Must a point cross three sigma for a signal?

No; sequence rules can signal earlier.

Does adding rules reduce false alarms?

Generally it increases them while increasing sensitivity to some changes.

AI Glossary / Statistics and probability / Intermediate

Western Electric rules

Definition

Western Electric rules are a set of pattern checks used with statistical process control charts to flag possible changes in a process. They examine unusually distant points and sequences on one side of the center line. Their purpose is to prompt investigation of potential special causes. A rule firing is evidence to examine, not proof of a particular cause or a guarantee that a product fails its specification. The chart and its underlying assumptions must be appropriate first.

What it means

A process can begin shifting before an individual observation crosses its outer control limits. Looking at recent sequences can make a chart more sensitive to smaller sustained changes. That sensitivity has a cost: additional rules also produce more false alarms. For an AI service, meaningful monitoring therefore involves both statistical signals and a response procedure. The rules do not themselves determine whether a change reflects a software release, a new user population, measurement error or ordinary chance.

How it works

A commonly stated four-rule version signals when one point exceeds three standard deviations from the center, two of three consecutive points exceed two standard deviations on the same side, four of five exceed one standard deviation on the same side, or eight consecutive points fall on one side of the center. Use the standard deviation of the plotted statistic. For example, the variability of subgroup averages differs from the variability of individual observations. Establish suitable baseline limits and apply the chosen rules consistently. Software may offer additional trend rules or different conventions, so document the configuration rather than relying on the name alone.

Three practical examples

Example 1: A stable manufacturing process charts subgroup averages and investigates a sequence suggesting an upward shift. Example 2: A service team monitors a suitably modeled daily performance statistic after accounting for predictable traffic changes. Example 3: A laboratory checks a measurement process for sustained movement away from its established center. Each application needs comparable observations and an appropriate chart. Raw latency data with strong seasonality or dependence may require a different model or chart before these rules are informative.

Think of it this way

Think of an instrument panel that reacts both to one large movement and to a succession of smaller movements in the same direction. The panel tells an operator where to look. It cannot identify the mechanical fault by itself, and making it more sensitive also means investigating more signals that have no special cause.

Worked example

Suppose a chart's established center is 100 and the standard deviation of its plotted statistic is two. The upper one-, two- and three-sigma boundaries are 102, 104 and 106; their lower counterparts are 98, 96 and 94. Three successive points of 104.6, 101 and 104.4 have standardized positions 2.3, 0.5 and 2.2. Two exceed the upper two-sigma boundary, so the two-of-three rule signals even though none exceeds three sigma. By contrast, 104.6, 101 and 95.6 include one point above two sigma and one below minus two sigma; these are not two points on the same side. Separately, eight observations of 100.2, 100.4, 100.6, 100.3, 100.8, 100.5, 100.7 and 100.1 trigger the same-side rule despite all lying within one sigma. Under an ideal known-parameter normal model, a single point exceeds either three-sigma boundary with probability about 0.00270. That is not the overall false-alarm probability for all rules combined.

Illustrative case study

In this fictional case study, Harbor Inspection charts averages from consistent production subgroups. A two-of-three signal appears after a maintenance adjustment. The team follows a documented investigation procedure, checking instrument calibration, subgroup composition and maintenance records before changing the process. It records the signal and its eventual explanation. The team also reviews its alert workload, because repeated unnecessary adjustments could introduce variation instead of reducing it. The case illustrates investigation, not a demonstrated commercial outcome.

Common misconceptions and mistakes

Common mistakes include combining points on opposite sides for a same-side rule, using individual-observation variability for a chart of averages, and confusing control limits with customer specifications. Another mistake is assuming that no signal proves stability or that any signal proves a defect. Correlation, changing sample sizes and poorly estimated baselines can alter performance. Western Electric and Nelson rule sets should not be silently mixed: the familiar eight-point same-side rule is not interchangeable with every software package's run criterion.

Business use

Businesses can use these rules to organize monitoring and investigation when a suitable process-control model exists. Choose the monitored statistic, baseline and response owner before enabling alerts. Record which rules operate, assess the cost of false alarms and missed changes, and investigate signals with operational context. Revise limits for justified process changes through a documented procedure. Constantly resetting the center to recent observations can hide the very changes the chart is intended to detect.

Check your understanding

What is the example's upper two-sigma boundary?

Answer: 104.

Why does the first three-point sequence signal?

Answer: Two points exceed 104.

Why does the opposite-side sequence fail that particular rule?

Answer: Its two extremes are on different sides.

Can eight modest positive deviations signal?

Answer: Yes, under the stated eight-point rule.

Learn next

Learn next how subgroup construction determines the variability of a plotted average. Compare a three-sigma-only chart with an explicitly configured rule set, examining both detection speed and investigation workload. Study dependence and baseline estimation before applying textbook false-alarm probabilities to operational data.

Sources

Source: NIST/SEMATECH Engineering Statistics Handbook, What Are Variables Control Charts?, https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc32.htm. This primary reference documents the Western Electric rule patterns, limits and false-alarm tradeoff. The numerical sequences and normal-tail calculation are original and independently checked. The inspection case is fictional.

Previous: Welch's t-test Next: Wetware computer