Should left-skewed data always be transformed?

No. The analytical goal and model assumptions determine whether a transformation is useful.

Can all values be positive?

Yes. The five-score worked example contains only positive values and has a negative skewness coefficient.

AI Glossary / Statistics and probability / Advanced

Negatively skewed

Definition

A negatively skewed distribution is asymmetric with a tail extending farther toward lower values than toward higher values. It is also called left-skewed. Under a moment-based definition, negative skewness means a negative standardized third central moment, when that moment exists. The term describes one variable's distribution, not a relationship between two variables.

What it means

Many observations may cluster toward the high end while fewer extend into the low end. The tail determines the descriptive direction; negative does not mean that most observations are negative numbers. A collection of positive scores can be left-skewed. The direction also depends on how the measurement is coded: reversing the scale reverses left and right.

How it works

Inspect the distribution and state the skewness measure used. One common sample coefficient is g1 = m3 divided by m2 raised to the power 3/2, where mr averages the rth powers of deviations from the sample mean using denominator n. Some software reports the adjusted coefficient G1 = square root of n(n-1), divided by n-2, times g1. These are different numerical conventions.

Three practical examples

An assessment team notices that most participants score highly while a few scores extend much lower. It investigates the low-score tail before concluding that the assessment works equally well for every participant group.
A quality team studies a bounded performance score. Many readings near the upper bound and fewer low readings can create a left-skewed pattern, but the actual measurements must be examined rather than assuming the shape from the bound alone.
An AI evaluation report changes from accuracy to error rate. Because error rate is one minus accuracy, this reflection reverses the skewness sign. The underlying model evaluations remain exactly the same.

Think of it this way

Picture a crowd near the right side of a room with a smaller line stretching toward the left wall. The line gives the shape its leftward direction. Moving the whole crowd into another room does not change its shape, but viewing the same arrangement in a mirror reverses its direction.

Worked example

Consider an original five-score dataset: 1, 7, 8, 9, and 10. The mean is 35/5 = 7, while the median is 8. Relative to the mean, the deviations are -6, 0, 1, 2, and 3. Their squared values sum to 36 + 0 + 1 + 4 + 9 = 50, so m2 = 50/5 = 10.

The cubed deviations sum to -216 + 0 + 1 + 8 + 27 = -180, so m3 = -36. Consequently, g1 = -36 divided by 10 raised to the power 3/2, approximately -1.138. Using the adjusted convention multiplies this by the square root of 20 divided by 3, giving G1 approximately -1.697. Two software outputs can therefore differ substantially while both correctly describe their stated convention.

The low score of 1 contributes -216 to the cubed-deviation total, outweighing the combined positive contribution of 36. That explains the negative coefficient in this specific dataset. It does not establish that the low score is erroneous. If the score represents a valid participant outcome, removing it would change the population description being reported.

To see the effect, remove 1 only as an illustrative sensitivity calculation. The remaining scores 7, 8, 9, and 10 are symmetric around 8.5, so their third central moment and skewness coefficient are zero. An analyst should show why a record was excluded rather than silently replacing the original result with this more symmetric subset.

Next transform every original score using z = 11 - y. The transformed values are 10, 4, 3, 2, and 1, with mean 4 and median 3. Each deviation from the new mean is the negative of its original counterpart. Therefore m2 stays 10, m3 becomes +36, and g1 becomes approximately +1.138. Adding 100 to every original score, in contrast, changes its location but leaves all deviations and the negative skewness unchanged.

Finally, consider a separate dataset: 0, 10, 10, 10, 10, 16, and 16. Its median is 10 and its mean is 72/7, approximately 10.286. Nevertheless, its moment coefficient g1 is approximately -0.843: m2 is approximately 24.490 and m3 approximately -102.157. This counterexample shows that mean below median is a useful pattern in many familiar left-skewed distributions, not a universal test for negative third-moment skewness. The measure and the full shape both matter.

Illustrative case study

This fictional case concerns Alder Assessment, an imaginary training provider. Its analyst finds that two programs report different skewness values for the same five scores. A review confirms that one reports g1 and the other reports G1. The team labels both conventions explicitly and checks that neither program has omitted the low score.

The report includes the full distribution and investigates the low result's context. Staff resist deleting it merely to make the chart look more balanced. They also distinguish a pattern in five observations from a reliable statement about future participants. The case describes an analytical workflow, not an actual provider or a measured training outcome.

Common misconceptions and mistakes

Negative skewness is not negative correlation, a low mean, or a negative minimum. Nor does one negative sample coefficient prove that the underlying population has negative skewness; sampling variability remains relevant.

Moment-based skewness requires a nonzero variance and appropriate finite moments. A symmetric distribution with undefined moments does not automatically have a defined zero moment coefficient. Multimodal shapes and alternative skewness measures can also produce descriptions that need more explanation than a single sign.

Business use

Keep the tail visible when summarizing performance. A high median can coexist with a small set of much lower outcomes that deserve investigation. Report the units, sample size, coefficient convention, and handling of missing or excluded observations. Whether the low tail signals measurement problems, subgroup differences, or valid variation requires evidence beyond its shape.

Check your understanding

Why did reflection reverse the sign?

Answer: Cubed deviations changed sign while squared deviations did not.

Why is a mean above the median insufficient to rule out negative moment skewness?

Answer: The seven-score counterexample has that ordering but a negative third central moment.

Related terms

Median
Histogram
Reflection
Kurtosis
Left skewness
third central moment
Fisher-Pearson coefficient
tail

Learn next

Calculate g1 and G1 for the five scores, then repeat after reflection and translation. Label the convention and compare the results with a plot of the actual observations.

Reflection

Sources

NIST/SEMATECH, Measures of Skewness and Kurtosis: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm . David C. Howell, Statistical Methods for Psychology glossary, Negatively skewed: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . All datasets, calculations, and the Alder Assessment case are original illustrations.

Previous: Negative testing Next: Neighbor