A report selects the ten service desks with the worst waiting times, introduces a new reminder and celebrates their improvement the following month. The reminder might help. The before-and-after comparison alone cannot tell us how much of the change it caused.
An unusually poor measurement can combine a persistent problem with temporary circumstances and measurement variation. If desks are selected precisely because that first result was extreme, a later result may be less extreme even without the new measure. This is the concern behind regression towards the mean.
Bland and Altman’s BMJ statistical note explains why selecting an extreme group complicates the interpretation of subsequent measurements. It is a property of the selection and measurement process, not evidence that an intervention necessarily failed.
Inspect the starting point
In the fictional service-desk example, ask how the ten desks were chosen. Was the selection based on one exceptionally busy week or a pattern across many comparable periods? Did a temporary staff absence coincide with the first measurement? Were the same hours and kinds of enquiries counted later?
Those questions do not prove regression towards the mean explains the rebound. They identify why a claim such as “the reminder cut waiting times” needs more evidence than two numbers from a selected group.
What would make the comparison stronger?
Look for the evaluation’s design, including a suitable comparison group, how any groups were assigned, and whether several baseline observations were available. Check whether the report shows results for everyone eligible or only the most dramatic starting cases.
A useful reading note has three separate lines: the observed change, the proposed explanation, and the evidence that distinguishes that explanation from alternatives. If the third line is empty, keep the causal claim provisional.
Regression towards the mean is not a promise that every exceptional performance will reverse, nor a rule that good work is punished. It also does not mean a particular person is “due” a result. The practical lesson is to be cautious when an extreme starting measurement becomes both the reason for selecting cases and the baseline used to declare a success.

