Skewness and Kurtosis: Measuring a Distribution's Shape

Comparing a data set's mean and median gives a rough sense of skew, but it's an informal shortcut, not a measurement. Skewness and kurtosis are the formal, precise tools statisticians use to quantify exactly how a distribution's shape departs from symmetry and from the normal curve — numbers that turn "this looks a bit lopsided" into something you can actually calculate, compare, and report.

Skewness: Measuring Asymmetry

Skewness and kurtosis extend the same descriptive-statistics toolkit as mean, median, and mode and standard deviation a step further — where those measures describe a distribution's centre and overall spread, skewness and kurtosis describe its shape in more precise detail than a histogram alone conveys as a single comparable number.

This formula looks intimidating, but the core idea is simple: it's an average of cubed, standardized deviations from the mean. Cubing (rather than squaring, as in variance) preserves the sign of each deviation — values above the mean contribute positively, values below contribute negatively — which is exactly what lets skewness detect asymmetry that squaring would erase.

• Skewness ≈ 0: roughly symmetric distribution.
• Skewness > 0 (right/positive skew): a longer tail extending toward higher values — mean typically exceeds median. Income and house prices are classic examples.
• Skewness < 0 (left/negative skew): a longer tail extending toward lower values — mean typically falls below median. Age at retirement in a workforce with a hard minimum retirement age often looks like this.

As a rough interpretive guide: |skewness| below about 0.5 suggests approximate symmetry; between 0.5 and 1 suggests moderate skew; above 1 suggests substantial skew — though, like most such benchmarks in statistics, these are conventions, not hard mathematical boundaries.

Kurtosis: Measuring Tail Weight

Kurtosis is widely, and somewhat misleadingly, described as measuring a distribution's "peakedness." This is only part of the story — kurtosis is more precisely a measure of how much of a distribution's variance comes from infrequent, extreme deviations (heavy tails) versus frequent, modest deviations. A distribution can be described as having heavy or light tails relative to the normal distribution, and this is really what kurtosis quantifies, with peak shape as a secondary, often-overstated visual side effect.

This formula (excess kurtosis) is calibrated so that a normal distribution scores exactly 0:

• Excess kurtosis ≈ 0 (mesokurtic): tail weight similar to a normal distribution.
• Excess kurtosis > 0 (leptokurtic): heavier tails than normal — more frequent extreme values/outliers than a normal distribution would produce.
• Excess kurtosis < 0 (platykurtic): lighter tails than normal — fewer extreme outliers than a normal distribution would predict.

Worked Example

Problem: A small data set of annual bonuses ($'000s): 5, 6, 6, 7, 7, 7, 8, 45. Describe its shape.

Mean = 11.375. Median = 7 (the middle two values, sorted, are both 7). The mean sits well above the median — an immediate visual signal of right skew, driven by the single extreme value (45).

A full skewness calculation on this data would produce a substantially positive value (well above 1), confirming strong right skew — consistent with the mean-vs-median comparison, but giving a precise, comparable number rather than just a directional impression. This is a realistic shape for bonus or income data: most values cluster together, with a small number of much larger values pulling the mean and the skewness measure upward.

Why This Matters for Choosing Statistical Methods

Many common statistical techniques — t-tests, ANOVA, linear regression's residual assumptions — carry an underlying normality assumption. Checking skewness and kurtosis (alongside a visual histogram or Q-Q plot, as covered in exploratory data analysis) is a standard part of verifying that assumption before trusting a method's results. Strongly skewed data might call for a data transformation (like a log transform, which often reduces right skew substantially) or a non-parametric method less dependent on normality; high excess kurtosis warns that outliers may have more influence on mean-based statistics than a normal-distribution assumption would predict.

Real-World Uses

Finance: asset returns are famously leptokurtic (heavier-tailed than normal) — extreme market moves happen more often than a normal-distribution model predicts, a well-documented limitation of simple risk models that assume normality, and a major factor behind several historical financial crises where "normal" models badly underestimated tail risk.

Quality control: monitoring the skewness and kurtosis of a manufacturing process's output over time can flag a process drifting away from its expected, stable shape before it produces an unacceptable defect rate.

Survey and social science research: checking skewness before choosing between reporting a mean or a median, or before selecting a statistical test, is considered standard due diligence in careful applied research.

Common Mistakes

Treating kurtosis purely as "peakedness." This popular simplification is only loosely accurate — two distributions can have identical peak height and shape but noticeably different kurtosis, because kurtosis is really driven by the tails, not the centre. A picture of a "pointier" curve is a common but imperfect visual mnemonic for what's fundamentally a tail-weight measure.

Applying rigid skewness/kurtosis cutoffs universally. The rough benchmarks given above (|skew| > 1 = substantial) are genuinely just conventions, and different fields and textbooks use somewhat different thresholds. Treat them as a starting orientation, not a strict pass/fail rule.

Ignoring sample size when interpreting these statistics. As covered in the FAQ, both measures are unstable in small samples — reporting a precise skewness value to several decimal places from n=10 data overstates the precision that's actually available, and a formal significance test for skewness (comparing it against its own standard error) is more defensible than eyeballing the raw number alone for small samples.

Visualizing Skewness and Kurtosis

While the formulas above give precise numbers, a histogram remains the fastest and most intuitive way to actually see skewness and kurtosis in a data set — right skew shows up as a longer tail stretching rightward with the bulk of data piled on the left, and heavy kurtosis shows up as a taller central peak combined with visibly more extreme values scattered further from the centre than a smooth bell curve would predict. As with most descriptive statistics, the numeric measure and the visual should be used together, per the same discipline data visualization principles and exploratory data analysis recommend generally — a number alone can hide exactly the kind of shape detail a quick plot reveals immediately.

A Second Worked Comparison: Symmetric vs. Skewed

Problem: Compare two data sets with the exact same mean but different shapes. Set A (symmetric): 10, 15, 20, 25, 30. Set B (right-skewed): 10, 12, 13, 15, 50.

Both sets have a mean of exactly 20. Set A's median is also 20 — mean and median coincide, consistent with its symmetric shape and zero skewness. Set B's median is 13, far below its mean of 20 — immediately signalling substantial right skew, driven by the single value of 50. Same mean, dramatically different shape: this is precisely why relying on the mean alone, without also checking skewness (or at minimum comparing it to the median), can misrepresent what a data set actually looks like.