• Understand the basic idea of skewness and what it means for a distribution to be asymmetric
• Explain the difference between positive and negative skew in terms of the direction of the long tail
• Describe the relationship between mean, median and mode in a skewed distribution
Deeper study
• Understand the formula for skewness and the guidelines for interpreting it
• Judge the shape of a distribution by linking the tail direction and the positions of the mean and median in a histogram to the sign and size of the skewness
• Explain how skewness affects statistical analysis
• Apply data transformations to reduce skew
Real-world practice
• Use advanced transformations such as Box-Cox and Yeo-Johnson
• Choose confidence-interval and hypothesis-testing methods suited to skewed distributions
• Draw insights from skewness analysis of real-world data
• Recognize the limits of skewness and complement it with visualization and kurtosis
At a glance
Skewness is a standardized third-moment statistic that measures how far a data distribution departs from left–right symmetry (it is a defined term in the international standard ISO 3534-1, also called the coefficient of skewness).
The sign follows the direction of the long tail — positive means a long right tail (as with income, where mean > median), negative means a long left tail
The standard error of sample skewness is roughly √(6/n) — about 0.24 for n = 100 — so skewness values from small samples should be interpreted cautiously
A widely used rule of thumb treats |skewness| > 1 as strong skew (it is not an absolute criterion)
The mean plays a key role in computing skewness, so you need to understand what the mean is
Standard Deviation
Skewness is the third moment standardized by the standard deviation, so you need the concept of standard deviation
Median
Understanding how the mean and median relate in a skewed distribution requires the concept of the median
Basic: What Is Skewness?
Difficulty 3/5
About 25 min
Skewness measures whether a data distribution is symmetric or lopsided. It is a standard term in ISO 3534-1, the international vocabulary of statistics. Depending on which way the long tail stretches, a distribution is positively skewed (long right tail), negatively skewed (long left tail), or symmetric.
Picture the grade distribution of a class. If most students score in the 80s and a few score much lower, the tail stretches toward the low scores: the distribution is negatively skewed. If instead most students score in the 60s and only a few come close to full marks, the tail stretches toward the high scores: the distribution is positively skewed. If everyone is spread evenly around the average, the distribution is symmetric. Careful: the direction of skew is always judged by where the long tail points, not by where the bulk of the data sits.
Key points
•
A measure of the left–right symmetry of a distribution (ISO 3534-1 term: skewness, or coefficient of skewness)
•
Positive: positively skewed — long right tail (most of the data is bunched on the left)
•
Negative: negatively skewed — long left tail (most of the data is bunched on the right)
•
Near 0: probably close to symmetric — but zero skewness does not guarantee symmetry, so check a histogram too
•
For a single-peaked distribution, positive skew usually means mode < median < mean, and negative skew the reverse
A simple example
Income distributions are typically positively skewed. Most people fall in the lower-to-middle income range, but a small number of very high earners create a long right tail that pulls the mean upward.
Check your understanding
Answer:
Positive. The sign of skewness follows the direction of the long tail, not the side where the data is bunched. A long left tail gives negative skewness.
Answer:
For a smooth, single-peaked distribution, usually mode < median < mean. The large values in the right tail pull the mean up while affecting the median and mode much less. This ordering is not a law, though — it can fail for discrete distributions or distributions with several peaks.
Answer:
The skewness of a normal distribution is 0, because the normal distribution is perfectly symmetric. The converse does not hold, however — skewness of 0 does not by itself mean a distribution is symmetric.
Skewness is defined as the standardized third moment. The population skewness is γ₁ = E[(X−μ)³]/σ³; for a sample you use the sample skewness g₁ that estimates it, or G₁, which corrects for small-sample bias. Many statistics packages, including Excel, SPSS and SAS, use the corrected G₁ by default, but the default can differ from tool to tool (for example, the default in R's e1071 package is yet another estimator), so check the formula before comparing results from different software. A quick alternative based on the mode or median is Pearson's skewness coefficient. The first coefficient is (mean − mode)/standard deviation, but because the mode is hard to estimate reliably, the second coefficient, 3(mean − median)/standard deviation, is used more often in practice. For interpretation there is a widely used empirical guideline (Bulmer, 1979): |skewness| < 0.5 roughly symmetric, 0.5 to 1 moderately skewed, 1 or more highly skewed. It is a rule of thumb, not an absolute criterion.
Mathematical formula
γ₁ is the population skewness (μ is the population mean, σ the population standard deviation); g₁ is the sample skewness (mₖ = (1/n)Σ(xᵢ−x̄)ᵏ is the k-th sample central moment); G₁ is the sample skewness corrected for small-sample bias. The population definition (γ₁) and the sample estimators (g₁, G₁) give different values, and software packages differ in which one they compute by default.
Learn through examples
Example 1: Properties of a Positively Skewed Distribution
The typical pattern of income or house-price distributions
Skewness > 0: long tail on the right
(usually, for a single-peaked distribution) mode < median < mean
A few large values pull the mean upward
Answer:
A log or square-root transformation can improve symmetry.
Positively skewed distributions are very common in economic data. Because a few high earners raise the overall mean, the median is a better description of the typical level.
Example 2: Properties of a Negatively Skewed Distribution
The pattern of exam scores bunched near full marks, or of human age at death (note: lifetime and survival-time data for machine parts are usually positively skewed, the opposite pattern)
Skewness < 0: long tail on the left
(usually, for a single-peaked distribution) mean < median < mode
A few small values pull the mean downward
Answer:
A square transformation, a reflect-then-log transformation (e.g. log(max + 1 − x)), or the Yeo-Johnson transformation can improve symmetry.
Negatively skewed distributions appear in data with an upper bound. Most values cluster near the ceiling and only a few are low.
Example 3: Guidelines for Interpreting Skewness
Judging the shape of a distribution from its skewness value — the widely used empirical guideline (Bulmer, 1979)
|skewness| < 0.5: roughly symmetric
0.5 ≤ |skewness| < 1: moderately skewed
|skewness| ≥ 1: highly skewed
Answer:
The stronger the skew, the more you should consider nonparametric methods. Since the standard error of sample skewness is roughly √(6/n), the same value is less trustworthy from a smaller sample.
Skewness is an important tool for quantifying how symmetric your data is, and a pre-check for any analysis that assumes normality. Remember, though, that these thresholds are a rule of thumb and not an absolute criterion.
Example 4: How Skewness Affects Statistical Analysis
Why mean-based analyses are distorted by heavily skewed distributions
Mean: dragged toward the long tail, so it becomes a weaker representative of central tendency
Small-sample t-tests and confidence intervals: the normality assumption is shaken, so the actual significance level and coverage probability can deviate from their nominal values
Regression: heavily skewed residuals signal a violation of the normality assumption
Answer:
Report the median alongside the mean, improve symmetry with a log or similar transformation, or switch to alternatives such as nonparametric tests or the bootstrap. With a large enough sample, the central limit theorem makes inference about the mean fairly robust.
Skewness is a diagnostic that helps you decide which statistics and which methods to use. Make it a habit to check the skewness value together with a histogram before you start analysing.
Common mistakes
Mistake:
Confusing the sign (positive/negative) of skewness with its direction
Why is it wrong?
The word "skewed" sounds like it should describe where the body of the data leans, but the sign of skewness follows the tail, which causes confusion
Correct approach:
The sign of skewness always follows the long tail: positive means a long right tail, negative a long left tail
Mistake:
Overlooking how much a single extreme value can change the skewness
Why is it wrong?
It is easy to conclude that the whole distribution is skewed when a single extreme value is responsible
Correct approach:
Consider the influence of extreme values and use a histogram to see the overall shape of the distribution
Mistake:
Ignoring how unstable skewness is in small samples
Why is it wrong?
The standard error of sample skewness is roughly √(6/n), so with a small sample the estimate is unstable and unreliable
Correct approach:
Interpret skewness only when the sample size is large enough
Practice questions
Hint:
The sign of skewness follows the direction of the long tail, and its absolute value reflects the degree of skew.
Answer:
+2.3
Explanation:
A long right tail means a positive sign. 0.1 would be roughly symmetric and −1.2 corresponds to a long left tail, so the value that describes a heavily stretched right tail is +2.3.
Hint:
Think of a transformation that compresses large values relatively more.
Answer:
A log transformation (or a square-root transformation)
Explanation:
Log and square-root transformations compress large values more than small ones, shrinking the right tail. If the data contains zeros or negative values, shift it by adding a constant before transforming, or use the Yeo-Johnson transformation.
Hint:
Recall the empirical guideline and the standard error of sample skewness (about √(6/n)).
Answer:
No — it falls in the roughly symmetric range.
Explanation:
By the empirical guideline, |skewness| < 0.5 counts as roughly symmetric. Moreover, with n = 100 the standard error is about √(6/100) ≈ 0.24, so 0.4 lies within two standard errors of 0 and is hardly clear evidence of skew.
Skewness plays a central role in planning data transformations and choosing appropriate statistical methods. When a distribution is heavily skewed, the premises of mean-based inference (such as the t-test) are shaken, so you either improve symmetry with a transformation or choose an inference method that accommodates the skew.
The Box-Cox transformation is defined as y = (x^λ − 1)/λ for λ ≠ 0 and y = ln(x) for λ = 0, where λ is estimated from the data by maximum likelihood. It is only defined for x > 0, so if the data contains zeros or negative values you shift it by adding a constant first. The Yeo-Johnson transformation, published by Yeo and Johnson in Biometrika in 2000, is a generalization that can be applied to any real-valued data, negative values included.
For interval estimation with skewed distributions, robust methods such as the bootstrap percentile method are preferable to the normal approximation, and for hypothesis testing consider nonparametric methods such as the Wilcoxon rank-sum test or permutation tests.
Limits and complements: skewness estimates are less stable the smaller the sample, and they are highly sensitive to outliers. For a distribution with several peaks, a single skewness value cannot summarize the shape, so always interpret it together with a visualization such as a histogram and with the kurtosis. If you used a transformation, think about back-transforming to the original scale when interpreting the results.
Case study: Analysing User Behaviour on an Online Platform
Background
An online shopping platform wants to analyse its users' daily time-on-site data in order to segment users and offer tailored services.
The problem
When time-on-site is not normally distributed, the mean alone gives a misleading picture of how users actually behave.
The data
Daily time on site for 100,000 users: mean 45 minutes, median 25 minutes, mode 15 minutes, skewness 2.3
Method
Characterize the distribution by computing its skewness, normalize it with a log transformation, and segment users in a way that reflects the shape of the distribution.
Solution and results
1) Strong positive skew (skewness 2.3 > 1): most users stay briefly, a few stay a very long time
2) Mode (15 min) < median (25 min) < mean (45 min): the classic positive-skew pattern
3) After a log transformation the skewness drops to 0.2, close to a normal distribution
4) Having confirmed the shape via the transformation, segment on the original, easier-to-read scale using quantiles such as the median: light users (under 25 min), regular users (25–60 min) and heavy users (60 min or more)
The skewness analysis confirmed that a small group of heavy users pulls the mean up substantially.
Conclusion:
This supports a differentiated strategy: ① a streamlined interface for the majority of users (light/regular), and ② advanced features and personalized services for heavy users.
References
NIST/SEMATECH e-Handbook of Statistical Methods §1.3.5.11 — sample skewness and the correction factor G₁
Yeo, I.-K. & Johnson, R. A. (2000), "A new family of power transformations to improve normality or symmetry", Biometrika 87(4), 954–959
Bulmer, M. G. (1979), Principles of Statistics — empirical guideline for interpreting skewness
Box, G. E. P. & Cox, D. R. (1964), "An analysis of transformations", JRSS B 26(2)
