Statistical Process Control

Introduction and Context

Statistical Process Control (SPC) is a systematic approach for monitoring, controlling, and improving processes through the use of statistical methods. Its central purpose is not merely to inspect finished products, but to understand process behaviour while work is being performed. By distinguishing expected process variation from unusual variation, SPC enables organisations to prevent defects, reduce waste, improve consistency, and make decisions based on evidence.

Every process exhibits variation. Measurements such as diameter, weight, delivery time, temperature, and number of defects are never exactly identical. The managerial question is whether the observed variation is an inherent feature of the current process or evidence that something unusual has occurred. Control charts provide a disciplined method for answering this question.

SPC is a core element of quality planning and control because it links process knowledge, measurement, statistical analysis, corrective action, and continuous improvement. It is used in manufacturing, healthcare, logistics, financial services, software development, laboratories, and public administration.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain the meaning and sources of process variation.
2. Distinguish common-cause and special-cause variation.
3. Select suitable variable and attribute control charts.
4. Calculate and interpret control limits.
5. Recognise non-random patterns and investigate assignable causes.
6. Develop practical reaction plans.
7. Avoid overcontrol, tampering, and manipulation of results.
8. Evaluate the limitations and governance requirements of SPC.

1. Process Variation and Statistical Thinking

A process is a set of interacting activities that transforms inputs into outputs. Inputs include materials, equipment, methods, people, environmental conditions, information, and measurement systems. Since these inputs are not perfectly constant, process outputs vary.

Statistical thinking involves three propositions:

• Work is performed through processes.
• Variation exists in all processes.
• Understanding and reducing variation requires data and process knowledge.

A process may be stable but incapable of meeting customer requirements, or capable on average but unstable because unusual events occur. Stability and capability are therefore different concepts. SPC primarily addresses stability; capability analysis should normally be performed only after reasonable statistical control has been established.

1.1 Common-Cause Variation

Common causes are the numerous, usually small sources of variation that are inherent in the design and operation of a process. Examples include ordinary differences in raw material, normal machine vibration, routine operator differences, and unavoidable measurement noise.

When only common causes are present, the process is described as statistically stable or predictable. This does not mean that every result is identical. It means that future variation can be estimated from past behaviour, provided the process conditions remain materially unchanged.

Improving a common-cause process generally requires changing the system: redesigning equipment, improving methods, controlling inputs, training, reducing queue variation, or improving measurement capability. Asking individual operators to adjust the process after every minor fluctuation is usually ineffective and may increase variation.

1.2 Special-Cause Variation

Special causes, also called assignable causes, are unusual sources of variation that are not part of the normal process pattern. Examples include a broken tool, incorrect machine setting, contaminated material, a power interruption, a new supplier batch, an untrained employee, or a faulty sensor.

Special causes may produce an observation outside a control limit, but they can also create patterns within the limits, such as a sustained shift, a trend, or cyclic behaviour. The correct response is investigation, containment where necessary, removal of the cause, and verification that the process has returned to a predictable state.

1.3 The Funnel Effect and Tampering

A stable process naturally produces random movement around its average. If an operator adjusts the process whenever a measurement is slightly above or below the target, the adjustment may respond to routine noise rather than a real change. This practice is called tampering or overcontrol.

For example, repeatedly changing a filling-machine setting after each individual bottle measurement can produce larger variation than leaving the correctly adjusted machine alone. Adjustment should be based on an established rule and process knowledge, not on emotional reaction to every observation.

2. Control Charts: Purpose and Structure

A control chart displays process data in time order together with a centre line and statistically derived control limits. It is a time-sequenced decision tool, not simply a graph of specification compliance.

The centre line commonly represents the process mean, average proportion, average count, or another appropriate statistic. The upper and lower control limits estimate the range of routine process variation. They are not customer specifications.

For a statistic W, the general three-sigma chart structure is:

Here, μ_W is the expected value of the plotted statistic and σ_W is its standard deviation. The limits are appropriate when the statistical assumptions behind the selected chart are reasonable. For some charts, μ_W and σ_W are estimated from preliminary data.

Approximately 99.73% of observations from a stable, normally distributed process fall within ±3 standard deviations of the mean. This probability is a reference principle, not a guarantee, and the exact interpretation depends on the chart statistic and sampling structure.

2.1 Control Limits versus Specification Limits

Control limits describe what the process is doing. Specification limits describe what customers, designers, regulators, or contracts require. A process can be:

• stable and capable;
• stable but incapable;
• unstable but apparently meeting specifications temporarily; or
• unstable and incapable.

Specification limits must not be calculated from process data. Conversely, control limits must not be moved simply to fit specification limits. Confusing these concepts can conceal risk.

2.2 Rational Subgroups

A rational subgroup is a small sample collected so that within-subgroup variation represents short-term, ordinary variation, while differences between subgroups can reveal changes over time. For example, five consecutive parts taken every hour may form a rational subgroup if the purpose is to detect changes between production periods.

Poor subgrouping can hide important variation. If observations from different machines, shifts, or products are mixed indiscriminately, the chart may show excessive variation or misleading averages. Subgroup design should reflect the process and the suspected sources of change.

3. Variable Control Charts

Variable data are measured on a continuous scale, such as length, mass, pressure, time, concentration, or temperature. The principal variable charts are the X̄–R chart, X̄–S chart, and individuals–moving range chart.

3.1 X̄–R Charts

The X̄–R chart is commonly used for small subgroups, often with subgroup sizes between 2 and 10. The X̄ chart monitors subgroup averages, while the R chart monitors within-subgroup range.

For subgroup i with observations x_i1, x_i2, ..., x_in:

The average subgroup mean and average range are:

where k is the number of subgroups and n is the subgroup size. Estimated limits are:

The constants A₂, D₃, and D₄ depend on subgroup size and are obtained from standard SPC tables. The lower range limit is sometimes zero because a range cannot be negative. The R chart should be interpreted first: if within-subgroup variation is unstable, interpretation of the average chart is less reliable.

3.2 Worked Example: X̄–R Chart

A process is sampled in four subgroups of size four. The observations are:

Thus, X̄̄ = 50.25 and R̄ = 2. For subgroup size four, commonly used constants are A₂ = 0.729, D₃ = 0, and D₄ = 2.282. Therefore:

All subgroup ranges are between 0 and 4.564, so the range chart shows no signal. All subgroup averages are between 48.792 and 51.708, so the average chart also shows no three-sigma signal. This supports—but does not prove—that the process was stable during the sampled period.

3.3 X̄–S Charts

The X̄–S chart uses the subgroup standard deviation instead of the range. It is generally preferred for larger subgroups because the standard deviation uses all observations and provides a more efficient estimate of dispersion.

The sample standard deviation for subgroup i is:

The exact control-limit equations use constants such as A₃, B₃, and B₄, selected according to subgroup size. As with an X̄–R chart, variation should be evaluated before interpreting the subgroup-average chart.

3.4 Individuals and Moving-Range Charts

When data are naturally collected one observation at a time, an individuals–moving range (I–MR) chart is appropriate. The moving range between consecutive observations is:

The average moving range estimates short-term variation. A commonly used estimate of the individual-observation standard deviation is:

For a moving range of two observations, d₂ is approximately 1.128. The individuals chart is then commonly calculated as:

This method assumes that successive observations are comparable and that the moving range provides a meaningful estimate of short-term variation. Autocorrelation, seasonality, or changing operating conditions may make the standard interpretation inappropriate.

4. Attribute Control Charts

Attribute data classify units or count events. They are used when measurement on a continuous scale is unavailable, impractical, or not necessary.

A defective unit fails at least one requirement; a defect is an individual nonconformity. One unit can contain several defects, so defective units and defects must not be confused.

4.1 p Chart

Let d_i be the number of defective units in sample i, and n_i the sample size. The sample fraction defective is:

The overall fraction defective is:

For sample i, the approximate limits are:

If the lower limit is negative, it is set to zero because a proportion cannot be negative. The binomial model is the usual basis, requiring comparable classification and reasonably independent outcomes.

4.2 np Chart

For constant sample size n, the number defective is plotted directly:

The np chart is easier for operational users when every sample contains the same number of units.

4.3 c and u Charts

A c chart is used for counts of defects when the inspection unit and opportunity area are constant. Under a Poisson model:

A u chart is used when the number of units or opportunities varies. If c_i defects are found across n_i units:

The Poisson assumption requires that events are countable and that the opportunity structure is sufficiently comparable. Clustering of defects or changing inspection opportunities may require a different analytical approach.

5. Interpreting Control-Chart Signals

A point outside a control limit is a strong indication of a special cause, but it is not automatically proof. The observation should be verified and investigated using process knowledge. A point close to a limit is not, by itself, a failure.

Common supplementary rules include:

• one point beyond a three-sigma limit;
• two of three successive points beyond the same two-sigma zone;
• four of five successive points beyond the same one-sigma zone;
• eight successive points on one side of the centre line;
• six successive points steadily increasing or decreasing;
• a repeating cycle or systematic pattern.

Rules should be defined before data are examined. Applying many rules indiscriminately increases false alarms. The selected rules should reflect the cost of missed signals, the frequency of sampling, and the practical response capacity.

A control chart should be read chronologically. The analyst should ask: What changed? When did it change? Which machine, shift, material, operator, supplier, or environmental condition was involved? A signal is a prompt for learning and investigation, not an accusation.

6. Establishing a Reaction Plan

A reaction plan converts statistical evidence into controlled action. It should be agreed before routine monitoring begins and should identify responsibilities, escalation rules, and records.

A practical reaction plan contains the following steps:

1. Detect: Identify a defined chart signal or credible process abnormality.
2. Verify: Confirm the measurement, calculation, sample identity, and instrument status.
3. Contain: Protect customers and downstream processes when nonconforming output may have been produced.
4. Investigate: Examine recent changes in materials, methods, machines, people, environment, and measurement.
5. Correct: Remove or control the special cause using authorised procedures.
6. Document: Record the signal, evidence, decisions, actions, and affected product.
7. Resume: Restart or continue only under defined conditions.
8. Review: Determine whether limits, procedures, training, or process design require improvement.

A reaction plan must distinguish between an out-of-control signal and an out-of-specification result. An in-control process may still produce some output outside specifications if its natural spread is too wide. Conversely, an unstable process may produce conforming output by chance.

7. Avoiding Overcontrol and Data Manipulation

Overcontrol occurs when routine variation triggers unnecessary adjustment. It can be prevented by:

• defining adjustment rules in advance;
• requiring evidence of a special cause before intervention;
• using appropriate subgrouping;
• separating operator control from engineering improvement;
• training personnel in signal interpretation; and
• reviewing the effect of adjustments on subsequent variation.

Data manipulation includes selective sampling, removing inconvenient observations without documented cause, changing measurement methods to improve results, delaying data entry, or resetting limits without technical justification. Such practices undermine safety, customer confidence, and decision quality.

A deleted point may be excluded only when a documented, verifiable non-process event explains why it is not representative—for example, a confirmed instrument failure during the measurement. Even then, the event should remain visible in the record. Limits should be recalculated only after a justified process change or a verified change in the underlying process.

Quality governance should provide audit trails, access controls, version-controlled chart definitions, independent review of unusual exclusions, and protection for employees who report problems. Performance targets should not reward concealment of variation, because unrealistic targets can create incentives to manipulate data.

8. Implementation in Organisations

A successful SPC programme begins with process selection. Priority should be given to characteristics that are critical to safety, function, customer satisfaction, compliance, cost, or downstream performance.

Implementation typically involves:

1. defining the process and customer requirement;
2. confirming measurement-system adequacy;
3. selecting the characteristic and chart type;
4. designing rational subgroups and sampling frequency;
5. collecting reliable baseline data;
6. calculating and reviewing preliminary limits;
7. investigating special causes;
8. establishing standard work and reaction plans; and
9. periodically reviewing chart effectiveness.

Measurement-system analysis is essential. If the measurement system has excessive repeatability or reproducibility error, the chart may show measurement noise rather than process variation. A stable chart based on a poor measurement system can create false confidence.

Digital SPC platforms can automate data collection, charting, alerts, and traceability. Artificial intelligence may support anomaly detection, root-cause prioritisation, and predictive maintenance. However, AI does not remove the need for process understanding. Algorithms can amplify biased data, confuse correlation with causation, or generate excessive alerts. Human review, explainability, cybersecurity, and change control remain necessary.

9. Illustrative Case: Stabilising a Filling Process

Background

A beverage plant fills 500-millilitre containers on three production lines. Operators report that fill volumes fluctuate and that adjustments are made several times per shift. Customer complaints show occasional underfilled containers.

Problem

The organisation has been plotting individual fill measurements against the customer specification limits, but it has no control chart or reaction plan. Operators adjust the filler whenever one container appears high or low.

Analysis

A measurement-system check confirms that the scale is suitable. Five consecutive containers are sampled every hour, making an X̄–R chart appropriate. The R chart shows a stable within-sample spread, while the X̄ chart shows frequent adjustments followed by increased oscillation around the target. Investigation reveals that operators are reacting to normal individual variation rather than to a statistical signal.

Approach and Solution

The plant establishes a baseline using representative data, creates separate charts for each line, and defines the following reaction plan: verify any signal, hold potentially affected product, check the filler setting and scale, investigate material and pressure conditions, and adjust only when evidence supports a special cause. Operators are trained not to change settings for an isolated point within the limits.

Engineering then addresses a genuine special cause: unstable supply pressure on Line 2. A regulator and preventive-maintenance check are introduced. The process target is reviewed against the specification, and the sampling plan is retained for the critical filling characteristic.

Outcome

Unnecessary adjustments decline, the spread of fill volumes decreases, and underfill complaints reduce. The improvement does not result from making the chart look better; it results from separating routine variation from assignable causes and improving the process mechanism.

Learning Points

The case demonstrates that SPC is a management system as well as a statistical technique. Correct chart selection, reliable measurement, rational subgrouping, operator training, and an authorised reaction plan are all necessary. A control chart cannot compensate for poor process design or inappropriate incentives.

10. Limitations and Critical Considerations

Control charts are based on assumptions. Data may be autocorrelated, non-independent, seasonal, non-normal, clustered, or affected by changing sample sizes. A standard chart can therefore give misleading signals when its model does not represent the process.

Statistical limits also depend on the data used to estimate them. Too few observations produce unstable estimates. Combining multiple populations can create artificial variation, while separating data excessively can conceal a common problem. Charts should be stratified when process streams genuinely differ, but stratification should not be used to hide poor performance.

SPC is not a substitute for process capability analysis, design of experiments, preventive maintenance, supplier control, risk analysis, or compliance verification. It is one component of a broader quality-management system.

Chapter Summary

Statistical Process Control uses data collected in time order to distinguish routine process variation from unusual variation. Common causes are inherent in the current system and usually require systemic improvement; special causes are unusual events requiring investigation and controlled response. Control charts contain a centre line and statistically derived control limits, which must be distinguished from specification limits.

Variable charts include X̄–R, X̄–S, and individuals–moving-range charts. Attribute charts include p, np, c, and u charts. Correct selection depends on the type of data, subgroup structure, sample size, and statistical assumptions. Chart interpretation requires attention to points beyond limits and non-random patterns, while avoiding excessive rules that create false alarms.

Effective SPC requires measurement-system adequacy, rational subgrouping, documented reaction plans, process knowledge, ethical data governance, and disciplined avoidance of tampering. Digital and AI-enabled tools can improve timeliness and analysis, but they must remain subject to human judgement, transparency, and control. SPC is most valuable when it supports learning and process improvement rather than merely judging operators or reporting conformance.

Key Terms

Common-cause variation: Routine variation inherent in the current process system.

Special-cause variation: Unusual variation arising from an identifiable change or event.

Statistical Process Control: Use of statistical methods to monitor and improve process behaviour.

Control chart: A time-ordered graph with a centre line and control limits.

Control limit: A statistically calculated boundary used to identify unusual process behaviour.

Specification limit: A customer, design, regulatory, or contractual requirement.

Rational subgroup: A sample designed to represent short-term routine variation.

Variable data: Measured data recorded on a continuous numerical scale.

Attribute data: Count or classification data.

Defective: A unit that fails one or more requirements.

Defect: An individual nonconformity within a unit.

X̄–R chart: A chart for subgroup averages and subgroup ranges.

X̄–S chart: A chart for subgroup averages and standard deviations.

Individuals–moving-range chart: A chart for sequential individual observations and moving ranges.

p chart: A chart for the proportion defective.

np chart: A chart for the number defective at constant sample size.

c chart: A chart for the number of defects with constant opportunity.

u chart: A chart for defects per unit when opportunity varies.

Overcontrol: Unnecessary adjustment in response to routine variation.

Reaction plan: A predefined procedure for verifying, investigating, containing, and correcting a chart signal.

Review Questions

Short Answer Questions

1. What is the principal difference between common-cause and special-cause variation?
2. Why must control limits be distinguished from specification limits?
3. When is an np chart preferable to a p chart?
4. What is the purpose of a rational subgroup?
5. Define overcontrol and explain why it is harmful.

Long Answer Questions

1. Explain the structure, assumptions, and interpretation of an X̄–R chart.
2. Compare p, np, c, and u charts, including the type of data appropriate for each.
3. Describe the stages of an effective SPC reaction plan.
4. Explain how poor subgrouping and an inadequate measurement system can distort SPC conclusions.
5. Discuss the ethical and governance risks associated with manipulating control-chart data.

Analytical / Application Questions

1. A process has all points within control limits, but the process average is regularly close to the lower specification limit. What conclusions can be drawn, and what additional analysis is required?
2. An operator adjusts a machine after every individual measurement, and the chart subsequently shows greater oscillation. Diagnose the likely problem and recommend a control strategy.
3. A service organisation counts customer complaints each day, but the number of transactions varies substantially. Which chart would be appropriate, what assumptions should be checked, and how should a signal be investigated?

Multiple Choice Questions with Answers

Question 1

Which statement best describes common-cause variation?

A. It is always caused by operator error.

B. It is inherent in the current process system.

C. It occurs only when a point exceeds a specification limit.

D. It can always be removed by adjusting the machine immediately.

Correct Answer: B. It is inherent in the current process system.

Question 2

What is the primary purpose of a control chart?

A. To replace customer specifications

B. To rank employees by performance

C. To distinguish routine variation from unusual process behaviour

D. To guarantee that every unit meets specifications

Correct Answer: C. To distinguish routine variation from unusual process behaviour

Question 3

Which chart is generally appropriate for the proportion defective when sample sizes may vary?

A. p chart

B. np chart

C. c chart

D. X̄–R chart

Correct Answer: A. p chart

Question 4

A process is sampled in subgroups of four. The subgroup averages are 50, 51, 49, and 51. What is X̄̄?

A. 49.50

B. 50.00

C. 50.25

D. 51.00

Correct Answer: C. 50.25

Question 5

For an X̄–R chart, what should generally be examined first?

A. The specification limits

B. The range chart

C. The customer complaint rate

D. The operator’s performance ranking

Correct Answer: B. The range chart

Question 6

Which chart is suitable for counting defects per unit when the number of units inspected varies?

A. np chart

B. p chart

C. c chart

D. u chart

Correct Answer: D. u chart

Question 7

What is the most appropriate response to one unusual point outside a control limit?

A. Automatically delete the point

B. Immediately widen the limits

C. Verify the result and investigate a possible special cause

D. Adjust the process after every subsequent observation

Correct Answer: C. Verify the result and investigate a possible special cause

Question 8

Why can repeated adjustment after every small fluctuation increase process variation?

A. It responds to routine noise as though it were a real process change.

B. It permanently reduces the process mean.

C. It converts variable data into attribute data.

D. It makes specification limits narrower.

Correct Answer: A. It responds to routine noise as though it were a real process change.

Question 9

Which statement about specification limits is correct?

A. They are always calculated as three standard deviations from the mean.

B. They describe customer, design, or regulatory requirements.

C. They should be changed whenever a point exceeds a control limit.

D. They are identical to control limits in every stable process.

Correct Answer: B. They describe customer, design, or regulatory requirements.

Question 10

In an individuals–moving-range chart, the moving range between consecutive observations is calculated as:

A. The sum of the two observations

B. The square root of their product

C. The absolute difference between the observations

D. The average of all observations

Correct Answer: C. The absolute difference between the observations