Variable Control Charts: Means and Spread
NIST begins its mean-and-standard-deviation procedure by examining the standard-deviation chart for stability.

| Evidence or term | What it means here |
|---|---|
| Subgroup mean | The location statistic for a defined subgroup. |
| Range or standard deviation | Spread information used with the chosen subgroup design. |
| Control limits | Historical-process reference limits, distinct from product specification limits. |
What does the evidence explain?
Why examine subgroup spread before interpreting the mean chart?
NIST begins its mean-and-standard-deviation procedure by examining the standard-deviation chart for stability. The two plots describe different aspects of the observations: the subgroup average describes location, while the subgroup standard deviation describes dispersion. A mean plot alone does not show whether the within-subgroup variation used to establish its limits has remained stable.
In this setup, m preliminary subgroups each contain n observations. Calculate a standard deviation for each subgroup, then average those m standard deviations to obtain s-bar. The quantity is an average of subgroup standard deviations, not the standard deviation of all measurements after discarding their subgroup identities. Retaining the grouping is therefore part of reproducing the stated calculation.
Why does the standard-deviation estimate use c4?
The usual sample variance divides the sum of squared deviations from the sample mean by n − 1. Its unbiasedness as an estimate of population variance does not make its square root an unbiased estimate of population standard deviation. Under the normal-distribution model discussed by NIST, the expected sample standard deviation is c4 times the population standard deviation; c4 depends on subgroup size.
With the convention used here, the estimate of process standard deviation is s-bar divided by c4. The handbook warns that some sources define the reciprocal factor instead. Before importing a constant into a spreadsheet, check both the definition and whether the formula multiplies or divides by it. A number labelled c4 without its convention is insufficient to reproduce the estimate.
What determines the limits of the s chart?
In the stated equal-subgroup-size normal model, the s-chart center is s-bar. Its three-sigma limits are s-bar ± 3(s-bar/c4)√(1 − c4²). The square-root term concerns the variability of sample standard deviations. It is not the standard error of the subgroup mean, and substituting a mean-chart formula changes the statistic being assessed.
The center and limit calculation use the same preliminary subgroup design. Thus the subgroup size used to select c4 must correspond to the observations used to calculate each s. This formula is presented for understanding the source method; it does not establish an approved chart for a supplier or supply a product-specific limit. Any operational implementation must also handle the nonnegative support of a standard deviation correctly.
Why is the mean-chart width divided by the square root of subgroup size?
Using the source’s s-based estimate, the mean-chart center is the grand mean, and the limits are grand mean ± 3s-bar/(c4√n). The grand mean is the mean of all observations in this equal-size design. The estimated process standard deviation is s-bar/c4; the further division by √n produces the scale used for the subgroup average.
For a fixed estimate of process spread, a larger subgroup gives narrower limits for its average. That change concerns averages, not tighter individual-product tolerances. Do not take the mean-chart interval and label it the permitted range for every individual item. Nor should n be replaced with the total number of measurements across all preliminary subgroups: n is the number of observations contributing to each plotted mean.
When does the handbook use ranges instead of standard deviations?
For relatively small subgroups, NIST describes using the range as the dispersion statistic. A subgroup range is its largest observation minus its smallest. The handbook regards the range approach as satisfactory up to roughly ten observations per subgroup and prefers subgroup standard deviations for larger groups. This is a choice of statistical summary, not a required inspection sample size.
The source compares statistical efficiency: for n = 5 the listed relative efficiency is 0.955, while for n = 10 it is 0.850. Those values compare the range-based estimator with s/c4 under the stated model. They are not production yields, inspection coverage percentages or probabilities of accepting a batch. The comparison explains why convenience and information retention can lead to different choices as subgroup size increases.
How does an average range estimate process spread?
For normally distributed observations, the expected range of a subgroup is d2 times the process standard deviation. The factor d2 depends on the subgroup size. If the preliminary ranges are R1 through Rk, their arithmetic average is R-bar, and the source estimates process standard deviation by R-bar/d2.
Neither R-bar alone nor the largest observed range is the same estimator. Also distinguish k, the number of preliminary subgroups in this notation, from n, the number of observations in each subgroup. A report that gives only an average range without subgroup size leaves the conversion factor unresolved. This statistical estimate describes spread within the selected design; it is not a measurement uncertainty statement or evidence that a particular lot passed inspection.
How is the A2 factor matched to the subgroup?
For the range-based mean chart, NIST defines A2 = 3/(d2√n). The limits become grand mean ± A2R-bar. Its table gives A2 = 0.729 for n = 4 and 0.577 for n = 5. Selecting the row therefore requires the number of observations in each subgroup, not the number of subgroups or the number of units in the shipment.
For an arithmetic illustration only, let the grand mean be 50 and R-bar be 2 in a chosen measurement unit, with n = 5. Using the tabulated 0.577 gives mean-chart limits of 48.846 and 51.154. These invented calculation inputs describe no actual product, supplier or permissible tolerance. They demonstrate how the subgroup-size factor changes the width of the mean-chart limits.
What does the companion range chart calculate?
The R-chart center is R-bar. Its limits use D3R-bar and D4R-bar, with factors selected by subgroup size. The source derives the factors through the standard deviation of the range distribution, which is d3 times the process standard deviation. This is why the range plot uses a different pair of factors from the mean plot even when both come from the same measurements.
At n = 5 the source table gives D3 = 0 and D4 = 2.115. Continuing the labelled arithmetic illustration with R-bar = 2 gives a lower range limit of 0 and an upper limit of 4.230. The zero lower factor is a feature of the tabulated small-subgroup chart, not proof that zero observed variation is normal for every instrument. Retain the measurement resolution and raw observations when investigating such a result.
Why can a short estimated baseline alter chart performance?
NIST cautions that published false-alarm properties commonly assume known process parameters. In practice the mean and spread are often estimated from preliminary data. With a limited baseline, actual chart properties can differ from the nominal properties calculated as if the parameters were known. The estimated limits have an evidence basis that should remain distinguishable from fixed known parameters.
The handbook quotes a historical rule of thumb involving twenty-five samples of four observations under essentially unchanged conditions. It is not a universal contractual inspection plan, a regulatory requirement or a guarantee of statistical control after one hundred readings. Record what preliminary observations actually support the limits; do not turn an example of baseline development into acceptance of a shipment or a substitute for the agreed lot-sampling plan.
When can recalculating limits obscure the comparison?
A control chart compares current behavior with a historical basis. NIST warns that frequently changing its limits defeats that comparison. It describes reasons for revision such as substantially more unchanged-process data, a major process change, or a known event affecting the process or tool. These are reasons to assess the baseline, not an instruction to refit limits whenever an unwelcome point appears.
For a continuing record, distinguish observations assessed under the previous limits from those assessed after a justified revision. Keep the previous baseline and the reason for the change available. Otherwise a later reader cannot tell whether apparent improvement came from changed process behavior or simply a changed reference. The source gives an example involving thirty additional points; that example is not a universal minimum for every chart or process.
Which additional patterns does the displayed WECO scheme flag?
The handbook displays same-side rules including two of three successive points beyond two sigma, four of five beyond one sigma, and eight consecutive points on one side of the center line. These rules examine ordered patterns relative to chart zones. They cannot be reproduced from an unordered list of values or a single average for the reporting period.
The same display also lists six successive points trending up or down and fourteen alternating up and down. If a report uses supplementary rules, identify the selected rule set and preserve the observation order. Different pattern definitions are different monitoring procedures. A flag identifies a reason to investigate the sequence; it does not establish which machine, material or operator caused the pattern.
Does average run length promise a detection deadline?
Average run length, or ARL, expresses the average number of plotted points before a specified signal in the stated situation. For the basic normal-model three-sigma rule with no process change, the source uses a per-point signal probability of 0.0027 and ARL approximately 371. The reciprocal calculation is 1/0.0027, approximately 370.37; the source’s whole-number description is approximate.
An average is not a deadline by which every change will be found, nor a count of individual products when each point represents a subgroup. NIST separately asks how rapidly the chart detects specified mean shifts. That shifted-process performance is not obtained by reusing the unchanged-process false-alarm ARL. Reporting an ARL without the signal rule, process condition and plotted statistic leaves its operational meaning incomplete.
What trade-off accompanies supplementary run rules?
Supplementary rules can improve sensitivity to patterns while increasing false alarms. NIST contrasts the basic Shewhart rule’s approximate 371-point average false-alarm interval with an approximate 91.75-point interval after adding the discussed WECO rules. These are model-based descriptions of specified monitoring procedures, not guarantees for arbitrary production records or estimated limits.
A chart review should therefore ask which procedure generated the signal before comparing alarm rates between periods or suppliers. More flags may reflect a more sensitive rule set rather than worse process performance. Conversely, removing rules changes what is being monitored. The practical choice involves investigation effort as well as sensitivity; it does not alter the buyer’s specification or authorize release of nonconforming goods.
FAQ
Why examine subgroup spread before interpreting the mean chart?
NIST begins its mean-and-standard-deviation procedure by examining the standard-deviation chart for stability. The two plots describe different aspects of the observations: the subgroup average describes location, while the subgroup standard deviation describes dispersion. A mean plot alone does not show whether the within-subgroup variation used to establish its limits has remained stable.
What trade-off accompanies supplementary run rules?
Supplementary rules can improve sensitivity to patterns while increasing false alarms. NIST contrasts the basic Shewhart rule’s approximate 371-point average false-alarm interval with an approximate 91.75-point interval after adding the discussed WECO rules. These are model-based descriptions of specified monitoring procedures, not guarantees for arbitrary production records or estimated limits.
Sources and scope
Sources: NIST/SEMATECH 6.3.2 and 6.3.2.1, variables charts with their subgroup-size and distribution conditions. Numerical arithmetic illustrations are labelled and are not supplier results or product limits. No charts, linked data or third-party calculator results are reproduced. NIST PMC32 NIST PMC321