Lesson 5: Control chart anatomy, subgroups, and limit calculations

Green Belt Lesson 9 introduced the family of control charts and when to use each one. This lesson goes underneath the chart itself: what every control chart is built from, how the samples that feed it should be chosen, and exactly how its center line and limits are calculated, worked all the way through with real numbers for the two charts a Black Belt uses most, Xbar-R and I-MR, plus the p-chart limit formula.

The parts of a control chart

Every control chart, regardless of which statistic it plots, is built from the same three elements. The center line (CL) is the process average of the plotted statistic. The upper and lower control limits (UCL, LCL) sit at plus and minus three standard deviations of that statistic, not three standard deviations of the raw data, a distinction that matters because the plotted statistic (a subgroup average, a range, an individual value) has its own standard deviation. Between the center line and the limits, the space is divided into three zones, each one standard deviation wide: Zone C nearest the center line (within 1 sigma), Zone B (1 to 2 sigma), and Zone A (2 to 3 sigma), used by run-rule tests such as "2 of 3 consecutive points in Zone A or beyond, on the same side of the center line" to catch a process that is drifting before any single point actually crosses a limit.

Rational subgrouping

A control chart's limits are only meaningful if the variation used to calculate them is the variation the chart is meant to detect. Rational subgrouping is the practice of choosing which measurements go together in a subgroup so that the variation within a subgroup reflects only common-cause (routine, expected) variation, while any special cause shows up as variation between subgroups. In practice this usually means grouping measurements taken close together in time and under the same conditions: five consecutive units off one production run, not one unit from each of five different shifts. If a subgroup accidentally spans a shift change, a crew change, or a material lot change, that special cause gets baked into the within-subgroup variation, inflating the control limits and hiding the very signal the chart exists to catch.

At GCHS, a subgroup of five inspection turnaround times pulled from the same inspector on the same day is rationally subgrouped; five turnaround times, one from each of five different inspectors on five different days, is not, because it mixes inspector-to-inspector and day-to-day differences into what should be routine, within-subgroup noise.

Subgroup size, sampling frequency, and the impact of variation on limits

Two practical choices shape every chart before a single point is plotted:

• Subgroup size (n): a larger subgroup narrows the control limits (because the standard deviation of an average shrinks as n grows, proportional to 1/√n), making the chart more sensitive to small shifts, but it costs more to sample and delays the first point until n units are collected. A subgroup of n = 1 (no subgrouping at all) calls for an I-MR chart instead of an Xbar-R chart.
• Sampling frequency: sampling more often catches a shift sooner but costs more; sampling less often is cheaper but leaves a process out of control longer before anyone notices. The right frequency balances the cost of a missed shift against the cost of sampling.

Because the control limits are calculated directly from the observed within-subgroup variation (Rbar or the average moving range), any change to subgroup composition, size, or measurement conditions changes the limits themselves, not just the data being plotted against them; this is why subgrouping decisions are made deliberately, before data collection starts, rather than adjusted after the fact to make a chart look better.

Xbar-R chart: calculations

The Xbar-R chart pairs a chart of subgroup averages with a chart of subgroup ranges, and both sets of limits are built from Rbar, the average of the subgroup ranges, using tabulated constants that depend on subgroup size n:

Published tables round the last digit slightly differently (D4 for n = 3 is printed as 2.574 or 2.575, and for n = 5 as 2.114 or 2.115); the worked limits below come out the same either way.

Xbar chart: UCL = Xbarbar + A2 x Rbar, LCL = Xbarbar - A2 x Rbar. R chart: UCL = D4 x Rbar, LCL = D3 x Rbar (which is 0 for n = 2 through 6, since D3 is 0 across that whole range).

Worked example. A GCHS team tracks punch-list rework turnaround time (days) in 5 subgroups of n = 5. Subgroup averages: 8.2, 7.9, 8.4, 8.0, 8.5 (Xbarbar = 41.0 / 5 = 8.2). Subgroup ranges: 2.1, 1.8, 2.4, 1.9, 2.3 (Rbar = 10.5 / 5 = 2.1). Using A2 = 0.577 and D4 = 2.114 for n = 5:

Xbar chart: UCL = 8.2 + (0.577 x 2.1) = 8.2 + 1.212 = 9.41; LCL = 8.2 - 1.212 = 6.99

R chart: UCL = 2.114 x 2.1 = 4.44; LCL = 0 x 2.1 = 0

I-MR chart: calculations

When measurements arrive one at a time, with no natural subgroup (n = 1), an Individuals and Moving Range (I-MR) chart replaces Xbar-R. The moving range MR is the absolute difference between each consecutive pair of individual values, MR_i = |x_i - x_(i-1)|, and its average, MRbar, stands in for Rbar. NIST's derivation gives the individuals limits as Xbar ± 3 x (MRbar / d2), and since d2 for n = 2 (a moving range is always computed from 2 points) is 1.128, 3/1.128 rounds to 2.66, giving the commonly quoted form:

I chart: UCL = Xbar + 2.66 x MRbar; LCL = Xbar - 2.66 x MRbar

MR chart: UCL = 3.267 x MRbar; LCL = 0 (D3 is 0 for n = 2)

Worked example. Daily defect counts on a five-day punch list: 12, 15, 11, 14, 10 (Xbar = 62 / 5 = 12.4). Moving ranges: |15-12|=3, |11-15|=4, |14-11|=3, |10-14|=4 (MRbar = 14 / 4 = 3.5).

I chart: UCL = 12.4 + (2.66 x 3.5) = 12.4 + 9.31 = 21.71; LCL = 12.4 - 9.31 = 3.09

MR chart: UCL = 3.267 x 3.5 = 11.43; LCL = 0

p-chart limits

For attribute data, the proportion of defective units in samples that can vary in size, the p-chart plots p, the fraction defective in each sample, against limits built from the overall average fraction defective, pbar, and the sample size n:

UCL = pbar + 3 x √(pbar(1-pbar)/n); LCL = pbar - 3 x √(pbar(1-pbar)/n)

Because n can differ from sample to sample (for example, the number of homes inspected each week varies), the limits widen or narrow with each sample's own n rather than staying fixed, which is why a p-chart's limits are often drawn as a stair-step rather than a straight line. If a calculated LCL comes out negative, it is set to 0, since a fraction defective cannot be negative.

When to recalculate control limits

Control limits are not recalculated every time new data comes in; doing so would let the limits chase the data and mask the very shifts a chart is meant to detect. Limits are recalculated only when there is a documented, process reason to do so: after a verified special cause has been identified, removed, and enough new in-control data has been collected to re-estimate sigma; after a deliberate process change, such as a new inspection method, that is expected to shift the baseline; or on a planned periodic review (for example, annually) to confirm the original limits still reflect current common-cause variation. Recalculating limits just because a chart is running "out of control" too often, without addressing the underlying cause, defeats the purpose of having limits at all.