Answer

What does kurtosis actually tell you about your data?

A frequently asked statistics question · September 21, 2026

The short answer

Kurtosis measures how prone a distribution is to extreme values. Compared with a normal distribution (excess kurtosis of 0), positive values mean heavier tails and more outliers, and negative values mean lighter tails. It is not a measure of how peaked the curve looks.

What kurtosis measures

Kurtosis is a summary of a distribution's tails: how much of the variability comes from rare, extreme values rather than from values near the middle. A distribution with high kurtosis produces outliers more often than a normal distribution with the same standard deviation would.

Technically, it is the average of the fourth power of the standardized values, mean(z^4). Raising values to the fourth power lets extreme observations dominate the result, which is why kurtosis is really a tail statistic. The normal distribution has a kurtosis of 3.

How to read the number

Most people work with excess kurtosis, which is kurtosis minus 3, so the normal distribution scores 0.

Near 0 (mesokurtic): tails similar to a normal distribution.
Positive (leptokurtic): heavier tails, so more extreme values than a normal distribution. Examples include the Laplace distribution (excess kurtosis 3) and t distributions with few degrees of freedom.
Negative (platykurtic): lighter tails and fewer extreme values. The uniform distribution has an excess kurtosis of -1.2.

Check which convention your software uses before you interpret a value. For example, Excel's KURT and SPSS report excess kurtosis, while Stata's summarize, detail reports raw kurtosis, where the normal distribution is 3.

What you can say about a graph

In a histogram, high kurtosis shows up as a thin stretch of observations far from the center. It is easier to judge in a normal Q-Q plot: with heavy tails, the points at both ends curve away from the reference line, the left end below it and the right end above it. With light tails, the ends bend the other way.

Skewness is a separate question. Skewness describes asymmetry, while kurtosis describes tail weight in both directions combined. A distribution can be symmetric with heavy tails, or skewed with light ones, so look at both.

Common mistakes

Calling it peakedness. Many textbooks say kurtosis measures how sharp or flat the peak is. Statisticians have argued against that description (for example Westfall, 2014), because distributions with the same kurtosis can have very different peaks. Think tails, not peaks.
Trusting it in small samples. Sample kurtosis is noisy and biased. Simulating samples of 30 from a perfectly normal distribution gives estimates that average about -0.4 and vary with a standard deviation of about 0.7, and a single extreme point can shift the value by several units.
Treating rules of thumb as tests. Cutoffs such as 'between -2 and +2 is acceptable' are conventions, not statistical tests. A plot and the context of your analysis tell you more.
Ignoring why the tails are heavy. Heavy tails can come from a data entry error, a mix of different groups, or a real process that produces extremes. Find out which before deciding what to do.

See it in R

This simulation compares three distributions with known excess kurtosis. The Laplace distribution here is built from two exponentials, so it needs no extra packages.

set.seed(42)
n <- 100000

normal  <- rnorm(n)             # excess kurtosis 0
laplace <- rexp(n) - rexp(n)    # heavy tails: excess kurtosis 3
uniform <- runif(n)             # light tails: excess kurtosis -1.2

excess_kurtosis <- function(x) {
  z <- (x - mean(x)) / sd(x)
  mean(z^4) - 3
}

round(sapply(list(normal = normal, laplace = laplace, uniform = uniform),
             excess_kurtosis), 2)
# Roughly: normal 0, laplace 3, uniform -1.2

qqnorm(laplace); qqline(laplace)   # the ends curve away from the line

A sample of 100,000 is only for illustration. With a few dozen observations, treat the estimate as a rough guide and rely on the plot.

Related tools and guides

Working with your own data?

General answers only go so far. Send us your situation and we'll reply by email within two business days.

Ask your question

Written by AskStats with AI assistance. This is general information, not advice for your specific data or study. When the results matter, check your approach with a qualified statistician.