An online normality test checks whether your sample data follows a normal (Gaussian) distribution by computing a statistical test statistic and a p-value. You paste or upload your numbers, the tool runs one or more standard tests (Shapiro-Wilk, Anderson-Darling, Kolmogorov-Smirnov, Jarque-Bera), and it returns a verdict: if the p-value is below your chosen significance level (usually 0.05), the data is considered not normal. These tools are free, run entirely in your browser, and are the fastest way to check the normality assumption before doing t-tests, ANOVA, or linear regression.
Formula
The Shapiro-Wilk test is the most widely recommended normality test for small to medium samples (n < 5000). Its statistic W is calculated as:
W = (Σ ai x(i))² / Σ (xi − x̄)²
where:
- x(i) are the ordered sample values (smallest to largest)
- ai are constants derived from the expected values of the order statistics of a normal distribution, computed from the sample size n
- x̄ is the sample mean
- W ranges from 0 to 1; values close to 1 indicate normality
The p-value is obtained by transforming W to a standard normal variable using the mean and variance of W under the null hypothesis. The Anderson-Darling statistic A² measures the weighted squared distance between the empirical cumulative distribution function and the normal CDF, giving more weight to the tails:
A² = −n − (1/n) Σ (2i−1) [ln(F(x(i))) + ln(1−F(x(n+1−i)))]
where F is the normal cumulative distribution function with the sample mean and variance. Skewness measures asymmetry (0 for a symmetric distribution), and excess kurtosis measures tail heaviness (0 for a normal distribution).
Worked Example
Using the default values: 5.1, 4.9, 5.6, 4.8, 5.2, 5.0, 5.4, 4.7, 5.3, 5.5 (n = 10).
Step 1: Calculate the mean: x̄ = (5.1+4.9+5.6+4.8+5.2+5.0+5.4+4.7+5.3+5.5) / 10 = 51.5 / 10 = 5.15
Step 2: Calculate the sum of squared deviations from the mean: Σ(xi − x̄)² = (5.1−5.15)² + (4.9−5.15)² + (5.6−5.15)² + (4.8−5.15)² + (5.2−5.15)² + (5.0−5.15)² + (5.4−5.15)² + (4.7−5.15)² + (5.3−5.15)² + (5.5−5.15)² = 0.0025 + 0.0625 + 0.2025 + 0.1225 + 0.0025 + 0.0225 + 0.0625 + 0.2025 + 0.0225 + 0.1225 = 0.825
Step 3: The Shapiro-Wilk coefficients ai for n=10 (from published tables) are approximately: 0.5739, 0.3291, 0.2141, 0.1224, 0.0399, −0.0399, −0.1224, −0.2141, −0.3291, −0.5739. Ordered data: 4.7, 4.8, 4.9, 5.0, 5.1, 5.2, 5.3, 5.4, 5.5, 5.6.
Step 4: Compute the weighted sum: 0.5739×4.7 + 0.3291×4.8 + 0.2141×4.9 + 0.1224×5.0 + 0.0399×5.1 + (−0.0399)×5.2 + (−0.1224)×5.3 + (−0.2141)×5.4 + (−0.3291)×5.5 + (−0.5739)×5.6 = 2.6973 + 1.5797 + 1.0491 + 0.6120 + 0.2035 − 0.2075 − 0.6487 − 1.1561 − 1.8101 − 3.2138 = −0.8946
Step 5: W = (−0.8946)² / 0.825 = 0.8003 / 0.825 = 0.9701. The p-value for W = 0.9701 with n=10 is approximately 0.912, matching the calculator output. Since p > 0.05, we fail to reject normality.
Assumptions
- The sample is a random, independent sample from the population
- The data are continuous (not counts or categories)
- The test compares the sample to a normal distribution with the same mean and variance as the sample
- No outliers are present that are recording errors (the test is sensitive to them)
Which Test Should You Use Online?
Different online calculators implement different tests, and the choice matters for your sample size. For samples under about 50, the Shapiro-Wilk test has the highest statistical power — it is most likely to detect non-normality when it truly exists. For samples between 50 and 5000, the Anderson-Darling test is often preferred because it is sensitive to deviations in the tails, which matter for hypothesis tests. For very large samples (thousands of points), the Kolmogorov-Smirnov test becomes too sensitive: it will flag tiny, practically irrelevant deviations as significant. The Jarque-Bera test, based on skewness and kurtosis, works well for large samples but is unreliable for n under about 100. When you use an online tool, check which test it runs. Many good calculators let you pick the test or run several at once. If the results disagree, look at the sample size: for small samples trust Shapiro-Wilk; for large samples trust Anderson-Darling or a visual check.
How to Read the Output Correctly
The key output is the p-value. The null hypothesis is “the data are normally distributed.” A p-value less than 0.05 (the common threshold) means you reject normality — the data are significantly non-normal. A p-value above 0.05 means you do not have enough evidence to say the data are non-normal; it does not prove they are normal. This is a subtle but crucial point: failing to reject normality is not the same as confirming it. With a small sample (n < 10), even strongly non-normal data can produce a high p-value simply because the test lacks power. Conversely, with a huge sample (n > 1000), trivial deviations from normality will produce a tiny p-value. Always look at a histogram or Q-Q plot alongside the p-value. Many online tools include these plots automatically. A Q-Q plot shows your data quantiles against theoretical normal quantiles; if the points roughly follow the diagonal line, normality is reasonable regardless of the p-value.
Common Mistakes and Misconceptions
The most frequent error is testing normality on the raw data when the statistical method only requires normality of the residuals. For example, in linear regression, the outcome variable itself does not need to be normal — the residuals (the differences between observed and predicted values) do. Testing the raw outcome will often show non-normality even when the model is perfectly valid. A second mistake is ignoring the sample size. For n under about 15, normality tests are nearly useless; use a Q-Q plot instead. For n over a few hundred, the tests are too sensitive; focus on the shape of the distribution rather than the p-value. A third error is assuming that non-normal data forces you to abandon parametric tests. The central limit theorem means that for sample means, the sampling distribution is approximately normal for n ≥ 30 even if the population is not. For t-tests and ANOVA with moderately sized samples, mild non-normality is usually not a problem. Only severe skewness or heavy outliers warrant a transformation (log, square root) or a non-parametric alternative like the Mann-Whitney U test.
How to Use an Online Normality Test Step by Step
- Copy your data from a spreadsheet or CSV file. Ensure there are no text labels, commas as decimal separators, or empty cells.
- Paste the numbers into the input box. Most tools accept space-separated, comma-separated, or newline-separated values.
- Choose the significance level (usually 0.05). Some tools let you set 0.01 or 0.10.
- Select the test(s) you want. If unsure, choose Shapiro-Wilk for n < 50 or Anderson-Darling for larger samples.
- Click “Calculate” or “Test.” The tool will output the test statistic, the p-value, and often a histogram and Q-Q plot.
- Interpret: if p < 0.05, the data deviates from normal; if p ≥ 0.05, you can proceed assuming normality for most practical purposes, but check the plot.
Limitations of Online Normality Tests
Online tools are convenient but have real limitations. They cannot handle missing data — you must remove or impute blanks first. They assume your data are a simple random sample; if your data are clustered, repeated measures, or time series, the independence assumption is violated and the normality test is invalid. They also cannot tell you whether the non-normality matters for your specific analysis. A distribution that is slightly skewed but has no outliers may be fine for a t-test, while a symmetric distribution with a single extreme outlier can break the same test. Finally, no online calculator can replace domain knowledge: some variables (reaction times, incomes, antibody titers) are known to be log-normal, so testing for normality is often the wrong question — you should test whether the log-transformed values are normal instead. Always report which test you used and the sample size when you mention a normality test in your work.
Frequently Asked Questions
What is the best normality test for a small sample? For n under 50, the Shapiro-Wilk test is the most powerful and is the standard choice. For n under 10, even Shapiro-Wilk has low power, so rely on a Q-Q plot and your judgment rather than the p-value alone.
Can I use a normality test for any type of data? No. The tests assume continuous data. For discrete data (like Likert scales, counts, or percentages), normality tests are not appropriate. For ordinal data, use a non-parametric test directly without checking normality.
What does a p-value of exactly 0.05 mean? By convention, p = 0.05 is the boundary. If your p-value is exactly 0.05, you are at the threshold. Most statisticians would treat it as borderline and recommend looking at the Q-Q plot and considering the sample size before deciding. Some fields use 0.01 as a stricter cutoff.
Why do different online calculators give different p-values? Different tools use different tests (Shapiro-Wilk vs. Anderson-Darling vs. Kolmogorov-Smirnov), and even for the same test, approximations for the p-value can vary slightly. Also, some tools round the input data or use different algorithms for the test statistic. If the discrepancy is large, check the sample size and the specific test used.
Do I need to test normality before every t-test? Not strictly. For independent samples t-tests with n ≥ 30 per group, the central limit theorem makes the test robust to non-normality. For small samples or when you have strong skewness or outliers, testing normality (or using a non-parametric test) is wise. The bigger assumption is equal variances, which you should check separately.
Is it okay to transform data to make it normal? Yes, common transformations are log, square root, and Box-Cox. However, transforming changes the interpretation of your results — you are now testing the transformed values. Also, transformation does not fix outliers; it only reduces skewness. If you transform, state it clearly in your methods.
Related
Formula used
list data | Sample values | 5.1, 4.9, 5.6, 4.8, 5.2, 5.0, 5.4, 4.7, 5.3, 5.5 | out shapiro | Shapiro-Wilk p-value | - | 0.912 | 4 out anderson | Anderson-Darling statistic | - | 0.214 | 4 out skew | Skewness | - | -0.156 | 3 out kurt | Excess kurtosis | - | -0.412 | 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 Results: How to Analyze Them
- Normality Test Q-Q Plot: How to Read It
- Normality Slay the Spire?
- Normality Test for Paired t Test: How to Check