Written and maintained by CASRAI Editorial Board
Last updated
Winsorizing is a way of handling extreme values in a dataset without discarding them: instead of removing an outlier, you replace it with the value at a specified percentile of the distribution — commonly the 5th and 95th percentiles, though 1st/99th and other cutoffs are used depending on how aggressive the correction needs to be. A value below the lower cutoff is set equal to the lower cutoff; a value above the upper cutoff is set equal to the upper cutoff. The sample size never changes. This guide covers what Winsorizing actually does, how it differs from trimming (also called truncation) and from outright deletion of outliers, when each choice is defensible, and why choosing a method after seeing which one produces a significant result is a real methodological risk — not just a style preference.
What Winsorizing Does
Winsorizing is named for statistician Charles Winsor, who worked with John Tukey on early robust-statistics methods. The mechanics are simple: pick a percentile threshold (say, 5% and 95%), find the values of the distribution at those two points, and reassign every observation below the 5th percentile to equal the 5th-percentile value, and every observation above the 95th percentile to equal the 95th-percentile value. Everything between the two cutoffs is untouched. Because no observations are dropped, the sample size (n) is identical before and after — only the values at the extremes change. This is what distinguishes Winsorizing from every other outlier-handling approach: it reduces the influence of extreme values on statistics like the mean, variance, and regression coefficients, while preserving every row of data and every degree of freedom that depends on n.
A closely related but distinct technique is the Winsorized mean (or trimmed/Winsorized estimator more generally): compute the mean after Winsorizing the data, rather than reporting the raw sample mean. This is a common way to get a measure of central tendency that is less sensitive to extreme values than the ordinary mean, without switching to the median or losing sample size the way trimming does.
Winsorizing vs. Trimming (Truncation)
Trimming (also called truncation) is the more aggressive relative of Winsorizing. Instead of replacing values beyond a threshold, trimming removes them from the dataset entirely. A 5% trim removes the bottom 5% and top 5% of observations outright, shrinking n by that amount. Winsorizing and trimming often use the same threshold language (“Winsorized at the 5th/95th percentile” vs. “trimmed at 5%”), which is exactly why the two get confused — the cutoff logic is identical, but what happens to the flagged observations is not.
| Question | Winsorizing | Trimming / Truncation |
|---|---|---|
| What happens to a flagged value | Replaced with the threshold value | Removed from the dataset |
| Effect on sample size | Unchanged | Reduced by the trimmed proportion |
| Effect on degrees of freedom | None | Reduces df, which matters more in small samples |
| How aggressively it limits outlier influence | Caps influence at the threshold value | Eliminates the observation’s influence entirely |
| Information retained | All original observations contribute (in capped form) | Discarded observations contribute nothing |
Because Winsorizing keeps every observation, it’s generally the less aggressive, more data-preserving choice of the two — useful when the goal is to limit the leverage of a small number of extreme points without shrinking an already-limited sample. Trimming is more defensible when a value isn’t just extreme but is suspected to be invalid or non-representative of the population you’re trying to measure (a data-entry error, a sensor fault, a respondent who clearly didn’t understand the question) — in that case, the value doesn’t deserve even a capped influence on the analysis.
Winsorizing vs. Deleting Outliers Outright
Ad hoc deletion — removing individual outliers case by case, without a pre-specified percentile rule — is different again from trimming. Trimming applies a fixed, symmetric proportion (top and bottom 5%, for example) regardless of how many “genuine” outliers actually exist in a given sample. Ad hoc deletion instead relies on a judgment call about which specific points look wrong, often using a rule like the interquartile range (IQR) method (flagging anything beyond 1.5 × IQR from Q1/Q3) or a z-score cutoff, then dropping only the points that trip the rule. This is the most flexible of the three approaches and also the easiest to apply inconsistently, because the researcher is making a separate decision about each flagged point rather than committing to one rule in advance.
When Each Choice Is Defensible
- Winsorizing fits best when extreme values are plausible, real measurements (a genuinely high response time, a real but unusually large donation amount) that you don’t want to discard, but whose magnitude would otherwise dominate a mean, variance, or regression coefficient. It’s common in finance, survey research, and any analysis using ordinary least squares, where a handful of extreme points can have outsized leverage.
- Trimming fits best when you expect a genuine contaminating fraction in the data — some proportion of observations that don’t belong to the population of interest — and you’re comfortable losing that fraction of your sample to remove their influence entirely rather than cap it.
- Ad hoc deletion fits best when a specific value is identifiable as an error rather than a legitimate extreme observation (an implausible age, a negative value in a variable that cannot be negative, a duplicate row). It is not a substitute for a principled outlier-handling rule applied to values that are extreme but plausible.
None of the three is universally “correct.” The right choice depends on why the extreme values exist in your data and what your downstream analysis (a mean, a regression coefficient, a correlation) is sensitive to — see CASRAI’s descriptive statistics guide for how outliers affect measures of central tendency and spread more broadly, and the repeated-measures ANOVA assumptions guide for a worked example of outlier-checking as a formal assumption test rather than a general-purpose cleanup step.
The Researcher-Degrees-of-Freedom Risk: Deciding After You’ve Seen the Result
The methodological hazard isn’t Winsorizing itself — it’s choosing which outlier-handling method to use, or which threshold to apply, after already knowing which choice produces a statistically significant result. This is a specific instance of what Simmons, Nelson, and Simonsohn’s widely cited 2011 paper “False-Positive Psychology” calls researcher degrees of freedom: any point in an analysis where a seemingly reasonable, defensible choice could have gone more than one way, and the analyst picks the version that works, inflates the true false-positive rate well above the nominal 5% threshold most significance tests assume. Outlier handling is one of the most common such choices, because Winsorizing at 1% vs. 5%, trimming vs. Winsorizing, and IQR-based vs. z-score-based flagging can each produce a different p-value from the identical underlying data — and none of the options is obviously wrong on its face.
This is the same failure mode CASRAI’s garden of forking paths and p-hacking entries describe, and it’s why questionable research practices guidance treats undisclosed analytic flexibility as a problem independent of whether any single choice was individually defensible. The fix isn’t to avoid Winsorizing or any other outlier-handling method — it’s to fix the method and threshold before looking at how each choice affects the result, ideally in a pre-registration or pre-analysis plan, so the choice can’t be reverse-engineered from the outcome. Where a pre-specified plan isn’t available or a reviewer asks how sensitive the result is to the outlier rule, reporting a robustness check — the same analysis run under a couple of alternative outlier-handling choices — shows whether the finding depends on that one decision.
Reporting Transparency: What to Disclose
Whichever method is used, the methods section should state, at minimum:
- That an outlier-handling step was applied at all. Silently Winsorizing or trimming data without mentioning it in the methods section is itself a transparency failure, independent of whether the choice was reasonable.
- Which method was used — Winsorizing, trimming, or ad hoc deletion — stated by name, not implied.
- The exact threshold — e.g., “Winsorized at the 5th and 95th percentiles” or “trimmed at 1%” — not a vague description like “extreme values were adjusted.”
- How many observations were affected, where practical — the count or proportion of values that were changed (Winsorizing) or dropped (trimming), so a reader can judge whether the correction was minor or substantial.
- Whether the threshold was decided in advance or in response to the data, and if the latter, disclosing that a robustness check under alternative thresholds was also run.
This level of disclosure is what lets a peer reviewer, a replicator, or a reader distinguish a defensible, pre-specified outlier-handling choice from one selected because it happened to produce the result the analysis was looking for.
Frequently Asked Questions
What percentile is typically used for Winsorizing?
5% and 95% (a “5% Winsorization”) is the most common convention, with 1% and 99% used for a lighter-touch correction on larger datasets. There is no single mandated threshold — the right cutoff depends on how much of the distribution’s tail genuinely needs correcting, and the chosen threshold should be reported explicitly rather than left as an unstated default.
Is Winsorizing the same as trimming?
No. Both use a percentile cutoff, but Winsorizing replaces flagged values with the threshold value and keeps every observation, while trimming removes flagged values from the dataset and reduces the sample size.
Does Winsorizing bias the result?
Winsorizing changes the value of specific observations, so it necessarily changes downstream statistics like the mean and variance compared to the untouched data — that’s the intended effect, since the goal is to reduce the leverage of extreme values. Whether that change counts as a “bias” or a “correction” depends on whether the extreme values were representative of the population being studied. The practice becomes a genuine problem only when the method or threshold is chosen after the fact specifically because it produces a preferred result, rather than being fixed in advance.
Should the outlier-handling method be decided before or after looking at the data?
Before, wherever possible. Deciding the method and threshold in a pre-registration or pre-analysis plan — before running the substantive analysis — is the strongest protection against the researcher-degrees-of-freedom problem described above. When that isn’t feasible, disclosing the choice explicitly and reporting a robustness check under at least one alternative threshold is the next-best option.
Do Winsorizing and trimming require reporting a robustness check?
Not always, but it strengthens a paper considerably when the outlier-handling threshold was not pre-registered. Showing that the main result holds (or noting clearly if it doesn’t) under a different reasonable threshold demonstrates the finding isn’t an artifact of one specific outlier-handling decision.
