Is a special cause always harmful?

No; an unusual beneficial change can also qualify.

Does a point outside a control limit identify its cause?

No; it is a signal for investigation.

Can a stable process be unsatisfactory?

Yes; its routine performance may not meet requirements.

Should a confirmed lasting improvement use the old limits forever?

Not necessarily; a justified new baseline may be needed.

AI Glossary / Statistics and probability / Intermediate

Special cause

Definition

A special cause is a source of process variation associated with a particular circumstance or change beyond the routine common-cause variation of the process being studied. It is also called an assignable cause. Examples can include a changed material batch, an equipment fault or a specific alteration in operating conditions. A statistical signal may prompt a search for such a cause, but the signal itself does not identify the cause or prove that one has occurred.

What it means

Processes vary even when nothing unusual has happened. Common causes belong to the ordinary system; special causes indicate circumstances that differ from that established pattern. This distinction helps determine whether to investigate a specific event or improve the broader process. It is relative to the defined process and baseline. A condition that was unusual can become routine after a deliberate change, so the classification is not an inherent moral judgment about a person, event or measurement.

How it works

Establish an appropriate baseline and monitoring method for the process, including the measurement system, sampling scheme and charted statistic. Apply prespecified control-chart rules or another justified signal procedure. When a signal occurs, investigate relevant records and operating conditions, verify the measurement and seek evidence for a specific explanation. If a cause is confirmed, address it and assess the result. Re-establishing limits after a lasting process change should follow a reasoned procedure rather than automatically absorbing every unusual observation into a new baseline.

Three practical examples

Example 1: A production measurement shifts after a worn tool is replaced incorrectly. Example 2: An AI service's response-time pattern changes when a particular deployment alters its request routing. Example 3: A data-quality monitor signals a sudden change after an upstream field definition is modified. These circumstances are candidate special causes, not conclusions obtained solely from a chart. The investigation must connect the observed change to supporting evidence and distinguish genuine process change from measurement or sampling changes.

Think of it this way

Think of a familiar background hum interrupted by a new rattle. The hum represents the ordinary system's variability, while the rattle prompts a search for something different. Hearing it does not immediately tell you which component changed. The useful response is a focused investigation, followed by action supported by what is found, rather than adjusting every component whenever the sound varies slightly.

Worked example

Suppose individual measurements from an established baseline are independent normal observations with known mean 100 and standard deviation two. A simple three-standard-deviation rule gives limits 94 and 106. A new value of 107 is 3.5 standard deviations above the mean and triggers the rule. Under the unchanged baseline model, the probability of any one observation falling outside the three-standard-deviation limits is approximately 0.0027. Across 100 independent observations, the probability of at least one such signal is 1 minus (1 minus 0.0027) to the power 100, approximately 0.237. Thus a signal warrants investigation but is not infallible proof of a special cause. The calculation assumes known parameters and independent normal individual measurements; estimated limits, other distributions and additional rules change the operating behavior.

Illustrative case study

In this fictional case study, Maple Support monitors an AI service's response times using a chart chosen for its data. A signal appears shortly after a routing update. Engineers check workload mix, timestamps and infrastructure records before attributing it to the deployment. They find a configuration change that affects a subset of requests, correct it and continue monitoring. The incident report separates the initial statistical alert from the later evidence identifying the cause, making both the investigation and the corrective action reviewable.

Common misconceptions and mistakes

Common mistakes include blaming an individual whenever a chart signals, treating every outlier as a confirmed special cause, and assuming all points inside control limits prove stability. Patterns within the limits can also be relevant under the chosen rules. Another error is changing the process in response to every ordinary fluctuation, which can add unnecessary variation. Special causes can produce improvements as well as deterioration. A favorable unexpected shift should be understood before it is adopted as a sustainable improvement.

Business use

Businesses use the distinction to direct investigation and avoid unproductive reactions to routine noise. A useful monitoring plan names who reviews a signal, which records are checked and how findings are documented. It also recognizes alert burden: monitoring many metrics or applying many rules changes the frequency of false alarms. Control limits describe a baseline process, while specification limits describe requirements. A process can remain statistically stable and still fail to meet customer needs, requiring systematic improvement rather than a search for a single exceptional incident.

Check your understanding

What are the illustrative limits in the example?

Answer: 94 and 106.

Does 107 trigger the stated rule?

Answer: Yes.

Why is a signal not absolute proof?

Answer: Even the unchanged baseline can generate a value outside the limits.

What distinguishes the initial alert from a confirmed special cause?

Answer: Additional investigation that supports a specific explanation.

Related terms

Control chart
Statistical process control
Related terms include common cause
assignable cause
process stability
root-cause investigation and false alarm

Learn next

Learn next how chart choice depends on the measured quantity and sampling structure. Compare signals from individual values, subgroup averages and attribute counts. Then examine the consequences of autocorrelation, multiple monitoring rules and baseline estimation, and plan how a confirmed process change will be incorporated into future monitoring.

Sources

Sources: NIST/SEMATECH Engineering Statistics Handbook, Glossary, https://www.itl.nist.gov/div898/handbook/glossary.htm; What are Control Charts?, https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc31.htm. These primary institutional references distinguish common and special causes and discuss chart signals. The normal-probability calculation is original and independently checked; the service incident is fictional.

Previous: Spearman rank correlation Next: Specialization