Does trimming permanently delete records?

It should be an analysis operation with the originals preserved.

Is a 10% trim always two values?

No; the sample size and rounding rule determine the count.

Are trimmed samples automatically representative?

No; they deliberately exclude selected tail observations.

AI Glossary / Statistics and probability / Intermediate

Trimmed samples

Definition

Trimmed samples are samples from which a specified portion of the most extreme observations has been excluded for a particular analysis. A common approach removes an equal proportion from the lower and upper ends after sorting. The retained observations support a trimmed statistic, such as a trimmed mean, while the original data should remain available.

What it means

An ordinary average can be strongly affected by a few unusually large or small values. Trimming reduces the influence of the tails by defining a summary based on a central portion of the observed distribution.
The excluded observations are not automatically errors. They may be legitimate extremes, and their importance depends on the question. Trimming changes the information used and can change the population quantity that the analysis estimates.

How it works

Specify the rule before examining which choice produces the preferred conclusion. State whether the trimming percentage applies to each tail or to the total sample. A 10% trim from each tail removes approximately 20% overall, not 10% overall.
Sort observations for the variable being summarized. For a simple symmetric rule, remove k smallest and k largest values, leaving n - 2k observations. The rule must leave enough values for the intended statistic.
Document how fractional counts are handled. If a per-tail proportion times sample size is not an integer, implementations may round, truncate, or use fractional weighting. For example, SciPy's trim_mean uses a count rounded down, so small samples may have less actual trimming than the nominal proportion suggests.
Keep the original records and identify the analysis-specific retained subset. In multivariable data, trimming on one variable does not mean every variable should be separately sorted and reassembled. Doing so would destroy relationships between measurements from the same unit.
Decide how ties and missing values are handled. A fixed-count rule may remove only some observations sharing an endpoint value; an inclusive threshold rule can remove a different number. These choices should be reproducible rather than implicit.
Distinguish trimming from winsorization. Trimming removes tail observations from a calculation; winsorization replaces their values with selected boundary values while retaining the original count. The two procedures can produce different summaries and require appropriate uncertainty methods.
Do not treat the retained central observations as an ordinary independent random sample from the original population. Selection depends on the sample's order statistics. An ordinary standard-error formula applied mechanically to the retained subset need not provide valid inference for a population trimmed mean.
Assess the information that trimming excludes. If extreme delays, losses, or failures are central to the decision, analyze them separately rather than allowing a central summary to conceal them. Report the rule and retained count alongside the result, and compare reasonable alternatives when sensitivity matters.

Three practical examples

A service analyst reports a prespecified trimmed mean response time alongside tail-delay measures.
A quality team examines central measurement behavior while retaining unusual observations for fault investigation.
A model-evaluation team compares a trimmed error summary with the ordinary mean to understand sensitivity to extreme cases.

Think of it this way

Think of describing the middle of a crowd after temporarily setting aside the shortest and tallest people. The remaining group can help describe typical heights, but it no longer tells the whole story about the crowd or how extreme its heights are.

Worked example

Consider an original ordered sample of ten measurements: 1, 2, 3, 4, 5, 6, 7, 8, 9, and 100. A rule that trims 10% from each tail removes one value at each end: 1 and 100.
The trimmed sample is 2, 3, 4, 5, 6, 7, 8, and 9. Eight observations remain, and their sum is 44, so the trimmed mean is 44/8 = 5.5. The ordinary mean of all ten observations is 145/10 = 14.5.
The actual removal is two of ten observations, or 20% in total. If the original 100 were a genuine expensive failure, the trimmed mean would not describe the average cost per original case. That quantity remains 14.5 for these data.
For comparison, replacing 1 with 2 and 100 with 9 gives a winsorized sample of ten values with mean 55/10 = 5.5. The numerical equality here is incidental; it does not make the samples or their uncertainty calculations equivalent.

Illustrative case study

Fictional case: Fairwater Support considers using a trimmed sample of ticket-resolution times for its routine dashboard. Analysts discover that the longest tickets are mostly complex accessibility requests, not corrupted records.
They retain the prespecified trimmed summary as one view of central performance, but add separate reporting for those longer cases and preserve the overall average. The resulting dashboard makes clear which users are excluded from the central summary instead of presenting the trimmed result as universal service performance.

Common misconceptions and mistakes

Trimming is not proof that the removed observations are invalid. It is not the same as deleting missing values or removing a known data-entry error. Equal percentages from both tails do not imply equal numerical distances from the center. A lower trimmed average is not necessarily a better business outcome. Repeatedly changing the trim percentage until a desired result appears undermines interpretation.

Business use

Trimmed samples can support robust descriptive comparisons when unusually large or small values dominate a central summary. They are most useful when the estimand, rule, and excluded-tail information are explicit. Retain auditability, check whether exclusions concentrate in important groups, and avoid replacing total-cost or reliability measures with a statistic that intentionally discards extremes.

Check your understanding

In the example, how many observations remain after trimming 10% per tail?

Answer: Eight.

Why is 5.5 not the original average cost if these are costs?

Answer: It averages only the retained observations.

What distinguishes winsorization?

Answer: It replaces extreme values instead of removing their positions.

Related terms

Trimmed statistics
Order statistic
Robust statistics
Outlier
Estimand
Trimmed mean
winsorization
order statistics

Learn next

Compare an ordinary mean, trimmed mean, and median on a skewed sample. Record exactly which observations each summary emphasizes and which business question it answers.

Sources

University of Vermont, Statistical Methods for Psychology glossary, Trimmed samples: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html SciPy, trim_mean documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.trim_mean.html SciPy, winsorize documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.mstats.winsorize.html

Previous: Trigram search Next: Trimmed statistics