Does common-cause variation disappear when a process is in control?

No. Statistical control means the process exhibits only its stable background variation, not zero variation.

Is a stable process necessarily good enough?

No. Its consistent average or spread may still fail requirements.

Can common causes be reduced?

Yes. Changes to the process system can reduce their combined contribution, provided the improvement is demonstrated rather than assumed.

AI Glossary / Statistics and probability / Intermediate

Common cause

Definition

A common cause is a source of variation that forms part of the usual operation of a process. In statistical process control, common causes collectively produce the background variation observed when the process is stable. They are distinguished from special causes associated with identifiable departures from that routine pattern. This entry uses the process-improvement meaning, not the separate causal-inference meaning of a variable that causes two other variables.

What it means

Even a well-run process does not produce identical results every time. Small differences in materials, measurement, working conditions and routine execution can contribute to its usual spread. Common-cause variation is a property of the existing process system, not necessarily an individual worker's mistake. A process influenced only by common causes can be statistically stable while still performing poorly against customer requirements. Stability describes consistency over time; capability concerns whether that consistent performance meets specifications.

How it works

Define the process and measurement, collect observations in meaningful time order and establish an appropriate reference period. Use a suitable control chart or other stability analysis to look for evidence of changes beyond the expected pattern. Investigate signals using process knowledge rather than assuming that every extreme point has an obvious cause. When the process appears stable but its variation is unacceptable, improvement usually requires changing the underlying system, such as equipment, methods or input consistency. Confirm the effect of changes with new evidence and revise the baseline deliberately rather than continually moving limits to follow the latest observations.

Three practical examples

Service delivery: study routine variation in handling time before interpreting every slower case as an exceptional failure.
Measurement systems: quantify ordinary repeatability and environmental contributions to the spread of repeated readings.
AI operations: monitor a consistently defined quality or latency measure while distinguishing ordinary fluctuations from changes in model version, workload or infrastructure.

Think of it this way

Think of small ripples on a pond under steady conditions. They are part of the normal pattern, while a sudden disturbance may create a different pattern requiring investigation. Reducing the ordinary ripples may require changing the conditions themselves rather than reacting to each individual wave.

Worked example

Consider a hypothetical stable measurement process with two independent additive sources of routine variation. One has variance 9 and the other variance 16, in squared measurement units. Total variance is 9 + 16 = 25, so the total standard deviation is 5. The standard deviations do not add to 7. Suppose a process change reduces the first component's variance from 9 to 4 while the second remains 16. Total variance becomes 20 and standard deviation becomes the square root of 20, about 4.472. That is a 20% reduction in variance but only about a 10.56% reduction in standard deviation. The calculation assumes independent additive components; correlated sources introduce covariance terms. The numbers illustrate system-level improvement, not evidence that particular observed sources have been identified.

Illustrative case study

Fictional case: Fir Support reviews daily variation in assistant response time. Managers initially change routing settings after each slightly slower day. Analysts establish a consistent workload definition and examine the time-ordered results. They find no convincing evidence of a changing pattern during the reference period, yet typical response times still miss the service target. The team investigates a system-wide bottleneck rather than treating every daily fluctuation as a separate incident. It evaluates a proposed infrastructure change against the baseline. This invented case illustrates the distinction between stable variation and acceptable performance.

Common misconceptions and mistakes

Common cause does not mean harmless, unchangeable or too small to matter. A stable process can repeatedly produce unacceptable outputs. Conversely, a control-chart signal does not identify the physical cause by itself, and the absence of a signal does not prove that every relevant condition is unchanged. Control limits are not customer specification limits. Autocorrelation, changing case mix and measurement changes can make a naive chart misleading. Adjusting a stable process in response to ordinary noise can create additional variation, so intervention rules should reflect the process and evidence.

Business use

Businesses benefit from distinguishing routine system variation from unusual incidents because the two often require different responses. Incident investigation may address a special event, while persistent background variation calls for broader improvement. For AI services, preserve version, traffic and measurement context alongside monitoring data. Compare like with like and avoid interpreting normal workload differences as model deterioration. Choose measures that reflect user needs, then assess both stability and capability. Monitoring should support decisions about improvement rather than become a daily ranking of staff based on noisy fluctuations.

Check your understanding

Why does the worked example add 9 and 16 rather than 3 and 4?

Answer: Under the stated independence and additive model, variances add. Standard deviations are square roots of variances and generally do not add directly. Correlated components would require covariance terms as well.

Related terms

Special cause
Statistical process control
Control chart
Variance components
Control limits
Specification limits
process stability
process capability
measurement variation
process improvement

Learn next

compare control limits with specification limits and distinguish a process shift from a stable but inadequate baseline. Study rational subgrouping and variance components before selecting a monitoring chart. For AI operations, examine how case mix and serial dependence affect the interpretation of apparent performance changes.

Sources

NIST/SEMATECH, e-Handbook glossary, common cause: https://www.itl.nist.gov/div898/handbook/glossary.htm NIST/SEMATECH, Process Variability: https://www.itl.nist.gov/div898/handbook/ppc/section1/ppc132.htm NIST/SEMATECH, Assessing Process Stability: https://www.itl.nist.gov/div898/handbook/ppc/section4/ppc45.htm

Previous: Committee machine Next: Common data model