Regression to the mean: why scores can rise without a programme
A group selected for low baseline scores can improve without an intervention. Work through the 50-to-60 example and check what your study design can establish.
AcademicFix Editorial Team · Sources reviewed
- Population mean at both tests
- 70
- Selected group’s baseline mean
- 50
- Expected follow-up mean
- 60
Fictional example. The 60 is a model expectation, not an observed score.
Work through the three baseline scoresYou select students who scored lowest on a test. At follow-up, the group averages ten points more. How much did the programme contribute? That rise alone cannot answer the question. If the tests are comparable, the population mean and spread stay the same, and the scores are positively but imperfectly correlated, the selected group’s mean can move closer to the population mean even without a programme. This is regression to the mean. Barnett et al., 2005 · Linden, 2013
Use this guide if you are planning or reporting a before-and-after study with a continuous score. The worked calculation illustrates one model; it does not estimate how much of a real change came from regression to the mean.
On this page
Why does selecting on baseline change the comparison?
A first score may reflect the skill you want to measure, that day’s circumstances and measurement error. Selecting only low scorers can include students whose first result was temporarily low. That temporary component need not recur in the same direction or by the same amount. Selecting unusually high scores can produce the reverse pattern. This is a statement about the group’s expected average: a student’s individual score can still fall. Barnett et al., 2005 · Linden, 2013
Taking everyone in a class does not select on the first score in the same way as including only students who scored 55 or below. Write down exactly which of these decisions your study made.
How could the average rise from 50 to 60 without a programme?
In this fictional example, the two scores have a joint normal distribution, both population means are 70, both standard deviations are 10 and their correlation is 0.50. There is no intervention. Under these assumptions, the conditional mean is 70 + 0.50 × (baseline score − 70). With unequal standard deviations in the joint normal model, the slope is the correlation multiplied by the ratio of follow-up to baseline standard deviation. Linden, 2013, equations 5–6 · Khan and Olivier, 2025, section 2
| Baseline score | Calculation | Expected follow-up |
|---|---|---|
| 45 | 70 + 0.50 × (45 − 70) | 57.5 |
| 50 | 70 + 0.50 × (50 − 70) | 60 |
| 55 | 70 + 0.50 × (55 − 70) | 62.5 |
| Mean: 50 | 60 |
The 10-point rise is the model’s expectation for these three baseline scores. There are no observed follow-up scores in this example. Do not subtract ten points from a real programme’s result. If the population mean changes, the spread differs between tests, or scores pile up at the floor or ceiling, the simple calculation may not describe your data. Linden, 2013 · Khan and Olivier, 2025
How can you design a comparison that answers the programme question?
- Write the eligibility rule before collecting data. Specify the test and the eligible score range. If feasible, plan several baseline measurements before selection to reduce the influence of fluctuations on any one day. Barnett et al., 2005 · Linden, 2013
- Apply the same rule to both groups. Where possible, randomly assign eligible students to the programme and comparison groups. Use comparable test conditions and timing. Comparing low-scoring programme participants with the whole school does not give the same comparison. Barnett et al., 2005 · Linden, 2013
- Estimate the difference between groups directly. In a suitable controlled design with a continuous outcome, you can prespecify an analysis that models follow-up using baseline score and group. Report the estimated difference and its confidence interval. Including baseline in a one-group analysis cannot replace a missing comparison group. Barnett et al., 2005, equation 3 · Vickers and Altman, 2001
Regression to the mean and a programme effect can coexist. Practice with the test, learning outside the programme and different testing conditions may also contribute. If your design cannot separate these explanations, report the observed change and the limits on its interpretation. Barnett et al., 2005 · Linden, 2013
What can you report if you only have one group?
Fictional reporting example: “The programme enrolled students who scored 55 or below on the baseline test. Their mean score rose at follow-up. Without a comparison group, we could not determine how much of the increase was due to the programme. Selection on baseline score and practice from taking the test a second time are possible explanations.”
In your own report, state the selection rule, number of participants analysed, time between tests and observed difference. Report how many follow-up scores are missing. For the next decision, see the missing data guide.
Three checks before interpreting a score change
Does a high correlation rule out regression to the mean? In the equal-spread example, the expected movement towards the mean shrinks as the correlation approaches 1. That is a model result; high correlation alone does not establish the cause of a change. Linden, 2013, equations 5–6 · Khan and Olivier, 2025, section 2
Must every student improve? No. The calculation is about expected scores, and individuals can end up above or below their own expectation. Barnett et al., 2005 · Linden, 2013
Does a statistically significant change show the programme worked? A test of change within one group does not separate a programme effect from other explanations. See p-values, effect sizes and confidence intervals.
Check your plan or results paragraph
- Did the first score determine eligibility? How did you choose the threshold?
- Were the testing conditions, scale and follow-up interval comparable?
- Did the comparison group come from the same eligible population?
- Does the report distinguish the observed change from an estimate of the programme effect?
Next, use the guide to choosing a statistical analysis to match the analysis to your research question and design.
Sources
- Barnett, van der Pols and Dobson (2005), Regression to the mean: what it is and how to deal with it
- Linden (2013), Assessing regression to the mean effects in health care initiatives
- Khan and Olivier (2025), Regression to the mean for bivariate distributions, section 2
- Vickers and Altman (2001), Analysing controlled trials with baseline and follow up measurements