Measurement Data Statistics
Paste your measured values and get the mean, standard deviation, skewness and kurtosis, the real ±3σ and ±4σ widths of your data, capability indices and a normality check. Free, no sign-up, and your numbers never leave your device.
Histogram with the normal curve
Normal Q-Q plot
Run chart (values in the order entered)
How to read the results
Mean and standard deviation. The mean is the average. The standard deviation s measures the spread; the page shows the sample version (divides by n − 1), which is the usual estimate of the spread of the process the values came from, and the population version (divides by n).
Skewness. Zero for a symmetrical distribution. Positive: a longer tail to the right. Negative: a longer tail to the left. The page shows the plain moment form, g1 = m3 / m23/2, and the sample-corrected version used by spreadsheet programs.
Kurtosis. Three for a normal distribution. Larger than three: heavier tails and a sharper peak. Smaller than three: lighter tails (a uniform distribution has 1.8). The moment form is m4 / m22; subtract 3 to get the excess kurtosis.
Real ±3σ and ±4σ. For a normal distribution, 99.73% of the values lie within ±3σ of the mean and 99.9937% within ±4σ. The table reads the percentiles that hold the same shares directly from your data (0.135% and 99.865% for 3σ; 0.0032% and 99.9968% for 4σ). The result is the width of your real process at that level, whatever the shape. The “equivalent σ” is that width divided by 2k.
Fitted curve. The same flexible curve (a generalised lambda distribution) that the tolerance calculator uses for the distribution “Custom” is fitted to your mean, standard deviation, skewness and kurtosis. It is drawn dashed on the histogram, and its ±3σ and ±4σ values are listed next to the measured and the normal ones. Where your data has too few values to measure the far tails, the fitted curve gives an estimate. It is a model: it is only as reliable as the skewness and kurtosis it is built from, and it cannot copy shapes with two peaks.
Run chart and stability. The values are plotted in the order you entered them, with the mean and the control limits mean ± 3 × (average moving range / 1.128). Points outside the limits, long runs on one side of the mean, steady rises or falls and a straight-line drift are flagged. All the other statistics assume one stable process, so look at this chart first. If the values are not in production order, the stability check has no meaning.
Jarque-Bera test. Combines skewness and kurtosis into one number. A p-value below 0.05 means the shape differs from a normal distribution by more than chance would explain. The Q-Q plot shows the same thing visually: points on the line mean normal data.
Pp, Ppk, Cp and Cpk. You type the lower and upper limits (LSL, USL). Pp = (USL − LSL) / 6σ and Ppk = min(USL − mean, mean − LSL) / 3σ. With σ = s, the standard deviation of all the values, these are the long-term indices Pp and Ppk. With σ = the short-term sigma (average moving range / 1.128, from the differences between neighbouring values) the same formulas give Cp and Cpk. For a stable process they are nearly equal. If Cp/Cpk is much better than Pp/Ppk, the process drifts or shifts over time. Cp/Cpk need the values in production order.
Limits you should know
- Skewness and kurtosis are very uncertain for small samples. Use at least 50 values, better 200 or more.
- The tail percentiles need many values: about 3,700 for ±3σ and about 160,000 for ±4σ. The page marks rows with too few values.
- The calculations assume the values are independent and come from one stable process. Values from a drifting process, or from several mixed sources, give misleading results.
- Up to 1,000,000 values are used. Larger lists are cut off after the first million.
Next step: use the mean, standard deviation, skewness and kurtosis in the tolerance stack-up calculator (choose the distribution “Custom” for a line). The methods page lists the formulas and public sources, and the tutorial explains the statistical methods from scratch.