Regression to the Mean: Why Extreme Results Often Change

Regression to the mean is the tendency for unusually high or low observations to be followed, on average, by less extreme observations when the measurements are imperfectly related. It is a statistical pattern, not a force that makes people ordinary or guarantees that good results must be punished by bad ones.

The idea matters whenever action follows an extreme result. A school introduces extra help after unusually low test scores, a team changes training after a poor performance, or a manager celebrates improvement after a disappointing month. Some change might have occurred without the new action. Understanding regression to the mean helps us ask what evidence distinguishes a genuine effect from the consequences of selecting an unusually extreme starting point.

Key takeaways

  • Extreme observations can contain both persistent differences and temporary variation.
  • Selecting an extreme first result can produce less extreme average follow-up results.
  • The pattern does not mean every individual moves towards the average.
  • Before-and-after improvement alone does not establish an intervention’s effect.
  • Appropriate comparisons and repeated measurements help interpretation.

What does regression to the mean mean in practice?

Suppose a measurement reflects an underlying level plus temporary influences. A student’s score may reflect knowledge and preparation, but also the particular questions, momentary concentration and ordinary variation in performance.

Among students selected because their first scores were exceptionally high, some are likely to have benefited from unusually favourable temporary influences. On a later comparable test, those influences need not be equally favourable. Their average score can move closer to the wider average without any loss of underlying knowledge.

The same reasoning works at the low end. A group selected after unusually poor results may include people experiencing unusually unfavourable conditions. A later improvement does not automatically mean that something new caused it.

The selection rule is central. We are not claiming that everyone’s results must improve or decline. We are examining what happens after choosing a group partly because one measurement was extreme.

A simple model with no change in ability

Consider an explicitly fictional exercise. Every participant has an underlying score of 50. On each occasion, an independent random influence adds either minus 10, zero or plus 10, with equal probability. Nobody learns, forgets or changes ability between occasions.

First scores are therefore 40, 50 or 60. Select only those who scored 60. Their high score necessarily included the plus-10 influence in this simplified model.

On the second occasion, each selected participant again has an equal chance of scoring 40, 50 or 60. The expected second score is 50. The selected group’s expected result falls by 10 points even though its underlying level has not changed.

Expected means the average across many repetitions of the model, not the guaranteed outcome in one small group. Some individuals score 60 again. A small sample might even have an unexpectedly high average.

Try the low-score group

Now select those who initially scored 40. Their expected second score is also 50. They improve by 10 points on average without receiving any intervention.

The model is deliberately simple. Real people have different underlying levels and more complicated sources of variation. Its purpose is to isolate the statistical mechanism so that improvement and decline cannot be attributed to changing ability.

Why real examples are less tidy

In reality, unusually high performance can reflect genuine differences as well as temporary advantages. The best-scoring group may remain above average on a later test while becoming less extreme than before.

Regression to the mean therefore does not predict that every outstanding performer becomes average. It predicts a conditional pattern whose size depends on the relationship between measurements and the way the group was selected.

Measures that are strongly related tend to preserve more of the initial ordering. If repeated measurements were perfectly related in the relevant model, there would be no unexplained variation producing the usual regression pattern.

But a strong correlation still does not make every individual’s next result certain. Group-level expectations and personal predictions remain different tasks, particularly when circumstances can change between observations.

A numerical prediction example

In a simple linear prediction model using standardised scores, the predicted second score is the correlation multiplied by the first standardised score. A standardised score expresses distance from a mean in standard-deviation units.

Suppose the correlation is 0.6 and a first result is two standard deviations above its mean. The model predicts a second result 1.2 standard deviations above the second measurement’s mean: 0.6 multiplied by 2.

The prediction remains above average but is less extreme. It does not say the person must obtain that exact score, and it does not establish why either score occurred. It is a model-based average prediction.

Care is needed if the two tests have different scales or distributions. Raw numerical changes can reflect a changed test mean or spread as well as the regression pattern. Standardisation helps make the comparison explicit, but it does not repair an inappropriate model.

Why before-and-after stories can mislead

Imagine a school selects students after one exceptionally low test and offers a new revision programme. Their next scores rise. The programme may be useful, but the rise alone does not identify how much improvement it caused.

Some students might have improved through ordinary practice, a different test, changing circumstances or regression to the mean. Several explanations can operate simultaneously. Recognising alternatives does not prove the programme failed.

Earth Site’s guide to correlation and causation explains the broader problem. An observed sequence is not enough: we need a defensible account of what would have happened without the intervention.

This is especially important when an intervention is introduced precisely because things look unusually bad. The trigger for action selects the extreme starting point that makes regression particularly relevant.

What a comparison group can show

Suppose eligible students are allocated fairly between a new programme and an appropriate comparison condition. Both groups are selected through the same low-score rule. Both may experience regression to the mean.

If the programme group improves more, that difference can provide evidence about the intervention, subject to the study’s design and uncertainty. The comparison does not require pretending that the starting scores were permanent measures of ability.

For a fictional illustration, the programme group’s average rises by 12 points while the comparison group’s rises by 8. The difference in changes is 4 points. The relevant claim is not that the programme necessarily caused all 12 points of improvement.

Nor is the 4-point estimate automatically definitive. Random variation, unequal dropout, measurement differences and the allocation process still need examination. Arithmetic is only one part of causal evaluation.

Why matching requires care

If random allocation is unavailable, researchers may compare similar groups or use other designs. But matching only on an extreme observed score may not make groups identical in their underlying characteristics.

The design must address why participants entered each group and what else changed. A convenient comparison is not necessarily a valid counterfactual, meaning an estimate of what would have happened under the alternative condition.

Regression is not the gambler’s fallacy

The gambler’s fallacy assumes that an independent random process must compensate for recent outcomes. After several heads, someone may believe a fair coin is now more likely to produce tails. The coin has no such memory.

Regression to the mean does not require compensation. In the simple score model, the next random influence has the same probabilities as before. The apparent movement arises because we selected people whose first influence was unusually favourable or unfavourable.

Similarly, after selecting unusually high totals from repeated coin-toss sequences, later independent sequences are expected to be less extreme on average. No individual toss changes its probability to correct the earlier result.

The distinction is between a conditional group comparison and a claim that chance owes someone a reversal. Keeping that difference clear prevents a useful statistical idea from becoming a superstition.

Measurement error is not the only source of variation

A measurement can be accurate for the occasion and still vary across occasions. A student’s tiredness or a team’s opponent can produce real temporary differences, not merely a faulty instrument.

Earth Site’s explanation of measurement uncertainty and real-world variation helps separate these ideas. Instrument noise is one possible contributor, but genuine short-term variability can also make repeated results imperfectly related.

Repeating measurements before selecting a group can reduce the influence of one unusual occasion. An average across several comparable observations may better represent persistent performance than a single extreme result.

Nevertheless, repetition does not remove every problem. If conditions systematically change, or all measurements share the same bias, a larger collection can remain misleading. The purpose and comparability of the observations still matter.

Why the relevant mean must be specified

The phrase the mean can sound as though every person is pulled towards one worldwide average. That is not the idea. The relevant comparison depends on the population and model used.

An advanced class and a beginner class may have different distributions. A person selected as extreme within one group should not automatically be compared with an unrelated population. Age, task and context can affect what counts as unusual.

Changes in a whole population’s average can also coexist with regression. If everyone learns between two tests, the later mean may rise. A high-scoring subgroup could become less extreme relative to that new mean even while its raw scores increase.

This is why interpreting only the direction of a raw score change can miss the underlying statistical relationship. Ask both where the person started and how the relevant distribution changed.

Percentages can magnify the story

A small starting value can make a later increase look enormous in percentage terms. If a fictional score rises from 10 to 20, the increase is 100 per cent, but that description alone says nothing about the cause or the scale of the assessment.

Earth Site’s guide to percentage calculations explains the denominator. For causal interpretation, the selection rule and a suitable comparison are equally important.

A headline combining an extreme low baseline with a large percentage improvement can therefore be doubly persuasive. The calculation may be correct while the implied explanation remains unsupported.

How to check a claim in five questions

  1. Was the group selected because its first result was unusually high or low?
  2. How variable or reliable is the measurement across occasions?
  3. Were the first and later measurements genuinely comparable?
  4. What happened to an appropriate comparison group?
  5. What evidence separates an intervention effect from other changes?

These questions do not require dismissing every success story. They help identify what additional information would make a claim stronger. Sometimes the answer will support a real improvement beyond regression; sometimes the evidence will remain inconclusive.

Conclusion

Regression to the mean is a warning about how we select and interpret observations. Extreme starting points often contain temporary influences that need not recur, so later change cannot automatically be credited to whatever happened in between. The practical response is not cynicism but better comparison: define the population, understand the measurement and ask what would probably have happened without the new action.

Frequently asked questions

Does everyone move towards the mean?

No. It is an average pattern for appropriately selected observations, not a rule determining each person’s next result.

Does regression prove an intervention did nothing?

No. A real effect can coexist with regression. A suitable design is needed to estimate the intervention’s contribution.

Is it the same as ordinary averaging?

No. It concerns relationships between selected extreme observations and other measurements, not simply the arithmetic operation of calculating a mean.

Can it happen with accurate measurements?

Yes. Genuine temporary variation can create imperfect relationships even when each occasion is measured accurately.

What is the simplest defence against misinterpretation?

Notice extreme-value selection and seek a fair comparison. Repeated baseline observations can also help reduce dependence on one unusual result.

Scroll to Top