A report says it has winsorized the data before calculating an average. The extreme observations have not necessarily been deleted. Instead, specified tail values have been replaced with boundary values. The row count can stay the same while the numbers used in the calculation change.
NIST’s definition of the winsorized mean describes replacing observations in the tails with values at chosen percentile boundaries. The thresholds and the convention used to determine them are part of the method.
A replacement example with visible arithmetic
Consider these seven invented measurements: 1, 4, 5, 6, 7, 8, and 39. Their sum is 70, so the ordinary mean is 10.
For this teaching example, specify the rule directly: replace the smallest value with the next smallest, and the largest with the next largest. The calculation now uses 4, 4, 5, 6, 7, 8, and 8. There are still seven values, but their sum is 42 and their mean is 6.
That is different from deleting the two end observations. Deletion would leave five values and a different denominator. It is also different from correcting a known recording error: the original 39 may be a perfectly valid measurement.
Read the adjustment as part of the headline
Ask whether the report shows the unadjusted result, what thresholds were chosen, and why that treatment suits the question. In a service-quality story, a very long delay may be important precisely because it is unusual. Capping its numerical influence does not make the experience disappear.
For a reproducible analysis, preserve the original data and identify which values changed. If observations are weighted or grouped, explain how the rule interacts with those choices. A statement that data were “cleaned” is not an adequate substitute.
Winsorization can make an average less sensitive to extreme values. Whether that is helpful depends on the question, and readers need the replacement rule to judge the result fairly.

