Choose your reading route
eduKateSG · Why Mathematics? · Winsorised Variance, Tail Replacement and Stable Spread
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 4–7
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 8–13
Check assumptions
See where design and calibration matter.Route 4 · Chapters 14–18
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 19–27
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important when a few tail values dominate a variance calculation? Winsorised variance replaces observations beyond declared lower and upper cutoffs with the cutoff values, then calculates variance on the modified sample.
It combines percentiles, data transformation, squared distance and transparent sensitivity analysis. The method can stabilise a spread estimate, but only when the replacement rule is declared and the raw data remain available.
Did you know? Winsorising does not remove a row. It keeps the sample size and replaces tail values, which is mathematically different from trimming.
Full chapter index
Chapters 1–3 · Understand the question
Chapters 4–7 · Reproduce the example
Chapters 8–13 · Check assumptions
Chapters 14–18 · Explore applications
Chapters 19–27 · Study and verify
Variance squares every distance from the mean. A value ten times farther away contributes one hundred times the squared distance.
Winsorising limits that tail contribution by capping observations before the ordinary variance formula is applied.
NIST defines tail percentages P1 and P2. Values below the chosen lower percentile are set equal to that percentile value, and values above the upper percentile are set equal to its value.
The standard sample variance is then calculated on the Winsorised observations.
| Raw data | Winsorised data | Winsorised mean |
|---|---|---|
| 1,2,3,4,100 | 2,2,3,4,4 | 3 |
Take 1, 2, 3, 4 and 100. For a simple 20% replacement in each tail using the observed one-value cutoffs, replace 1 by 2 and 100 by 4.
The Winsorised sample is 2,2,3,4,4. Its mean is 3.
| Squared-deviation sum | Denominator | Winsorised variance |
|---|---|---|
| 4 | 4 | 4/4 = 1 |
Distances from 3 are −1,−1,0,1,1. Squared distances sum to 4.
Divide by n−1=4 to obtain Winsorised variance 1. The Winsorised standard deviation is therefore 1 in the original unit.
“Winsorised variance” is incomplete without lower and upper percentages. Equal tails are common but not mandatory.
Also state how percentiles are interpolated and how a percentage maps to an integer number of observations in small samples.
Store the original value, replacement indicator and Winsorised value in separate columns. Never overwrite the only copy of the data.
This audit trail shows exactly which observations influenced the result through a cutoff rather than their original magnitude.
For the raw sample, the mean is 22 and the squared-deviation sum is 7,610, giving sample variance 1,902.5. Winsorising reduces it to 1.
That dramatic change is a prompt to investigate the tail, not proof that either summary alone tells the full story.
The same transformed sample has Winsorised mean 3. The Winsorised-mean guide develops the location statistic.
Variance must be centred on the Winsorised sample’s mean, not automatically on the raw mean.
Trimming discards chosen tail observations and changes the effective sample size. Winsorising keeps rows and caps their values.
The denominator, centre and inferential adjustments can therefore differ. Do not use the words interchangeably.
A larger percentage limits more observations and usually reduces tail influence. It can also erase genuine variation.
Choose the percentage before looking for a preferred answer, or report a sensitivity table across defensible choices.
Some applications have one physically bounded tail and one open tail. Unequal P1 and P2 may then be considered.
The asymmetry needs a subject-matter reason and clear reporting, because it can shift both the transformed mean and variance.
If many values equal a cutoff, software packages can cap the same data yet report slightly different thresholds because percentile definitions vary.
Record the algorithm or library version when exact reproducibility matters.
Variance has squared units; its square root returns the original unit. Multiplying every observation by c multiplies Winsorised variance by c squared if the cutoffs transform consistently.
Adding a constant leaves variance unchanged.
With five values, changing one tail observation represents 20% of the sample. A single replacement decision is therefore substantial.
Small-sample results should show the transformed data rather than hide behind a smooth percentage label.
Define the analysis set before finding cutoffs. Do not let missing values silently alter the tail count.
For repeated measurements, batches or strata, decide whether cutoffs are global or group-specific; those choices answer different questions.
Students can compare the raw and Winsorised dot plots for 1,2,3,4,100. Move 100 to 1,000 and observe that the Winsorised result stays fixed under the same cutoff rule.
Then move 4 and see when the upper cutoff itself changes.
Sort a copy, calculate cutoff cells, use MIN and MAX to cap each value, then compute the sample variance of the transformed column.
Add counts of lower and upper replacements and display both cutoffs.
Use an explicit quantile method, apply lower and upper clipping, and calculate sample variance with the intended degrees-of-freedom setting.
Unit-test the worked data for transformed values 2,2,3,4,4, mean 3 and sample variance 1.
Standard errors for ordinary variance do not automatically apply after adaptive tail replacement. Bootstrap procedures must repeat the full cutoff selection inside each resample.
Dependence, small samples and chosen tail fractions all affect uncertainty.
Winsorising does not prove tail observations are erroneous. It can conceal mixtures, data-entry faults or genuine extremes if used mechanically.
It also does not make every distribution normal or every variance comparison valid.
Winsorised summaries appear in robust statistics, finance, quality analysis and experimental data review. The transferable skill is documenting a transformation and testing how conclusions respond to it.
This is useful in data science because preprocessing choices can matter as much as the final model.
The trimmed-mean guide contrasts removal with replacement. The biweight-scale guide uses adaptive residual weights rather than hard caps.
Students should list the raw values, cutoffs, replacements, transformed mean, squared deviations and denominator.
Parents can ask what scientific reason supports the tail percentage and whether the raw plot is still shown.
Use 0,2,3,5,50 and replace one value in each tail. Compare raw and Winsorised variance. Repeat with only an upper-tail cap and explain the shift in centre.
Convert every value from metres to centimetres and predict the variance factor.
### Is Winsorising the same as deleting outliers? No. Values are capped and rows remain.
### Is the ordinary variance formula still used? Yes, but on the transformed sample under the documented convention.
### Which tail percentage is correct? There is no universal percentage; justify it and check sensitivity.
### Can I report only the Winsorised result? Show the raw distribution and replacement rule too.
Read the official NIST Winsorised-variance documentation, the NIST Winsorised-standard-deviation documentation and the NIST measures-of-scale handbook. The worked transformations and both sample variances were independently calculated.
Explore more through the Mathematics Learning Hub and How Mathematics Works.
Winsorised variance makes a tail policy visible in the arithmetic. Its value comes from that transparency: cap deliberately, preserve the raw evidence and examine how the conclusion changes.
Extended practice laboratory
Practice case 1: Choose tail percentages
Setup: Use the article’s Winsorised-variance example and change one declared input for case 1. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 2: Sort the data
Setup: Use the article’s Winsorised-variance example and change one declared input for case 2. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 3: Find lower cutoff
Setup: Use the article’s Winsorised-variance example and change one declared input for case 3. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 4: Find upper cutoff
Setup: Use the article’s Winsorised-variance example and change one declared input for case 4. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 5: Replace lower tail
Setup: Use the article’s Winsorised-variance example and change one declared input for case 5. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 6: Replace upper tail
Setup: Use the article’s Winsorised-variance example and change one declared input for case 6. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 7: Calculate Winsorised mean
Setup: Use the article’s Winsorised-variance example and change one declared input for case 7. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 8: Calculate squared deviations
Setup: Use the article’s Winsorised-variance example and change one declared input for case 8. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 9: Divide by n minus one
Setup: Use the article’s Winsorised-variance example and change one declared input for case 9. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 10: Take the square root
Setup: Use the article’s Winsorised-variance example and change one declared input for case 10. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 11: Compare ordinary variance
Setup: Use the article’s Winsorised-variance example and change one declared input for case 11. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 12: Compare trimmed variance
Setup: Use the article’s Winsorised-variance example and change one declared input for case 12. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 13: Change the tail fraction
Setup: Use the article’s Winsorised-variance example and change one declared input for case 13. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 14: Use unequal tails
Setup: Use the article’s Winsorised-variance example and change one declared input for case 14. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 15: Handle ties at cutoffs
Setup: Use the article’s Winsorised-variance example and change one declared input for case 15. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 16: Change percentile convention
Setup: Use the article’s Winsorised-variance example and change one declared input for case 16. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 17: Use small n
Setup: Use the article’s Winsorised-variance example and change one declared input for case 17. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 18: Preserve replaced values
Setup: Use the article’s Winsorised-variance example and change one declared input for case 18. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 19: Plot raw and Winsorised data
Setup: Use the article’s Winsorised-variance example and change one declared input for case 19. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 20: Bootstrap carefully
Setup: Use the article’s Winsorised-variance example and change one declared input for case 20. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 21: Audit information loss
Setup: Use the article’s Winsorised-variance example and change one declared input for case 21. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 22: Check domain tolerances
Setup: Use the article’s Winsorised-variance example and change one declared input for case 22. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 23: Avoid outlier-deletion language
Setup: Use the article’s Winsorised-variance example and change one declared input for case 23. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.
Practice case 24: Write a stable-spread conclusion
Setup: Use the article’s Winsorised-variance example and change one declared input for case 24. Task: Recalculate from raw values, preserve units and state which mathematical or design condition changed.
Check: Check the relevant invariant, separate the number from its claim, and name the next diagnostic or source of evidence. Rebuild the relevant counts before interpreting the coefficient, preserve the stated design, and write one sentence separating the numerical result from the claim it can support. Then vary one defensible assumption, calculate again, and explain whether the conclusion changes because of new data, a new reference model, or a new decision purpose. Finish by naming the unit, tolerance, and evidence needed for transfer. Record one graph or table that would expose failure hidden by the summary.