Unlike the trimmed mean, which deletes extreme values, the Winsorized mean replaces them with the nearest remaining values.

To compute the $20\%$ Winsorized mean, the two smallest values ($3, 5$) are replaced by $10$, and the two largest values ($35, 90$) are replaced by $22$.

The $20\%$ Winsorized mean of this dataset is $16.1$.

The Winsorized mean is less sensitive to the extreme outlier $90$ than the original arithmetic mean.