Winsorized

Winsorized is a statistical term describing data that have been subjected to winsorization, a technique intended to reduce the influence of extreme values. In winsorization, values beyond specified percentile cutoffs are replaced by the nearest values within those cutoffs. For example, in a 5% winsorization, observations below the 5th percentile are set to the 5th percentile value, and those above the 95th percentile are set to the 95th percentile value. The aim is to soften the impact of outliers without discarding data.

The procedure requires choosing a cutoff level, often 1%, 5%, or 10%, applied symmetrically to both tails. It is commonly used before computing statistics such as the mean, variance, correlations, or regression coefficients, to yield more robust estimates when distributions are heavy-tailed or contain outliers.

Winsorization differs from trimming. Trimming removes extreme observations from the dataset, while winsorization keeps all observations but caps their influence by substitution. Because the original data are altered, winsorized statistics can be biased relative to their unmodified counterparts, with the bias depending on the cutoffs and the underlying distribution.

Used within robust statistics and data cleaning, winsorization offers a compromise between fully nonparametric approaches and traditional parametric estimators. Its appropriateness depends on analysis goals and the acceptable balance between bias and variance. The technique is named after Charles P. Winsor, who introduced the approach to limit the impact of extreme values.