Winsorized Mean

Feeling whimsical? Then stay away from the winsorized mean.

A winsorized mean is kind of like taking a normal mean (i.e. “average” of a set of numbers), except you replace the tail-ends of the set of numbers to make the mean less sensitive to outliers.

If you’ve ever taken an average that includes a large outlier, you might have been annoyed at how inaccurate it seemed. For instance, if you had 10 dogs and 8 of them drank one bowl of water a day, 1 of them drank half a bowl of water a day, and 1 had a condition that caused it to drink 10 bowls of water per day, the average bowls of water your dogs consumed would be skewed. The winsorized mean of this data would replace that one dog consuming 10 bowls of water per day with more moderate data, and the mean would be much closer to the median, which is one bowl of water per day per doggo.

A winsorized mean is a type of weighted average, being weighted toward the less extreme parts of the data, since the ends are truncated and replaced with less extreme values. Usually 10 to 25 percent of both ends are replaced, depending on how wacky the data is.