Normality test results tell you whether your data likely came from a normal (Gaussian) distribution, usually via a p-value from tests like Shapiro-Wilk or Kolmogorov-Smirnov. To analyze them, compare the p-value to your chosen significance level (commonly 0.05): if p ≥ 0.05, you fail to reject normality; if p < 0.05, the data significantly deviates from normal. But a p-value alone is rarely enough—you should also inspect histograms, Q-Q plots, and skewness/kurtosis values to make a sound decision.
Formula
The core formula is not a single equation but the logic of hypothesis testing. For any normality test (Shapiro-Wilk, Anderson-Darling, etc.), you compute a test statistic (e.g., W for Shapiro-Wilk) and its associated p-value. The decision rule is:
- If p-value ≥ α (usually 0.05): fail to reject the null hypothesis that data are normal.
- If p-value < α: reject the null hypothesis, concluding the data are not normally distributed.
Where α is your significance level (probability of wrongly rejecting normality when it is true). The p-value itself depends on the test statistic and sample size. For Shapiro-Wilk, W is calculated from the ordered sample values and expected values under normality; the exact formula is complex and typically handled by software.
Worked Example
Suppose you have a sample of 30 measurements and run a Shapiro-Wilk test in your statistical software. The output gives W = 0.95 and p = 0.20. Using the decision rule with α = 0.05:
- Compare p-value to α: 0.20 ≥ 0.05.
- Since p is greater than α, you fail to reject the null hypothesis.
- Conclusion: There is not enough evidence to say the data deviate from normality; you may proceed with parametric tests that assume normality.
The calculator above replicates this decision using your inputs.
Assumptions
- Observations are independent (no repeated measures or clustering unless accounted for).
- The test chosen is appropriate for the sample size (e.g., Shapiro-Wilk is recommended for n < 50, Kolmogorov-Smirnov for larger samples but with caveats).
- The significance level α is set before the test (commonly 0.05, but can be 0.01 for stricter control).
How Normality Tests Work
Normality tests compare your sample to the theoretical normal distribution. They calculate a statistic that measures how much the empirical distribution of your data differs from the expected normal curve. The p-value then quantifies the probability of observing such a difference (or more extreme) if the data truly came from a normal distribution. Different tests use different metrics:
- Shapiro-Wilk: Based on the correlation between your data and the corresponding normal scores. Powerful for small to medium samples (n < 50).
- Kolmogorov-Smirnov (Lilliefors correction): Compares the empirical cumulative distribution function to the normal CDF. Less powerful than Shapiro-Wilk for small samples.
- Anderson-Darling: A modification of Kolmogorov-Smirnov that gives more weight to the tails. Often used in quality control.
- Jarque-Bera: Based on skewness and kurtosis. Works best for larger samples (n > 100).
How to Choose the Right Test
Choosing a normality test depends on sample size and the nature of your data:
- For small samples (n < 50), Shapiro-Wilk is generally the most powerful and recommended.
- For moderate to large samples (n ≥ 50), Anderson-Darling or Kolmogorov-Smirnov (with Lilliefors correction) are common; however, with very large samples (n > 1000), even trivial deviations become significant, so graphical checks become more important.
- If you have ties or discrete data, exact tests may be needed; otherwise, the standard tests are approximate.
- Always check the test’s assumptions—some tests require a fully specified distribution (mean and variance known), which is rare in practice; use versions that estimate parameters from data.
Common Problems and Misconceptions
Many analysts misinterpret normality test results. A common mistake is to conclude that data are “normal” when p ≥ 0.05, but this only means you cannot detect a deviation—it does not prove normality. Conversely, with large samples, p < 0.05 may flag negligible deviations that are practically irrelevant. Also, some people confuse the null hypothesis: the null is that data are normal, so a low p-value rejects normality, not the opposite. Another issue is using the wrong test for the sample size; for example, Kolmogorov-Smirnov without Lilliefors correction is overly conservative for estimated parameters. Finally, normality tests are sensitive to outliers; a single extreme value can cause rejection even if the rest of the data are normal. Always plot your data alongside the test.
Interpreting Results with Graphs
Graphical methods complement numerical tests and often provide clearer insight:
- Histogram: Overlay a normal density curve. Look for symmetry, bell shape, and no extreme outliers.
- Q-Q plot: Plot your data quantiles against theoretical normal quantiles. If points fall roughly along the 45-degree line, normality is plausible. Systematic deviations (curvature, S-shape) indicate skewness or heavy tails.
- Boxplot: Check for symmetry and outliers. A skewed boxplot suggests non-normality.
- Skewness and kurtosis: Values near 0 for skewness and near 3 for kurtosis (or excess kurtosis near 0) indicate normality. Some software reports excess kurtosis (subtracting 3).
If graphs show clear non-normality, the p-value is usually low. If they look normal but p is low, investigate outliers or a large sample size amplifying small deviations.
What to Do If Data Are Not Normal
Rejecting normality does not mean your analysis is doomed. Options include:
- Transformations: Apply log, square root, Box-Cox, or other transforms to make data more normal.
- Non-parametric tests: Use Mann-Whitney U, Kruskal-Wallis, Wilcoxon signed-rank, or Spearman correlation, which do not assume normality.
- Robust methods: Use trimmed means, bootstrapping, or robust standard errors.
- Proceed with parametric tests if sample size is large: Due to the Central Limit Theorem, means are approximately normal even for non-normal data when n is large (typically n > 30 or 40), so t-tests and ANOVA may still be valid.
Always consider the practical significance of the deviation and the robustness of your planned analysis.
Frequently Asked Questions
What does a p-value of 0.07 in a normality test mean?
It means that, assuming the data are normally distributed, there is a 7% chance of obtaining a test statistic as extreme as yours. Since 0.07 > 0.05, you fail to reject normality at the 5% significance level. The data do not show strong evidence of non-normality, but the result is marginal—consider graphical checks and the sample size.
Can I use a t-test if my data are not normally distributed?
Yes, if your sample size is large enough (usually n > 30) due to the Central Limit Theorem, the t-test is robust to non-normality. For small samples, prefer non-parametric alternatives like the Mann-Whitney U test. Always check for outliers and consider transformations.
Why does my normality test reject normality even though the histogram looks normal?
This often happens with large samples (n > 1000) where even tiny deviations from normality become statistically significant. The test detects a difference that may be practically irrelevant. Examine the Q-Q plot and effect size; if deviations are minor, you can still use parametric methods.
Which normality test is best for small samples?
Shapiro-Wilk is generally recommended for small samples (n < 50) because it has higher power than Kolmogorov-Smirnov or Anderson-Darling for detecting deviations in small data sets. For very small samples (n < 10), no test has much power, so rely on graphical methods and domain knowledge.
How do I report normality test results in a paper?
Report the test name, the test statistic, degrees of freedom if applicable, and the p-value, along with the conclusion. For example: “The Shapiro-Wilk test indicated that the data were normally distributed (W = 0.98, p = 0.34).” If you used transformations, mention them. Also include a Q-Q plot or histogram as visual evidence.
What is the difference between Shapiro-Wilk and Kolmogorov-Smirnov?
Shapiro-Wilk is based on the correlation between the sample and normal scores and is more powerful for small to medium samples. Kolmogorov-Smirnov compares the empirical CDF to the normal CDF and is less powerful for small samples but can be used for any sample size; however, it requires the mean and variance to be known unless the Lilliefors correction is applied. In practice, Shapiro-Wilk is preferred for n < 50.
Related
Formula used
in n | Sample size | - | 30 | 3 in w | Shapiro-Wilk W statistic | - | 0.95 | 0 in swp | Shapiro-Wilk p-value | - | 0.2 | 0 out dec | Decision at α = 0.05 | - | if(swp >= 0.05, "Fail to reject normality", "Reject normality") | 0
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 meq/mL Conversion Calculator1.0000 meq/mL
- Normality to g/L Converter & Formula29.220 g/L
- Normality Test Q-Q Plot: How to Read It
- Normality Slay the Spire?
- Normality Test for Paired t Test: How to Check
- Normality SI Unit & How to Calculate It0.4000 eq/L