Capping Extreme Values with Winsorization

What Winsorization Does

An extreme value may be a valid observation rather than a recording error. Deleting it is not the only way to limit its influence on summaries or a model. Winsorization replaces values beyond specified lower and upper cutoffs with the cutoffs themselves. It retains the observations, but changes the recorded values at the tails.

The cutoffs are often chosen as quantiles. For example, winsorizing at the 5th and 95th percentiles caps values below the 5th-percentile cutoff at that lower cutoff, and values above the 95th-percentile cutoff at that upper cutoff. This is an editing rule, not a diagnosis: a value being capped does not establish that it was erroneous or unimportant.

A quantile convention must be stated because different conventions can produce different cutoffs, especially in small samples. R’s default for quantile() is type 7. For \(n\) sorted, non-missing observations \(x_{(1)}\le \cdots \le x_{(n)}\), type 7 calculates the quantile at probability \(p\) using \(h=1+(n-1)p\). If \(h\) is not an integer, it linearly interpolates between the observations at positions \(\lfloor h\rfloor\) and \(\lceil h\rceil\).

For example, if \(h=2.5\), the quantile is halfway between the second and third sorted observations. As a result, a cutoff need not equal an observed value, and the number of observations changed need not equal a simple percentage of the sample. Ties and interpolation both matter.

Calculating and Applying the Cutoffs in R

For a numeric vector x, the general R procedure is to calculate both cutoffs and clamp every observed value to that interval. For example, with probabilities 0.05 and 0.95, use quantile(x, probs = c(0.05, 0.95), na.rm = TRUE, type = 7) to obtain the cutoffs, then pmax(lower, pmin(x, upper)) to create the winsorized vector. The inner operation caps high values at the upper cutoff; the outer operation raises values below the lower cutoff.

With na.rm = TRUE, missing values are excluded when calculating the cutoffs. The clamping expression leaves missing entries missing. Check that the input is numeric, that the requested probabilities are in increasing order and between zero and one, and that the resulting lower cutoff is no greater than the upper cutoff. Inspect the number of non-missing values and the cutoffs before interpreting the result.

Why the Mean Changes and What Is Preserved

Winsorization changes the mean by changing the values it caps, not by changing the number of observations. The following identity makes that distinction explicit.

Here, the sums run over the non-missing observations, \(n\) is their number, \(\bar{x}\) is their original mean, and \(\bar{w}\) is their winsorized mean. To justify the identity, write the winsorized mean as \(n^{-1}\sum_i w_i\) and subtract the original mean \(n^{-1}\sum_i x_i\). For each observation below \(L\), the difference \(w_i-x_i\) is \(L-x_i\); for one above \(U\), it is \(U-x_i\); and for one between the cutoffs, it is zero. Adding these differences and dividing by \(n\) gives the formula.

If \(L\) is above every value it replaces and \(U\) is below every value it replaces, the terms in the first sum are positive and the terms in the second are negative. Thus lower-tail capping raises the mean, while upper-tail capping lowers it. The net direction depends on their combined changes; it is not guaranteed that the mean moves up or down overall.

Winsorization keeps the same rows and, when missing values are left untouched, preserves which rows are missing for that variable. Because the transformation is nondecreasing, it does not reverse the ordering of two values. It can, however, make different values equal by replacing them with the same cutoff. It therefore changes tail magnitudes, the range, and often the mean and variance. Relationships with other variables can change as well, so retaining rows does not mean retaining all the information in their original values.

A median or another central summary may remain the same when the values determining it are not capped. It is not safe to assume that it always remains unchanged: the result depends on the sample, quantile convention, and cutoffs. The correct practice is to compute and compare the summaries rather than infer the effect from the name of the operation.

Keep the Rule Reproducible

The cutoff probabilities and quantile convention are part of the data-preparation rule and should be recorded along with the resulting cutoffs. This makes the edit auditable: another analyst can distinguish a value that was observed from one that was capped. It also helps prevent a subtle inconsistency when the rule is used beyond the data on which it was developed.

For a predictive model, if cutoffs are estimated from a training sample, preserve those cutoff values and apply them to later data. Recalculating quantiles separately on new data creates a different transformation, so a given raw value can be mapped to a different capped value depending on which data set it appears in. Missing values should also be handled according to a stated rule rather than silently converted to a cutoff.

Winsorization is useful when the analysis calls for limiting extreme magnitudes while retaining the rows, but it is not a universal remedy for influential observations. Capping a valid large claim, for example, changes its amount and may affect a model’s interpretation. Before using the transformed variable, report the cutoff probabilities, the quantile convention, how many values were changed, and relevant before-and-after summaries. Keep the original variable available when appropriate so the edit can be reviewed.