattern.

For example, several assessments may reveal that a student selected after one exceptionally low score ordinarily performs at a higher level.

Using multiple baseline measurements may improve characterization of the starting condition and help identify fluctuations or existing trends.

Nevertheless, repeated baselines do not automatically eliminate regression to the mean. The method used to select participants and the statistical properties of the observations still matter.

Researchers should prespecify how baseline observations will be used rather than selecting the combination that produces the largest apparent improvement.

Can Statistical Adjustment Eliminate Regression to the Mean?

Appropriate statistical methods can help address regression to the mean, particularly in well-designed controlled studies.

Baseline-adjusted analyses, including analysis of covariance in suitable randomized trials, can improve precision and account for initial outcome differences.

However, adjusting posttest outcomes for baseline scores does not automatically identify an intervention effect in an uncontrolled study.

When treatment assignment depends on baseline performance, the selection mechanism and assumptions underlying the analysis require particular attention.

Statistical adjustment cannot substitute for a credible counterfactual merely by including the baseline score as a covariate.

What If Participants Are Selected Using a Threshold?

Selection thresholds are common in educational and clinical interventions. Students below a proficiency cutoff may receive additional instruction, while individuals above a symptom threshold may receive treatment.

Such rules can produce regression to the mean among selected participants.

In some settings, a carefully designed regression discontinuity study can exploit a treatment-assignment cutoff to estimate a local causal effect. However, this requires appropriate assumptions, including continuity of potential outcomes around the threshold and the absence of problematic manipulation of the assignment variable.

Simply comparing participants' scores before and after crossing a threshold is not a regression discontinuity design.

The distinction illustrates why the mechanism of treatment assignment matters as much as the observed score change.

How Regression to the Mean Can Mimic an Educational Intervention Effect

The observed improvement is 52 − 40 = 12 points.

Under the simplified hypothetical model, 8 points of improvement would be expected from regression to the mean alone, because 48 − 40 = 8.

The observed follow-up mean exceeds that model-based expectation by 4 points. However, the 4-point difference cannot automatically be interpreted as the intervention effect. The model may be misspecified, and other influences such as ordinary coursework, maturation, or concurrent educational activities may also contribute.

A stronger study might randomly assign eligible students to the supplementary program or an appropriate comparison condition after applying the same eligibility criteria.

If both groups experience similar regression-to-the-mean effects, their between-group contrast can provide more credible evidence about the intervention's additional contribution.

The example demonstrates why the entire 12-point improvement should not be attributed to the program merely because it occurred after implementation.

Common Misunderstandings About Regression to the Mean

How to Protect Your Study From Misleading Regression-to-the-Mean Effects

Begin by examining how participants entered the study. If eligibility depended on unusually high or low baseline values, regression to the mean deserves explicit consideration.

Next, investigate the reliability and temporal variability of the outcome. Determine whether the initial score could plausibly reflect temporary influences rather than a stable underlying condition.

Finally, select a design that provides credible evidence about what would have happened without the intervention.

When interpreting results, distinguish the observed change from the estimated intervention effect. A large improvement following an extreme baseline may be meaningful, but its causal interpretation requires more than the direction and magnitude of the difference.

The choice of baseline measurement should also be justified before outcome results are examined. Selecting the most extreme available baseline after seeing the follow-up can exaggerate apparent improvement.

Before Interpreting Improvement From an Extreme Baseline, Check:

Frequently Asked Questions About Regression to the Mean

Improvement From an Extreme Starting Point Is Not Automatically an Intervention Effect

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.