Normality of data means that a dataset follows a bell-shaped, symmetric distribution where most values cluster around the mean, with fewer values farther away. You check it using graphical methods (histograms, Q-Q plots) and statistical tests (Shapiro-Wilk, Kolmogorov-Smirnov, Anderson-Darling). This matters because many statistical tests, like t-tests and ANOVA, assume normality to produce valid results.
Formula
The calculator above gives you the expected skewness and excess kurtosis for a perfectly normal distribution with your sample size. The formulas are:
Standard error of skewness: SES = sqrt(6/n)
Standard error of kurtosis: SEK = sqrt(24/n)
Where:
- n = sample size (number of observations)
- sqrt = square root
These are the standard errors you compare your sample’s skewness and kurtosis against. If your sample skewness is more than about 2 SES away from zero, or your kurtosis is more than about 2 SEK away from zero, the data likely departs from normality.
Worked Example
Using the defaults: sample mean = 100, standard deviation = 15, sample size = 30.
For n = 30:
SES = sqrt(6/30) = sqrt(0.2) = 0.4472
SEK = sqrt(24/30) = sqrt(0.8) = 0.8944
So for a sample of 30, a skewness value between -0.894 and +0.894 (2 × 0.447) would be considered consistent with normality. For kurtosis, values between -1.789 and +1.789 (2 × 0.894) would be acceptable. The calculator shows 0 for both because the expected value for a normal distribution is zero skewness and zero excess kurtosis.
Assumptions
- The data are a random sample from a larger population.
- Observations are independent of each other.
- The sample size is used to estimate the precision of skewness and kurtosis.
What Is Normality and Why Does It Matter?
Normality refers to the normal distribution, also called the Gaussian distribution. It is described by two parameters: the mean (center) and standard deviation (spread). Many common statistical procedures—t-tests, ANOVA, linear regression, Pearson correlation—rely on the assumption that the residuals or the sampling distribution of the mean are normally distributed. When this assumption holds, p-values and confidence intervals are accurate. When it is violated, your conclusions may be wrong.
However, the Central Limit Theorem provides some relief: for sample sizes above about 30, the sampling distribution of the mean becomes approximately normal even if the raw data are not. This means you often do not need the raw data to be normal—only the sampling distribution of the statistic you are testing.
How to Check Normality: Graphical Methods
Graphical methods are the first step because they show the whole picture, not just a single number.
- Histogram: Overlay a normal curve. If the bars roughly follow the bell shape, normality is plausible. Small samples can look irregular even when normal.
- Q-Q plot (quantile-quantile): Plot your sorted data against the expected quantiles of a normal distribution. If the points fall roughly along the 45-degree line, the data are normal. Systematic deviations (curves, S-shapes) indicate skewness or heavy tails.
- Boxplot: A symmetric box with whiskers of roughly equal length suggests normality. Outliers shown as dots beyond the whiskers indicate departures.
Graphs are subjective, so they are best paired with a statistical test.
Statistical Tests for Normality
Several formal tests exist, each with strengths and weaknesses.
- Shapiro-Wilk test: The most powerful test for normality, especially for small samples (n < 50). It tests whether the data come from a normal distribution. A p-value below 0.05 suggests non-normality.
- Kolmogorov-Smirnov test: Compares the empirical distribution to a normal distribution with estimated parameters. It is less powerful than Shapiro-Wilk but works for larger samples.
- Anderson-Darling test: A modification of the K-S test that gives more weight to the tails. It is sensitive to departures at the extremes.
- Jarque-Bera test: Uses skewness and kurtosis to test normality. It is a good option for large samples.
These tests have a limitation: with large sample sizes (n > 1000), they can detect trivial departures from normality that have no practical impact. With small samples, they may miss real departures. Always interpret test results alongside graphs.
Skewness and Kurtosis: The Numbers Behind the Shape
Skewness measures asymmetry. A perfectly symmetric distribution has skewness of 0. Positive skewness means a long right tail; negative skewness means a long left tail. Kurtosis measures the heaviness of the tails. Excess kurtosis (kurtosis minus 3) is 0 for a normal distribution. Positive excess kurtosis means heavier tails and a sharper peak; negative means lighter tails and a flatter peak.
As a rule of thumb, if skewness is between -1 and +1 and excess kurtosis is between -1 and +1, the data are often considered acceptable for procedures that assume normality. The calculator above gives you the more precise standard errors to judge whether your observed values are statistically different from zero.
Common Mistakes and Misconceptions
- Mistaking the test for the data: A normality test on raw data is not the same as checking residuals from a regression. For regression, check the residuals, not the original outcome variable.
- Ignoring sample size: With n = 10, the Shapiro-Wilk test rarely rejects normality even if the data are skewed. With n = 5000, it almost always rejects even for minor deviations. Use graphs to judge practical significance.
- Assuming non-normal data means you cannot use parametric tests: The Central Limit Theorem often makes this unnecessary for means with moderate sample sizes. Also, robust methods (e.g., Welch’s t-test, bootstrap) handle non-normality well.
- Transforming data without reason: Log, square root, or Box-Cox transformations can fix skewness, but they change the interpretation of your results. Only transform if it makes substantive sense.
- Checking normality on pooled data when groups differ: If you have two groups with different means, the combined data will look non-normal even if each group is perfectly normal. Check each group separately.
What to Do If Data Are Not Normal
If your data fail normality checks, you have several options:
- Use non-parametric tests: Mann-Whitney U (instead of independent t-test), Wilcoxon signed-rank (instead of paired t-test), Kruskal-Wallis (instead of ANOVA), Spearman correlation (instead of Pearson). These do not assume normality.
- Transform the data: Log transformation for right-skewed positive data, square root for count data, or Box-Cox for general cases. Re-check normality after transformation.
- Use robust methods: Trimmed means, bootstrapping, or permutation tests that do not rely on distributional assumptions.
- Proceed anyway if the sample is large: For n > 30-40, the t-test and ANOVA are robust to moderate non-normality, especially if group sizes are similar.
The choice depends on your sample size, the severity of the departure, and the specific test you plan to run.
Frequently Asked Questions
Is normality the same as a bell curve? Yes, a normal distribution has a bell-shaped curve. But not every bell-shaped curve is normal—some are t-distributions or logistic distributions. Normality specifically means the data follow the mathematical function defined by the mean and standard deviation.
What sample size do I need to ignore normality? A common rule is n ≥ 30 due to the Central Limit Theorem. However, if the data are severely skewed or have extreme outliers, even n = 100 may not be enough. For hypothesis tests on means, larger samples make the test robust; for variance tests, normality matters more.
Should I test normality before or after removing outliers? Test first on the raw data. If outliers are genuine (not data entry errors), removing them just to pass a normality test is data manipulation. If they are errors, correct them and re-check.
What is the difference between the Shapiro-Wilk and Kolmogorov-Smirnov tests? Shapiro-Wilk is more powerful for small to moderate samples and is the recommended default. Kolmogorov-Smirnov is less sensitive to deviations in the tails and is better suited for large samples, but it requires the distribution parameters to be known or estimated, which weakens the test.
Can I use a t-test if my data are not normal? Yes, if your sample size is large enough (roughly n > 30) and the skewness is not extreme. The t-test compares means, and the sampling distribution of the mean becomes normal regardless of the raw data’s distribution. For small samples with clear non-normality, use a non-parametric alternative.
Does normality apply to the data or the residuals? For regression and ANOVA, the assumption is that the residuals (errors) are normally distributed, not the raw outcome variable. Always check the residuals after fitting the model, not the original data.
Related
Formula used
in m | Sample mean | - | 100 | 0 in sd | Sample standard deviation | - | 15 | 0 in n | Sample size | - | 30 | 2 out skew | Skewness (approx.) | - | 0 | 3 out kurt | Excess kurtosis (approx.) | - | 0 | 3
Each line is one input (in, sel, list) or one result (out: label, unit, formula, decimals). How the calculators work
Method: the standard formula for this calculation with the constants and conversion factors written on the page, rounded as each result states. See method and constants, the disclaimer or report an error.
More in this section
- Normality to g/L Converter & Formula29.220 g/L
- Normality to meq/mL Conversion Calculator1.0000 meq/mL
- Normality Test Q-Q Plot: How to Read It
- Normality Test Results: How to Analyze Them
- Normality Slay the Spire?
- Normality Test for Paired t Test: How to Check