Wait, what? A process can produce a value that is perfectly acceptable and still be drifting toward failure. It can also produce an alarming-looking value and still be behaving exactly as it always has.

That is the problem statistical process control, or SPC, is built to solve. It does not begin by asking, “Is this number good or bad?” It begins with a harder question: “Is the process behaving like the same process?”

The one-sentence answer

Statistical process control uses data collected in time order to distinguish routine process variation from evidence that the underlying process has changed, so people can investigate the right signals instead of reacting to every fluctuation.

Quick Read

• Control charts are about process behaviour, not pass/fail inspection.
• Control limits are estimated from process behaviour; specification limits come from requirements.
• Common-cause variation belongs to the current process; special-cause variation suggests something has changed.
• Shewhart charts are good at detecting relatively large shifts; CUSUM and EWMA charts can be more sensitive to smaller sustained shifts.
• A process can be statistically stable but incapable of meeting requirements.
• A process can appear unstable because the chart model is wrong: autocorrelation, seasonality, bad subgrouping, changing measurement systems and data errors all matter.
• SPC is strongest when a signal triggers investigation, not automatic blame.

1. Why ordinary “good number / bad number” thinking fails

Suppose a delivery operation normally takes between 43 and 57 minutes. Today one delivery takes 59 minutes. Is something broken?

Not necessarily. A single value does not tell us whether the system changed. The useful question is whether the observed value is plausible under the process that generated previous values. A process can fluctuate because of many small influences: traffic, workload, operator sequence, temperature, queue length, measurement noise and dozens of other ordinary causes. If we treat every deviation as a crisis, we create what W. Edwards Deming famously warned against: tampering with a stable system.

SPC therefore adds time, context and expected variation to the decision.

2. The central distinction: common cause versus special cause

A stable process does not produce identical outputs. It produces a distribution of outputs. Some variation is built into the system as it currently exists. This is often called common-cause variation. If we want materially less of it, we usually have to change the system itself: equipment, workflow, training, environment, sequencing, inputs, software, incentives or design.

Special-cause variation is different. It points to a specific departure from the process’s normal pattern: a sensor fault, machine wear, a new supplier batch, a software release, an exceptional weather event, a staffing disruption, a mis-set parameter or another identifiable change.

This distinction matters because the remedies are different. If common-cause variation is mistaken for a special cause, people chase random noise. If a genuine special cause is mistaken for routine variation, the process may continue to deteriorate.

3. What a control chart actually contains

A basic control chart places observations, subgroup summaries or counts in time order. A centre line represents the process location. Upper and lower control limits describe a range that would be expected under the assumed stable process model.

The famous “three sigma” convention is not a magic law. It is a practical signal rule derived from assumptions about the chart statistic. Its purpose is to balance two competing costs: reacting too often to noise and reacting too slowly to real change.

The chart is therefore not a decorative graph. It is a sequential decision instrument.

4. Control limits are not specification limits

This is one of the most important ideas in the whole subject.

• Control limits describe what the process is statistically expected to do if its behaviour has not changed.
• Specification limits describe what a customer, engineer, regulation, contract or design requires.

A stable process can sit entirely outside specification. An unstable process can still produce many in-specification items. Stability and capability are related, but they are not the same question.

This is why inspecting individual values against a requirement cannot substitute for monitoring the process that generates them.

5. Phase I and Phase II: building the baseline, then monitoring it

NIST distinguishes two broad stages. In Phase I, historical data are examined to understand the process, identify obvious special causes, check assumptions and estimate initial limits. In Phase II, those limits are used for prospective monitoring.

This sounds simple, but it prevents a common mistake: calculating limits from a contaminated period and then treating those limits as if they represented normal behaviour. A baseline should be defensible. If the process changed halfway through the baseline period, one set of limits may hide two different systems inside one average.

6. Rational subgrouping: the hidden design decision

Many control charts do not plot every raw observation. They plot subgroup averages, ranges, standard deviations or proportions. How observations are grouped matters because the chart is trying to separate within-subgroup variation from between-subgroup change.

A rational subgroup contains observations produced under conditions that are meaningfully similar. For a manufacturing line, that might mean items produced close together in time. For a service process, it could mean cases handled under the same shift conditions. Poor subgrouping can conceal a shift or manufacture a false one.

Subgrouping is therefore not clerical. It encodes a theory of how the process varies.

7. The main families of control charts

Shewhart charts

Shewhart charts compare the current statistic with control limits. They are transparent, fast to read and effective for relatively large changes. Common forms include X-bar and R or S charts for subgrouped continuous measurements, and individuals/moving-range charts when observations arrive one at a time.

Attribute charts

When the outcome is a count or proportion rather than a continuous measurement, charts such as p, np, c and u charts may be appropriate. The statistical model must match how opportunities and denominators behave; otherwise the limits can be misleading.

CUSUM charts

A cumulative sum chart accumulates evidence over time. Small deviations that would look unremarkable individually can add up into a clear signal. This makes CUSUM useful when sustained small shifts matter.

EWMA charts

An exponentially weighted moving average gives the newest observation the greatest weight while retaining diminishing information from older observations. ISO 7870-6:2024 specifically addresses EWMA charts for detecting small to moderate shifts in the process mean.

No chart is “best” in isolation. The right chart depends on the data type, sampling design, plausible shift, response cost, process dynamics and decision horizon.

8. Why runs, trends and patterns matter

A point beyond a control limit is one kind of signal. Some organisations also use supplementary rules for long runs on one side of the centre line, monotonic trends, repeated near-limit points or other unlikely patterns.

But every extra rule increases sensitivity and increases the chance of false alarms. A mature SPC system does not collect rules like trophies. It chooses rules whose operating characteristics fit the process and the cost of missing or chasing a signal.

9. Autocorrelation: when yesterday leaks into today

Classical chart calculations often rely on observations or subgroup statistics behaving approximately independently. Real systems may not. A furnace temperature, network latency series, patient-flow queue or inventory level can be strongly autocorrelated.

NIST work on autocorrelated quality-control data shows why this matters: conventional charts can produce misleading signals when serial dependence is ignored. Possible responses include modelling the time dependence first and charting residuals, choosing a chart designed for dependent data, or redesigning sampling.

That creates an important bridge to time-series analysis. A control chart is not outside time-series reasoning; it is one specialised way of turning time-ordered variation into operational decisions.

10. Process capability: stable does not mean good enough

Once a process is reasonably stable, capability analysis asks whether its output distribution fits within required tolerances. Indices such as Cp and Cpk compress that relationship into summary numbers.

Those indices are useful only when their assumptions are understood. A capability index built from an unstable process, a badly estimated standard deviation, a nonrepresentative sample or a badly specified tolerance can produce a precise-looking answer to the wrong question.

This is a recurring lesson across statistics: measurement, sampling and model validity come before arithmetic.

11. SPC is not only for factories

The word “process” is broader than a production line. A process is any repeatable system that generates outcomes over time: order fulfilment, call handling, software deployment, laboratory turnaround, defect discovery, ticket resolution, energy use, data-pipeline latency, educational marking turnaround or administrative processing.

The same discipline applies: define the process, define the measure, preserve time order, understand the sampling mechanism, estimate expected behaviour and respond differently to routine variation and credible change.

12. A worked example: delivery time

Imagine a fulfilment centre tracks dispatch-to-door time for a stable service zone. Most values vary because of ordinary routing and queue conditions. Then a new warehouse scanning step is introduced.

If average delivery time rises by only a few minutes, any one delivery may still look normal. A Shewhart chart might take longer to signal a small persistent shift. An EWMA or CUSUM design may accumulate evidence more quickly. Once a signal occurs, the question is not “Who made the bad delivery?” It is “What changed in the system around the time the signal became credible?”

That change in question is one of SPC’s deepest contributions.

13. What can go wrong

• Recomputing limits every time a signal appears. This can make the chart absorb deterioration instead of detecting it.
• Mixing different processes. Day and night shifts, different products or different service classes may need stratification.
• Ignoring measurement-system change. A new sensor or scoring method can look like a process shift.
• Using specification limits as control limits. Requirements do not describe statistical behaviour.
• Forgetting autocorrelation or seasonality. Time structure can distort false-alarm rates.
• Reacting to every point. Over-adjustment can increase variability.
• Blaming individuals. Many signals are system signals.
• Using a chart without a response plan. Detection without investigation is surveillance, not control.

14. A practical response ladder

1. Confirm the data and measurement system.
2. Check whether the signal rule was actually triggered.
3. Look for a real process change near the signal time.
4. Separate known interventions from unexplained changes.
5. Check for batching, seasonality, autocorrelation or subgrouping problems.
6. Contain risk if the consequence of continued operation is high.
7. Investigate root causes rather than merely resetting the chart.
8. Document the intervention.
9. Establish a new baseline only when the process has genuinely changed and stabilised.

15. How SPC connects to the wider eduKate Library

SPC sits at the crossing point of measurement, statistics, operations and decision-making. Useful companion routes include:

• How Statistical Inference and Uncertainty Work
• How Time-Series Analysis Works
• How Data Visualisation Works
• How Robust Statistics and Outlier-Resistant Estimation Work
• How Decision Theory and Expected Utility Work
• How Measurement Error and Misclassification Work

16. Model limits and responsible interpretation

A control chart does not prove a cause. It detects patterns that are more or less compatible with a baseline model. A signal is an invitation to investigate. A lack of signal is not proof that nothing important changed. Very small shifts, sparse sampling, poor measurement or a badly chosen chart can hide change.

Likewise, statistical stability is not moral, legal, technical or commercial acceptability. A process can be consistently unsafe, consistently slow or consistently unfair. SPC tells us about behavioural stability; other systems define what outcomes are required.

17. Frequently asked questions

Is a control chart just a line chart with thresholds?

No. Its limits are derived from a statistical model of process variation, and the chart is interpreted using sequential signal rules.

Does a point outside the control limits mean a defective product?

No. It means the chart statistic is unusually far from the baseline expectation. The product may still be within specification.

Can a process be in control and still be poor?

Yes. A process can be statistically stable but centred badly, too variable or fundamentally incapable of meeting requirements.

When should EWMA or CUSUM be considered?

When small sustained shifts matter and the data structure supports their use. The choice should be based on expected shift size, false-alarm tolerance and process dynamics.

18. Authoritative sources and further reading

• NIST/SEMATECH Engineering Statistics Handbook, Process or Product Monitoring and Control: NIST Chapter 6.
• ISO 7870-1:2019, Control charts — General guidelines: ISO 7870-1.
• ISO 7870-2:2023, Shewhart control charts: ISO 7870-2.
• ISO 7870-6:2024, EWMA control charts for the process mean: ISO 7870-6.
• NIST, Statistical Process Control in Biochemical and Hematological Quality Control Data: NIST publication.

The deeper idea

SPC teaches a discipline that travels far beyond quality engineering: do not confuse variation with change. The world is noisy. Good systems learn what ordinary noise looks like, watch for evidence that the generating process has shifted, and reserve intervention for moments when intervention is justified.

That is how a chart becomes more than a graph. It becomes a way of thinking about systems over time.