Kurtosis in Statistics: Definition, Types, and Examples

When we think about describing a dataset, we usually reach for the mean and standard deviation. But these two measures only tell part of the story. What about the shape of a distribution — specifically, how much of the data is concentrated in the tails versus the center?

That’s where kurtosis comes in. Kurtosis is a statistical measure that describes the “tailedness” of a probability distribution. A high kurtosis value signals that a dataset has heavy tails and a sharp peak, meaning extreme values appear more often than a normal distribution would predict. A low kurtosis value, by contrast, indicates lighter tails and a flatter peak.

Understanding kurtosis matters in fields like finance, engineering, and data science, where rare but extreme events can have outsized consequences. Before you can manage risk effectively, you need to know whether your data is hiding surprises in its tails.

Why Is Kurtosis Important?

It reveals tail risk. The most consequential reason to care about kurtosis is what it tells you about extreme events. High kurtosis — known as leptokurtosis — means your distribution has heavier tails than a normal curve. In practical terms, this means outliers are not flukes; they are a structural feature of your data. In finance, stock return distributions are famously leptokurtic. Standard risk models that assume normality will consistently underestimate the probability of large losses, sometimes with devastating consequences.

It exposes flawed assumptions. Many statistical tests — including t-tests, ANOVA, and linear regression — assume normally distributed data or residuals. When kurtosis is significantly high or low, these tests can produce misleading results. Checking kurtosis before running an analysis is a simple safeguard that can prevent you from drawing false conclusions.

It shapes model performance in machine learning. Features with extreme kurtosis can distort model training. Algorithms sensitive to the scale and distribution of inputs — such as linear regression, logistic regression, and neural networks — may perform poorly when fed high-kurtosis features. Identifying and transforming these variables, through log transforms or other techniques, often leads to meaningfully better results.

It matters in quality control and engineering. In manufacturing, a process with high kurtosis in its output measurements suggests that defects or failures, while infrequent, are more extreme when they do occur. Engineers monitoring tolerance levels need this information to design safety margins that actually hold under real conditions.

It informs risk management across industries. Insurance companies, epidemiologists, and climate scientists all work with data where rare but extreme outcomes define the problem. A hurricane model that underestimates tail risk because it ignores kurtosis is not just statistically imprecise — it is dangerous. Kurtosis gives practitioners the vocabulary and the tool to ask the right question: not just how variable is this data, but how likely are the worst outcomes?

How Kurtosis Works

The Formula

Kurtosis is calculated as the fourth standardized moment of a distribution:$$\kappa = \frac{\frac{1}{n}\sum_{i=1}^{n}(x_i – \bar{x})^4}{\left(\frac{1}{n}\sum_{i=1}^{n}(x_i – \bar{x})^2\right)^2}$$

Each observation’s deviation from the mean is raised to the fourth power, then averaged. That result is divided by the square of the variance. Raising deviations to the fourth power is the key move here: it amplifies large deviations dramatically while making small ones nearly irrelevant. Kurtosis is, in essence, a measure of how much the extreme values — rather than the typical ones — are driving the shape of your distribution.

Excess Kurtosis

In practice, most software reports excess kurtosis, which simply subtracts 3 from the raw kurtosis value:$$\kappa_{excess} = \kappa – 3$$

Why 3? Because a perfect normal distribution has a raw kurtosis of exactly 3, subtracting 3 re-centers the scale so that a normal distribution scores zero, making it easier to interpret deviations. When someone reports a kurtosis value without qualification, it is worth checking which version they mean.

What Kurtosis Is Not

A persistent misconception is that kurtosis measures the peakedness of a distribution — how tall and sharp the center looks. This framing is misleading. Two distributions can have identical kurtosis but very different peak heights, and vice versa. What kurtosis genuinely captures is the concentration of data in the tails relative to the shoulders of the distribution. The peak shape is a byproduct, not the point.

Another common confusion: kurtosis and skewness are related but distinct. Skewness measures asymmetry — whether one tail is longer than the other. Kurtosis measures tail weight regardless of direction. A distribution can be perfectly symmetric and still have extreme kurtosis.

A Simple Intuition

Think of kurtosis as a measure of how “surprise-prone” your data is. A low-kurtosis dataset is predictable—values stay close to what you’d expect. A high-kurtosis dataset is one where most observations are unremarkable, but when something unusual happens, it is very unusual. That asymmetry between routine observations and rare extremes is precisely what kurtosis quantifies.

Types of Kurtosis

Leptokurtic Distributions (Excess Kurtosis > 0)

The term comes from the Greek leptos, meaning slender or narrow. Leptokurtic distributions have heavier tails and a higher, sharper peak than a normal distribution.

What this looks like in practice: most observations cluster tightly around the mean, but the tails stretch further than you would expect. Extreme values appear more often than a normal distribution would predict — not because the data is poorly collected or corrupted, but because heavy tails are a genuine feature of the underlying process.

Real-world examples:

• Financial returns. Daily stock price changes are the textbook leptokurtic dataset. Most days are quiet, but crashes and surges occur far more frequently than a normal model predicts. This is sometimes called fat-tail risk.
• Insurance claims. The vast majority of claims are routine and small. Occasionally, a catastrophic claim arrives that dwarfs everything else.
• Network traffic. Internet data flows are typically stable, punctuated by sudden, massive spikes.

A leptokurtic distribution demands caution. If you model it as normal, you will systematically underestimate the probability and magnitude of extreme outcomes.

Mesokurtic Distributions (Excess Kurtosis = 0)

Mesokurtic distributions have tail behavior that matches the normal distribution. The normal distribution is itself the defining mesokurtic case, with a raw kurtosis of exactly 3 and an excess kurtosis of zero.

This does not mean mesokurtic distributions are always normal — other distributions can also produce an excess kurtosis near zero while differing from the normal distribution in other ways. What it does mean is that the tail weight is neither unusually heavy nor unusually light.

Real-world examples:

• Heights and weights in a stable, homogeneous population tend to approximate mesokurtic behavior.
• Measurement errors in well-controlled laboratory settings often fall close to this benchmark.

Mesokurtic is the baseline. When your data lands here, standard statistical tools designed around the normal distribution are on firmer ground.

Platykurtic Distributions (Excess Kurtosis < 0)

From the Greek platys, meaning broad or flat. Platykurtic distributions have lighter tails and a lower, flatter peak than a normal distribution. Extreme values are less frequent than the normal distribution would suggest, and data is spread more evenly across the range.

Rather than a sharp central spike with long tails, think of a broad, rounded shape where values are distributed more uniformly.

Real-world examples:

• Uniform distributions. When every outcome in a range is equally likely — rolling a fair die, for instance — the result is strongly platykurtic.
• Student test scores with score compression. When a test is designed so that nearly everyone scores between 60 and 90, the resulting distribution tends to be platykurtic: there are few extreme scores in either direction.
• Bounded physical measurements. Any process with hard upper and lower limits tends to produce platykurtic output, because the tails are cut off by the constraints.

Platykurtic distributions are generally less dangerous to mismodel than leptokurtic ones — the risk of underestimating extremes is lower — but they can still cause problems if a method assumes more tail weight than actually exists.

Type	Excess Kurtosis	Tails	Peak	Extreme Events
Leptokurtic	> 0	Heavy	Sharp	More frequent than normal
Mesokurtic	= 0	Normal	Normal	As frequent as normal
Platykurtic	< 0	Light	Flat	Less frequent than normal

Why the Distinctions Matter

The three types are not academic categories — they carry direct consequences for how you model data and manage risk. A leptokurtic dataset punishes anyone who treats it as normal. A platykurtic dataset may make standard confidence intervals unnecessarily wide. Identifying which type you are working with is one of the first diagnostic steps in any serious statistical analysis, and it shapes every modeling decision that follows.

How to Interpret Kurtosis Values

Calculating kurtosis is straightforward. Interpreting it correctly is where most analysts run into trouble. A kurtosis value sitting in your output table means nothing on its own — context, scale, and the right reference points are what turn a number into an insight.

Start With the Right Scale

Before interpreting any kurtosis value, confirm whether your software is reporting raw kurtosis or excess kurtosis.

• Raw kurtosis: The normal distribution scores 3. Values above 3 indicate heavier tails; values below 3 indicate lighter tails.
• Excess kurtosis: The normal distribution scores 0. Positive values mean heavier tails; negative values mean lighter tails.

Most modern software — including Python’s SciPy, R, and SPSS — reports excess kurtosis by default, but this is not universal. Excel’s KURT function, for example, also returns excess kurtosis, but the documentation does not always make this obvious. Checking the documentation before interpreting output is a small habit that prevents significant errors.

A Practical Interpretation Guide

Once you know which scale you are working with, the following benchmarks offer a starting framework. These assume excess kurtosis:

Excess Kurtosis	Interpretation
Below −1	Strongly platykurtic; tails are notably lighter than normal
−1 to 0	Mildly platykurtic; close to normal, slight flatness
0	Mesokurtic; tail behavior matches the normal distribution
0 to 1	Mildly leptokurtic; modest tail weight, often acceptable
1 to 3	Moderately leptokurtic; worth investigating in risk-sensitive contexts
Above 3	Strongly leptokurtic; heavy tails, extreme values are a real concern

These thresholds are guidelines, not hard rules. The significance of any kurtosis value depends heavily on sample size, the domain you are working in, and what you plan to do with the data.

Sample Size Changes Everything

Kurtosis estimates are sensitive to sample size in ways that skewness and variance are not. With small samples, kurtosis estimates are highly unstable — a single outlier can dramatically inflate the value, and two datasets drawn from the same population can produce very different kurtosis readings by chance alone.

As a rule of thumb:

• With fewer than 50 observations, treat kurtosis estimates with considerable skepticism.
• With 50 to 300 observations, kurtosis is informative but should be interpreted alongside visual tools like histograms and Q-Q plots.
• With more than 300 observations, kurtosis estimates become reliably stable and can be interpreted more confidently on their own.

When sample size is small, no single number will tell you the full story. Plot your data.

Pair Kurtosis With Visual Diagnostics

A kurtosis value gains meaning when paired with visual tools. Two distributions can share the same kurtosis value while looking quite different, and two distributions that look similar can have meaningfully different kurtosis values. Numbers and plots check each other.

Histogram: A quick histogram reveals the overall shape of your distribution. A sharp, narrow peak with visibly long tails is a visual signature of high kurtosis. A broad, flat shape suggests platykurtosis.

Q-Q plot (quantile-quantile plot): This is the most powerful visual companion to kurtosis. A Q-Q plot compares your data’s quantiles against those of a theoretical normal distribution. If the points follow the diagonal line closely, your data behaves normally. If the tails curve away from the line — bending upward on the right and downward on the left — you have heavy tails and elevated kurtosis. The degree of that curve maps directly onto kurtosis magnitude.

Box plot: Unusually many points flagged as outliers beyond the whiskers is a practical, immediate signal of leptokurtosis.

Kurtosis in Context: Domain Matters

What counts as a concerning kurtosis value varies by field, and experienced practitioners calibrate their interpretation accordingly.

In finance, excess kurtosis of 3 to 5 in daily return data is common and expected. A value in that range is not alarming — it is the norm. Values above 10, however, may signal a data quality issue or an extraordinary market event worth investigating separately.

In manufacturing and quality control, even modest excess kurtosis above 1 or 2 can be significant, because it suggests that tail failures — defective parts, tolerance breaches — are occurring more often than a normal process would produce.

In the social sciences, where sample sizes are often moderate and distributions are naturally messier, kurtosis is typically used as one diagnostic among many rather than a decisive standalone measure.

In machine learning, the concern is less about the absolute value and more about whether high kurtosis is distorting model training. Features with excess kurtosis above 3 or 4 are often candidates for transformation.

Statistical Tests for Kurtosis

Beyond visual inspection and rule-of-thumb thresholds, formal statistical tests can assess whether observed kurtosis departs significantly from normality.

Jarque-Bera test: Combines kurtosis and skewness into a single test of normality. Widely used in econometrics and finance. It is sensitive to large samples — with enough data, even trivially small deviations from normality will register as statistically significant.

D’Agostino-Pearson test: Tests for departures from normality using both skewness and kurtosis. Generally more robust than Jarque-Bera for smaller samples.

Anscombe-Glynn test: Tests kurtosis specifically, in isolation from skewness. Useful when you want to assess tail behavior independently of asymmetry.

A statistically significant result from any of these tests means your data departs from normality in a measurable way — but it does not automatically tell you what to do about it. That judgment depends on what analysis you are running and how sensitive it is to non-normality.

Kurtosis vs. Skewness

Kurtosis and skewness are both moments of a distribution, and both describe shape rather than location or spread. They are frequently reported together, often confused with each other, and sometimes mistakenly treated as redundant. They are not. They measure fundamentally different things, and understanding the distinction makes you a sharper analyst.

What Each Measure Captures

Skewness measures asymmetry. It tells you whether one tail of a distribution is longer or heavier than the other, and in which direction. A perfectly symmetric distribution — whether normal, uniform, or otherwise — has a skewness of zero. Positive skewness means the right tail is longer; the bulk of values sit on the left, with a few large values pulling the mean rightward. Negative skewness is the mirror image.

Kurtosis measures tail weight relative to the center of the distribution. It is indifferent to direction. A distribution with high kurtosis has heavy tails on both sides — it does not favor left or right, it simply produces extreme values more often than a normal distribution would. Where skewness asks which way does the distribution lean, kurtosis asks how often does the distribution produce surprises, and how large are they.

The Third and Fourth Moments

The distinction maps cleanly onto the mathematics. Both skewness and kurtosis are standardized moments — they measure deviations from the mean, raised to a power, and divided by the standard deviation raised to the same power.

• Skewness is the third standardized moment. Raising deviations to the third power preserves their sign, which is why skewness captures direction.
• Kurtosis is the fourth standardized moment. Raising deviations to the fourth power eliminates sign entirely — negative deviations become positive — which is why kurtosis is symmetric with respect to direction and focuses purely on magnitude.

This mathematical difference explains why the two measures are complementary rather than redundant. They are asking different algebraic questions of the same data.

They Are Independent of Each Other

A critical point that trips up many analysts: skewness and kurtosis are statistically independent. Any combination of the two is possible in principle:

Skewness	Kurtosis	Example
Zero	Zero (mesokurtic)	Normal distribution
Zero	High (leptokurtic)	Symmetric fat-tailed distribution (e.g., Student’s t)
Zero	Low (platykurtic)	Uniform distribution
Positive	High	Many insurance claim distributions
Negative	High	Some asset return distributions during market stress
Positive	Low	Some bounded distributions with right-side compression

You cannot infer one from the other. A distribution can be perfectly symmetric and still have dramatic kurtosis. A highly skewed distribution can have kurtosis close to zero. Always measure both independently.

Different Risk Signals

In applied work, skewness and kurtosis flag different kinds of problems.

Skewness signals directional bias. If your model assumes symmetry and the data is skewed, your predictions will be systematically off in one direction. In income data, for instance, strong positive skew means the mean overstates what a typical person earns. In reliability engineering, a negatively skewed failure-time distribution means most components fail earlier than the mean suggests. Skewness tells you your center of gravity is not where you think it is.

Kurtosis signals extreme event risk. High kurtosis does not tell you which direction the next shock will come from — it tells you that when a shock arrives, it will likely be larger than your model expects. In finance, a leptokurtic return distribution means portfolio losses during a crash will exceed what a normal-distribution risk model forecasts. Kurtosis tells you your tails are wider than you think they are.

The practical difference: skewness is about being wrong on average; kurtosis is about being catastrophically wrong occasionally.

When Both Are Present

Real datasets frequently exhibit both skewness and excess kurtosis simultaneously, and the combination compounds the challenge of modeling them accurately.

Consider daily rainfall data. It is positively skewed — most days have little or no rain, while a few days have very heavy rainfall. It is also leptokurtic — extreme rainfall events are more frequent and more intense than a symmetric normal model would predict. A model that corrects for skewness but ignores kurtosis will still underestimate the worst storms. A model that accounts for kurtosis but ignores skewness will misplace the center of the distribution. Getting the full picture requires attending to both.

In machine learning, a feature that is both skewed and leptokurtic often needs two separate transformations: one to address asymmetry (a log or square root transform) and one to address tail weight (clipping or winsorization).

	Skewness	Kurtosis
What it measures	Asymmetry	Tail weight
Moment	Third	Fourth
Normal distribution value	0	0 (excess)
Positive value means	Right tail is heavier	Tails are heavier than normal
Negative value means	Left tail is heavier	Tails are lighter than normal
Main risk signal	Directional bias	Extreme event frequency
Are they independent?	Yes — knowing one tells you nothing about the other

Real-World Examples of Kurtosis

Finance: Fat Tails and Market Crashes

Finance is where kurtosis has received the most attention, and for good reason. The daily returns of stocks, indices, currencies, and commodities are all leptokurtic. Excess kurtosis values of 5 to 10 are common in daily return data for individual stocks; during periods of market stress, values can climb far higher.

The consequences of ignoring this are well documented. The Black-Scholes options pricing model, developed in the 1970s and widely adopted through the 1980s and 1990s, assumes that asset returns follow a normal distribution. In practice, this means the model systematically underprices options on extreme moves — the very contracts that pay out during crashes.

The 1987 Black Monday crash, during which the Dow Jones Industrial Average fell over 22% in a single day, was an event so far into the tail of a normal distribution that it should have been effectively impossible under that model. The normal distribution assigned it a probability so small it was functionally zero. The real distribution, with its fat tails, made it merely very unlikely. This distinction — between impossible and unlikely — is precisely what kurtosis captures, and the financial system was not priced for it.

Value at Risk (VaR) models that assume normality have repeatedly underestimated losses during crises for the same reason. The 2008 financial crisis produced daily losses at major banks that normal-distribution VaR models predicted should occur once every several thousand years — yet they occurred repeatedly within weeks. Leptokurtosis was not a footnote; it was the story.

Insurance: Catastrophic Claims

Insurance is fundamentally a business of tail risk, and kurtosis sits at the center of how that risk is priced and reserved against.

For most lines of insurance — auto, health, homeowners — the bulk of claims are routine and fall within a predictable range. But the distribution of claim sizes is strongly leptokurtic. The rare catastrophic claim — a multi-vehicle accident, a major house fire, a severe liability judgment — is far more extreme than a normal distribution would suggest, and these tail events drive a disproportionate share of total losses.

Reinsurance pricing, which covers insurers against their own extreme losses, depends critically on accurate tail modeling. An insurer that assumes normally distributed claims will set aside insufficient reserves for tail events. When those events arrive — as they inevitably do — the shortfall can threaten solvency. The actuarial models used in serious insurance work account explicitly for excess kurtosis, often using heavy-tailed distributions such as the Pareto or log-normal to fit claim data more accurately.

Natural catastrophe modeling makes this even more pronounced. Hurricane losses, earthquake damage, and flood claims are all characterized by heavy tails: most years produce manageable losses, but a single catastrophic season can exceed the cumulative losses of the previous decade.

Manufacturing and Quality Control

In manufacturing, the goal is a process that produces output tightly clustered around a target specification. Normal distributions are the assumed ideal. But when a process develops faults — worn tooling, temperature variation, raw material inconsistency — the output distribution often becomes leptokurtic before the shift is visible in standard control charts.

The practical consequence is that defects and tolerance violations cluster not just around the mean, but in the tails — and those tail defects are often the most severe. A machined component that is slightly off-spec is a quality issue; one that is dramatically off-spec can cause downstream failures in assembly or field use.

Statistical process control methods that track only mean and variance can miss this entirely. A process can have a stable mean and standard deviation while its kurtosis climbs, signaling that something is changing in the extremes even as the center holds. Adding kurtosis monitoring to process control dashboards provides an early warning of deterioration that variance-only methods miss.

Medicine and Clinical Research

Clinical trial data frequently exhibits non-normal distributions, and kurtosis is one reason why. Biomarker measurements, response times, and physiological measurements in patient populations often have heavier tails than researchers expect, reflecting genuine biological variability at the extremes.

This matters for two reasons. First, statistical tests used to assess treatment effects — t-tests, ANOVA — assume normality. When the underlying data is strongly leptokurtic, these tests can produce unreliable p-values, either overstating or understating significance. Researchers who check kurtosis before choosing their analysis method are on firmer ground than those who apply parametric tests by default.

Second, in drug safety monitoring, adverse events are by definition tail occurrences. A clinical trial that models adverse event rates assuming normality will underestimate the frequency and severity of rare but serious side effects. Regulatory agencies increasingly expect sponsors to characterize the full distribution of outcomes, not just the mean and variance.

Hydrology and Climate Science

River flow data and rainfall measurements are classic examples of leptokurtic real-world phenomena. Average flow rates may be stable and predictable across most of the year, but flood events — the tail of the distribution — are both more frequent and more severe than a normal model predicts.

Flood frequency analysis, which underpins the design of dams, levees, bridges, and drainage systems, relies on fitting statistical distributions to historical flow data. Engineers speak of the “100-year flood” or “500-year flood” — return periods calculated from the tail of the fitted distribution. If that distribution assumes normal or near-normal tail behavior when the true distribution is heavily leptokurtic, the 100-year flood may actually arrive every 40 or 50 years. Infrastructure designed to those underestimated standards will fail ahead of schedule.

The consequences are not abstract. Levee failures, bridge washouts, and urban flooding events have all been linked in part to hydrological models that underestimated tail risk by assuming thinner tails than the historical data actually supported.

Cybersecurity and Network Traffic

Internet traffic volume and cybersecurity event data are both strongly leptokurtic. Normal traffic follows predictable patterns, but distributed denial-of-service attacks, data exfiltration events, and system intrusions manifest as extreme spikes — values far into the tail of the traffic distribution.

Anomaly detection systems that model traffic as normally distributed will set alert thresholds too conservatively in some areas and too permissively in others. Systems designed with heavy-tailed models, which account for the leptokurtic nature of real traffic data, set thresholds that better distinguish routine variation from genuine threats.

The same logic applies to the size of security incidents. Most breaches are contained; a small number are catastrophic. The distribution of breach severity is leptokurtic, and organizations that plan their incident response capacity based on average breach size will be structurally underprepared for the tail events that cause the most damage.

Advantages and Limitations of Using Kurtosis

How to Calculate Kurtosis in Popular Software

Microsoft Excel

Function:KURT()

Syntax:

excel

=KURT(number1, [number2], ...)

Example:

excel

=KURT(A2:A11)

Important Notes:

• Excel’s KURT function calculates sample excess kurtosis (Fisher’s definition), so a normal distribution returns approximately 0, not 3.
• Maximum of 255 data values.
• Requires at least 4 data points; returns #DIV/0! otherwise.
• For population kurtosis, you’ll need to adjust the result manually.

Python (SciPy)

Function:scipy.stats.kurtosis()

Syntax:

Python

from scipy.stats import kurtosis
kurtosis(data, axis=0, fisher=True, bias=True)

Example:

Python

import numpy as np
from scipy.stats import kurtosis

data = np.random.normal(size=1000)
result = kurtosis(data, fisher=True, bias=False)
print(result)  # Returns excess kurtosis (~0 for normal)

Key Parameters:

• fisher=True → Returns excess kurtosis (normal = 0.0)
• fisher=False → Returns Pearson kurtosis (normal = 3.0)
• bias=False → Applies bias correction using k-statistics

R

Packages:e1071, datawizard, or moments

Using e1071 (most common):

r

library(e1071)
kurtosis(x, na.rm = TRUE, type = 2)

Using datawizard:

r

library(datawizard)
kurtosis(x, type = "2")

Important: R defaults to type = 3 (Minitab-style), but Excel, SPSS, and SAS use type = 2. To match results across platforms, explicitly set type = 2 in R.

Type	Method	Software Equivalent
1	Classic (g2)	—
2	Bias-corrected (G2)	SAS, SPSS, Excel
3	Adjusted (b2)	Minitab

MATLAB

Function:kurtosis()

Syntax:

matlab

k = kurtosis(X)
k = kurtosis(X, flag)

Example:

matlab

data = randn(100,1);
k = kurtosis(data, 0);  % Bias-corrected

Key Behavior:

• MATLAB returns Pearson kurtosis (normal = 3), not excess kurtosis.
• flag = 1 → Biased estimate
• flag = 0 → Bias-corrected estimate (recommended for sample data)
• Requires at least 4 elements for bias correction.

SPSS

Menu Path:

plain

Analyze → Descriptive Statistics → Descriptives → Options → Check "Kurtosis"

Or via Syntax:

spss

DESCRIPTIVES VARIABLES=var1
  /STATISTICS=KURTOSIS.

Behavior:

• SPSS calculates sample excess kurtosis by default (equivalent to R’s type = 2).

SAS

Procedure:PROC MEANS or PROC UNIVARIATE

Syntax:

sas

PROC MEANS DATA=dataset KURTOSIS;
  VAR variable;
RUN;

Or with more detail:

sas

PROC UNIVARIATE DATA=dataset;
  VAR variable;
RUN;

Behavior:

• SAS calculates sample excess kurtosis (equivalent to R’s type = 2).

FAQs