Are control limits confidence intervals?

They serve different purposes: control limits monitor process statistics, while confidence intervals quantify uncertainty about parameters.

Must limits be horizontal?

No; changing sample size or other modeled conditions can produce varying limits.

Can an improvement trigger a signal?

Yes. An unexpectedly favorable shift can also warrant investigation and confirmation before adopting a new baseline.

AI Glossary / Statistics and probability / Intermediate

Control limits

Definition

Control limits are statistical reference boundaries used in a control chart to identify observations or patterns that warrant investigation under an assumed stable-process model. They are calculated for the specific statistic being monitored, using baseline information or justified reference parameters. They describe expected process variation for monitoring purposes; they are not customer specifications, operational targets, or guarantees that all future observations will remain inside them.

What it means

An upper control limit and lower control limit surround a reference center on many charts. A point outside a limit may indicate that the process has changed, but it can also occur by chance under the model. Some charts use one-sided limits, and additional rules can signal unusual sequences even when individual points remain inside the boundaries.
The limits answer a stability question. A customer tolerance answers whether an output meets a requirement. A process can remain within its control limits while producing unacceptable outputs, or generate a control-chart signal while individual outputs still meet specifications.

How it works

Choose the monitored statistic and a model for its variation. A chart of individual values, subgroup means, defect proportions, or defect counts needs limits appropriate to that quantity. For some charts, limits are a center plus or minus a chosen multiple of the statistic's standard deviation. Other charts use probability-based or specialized limits.
Estimate parameters from a suitable baseline and investigate evidence that the baseline itself is unstable. Avoid combining different processes indiscriminately. If sample sizes change, the variability of the plotted statistic often changes too, so its control limits may need to vary across observations.
For a proportion chart with a stable reference proportion p and subgroup size n, conventional three-sigma limits are p plus or minus 3 sqrt[p(1-p)/n], bounded to the feasible interval from zero to one. These are approximate limits derived from a binomial variance, not exact probability boundaries in every case.
Select signal rules and understand their false-alarm and detection properties. Parameter estimation, skewness, small counts, dependence, and extra pattern rules affect those properties. Treat a signal as a reason for investigation rather than a verified diagnosis.
Preserve limits during prospective monitoring unless there is a documented reason to revise the baseline or model. Recalculating them merely because an unwelcome point appears can normalize the very change that should be investigated. Keep the calculation method and baseline period visible in documentation.

Three practical examples

A quality team uses sample-size-specific limits for the proportion of inspected items that fail a criterion.
A service team distinguishes statistical limits for average handling time from its contractual response target.
An AI monitoring team revisits limits after a verified change in the evaluation population rather than silently mixing the old and new baselines.

Think of it this way

Think of control limits as reference lines on a process's behavior map. They mark observations that look unusual relative to the map, not a fence that reality cannot cross. A separate set of requirements determines whether the observed performance is acceptable to the people using the process.

Worked example

Suppose a simplified binomial monitoring model uses a known stable error proportion p = 0.04. For subgroups of one hundred independently assessed items, the standard deviation of the sample proportion is sqrt(0.04 x 0.96/100), approximately 0.019596.
The conventional upper limit is 0.04 + 3 x 0.019596 = 0.098788, or about 9.88%. The unbounded lower calculation is negative, so the lower limit is set to zero. Ten errors among one hundred items give a plotted proportion of 10%, above that upper limit.
For subgroups of four hundred, the standard deviation falls to approximately 0.009798. The limits become about 0.010606 and 0.069394, or 1.06% and 6.94%. An 8% error proportion is inside the one-hundred-item limits but above the four-hundred-item upper limit. Larger subgroups give more precise proportion estimates, making the same departure more conspicuous.
These limits illustrate the formula, not an exact 0.27% false-alarm probability. The counts are discrete, the first subgroup has a small expected error count, and a real baseline proportion would usually be estimated. Exact binomial probability limits or another suitable method may be preferable for a particular monitoring requirement.

Illustrative case study

Fictional case: Fern Processing reviews a dashboard that uses the same percentage limits every day despite large differences in the number of reviewed cases. Analysts show that the plotted proportions do not have equal sampling variability.
The team introduces a documented sample-size-aware monitoring method, checks whether cases are independent enough for the chosen model, and keeps the service target separately labeled. It also establishes a review process for signals. Managers learn that being inside the statistical limits is not evidence that the error rate is good enough, and that a signal calls for investigation rather than automatic blame.

Common misconceptions and mistakes

Do not call control limits the maximum permissible product values. Do not substitute the standard deviation of individual observations into a chart of means without accounting for subgroup size. A three-sigma formula does not guarantee the same tail probability for every distribution. Limits estimated from a short or unstable baseline can be unreliable, while excessive retrospective adjustment can conceal persistent process changes.

Business use

Clear control limits make monitoring decisions reproducible. Record the baseline, chart type, parameter estimates, subgroup sizes, and response rules. Explain the distinction between stability and business acceptability to dashboard users. Reassess the model when the process, measurement system, or case mix changes, and evaluate whether the chosen limits provide a useful tradeoff between false alarms and delayed detection.

Check your understanding

Why are the four-hundred-item limits narrower?

Answer: The sample proportion has less variability with the larger subgroup under the model.

Does a 9% result meet customer requirements because it is inside one set of limits?

Answer: That cannot be determined without the requirements.

Should a surprising point automatically be removed?

Answer: No; investigate and document it.

Related terms

Control chart
Statistical process control
Specification limits
upper control limit
lower control limit
subgroup size
false alarm
process stability

Learn next

Compare three-sigma and exact binomial limits for small expected counts. Then examine how estimated baselines, autocorrelation, and additional pattern rules change the performance of a monitoring procedure.

Sources

NIST/SEMATECH e-Handbook, Proportions Control Charts: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc332.htm NIST/SEMATECH e-Handbook, What are Control Charts?: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc31.htm

Previous: Control group Next: Control theory