Regression to the Mean in Biomarker and Intervention Studies
Key Takeaways
- An unusually high or low biomarker value is often followed by a value closer to the population average, even when no biological change has occurred. [1] [2]
- The effect is strongest when measurements are variable and participants are selected because their baseline values are extreme. [1] [3]
- A before-and-after improvement in one selected group cannot by itself separate an intervention effect from regression to the mean, measurement variation, or change over time. [1] [3]
- Concurrent randomized comparison groups, repeated baseline measurements, and appropriate baseline-adjusted analyses make intervention effects easier to distinguish from statistical fluctuation. [4] [5]
Biomarkers fluctuate because of biological variation, sampling conditions, laboratory variation, and measurement error. When a study begins with an unusually extreme result, some of that extremity is likely to be temporary. A later measurement can therefore look improved even if the participant's underlying state has not changed. This recurring pattern is called regression to the mean. [1] [2] [3]
Who This Is Useful For
This page is useful for readers evaluating studies that measure blood pressure, lipids, inflammatory markers, metabolic measures, physical performance, biological-age estimates, or other outcomes before and after an intervention. It is especially relevant when participants entered a study because a first measurement was unusually high or low, or when a report presents improvement in a single group without a concurrent comparison group. [1] [3]
The Basic Mechanism
An observed biomarker value can be thought of as a relatively stable component plus temporary variation. If the first value is extreme partly because temporary variation pushed it upward, that same upward contribution is unlikely to recur in exactly the same way at the next measurement. The second value will, on average, be less extreme. The same logic applies in the opposite direction to an unusually low first value. [1] [2]
Regression to the mean is therefore a statistical expectation, not a force that makes every individual result move toward normal. Some repeated values move farther from the mean, and genuine biological change can occur at the same time. The problem is that a simple pre/post difference does not identify how much of the observed movement came from each source. [1] [3]
Why Selection on an Extreme Baseline Matters
The effect becomes especially important when eligibility depends on crossing a biomarker threshold. People just above a high-risk cut-off may include some whose typical values are lower but whose screening measurement happened to be high. On retesting, the group average can fall even without an effective intervention. More measurement variation and more extreme selection generally produce more regression to the mean. [1] [2] [3]
How It Appears in Common Study Designs
| Study Situation | What May Be Observed | What the Observation Alone Cannot Establish |
|---|---|---|
| Extreme value followed by retesting | The second value is closer to the usual range | That the underlying biological state improved [1] |
| Single-group pre/post intervention | The group mean improves after treatment | How much change came from treatment rather than statistical fluctuation or other time-related influences [3] |
| Randomized controlled trial | Both groups may move toward the mean | A treatment effect from within-group change alone; the randomized between-group contrast is required [5] [6] |
| Observational change analysis | Baseline status is associated with subsequent change | A causal effect, because change scores create additional interpretive problems in non-randomized data [7] |
Why an Uncontrolled Before-and-After Result Is Ambiguous
Suppose a study recruits people whose biomarker is above a threshold, gives everyone an intervention, and reports a lower group mean at follow-up. The reduction is compatible with an intervention effect, but it is also compatible with regression to the mean. Secular trends, changes in co-interventions, altered measurement conditions, and natural biological change can add further explanations. Without a suitable comparison, these components cannot be cleanly separated. [1] [3]
A larger pre/post change among participants with the most extreme baseline values is not decisive evidence that those participants responded best. Baseline values and change scores are mechanically related because the baseline value is part of the subtraction used to calculate change. This can create or exaggerate an apparent relationship between starting level and response. [5] [7]
What Randomization and Comparison Groups Add
In a well-conducted randomized trial, regression to the mean can occur in both intervention and control groups. The treatment effect is estimated by comparing the groups, not by asking whether the intervention group changed significantly on its own. Separate within-group significance tests do not test whether the groups differ from each other and can therefore be misleading. [5] [6]
Randomization does not stop biomarkers from fluctuating. It makes the groups comparable in expectation, so shared regression to the mean and other background changes are less likely to be mistaken for the treatment contrast. The interpretation still depends on attrition, adherence, measurement quality, and whether the planned between-group analysis was followed. [5] [6]
Design and Analysis Features That Help
- Repeated baseline measurements: Averaging more than one pre-intervention measurement can reduce the influence of a single unusually high or low value and improve precision. [1] [4]
- A concurrent comparison group: A control group observed over the same period provides information about changes that occur without the intervention under study. Random allocation strengthens the causal interpretation of the between-group difference. [3] [5]
- Baseline-adjusted analysis: For continuous outcomes in randomized trials, an analysis of follow-up values adjusted for baseline, commonly using analysis of covariance, can handle chance baseline imbalance and is often more precise than an unadjusted change-score analysis. [5]
- Prespecified methods: The choice between follow-up, change-score, and repeated-measures analyses should be based on the study question and specified before results are examined. [5]
- Reliable measurement: Lower random measurement variation reduces the component of apparent change attributable to an extreme noisy baseline value. [1]
Questions to Ask When Reading a Study
- Were participants selected because the baseline biomarker was unusually high or low?
- Was eligibility based on one measurement or an average of repeated measurements?
- How variable and repeatable is the biomarker under the study's measurement conditions?
- Was there a concurrent comparison group, and was allocation randomized?
- Does the reported conclusion rely on within-group pre/post change or on a direct between-group comparison?
- Was the primary analysis adjusted for the baseline outcome as prespecified?
- Are participants described as “responders” partly because they began with extreme values?
These questions do not produce a universal statistical verdict, but they reveal whether a study design contains enough information to distinguish an intervention effect from ordinary variation. [1] [5] [7]
What This Does Not Mean
- It does not mean every improvement after an extreme baseline value is artificial; real change and regression to the mean can occur together. [1]
- It does not mean the population mean is biologically ideal or that every person's values must move toward it.
- It does not mean pre/post measurements are uninformative; their interpretation depends on selection, comparison, measurement reliability, and analysis. [4] [5]
- It does not mean subtracting baseline from follow-up is always invalid; the inferential problem depends on the design and the causal question, especially in observational studies. [7]
Practical Interpretation Examples
- If participants with the highest baseline biomarker values improve the most: Part of that pattern may arise because extreme starting values contain more temporary variation and because baseline is mathematically coupled to the change score. [1] [7]
- If a pilot study reports improvement in its only group: The result can establish that values changed after enrolment, but not how much of the change was caused by the intervention. [3]
- If both arms of a randomized trial improve: Shared improvement does not show that both treatments worked; regression to the mean and other background changes may affect both arms, while the prespecified between-group contrast estimates the comparative effect. [5] [6]
Summary
Regression to the mean is a predictable feature of repeated, imperfectly correlated measurements. It is most consequential when studies select extreme baseline values and then interpret later movement toward the average as intervention benefit. Stronger interpretation comes from reliable repeated measurement, concurrent randomized comparisons, and analyses that match the design and causal question. [1] [4] [5]
References
- Barnett, A. G., van der Pols, J. C., & Dobson, A. J. (2005). Regression to the mean: what it is and how to deal with it. International Journal of Epidemiology. https://academic.oup.com/ije/article/34/1/215/638499
- Bland, J. M., & Altman, D. G. (1994). Some examples of regression towards the mean. BMJ. https://www.bmj.com/content/309/6957/780
- Davis, C. E. (1976). The effect of regression to the mean in epidemiologic and clinical studies. American Journal of Epidemiology. https://pubmed.ncbi.nlm.nih.gov/984023/
- Frison, L., & Pocock, S. J. (1992). Repeated measures in clinical trials: analysis using mean summary statistics and its implications for design. Statistics in Medicine. https://pubmed.ncbi.nlm.nih.gov/1485053/
- Vickers, A. J., & Altman, D. G. (2001). Analysing controlled trials with baseline and follow up measurements. BMJ. https://www.bmj.com/content/323/7321/1123
- Bland, J. M., & Altman, D. G. (2011). Comparisons within randomised groups can be very misleading. BMJ. https://www.bmj.com/content/342/bmj.d561
- Tennant, P. W. G., et al. (2022). Analyses of ‘change scores’ do not estimate causal effects in observational data. International Journal of Epidemiology. https://academic.oup.com/ije/article/51/5/1604/6294759
This content is provided for educational purposes only and does not constitute medical advice.