Statistics Reference
Skewness and Kurtosis Explained
Two datasets can share a mean and standard deviation and still behave completely differently — one symmetric and tame, the other lopsided with wild outliers. Skewness and kurtosis are the numbers that capture that difference. This guide computes both by hand on five values, using the same formulas as the site's skewness and kurtosis calculator, and shows how to read the results.
The Interpretation Table First
| Statistic | Value | Reading |
|---|---|---|
| Skewness | |G1| < 0.5 | Approximately symmetric |
| 0.5 – 1 | Moderately skewed (sign gives direction: + right, − left) | |
| > 1 | Highly skewed | |
| Excess kurtosis | < −0.5 | Platykurtic — lighter tails than normal, extremes rarer |
| −0.5 – 0.5 | Mesokurtic — near-normal tails | |
| > 0.5 | Leptokurtic — heavier tails, outlier-prone |
The cutoffs are reading conventions, not laws — the same ones the calculator's plain-language labels use. Excess kurtosis is defined so a normal distribution scores exactly 0 (some sources report raw kurtosis, where normal scores 3 — check before comparing numbers across tools).
Worked Example: Five Values, Every Step
Data: 2, 2, 3, 4, 14 — four modest values and one extreme, a miniature right-skewed distribution.
- Mean and deviations: x̄ = 25/5 = 5; deviations −3, −3, −2, −1, 9.
- Central moments (average powered deviations): m₂ = 104/5 = 20.8; m₃ = 666/5 = 133.2; m₄ = 6740/5 = 1348.
- Population skewness: g₁ = m₃ / m₂^1.5 = 133.2 / 94.86 ≈ 1.404.
- Sample skewness: G₁ = g₁ × √(n(n−1))/(n−2) = 1.404 × 1.491 ≈ 2.093 — highly skewed right.
- Population excess kurtosis: g₂ = m₄/m₂² − 3 = 1348/432.64 − 3 ≈ 0.116.
- Sample excess kurtosis: G₂ = ((n+1)g₂ + 6)(n−1)/((n−2)(n−3)) ≈ 4.463 — strongly leptokurtic.
Entering the five values into the calculator reproduces every number, and the pairing is typical: one heavy right tail inflates skewness and kurtosis together. These are the adjusted Fisher–Pearson formulas — the same ones behind Excel's SKEW and KURT — so the article, the tool, and your spreadsheet agree.
Sample vs Population Variants, Honestly
The moment versions g₁ and g₂ describe the data in hand as a closed population. The adjusted versions G₁ and G₂ estimate the shape of the larger distribution the data was sampled from, correcting the systematic underestimate that small samples produce (extreme tail values are precisely what small samples tend to miss). The example above shows how dramatic the gap can be at n = 5: g₂ = 0.12 looks innocently near-normal while G₂ = 4.46 correctly flags that a fifth of the data is an extreme.
Practical rules: quote G₁/G₂ for sampled data; expect software differences (Excel/SPSS/SAS report the adjusted versions, while scipy and several R functions default to the moment versions); and treat shape estimates from fewer than ~20 values as indications, not measurements.
What Shape Changes in Practice
- Summary choice. Skew pulls the mean toward the long tail, away from the typical case — the reason incomes are reported by medians. Cross-check any suspect dataset with the descriptive statistics calculator: a mean far from the median is skewness talking.
- Normal-based rules. Heavy tails break the 68-95-99.7 percentages and make z-score outlier cutoffs fire too rarely — extremes are more common than the normal model claims.
- Visual confirmation. A box plot of the worked-example data shows the story instantly: a compact box with one far-out point. Numbers and pictures should agree.
Try the Skewness and Kurtosis Calculator
Sample and population skewness and excess kurtosis with central moments, interpretation labels, and the exact formulas — the worked example automated.
Frequently Asked Questions
What do skewness and kurtosis measure, in one sentence each?
Skewness measures asymmetry - whether one tail of the distribution stretches farther than the other, and in which direction. Kurtosis (reported as excess kurtosis) measures tail weight - whether extreme values occur more or less often than a normal distribution predicts. Together they describe shape, the part of a dataset that mean and standard deviation cannot see.
What counts as a lot of skewness?
The widely taught rule of thumb: |skewness| under 0.5 is approximately symmetric, 0.5 to 1 moderate, above 1 high. Context matters though - in small samples, sizable skewness arises by chance even from symmetric populations, so a 0.8 from 15 observations is weak evidence while the same 0.8 from 500 observations is compelling.
What is the difference between positive and negative skew?
Positive (right) skew means the long tail points toward high values: incomes, insurance claims, response times - mostly modest values with a few huge ones, and the mean sits above the median. Negative (left) skew points the tail toward low values: scores on an easy exam, age at retirement - and the mean sits below the median.
Does high kurtosis mean a sharp peak?
That old textbook description is misleading. Kurtosis is computed from fourth-powered deviations, so it is dominated by the tails - the center of the distribution contributes almost nothing. High excess kurtosis (leptokurtic) means heavy tails and outlier-proneness; low (platykurtic) means extremes are rarer than normal. Two distributions with identical peaks can have very different kurtosis.
Why are there sample and population versions of these statistics?
Small samples systematically understate shape, because extreme tail values are exactly what a small sample tends to miss. The sample-adjusted versions (G1 for skewness, G2 for excess kurtosis - what Excel's SKEW and KURT report) inflate the raw moment versions (g1, g2) to correct the bias. The worked example shows the gap dramatically: g2 = 0.12 becomes G2 = 4.46 on five values.
What should I do if my data is strongly skewed or heavy-tailed?
First, report medians and IQRs rather than means and standard deviations, since the latter are dragged by the tail. Second, be wary of normal-based shortcuts - the empirical rule's percentages and z-score cutoffs understate extremes on heavy-tailed data. Third, consider transformations (log for right-skewed positives) or robust/nonparametric methods for inference.