Written and maintained by CASRAI Editorial Board
Published · Last updated
On this page
A left-skewed distribution has a long tail on the left (low values) and a negative skewness; a right-skewed distribution has a long tail on the right (high values) and a positive skewness. Skewness describes the asymmetry of a distribution around its mean. A perfectly symmetric distribution — the classic bell curve — has a skewness of zero: values are spread evenly on either side of the mean, and the mean, median, and mode all fall in the same place. Real data rarely does this. When one side of a distribution has a longer, thinner tail than the other, that distribution is skewed, and the direction of that longer tail is what gives skewness its name.
This guide covers how to tell left-skewed from right-skewed data, the single most common point of confusion in that distinction, the practical diagnostic that mean, median, and mode ordering gives you, how skewness is actually measured, and what to do about it in analysis. For the broader set of shape, spread, and central-tendency statistics this sits inside, see CASRAI’s guides to descriptive statistics and histograms; for how skewness interacts with the normality assumption behind parametric tests, see normality of distribution.
Left-skewed vs. right-skewed: it’s named for the tail, not the bulk
This is the fact that trips up more people than any other part of skewness, and it is worth stating plainly before anything else: a distribution is named for the direction its long tail points, not for where most of the data sits. The bulk of the observations in a right-skewed distribution actually sit toward the left (lower) end of the scale — it is the thin trailing tail on the right that gives the distribution its name.
Right-skewed (positively skewed)
Picture a peak toward the lower end of the x-axis with a long, thinning tail stretching out to the right. Most values cluster at the low-to-moderate end, and a relatively small number of unusually large values stretch the distribution’s right side out. Household income is the textbook example: most households earn within a moderate range, but a small number of very high incomes pull a long tail to the right, with no natural upper bound.
Left-skewed (negatively skewed)
The mirror image: a peak toward the higher end of the x-axis with a long tail stretching out to the left, toward smaller values. Most observations sit at the high end, and a relatively small number of unusually low values pull a long tail to the left. Age at death in a population with good access to healthcare is a common example — most deaths cluster at older ages, with a comparatively thin tail of deaths at younger ages pulling the distribution left.
A quick way to keep the two straight without redrawing the picture every time: point your finger from the peak of the distribution toward the thin tail. Whichever direction your finger points — left or right — is the distribution’s skew.
Mean, median, and mode: the practical diagnostic
Skewness has a direct, predictable effect on how the three common measures of central tendency line up relative to one another, because the mean is pulled toward the long tail while the median and mode are much more resistant to it. This ordering is often the fastest way to diagnose skew from summary statistics alone, without plotting anything.
- Right-skewed data: usually mean > median > mode. The small number of unusually large values in the tail pull the mean upward, since the mean is calculated from every value’s magnitude. The median, which only depends on rank order, moves much less, and the mode — the single most frequent value — tends to sit closest to the peak, furthest from the tail.
- Left-skewed data: usually mean < median < mode, the mirror-image ordering. The tail of unusually small values pulls the mean downward, while the median and mode again sit closer to where most of the data actually is.
- Symmetric data: mean = median = mode (or very close to it, allowing for sampling noise), which is one reason a normal distribution makes the mean such a clean, representative summary of a dataset.
This is also why skewness matters practically, not just descriptively: in a skewed distribution, the mean is being pulled away from where most of the data actually sits, which is exactly what makes it a less representative “typical value” than the median. CASRAI’s guide to descriptive statistics covers this same point from the central-tendency side — why the median, not the mean, is the standard choice for reporting a skewed variable’s “typical” value.
Measuring skewness
Eyeballing a histogram is often enough to identify the direction of skew, but reporting a numeric skewness statistic lets you quantify how skewed a distribution is and compare that across datasets or variables.
The Fisher-Pearson moment coefficient of skewness
The most common formal measure is the Fisher-Pearson moment coefficient of skewness, built from the third standardized moment of the distribution:
g₁ = m₃ / m₂3/2
where m₂ and m₃ are the second and third sample central moments (the average squared and cubed deviations from the mean, respectively). A positive value indicates a right-skewed (positive) distribution; a negative value indicates left-skewed (negative). The NIST/SEMATECH e-Handbook gives the same coefficient and notes that the skewness of a normal distribution is 0. Because g₁ is biased in small samples, an adjusted version is also in use:
G₁ = [√(n(n−1))/(n−2)] × g₁
The adjustment factor is about 1.49 at n = 5, 1.19 at n = 10 and 1.02 at n = 100, so the two estimators converge as the sample grows. NIST gives no numeric cutoff for “too skewed”, and notes that multimodal data can affect the sign of the statistic. Which estimator you get depends on the software, as the next section shows.
Which skewness does your software report?
Tools do not agree on a single default, so a “skewness” value from one program may not match another’s for the same data. The data table below the article compares the estimators and their values for one ten-value sample. The documented behaviour is:
- Microsoft Excel:
SKEWreturns the adjusted sample skewness (G₁) andSKEW.Pthe population skewness (g₁). Both need at least three data points, otherwise they return #DIV/0!, and text or empty cells in a reference are ignored. - SciPy:
scipy.stats.skewdefaults tobias=True, the biased g₁;bias=Falseapplies the correction for statistical bias. - pandas:
Series.skew()andDataFrame.skew()return unbiased skewness, normalised by N−1. - R, package e1071:
skewness()offers three types after Joanes and Gill (1998): type 1 is g₁, type 2 is the estimator used by SAS and SPSS, and type 3 (the default) is the one used by Minitab and BMDP.
Practical consequence: state the tool and option when you report skewness, and do not assume a value from R matches one from SPSS or Excel on a small sample.
Worked example: one sample, three estimators
Microsoft’s SKEW documentation uses the ten values 3, 4, 5, 2, 3, 4, 5, 6, 4, 7. The mean is 4.3. The population estimator g₁ (Excel SKEW.P) is 0.303193, and the adjusted G₁ (Excel SKEW) is 0.359543, which is g₁ multiplied by √(10×9)/8 = 1.1859. The e1071 type 3 estimator, b₁ = g₁((n−1)/n)3/2, is 0.258871. All three are positive, so the sample is mildly right-skewed, but the magnitudes differ by about 0.1, which is enough to cross a rule-of-thumb band boundary in a small sample. For this sample the estimator choice changes the number but not the sign.
When the mean–median–mode ordering fails
The ordering rule above is a heuristic, not a theorem. von Hippel (2005, Journal of Statistics Education 13(2)) found that 14 of 18 textbooks surveyed state the rule, and shows it fails with surprising frequency, particularly for discrete data and for distributions whose skew comes from a long tail versus a heavy tail. His examples include the 2002 General Social Survey count of adults per household, which has skewness 1.11 yet a mean to the left of the median; Poisson distributions, where the mean is below the median for over 30% of parameter values; and a Weibull distribution with shape parameter 3.44 whose skewness is only 0.04 even though the ordering is violated. Use the ordering together with a histogram and the skewness statistic, and never as the only test.
Can skewness be zero for an asymmetric distribution?
Yes. Skewness is one number summarising the third moment, so asymmetry can cancel. NIST notes that multimodal data can affect the sign of the statistic, and von Hippel’s Weibull example shows a skewness near zero for a distribution that is not symmetric. A skewness near 0 is therefore consistent with symmetry but does not prove it; check the plot.
Related CASRAI guides
Pearson’s median skewness coefficient
A simpler, more intuitive alternative directly reflects the mean/median relationship covered above:
Skewness = 3(mean − median) / standard deviation
This version is easier to compute by hand and to explain to a non-statistical audience, since it follows directly from the mean-median-mode ordering logic, though the moment-based coefficient above is more commonly reported in statistical software output and published research.
Rough interpretive bands
A widely used, though informal, convention (usually attributed to Bulmer, 1979) for interpreting the magnitude of a skewness statistic:
- |skewness| < 0.5: fairly symmetric
- 0.5 – 1: moderately skewed
- > 1: highly skewed
Treat these bands as a convenient rule of thumb, not a fixed statistical standard — different textbooks and fields draw the lines slightly differently, and the standard error of a skewness estimate shrinks as sample size grows, so the same numeric skewness value carries more evidential weight in a large sample than a small one. As with correlation-coefficient magnitude bands (see CASRAI’s guide to the correlation coefficient for the same caveat applied there), always look at the actual histogram alongside the number rather than trusting the statistic in isolation.
Distributions that are characteristically skewed
Several variable types recur often enough in research and applied statistics that their typical skew direction is worth knowing in advance, as a sanity check on your own data. These are general patterns well established in statistical practice, not findings from any single specific study.
Commonly right-skewed
- Income and wealth — bounded at (or near) zero with no practical upper bound, and a small share of very high values.
- Reaction times — a hard floor near the fastest physically possible response, with occasional much slower responses stretching the tail rightward.
- Hospital length of stay — most stays are short, with a smaller number of much longer stays.
- Citation counts — most papers receive relatively few citations, while a small number receive very many.
Commonly left-skewed
- Exam scores with a ceiling effect — when a test is easy relative to the tested population, scores bunch up near the maximum possible score, with a thinner tail of lower scores pulling the distribution left.
- Age at death in populations with good healthcare access — most deaths cluster at older ages, with a comparatively thin tail of deaths at younger ages.
These are tendencies, not guarantees — the actual shape of any given dataset should always be checked directly with a histogram rather than assumed from the variable’s category alone.
Why skewness matters for analysis
Skewness is not just a descriptive curiosity; it has direct consequences for which statistics and tests are appropriate.
Related: Geometric Mean — The geometric mean is the correct average for growth rates, ratios, and log-normally distributed data — this guide covers the formula, when it beats the arithmetic mean, the geometric standard deviation, a common back-transformation error, and the related harmonic mean.
- The mean stops being representative. As covered above, a skewed distribution’s mean is pulled toward the tail, away from where most observations actually sit — reporting only a mean for skewed data can misrepresent the “typical” case.
- Parametric test assumptions can be affected. Many common parametric tests (t-tests, ANOVA, ordinary least squares regression) assume approximately normal residuals or sampling distributions. Substantial skewness in the underlying data is one of the things that can violate that assumption, particularly in smaller samples.
- The Central Limit Theorem caveat still applies. It’s worth being precise here, because this is a frequent source of confusion: the CLT describes the sampling distribution of a statistic like the mean, not the shape of the raw data itself. A skewed variable does not become less skewed as sample size grows — what changes is how closely the distribution of possible sample means approximates normal, which is why parametric tests on means often remain reasonably robust to skewness at larger sample sizes even though the raw data stays skewed. CASRAI’s guide to normality of distribution covers this distinction in full, including what parametric tests actually assume and how to check the assumption directly.
What to do about skewed data
Report the median and IQR instead of (or alongside) the mean
For a skewed variable, the median and interquartile range (IQR) describe the “typical” value and spread without being distorted by the tail the way the mean and standard deviation are. This is standard practice for variables known to be routinely skewed — income, length of stay, and reaction time among them — and is covered in more depth in CASRAI’s guide to descriptive statistics.
Transform the variable
A mathematical transformation can reduce skewness enough to make parametric methods more appropriate, at the cost of analyzing (and having to interpret) the transformed scale rather than the original one:
- Log transformation — the most common choice for right-skewed, strictly positive data (it compresses large values proportionally more than small ones); results are interpreted in log units or back-transformed, and coefficients from a log-transformed outcome are often interpreted multiplicatively (e.g., as a percentage change) rather than additively.
- Square-root transformation — a gentler compression than log, often used for right-skewed count data.
- Box-Cox transformation — a family of power transformations (of which log and square-root are special cases) that estimates the specific transformation parameter that best normalizes a given variable, rather than assuming log or square-root in advance.
The interpretive cost is real: a transformed variable’s mean, differences, and model coefficients no longer have the same direct real-world units as the original variable, and results have to be explained (or back-transformed) with that in mind.
Use a non-parametric alternative
Rank-based non-parametric tests (the Mann-Whitney U test, Wilcoxon signed-rank test, Kruskal-Wallis test, and Spearman’s rho in place of Pearson’s r) don’t assume normally distributed data and are correspondingly less affected by skewness, at some cost in statistical power relative to their parametric counterparts when the normality assumption does actually hold.
Use a generalized linear model with an appropriate distribution
Rather than transforming the outcome variable to force it toward normality, a generalized linear model (GLM) can model the outcome using a probability distribution that already matches its actual shape — a Gamma or log-normal distribution for right-skewed positive continuous data, or a Poisson or negative binomial distribution for right-skewed count data, for example. This keeps the outcome on its original, directly interpretable scale while still respecting its actual distributional shape.
Skewness vs. kurtosis: two different properties, often conflated
Skewness and kurtosis are both computed from standardized moments of a distribution (the third and fourth moments, respectively), which is likely why they get mixed up so often — but they describe different things entirely.
- Skewness measures asymmetry: whether one tail is longer than the other.
- Kurtosis measures tail weight (how much of a distribution’s variance comes from infrequent, extreme values) relative to a normal distribution, independent of whether the distribution is symmetric. A distribution can be perfectly symmetric and still have heavy or light tails compared to normal.
Excess kurtosis (the conventional reporting form, calibrated so a normal distribution scores 0) is described as leptokurtic when positive (heavier tails and a sharper peak than normal — more extreme outliers than a normal distribution would produce) and platykurtic when negative (lighter tails and a flatter peak). A symmetric distribution can be strongly leptokurtic or platykurtic; a skewed distribution can have normal, light, or heavy tails. The two statistics are reported together precisely because neither one tells you about the other.
Frequently asked questions
What does it mean for a distribution to be left-skewed?
A left-skewed (negatively skewed) distribution has a longer, thinner tail on its left (lower-value) side. Most observations cluster toward the higher end of the scale, and the mean is pulled below the median by the smaller number of unusually low values in the tail.
What does it mean for a distribution to be right-skewed?
A right-skewed (positively skewed) distribution has a longer, thinner tail on its right (higher-value) side. Most observations cluster toward the lower end of the scale, and the mean is pulled above the median by the smaller number of unusually high values in the tail.
Is right skew the same as positive skew?
Yes — “right-skewed” and “positively skewed” describe the same distribution shape and are used interchangeably, as are “left-skewed” and “negatively skewed.”
How can I tell if data is skewed just from the mean, median, and mode?
Usually: mean > median > mode points to right skew, mean < median < mode to left skew, and three roughly equal values to a roughly symmetric distribution. The ordering works because the mean is pulled toward the tail, but it can fail, especially for discrete data (see the section above), so confirm with a histogram or the skewness statistic.
What is considered a high skewness value?
A commonly used rule of thumb treats |skewness| below 0.5 as fairly symmetric, 0.5 to 1 as moderately skewed, and above 1 as highly skewed — but these bands are informal conventions, not a fixed statistical standard, and should be interpreted alongside a histogram rather than in isolation.
Is skewness the same as kurtosis?
No. Skewness measures asymmetry (whether one tail is longer than the other); kurtosis measures tail weight (how heavy-tailed a distribution is relative to normal), independent of whether it’s symmetric. A distribution can be skewed with normal tail weight, or symmetric with unusually heavy or light tails.
Which skewness formula do Excel, SPSS, R and Python use?
Excel’s SKEW is the adjusted G₁ and SKEW.P is g₁; SciPy defaults to the biased g₁; pandas returns the unbiased estimator; and R’s e1071 defaults to type 3 (b₁), with type 2 matching SAS and SPSS. See the table and the software section above.
Can skewness be zero if the distribution is not symmetric?
Yes; multimodal or mixed shapes can give a statistic near zero, so inspect the histogram.
How do I fix skewed data?
Report the median and IQR, transform the variable, use a rank-based test, or fit a model with a matching distribution, as described above.
Sources
Sources:
- NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.5.11 Measures of Skewness and Kurtosis: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm
- Microsoft, SKEW function: https://support.microsoft.com/en-us/office/skew-function-bdf49d86-b1ef-4804-a046-28eaea69c9fa
- Microsoft, SKEW.P function: https://support.microsoft.com/en-us/office/skew-p-function-76530a5c-99b9-48a1-8392-26632d542fcb
- SciPy, scipy.stats.skew: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.skew.html
- pandas, DataFrame.skew: https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.skew.html
- e1071 skewness (CRAN reference manual; Joanes and Gill, 1998, The Statistician 47:183–189): https://search.r-project.org/CRAN/refmans/e1071/html/skewness.html
- von Hippel, P. T. (2005), Mean, Median, and Skew, Journal of Statistics Education 13(2): https://jse.amstat.org/v13n2/vonhippel.html
Further reading: how to calculate a z-score — How to calculate a z-score, what standardization does (and doesn’t) do to a distribution, and how to correctly read a standard normal (z) table, including a full table.
Data: Sample skewness estimators by software, applied to one 10-value sample
Compares the sample-skewness estimator each tool or function documents (g1, G1 or b1), its default status and the value each estimator gives for the ten-value sample 3, 4, 5, 2, 3, 4, 5, 6, 4, 7 used in Microsoft's SKEW documentation. Estimator definitions come from the tools' documentation; values for g1 and G1 are the published Excel results and the b1 value is computed from the e1071 formula b1 = g1((n-1)/n)^(3/2).
| Tool / function | Estimator | Default behaviour | Skewness of sample (n=10) (dimensionless) |
|---|---|---|---|
| Excel SKEW | G1 (adjusted Fisher-Pearson) | Sample skewness; needs 3+ points | 0.359543 |
| Excel SKEW.P | g1 (population, m3/m2^1.5) | Population skewness; needs 3+ points | 0.303193 |
| SciPy stats.skew (bias=True) | g1 (biased) | Default | 0.303193 |
| SciPy stats.skew (bias=False) | G1 (bias-corrected) | Non-default | 0.359543 |
| R e1071 skewness type 1 | g1 | Non-default | 0.303193 |
| R e1071 skewness type 2 | G1 (SAS and SPSS) | Non-default | 0.359543 |
| R e1071 skewness type 3 | b1 = m3/s^3 (Minitab and BMDP) | Default (type = 3) | 0.258871 |
Sources
- Microsoft SKEW function (accessed 2026-11-14)
- Microsoft SKEW.P function (accessed 2026-11-14)
- SciPy scipy.stats.skew (accessed 2026-11-14)
- pandas DataFrame.skew (accessed 2026-11-14)
- e1071 skewness (CRAN reference manual) (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







