Normality Test Calculator (Anderson–Darling & Shapiro–Wilk)

Paste or upload a column of measurements to test whether it follows a normal distribution. You get the Anderson–Darling and Shapiro–Wilk statistics and p-values, skewness and kurtosis, a normal probability plot with a confidence band, a histogram with the fitted curve — and practical advice when the data are not normal.

Quality engineeringFreeNo sign-up

More examples:

How to use the normality test calculator

  1. Enter the data. Paste a column from Excel, type the values, or upload a .csv / .xlsx file. A header row is skipped automatically. If the file has several columns you can test one column or all cells together.
  2. Choose α. 0.05 is the usual significance level. Use 0.01 if you want to act only on strong evidence, 0.10 to be more sensitive.
  3. Read the verdict, then the plot. The p-values tell you whether a departure from normality is statistically significant; the probability plot tells you what kind of departure it is and whether it matters.

What the tests do

Both tests start from the same null hypothesis, H₀: the data come from a normal distribution with unknown mean and standard deviation. A small p-value (below α) means data like yours would be unlikely under H₀, so normality is rejected. A large p-value does not prove normality — it only means no departure was detected with the data available.

Anderson–Darling

The Anderson–Darling statistic measures the distance between the empirical distribution of the sorted data and the fitted normal distribution, with extra weight in the tails — exactly where capability and PPM predictions are made:

A² = −n − (1/n) · Σ (2i − 1) · [ln F(x₍ᵢ₎) + ln(1 − F(x₍ₙ₊₁₋ᵢ₎))]

where F is the normal CDF with the sample mean and standard deviation. Because those two parameters are estimated, the statistic is adjusted for sample size, A*² = A² · (1 + 0.75/n + 2.25/n²), and the p-value is taken from the D’Agostino & Stephens (1986) approximation. This is the test used by default in most SPC software and referred to in the AIAG manuals.

Shapiro–Wilk

The Shapiro–Wilk statistic is essentially the squared correlation between the ordered data and the values expected from a normal sample:

W = (Σ aᵢ · x₍ᵢ₎)² / Σ (xᵢ − x̄)²

W is close to 1 for normal data and falls as the probability plot bends. This calculator implements Royston’s algorithm AS R94 (1995), valid for 3 ≤ n ≤ 5000 — the same algorithm used by R and SciPy. Shapiro–Wilk is generally the most powerful general-purpose normality test for small and medium samples; Anderson–Darling is particularly sensitive in the tails. Reporting both guards against the blind spots of either.

Skewness and kurtosis

StatisticNormal valueWhat a departure looks like
Skewness G₁0Positive: long tail to the right (flatness, runout, cycle times). Negative: long tail to the left.
Excess kurtosis G₂0Positive: heavy tails or outliers. Strongly negative: flat or two-peaked data, often two mixed streams.

The values are the bias-adjusted sample statistics (the same as Excel’s SKEW and KURT). Each is shown with its standard error; values beyond about ±2 standard errors are worth a look.

Reading the normal probability plot

PatternMeaningTypical cause
Points on the line, inside the bandConsistent with normalStable process, symmetric characteristic
Smooth curve (banana shape)SkewedCharacteristic bounded at zero or at a physical limit
S-shapeTails too heavy or too lightOutliers; sorted or truncated data
Two parallel segments with a jumpBimodalTwo cavities, machines, shifts or lots mixed
Vertical stacks of pointsTiesGauge resolution too coarse
One or two points far off the endOutliersMeasurement error or special cause

Worked example

The built-in example “Flatness (skewed)” contains 40 flatness readings in millimetres with mean 0.0203 and standard deviation 0.0129.

Anderson–Darling: A² = 1.906, adjusted A*² = 1.906 × (1 + 0.75/40 + 2.25/40²) = 1.945. Since A*² ≥ 0.6, p = exp(1.2937 − 5.709 × 1.945 + 0.0186 × 1.945²) ≈ 0.00006 (displayed as “< 0.0001”).

Shapiro–Wilk: W = 0.8445, p ≈ 0.00007. Skewness is 1.62 and excess kurtosis 2.72.

Both tests reject normality at α = 0.05 and the probability plot is curved — as expected, because flatness cannot be negative. A normal-theory Cpk would be misleading here; a transformation or a lognormal / Weibull model is the right approach.

What to do when data are not normal

  1. Check for outliers and data errors. One mistyped value can fail a test on its own.
  2. Check stability. A process that shifted during sampling is a mixture of distributions. Plot the readings in time order on an I-MR chart.
  3. Look for mixed streams. Separate cavities, spindles, fixtures and suppliers with the histogram maker and test each stream.
  4. Check the measurement resolution. Heavily rounded data fail normality tests even when the process is normal.
  5. Use a suitable model. For naturally skewed characteristics use a Box–Cox or Johnson transformation, or fit a lognormal or Weibull distribution, and compute capability with the percentile method instead of the normal formulas in the process capability calculator.

Sample size matters in both directions: below about 20 values the tests rarely reject anything, and above a few thousand they reject for departures too small to matter. In both cases the plot deserves more weight than the p-value.

Frequently asked questions

How do I interpret the p-value of a normality test?

The null hypothesis is that the data come from a normal distribution. If the p-value is below your significance level (usually 0.05), reject normality: the data show a statistically significant departure. If the p-value is above it, you cannot reject normality, which means the data are consistent with a normal distribution but does not prove it, especially with a small sample.

Should I use Anderson–Darling or Shapiro–Wilk?

Shapiro–Wilk is generally the most powerful general-purpose normality test for small to medium samples. Anderson–Darling gives extra weight to the tails, which matters for capability and PPM estimates, and is the default in most SPC software. They usually agree; when they do not, the data are borderline and the probability plot should decide. This calculator reports both.

How many data points do I need for a normality test?

Shapiro–Wilk can be computed from 3 values and Anderson–Darling p-values are reliable from about 8, but tests on fewer than about 20 values have very little power and will pass almost anything. For capability studies, 50 to 100 values give a meaningful test. With several thousand values the tests become oversensitive and reject for trivial departures.

Why do my data fail the normality test when the histogram looks bell-shaped?

Common reasons are one or two outliers, heavily rounded measurements that create many tied values, a very large sample in which tiny departures become statistically significant, or slightly heavy tails that a histogram hides. The normal probability plot shows which of these applies far more clearly than a histogram.

What are skewness and kurtosis for normal data?

Both are zero for a normal distribution when kurtosis is expressed as excess kurtosis. Positive skewness means a longer right tail, negative skewness a longer left tail. Positive excess kurtosis means heavier tails than normal; strongly negative values indicate a flat or two-peaked distribution, often caused by mixing two process streams.

Can I calculate Cpk if my data are not normal?

Not with the standard formulas, because Cp, Cpk and the expected PPM assume a normal distribution. First look for an assignable reason such as outliers, instability or mixed streams. If the characteristic is naturally non-normal, transform the data (for example Box–Cox) or fit another distribution and use the percentile method, or report the observed fraction non-conforming.