10. Process Capability and Statistical Process Control

Introduction and Context

Six Sigma improvement does not end when a process average has been adjusted or a cause of variation has been removed. A process must also be evaluated against customer requirements and monitored over time. Two complementary questions are central:

1. Is the process capable of meeting specifications when it is stable?

2. Is the process remaining stable and predictable during operation?

Process capability analysis addresses the first question. Statistical process control (SPC) addresses the second. Capability indices such as Cp, Cpk, Pp, and Ppk convert process variation and centering into interpretable measures of performance. Control charts distinguish routine, inherent variation from unusual or assignable variation. Used together, these methods support evidence-based decisions about process improvement, control plans, supplier performance, equipment qualification, and risk reduction.

Capability analysis is meaningful only when the process is sufficiently stable and the measurement system is adequate. A highly capable process can deteriorate if it is not controlled; conversely, a stable process may consistently produce nonconforming output if its natural variation is too large or its mean is poorly centered. This chapter therefore treats capability and control as connected but distinct analytical activities.

Learning Objectives

After studying this chapter, the learner should be able to:

• distinguish common-cause and special-cause variation;
• explain the purpose, assumptions, and interpretation of Cp, Cpk, Pp, and Ppk;
• relate capability indices to sigma level and defect performance;
• select appropriate variable, attribute, and specialized control charts;
• evaluate capability when data are non-normal or autocorrelated;
• interpret control-chart signals without overreacting to random variation;
• apply capability and SPC methods in operational and managerial decisions; and
• identify risks associated with poor sampling, unstable processes, inadequate measurement systems, and misuse of indices.

The Logic of Stability Before Capability

A specification describes what the customer, regulator, design authority, or internal user considers acceptable. A control limit describes what the process is expected to produce when only common causes are operating. These are not interchangeable concepts. Specification limits are externally or functionally determined; control limits are calculated from process data.

For a characteristic with a lower specification limit (LSL) and upper specification limit (USL), the specification interval represents acceptable product performance. The process distribution represents actual performance. Capability analysis compares the spread and location of this distribution with the specification interval.

A process is stable when its statistical behavior is reasonably predictable over time. Stability does not mean that every observation is identical. It means that variation follows a consistent pattern attributable to common causes. A process may be stable but incapable, capable but temporarily unstable, both stable and capable, or neither.

The recommended sequence is:

1. define the quality characteristic and specification basis;

2. verify the measurement system;

3. collect data in rational subgroups or an appropriate time sequence;

4. examine the process using suitable control charts;

5. investigate and remove special causes when indicated;

6. assess capability using data from the stable process; and

7. establish ongoing monitoring and response rules.

If special causes remain, the estimated standard deviation may represent a mixture of different operating conditions. Capability indices calculated from such data can be misleading because they describe neither a reliable short-term process nor a meaningful long-term process.

Variation, Specifications, and the Capability Concept

Let the process characteristic have mean μ and standard deviation σ. Under the conventional approximately normal model, the natural process spread is often represented as approximately six standard deviations, or 6σ, from the lower to the upper side of the process distribution. This is a descriptive convention, not a universal law. The appropriateness of the model depends on the data and the process.

A process can fail specifications for two broad reasons:

• excessive variation, in which the process distribution is too wide; or
• poor centering, in which the mean is too close to one specification limit.

Cp measures potential capability based on spread, while Cpk incorporates centering. Their long-term counterparts, Pp and Ppk, use overall variation rather than within-subgroup variation. The distinction is important in Six Sigma because short-term and long-term estimates answer different questions.

Cp: Potential Capability Based on Within-Subgroup Variation

For a two-sided specification, the potential capability index is:

Cp = (USL − LSL) / (6σwithin)

where σwithin estimates short-term, within-subgroup variation. Cp compares the width of the specification interval with the width of the process distribution. It does not consider whether the process mean is centered between the specifications.

For example, Cp = 1.33 indicates that the specification width is 1.33 times the estimated six-sigma process width. This suggests reasonable potential capability if the process is centered and the assumptions are appropriate. However, Cp alone cannot establish that the process is actually producing conforming output.

Cpk: Actual or Centered Capability

Cpk accounts for the distance between the process mean and each specification limit:

Cpu = (USL − μ) / (3σwithin)

Cpl = (μ − LSL) / (3σwithin)

Cpk = min(Cpu, Cpl)

The smaller component identifies the specification side presenting greater risk. If Cp is high but Cpk is substantially lower, the process is relatively narrow but off-center. If Cp and Cpk are approximately equal, the process is close to centered, subject to estimation error.

A one-sided specification requires a one-sided capability measure such as Cpu or Cpl rather than an inappropriate two-sided Cp.

Pp and Ppk: Overall Performance Indices

Pp and Ppk use the overall standard deviation, σoverall, calculated from all relevant observations:

Pp = (USL − LSL) / (6σoverall)

Ppk = min[(USL − μ)/(3σoverall), (μ − LSL)/(3σoverall)]

Pp describes performance based on total observed variation over the study period. Ppk reflects both total variation and centering. Overall variation may include shifts, drifts, changes between lots, operators, machines, or time periods.

The relationship between the indices is diagnostically useful. When Pp is materially lower than Cp, long-term sources of variation are inflating performance spread. When Cpk is lower than Cp, centering is a problem. When Ppk is lower than Cpk, between-subgroup or time-related variation is important. These comparisons should not be interpreted mechanically; the sampling design and stability evidence must be considered.

Interpreting Capability Values

A value of 1.00 means that the estimated six-sigma spread just fits within the specification interval, assuming the relevant model. It does not guarantee zero defects. Common organizational criteria use values such as 1.33 for routine capability and higher values for critical characteristics, but the appropriate threshold depends on risk, customer requirements, regulatory expectations, measurement uncertainty, and the cost of failure. No index should be treated as universally acceptable without context.

Capability indices are estimates, not exact properties. Confidence intervals may be required when sample sizes are small or decisions are consequential. A reported Cpk of 1.34 and a lower confidence bound below the organizational requirement should not be treated as equivalent to a precisely established Cpk above that requirement.

Sigma Level and Defect Performance

Sigma level expresses the distance between the process mean and the nearest specification limit in standard-deviation units. The short-term or within-process Z value is:

Zbench = min[(USL − μ)/σwithin, (μ − LSL)/σwithin]

For a two-sided characteristic under the conventional interpretation, Zbench = 3Cpk. Thus, Cpk = 1.00 corresponds to a nearest specification distance of approximately 3 within standard deviations, and Cpk = 1.33 corresponds to approximately 4 within standard deviations.

The frequently cited Six Sigma benchmark involves additional conventions, including a distinction between short-term and long-term performance and a commonly used 1.5-sigma shift assumption. That convention should not be confused with a direct physical law or used to convert every Ppk value automatically into a defect rate. Defect probabilities depend on distributional assumptions, specification structure, estimated parameters, tail behavior, autocorrelation, and the definition of an opportunity.

For practical reporting, organizations should state whether sigma is based on within or overall variation, whether the characteristic is one-sided or two-sided, and how opportunities and defects are defined.

Non-Normal Capability Analysis

Normal capability calculations can be seriously misleading when the data are skewed, bounded, multimodal, heavy-tailed, or otherwise non-normal. Examples include cycle time, particle counts, reliability times, contamination levels, and financial or service measures. A histogram, probability plot, and substantive knowledge of the process should be used before applying normal capability indices.

Several approaches are available:

• Transformation: A logarithmic, Box–Cox, or other transformation may make the data more nearly normal. Capability is then assessed on the transformed scale, with interpretation translated carefully back to the original units.
• Percentile or quantile methods: The relevant lower and upper process percentiles can be compared directly with specifications without forcing a normal model.
• Distribution fitting: A theoretically or empirically appropriate distribution, such as Weibull, lognormal, gamma, or logistic, may be fitted. The estimated fraction beyond specifications is then calculated from the fitted distribution.
• Nonnormal capability indices: Some software provides indices based on fitted distributions or percentile distances. Their assumptions and parameter-estimation methods must be documented.
• Separate analysis of mixtures: Multimodal data may indicate different machines, products, suppliers, or operating states. Segmenting the process may be more appropriate than fitting one distribution.

A non-normal distribution is not automatically a data problem. It may be a real feature of the process or of the measurement scale. The analyst must distinguish model inadequacy from process heterogeneity. A transformation should not be used merely to obtain a desirable capability number.

Statistical Process Control as a Decision System

SPC is more than plotting data. It is a structured decision system in which observations are collected in a meaningful sequence, limits are calculated from representative data, signals are interpreted using defined rules, and actions are linked to likely causes.

Rational Subgroups

A rational subgroup contains observations produced under sufficiently similar conditions that within-subgroup variation estimates short-term or immediate variation. Differences between subgroups then reveal changes over time or across conditions. For example, consecutive units from one machine cycle may form a subgroup when the purpose is to detect shifts between cycles. If observations are collected across an entire shift and mixed together, the subgroup may conceal important short-term changes.

The sampling interval should reflect the process rhythm and the speed with which undesirable change must be detected. Sampling too slowly can allow large exposure to defects; sampling too frequently can create unnecessary intervention and false alarms.

Control Limits and Warning Limits

A control chart normally includes a center line and upper and lower control limits, often based on approximately three standard deviations of the chart statistic. Control limits are estimated from process data and should be recalculated only when justified by a documented change in the process or in the measurement system. Warning limits, commonly placed nearer to the center line, may support heightened attention but do not replace formal control limits.

A control limit is not a specification limit. An observation beyond a control limit signals an unlikely pattern under the assumed stable process; it does not by itself prove that the product is defective. Conversely, a point within control limits may still violate a specification.

Choosing the Appropriate Control Chart

Variables Charts

Variables charts are used for measured characteristics such as length, mass, temperature, pressure, voltage, or response time.

X̄–R charts are effective for small rational subgroups, commonly between two and ten observations. The X̄ chart monitors subgroup means, while the R chart monitors within-subgroup ranges. The range method is simple but less efficient for larger subgroups.

X̄–S charts use subgroup standard deviations and are generally preferred for moderate or larger subgroup sizes. The S chart provides a more efficient estimate of within-subgroup dispersion than the range when subgroup size increases.

Individuals and moving range (I–MR) charts are appropriate when observations are naturally collected one at a time or subgrouping is not meaningful. The moving range estimates short-term variation from successive observations. These charts require particular caution when data are autocorrelated.

Variables charts usually provide more information than attribute charts because they preserve the magnitude of each observation. If measurement is feasible and reliable, variables data are often preferable for process diagnosis.

Attribute Charts

Attribute charts are used when outcomes are counted or classified.

• p chart: fraction or proportion nonconforming, with sample size allowed to vary;
• np chart: number nonconforming, with constant sample size;
• c chart: number of defects per constant inspection unit when a Poisson-type model is reasonable;
• u chart: defects per unit when the number of inspection units varies.

A nonconforming unit is a unit that fails to meet requirements, whereas a defect is a specific nonconformity. One unit may contain several defects, so p and np charts should not be substituted for c and u charts without considering the data definition.

Attribute charts are useful for low-cost inspection, service errors, complaints, and pass/fail outcomes. However, they may be insensitive when defect rates are low or sample sizes are small. A process can also show a stable defect rate that is unacceptably high; stability is not equivalent to capability.

Specialized Charts

Some processes require methods beyond conventional Shewhart charts.

EWMA charts assign greater weight to recent observations and are effective for detecting small, persistent shifts in the mean. CUSUM charts accumulate deviations from a target and can detect small sustained changes efficiently. Both require careful selection of reference values, decision intervals, and initialization procedures.

Laney p′ and u′ charts may be useful when attribute data exhibit overdispersion or underdispersion relative to the binomial or Poisson assumptions. Rare-event charts address situations in which nonconforming events are infrequent and conventional charts would be inefficient. Multivariate charts, such as Hotelling’s T², monitor several correlated characteristics jointly, while generalized variance or related charts monitor dispersion.

For autocorrelated data, ordinary limits can generate excessive false signals or conceal change. Time-series models, residual charts, or appropriately designed sampling may be required. The analyst should not apply a specialized chart simply because it is sophisticated; the chart must match the data-generating process and the decision objective.

Reading Signals and Investigating Causes

Typical evidence of special-cause variation includes a point beyond a control limit, a sustained run on one side of the center line, a systematic trend, cycles, stratification, or an unusual pattern of alternation. Exact run rules vary among standards and organizations. The rule set must be selected in advance and applied consistently rather than chosen after inspecting results.

When a signal occurs, the response should include:

1. confirm the data, timestamp, measurement method, and chart calculation;

2. contain potentially affected material or service output when risk warrants;

3. examine changes in people, equipment, material, method, environment, and measurement;

4. use process knowledge and structured root-cause analysis;

5. verify the suspected cause with evidence rather than correlation alone; and

6. document the corrective action and confirm that the process returns to a predictable state.

Repeatedly adjusting a process after every minor fluctuation creates tampering. If the process is stable and centered, unnecessary adjustment increases variation. Control charts are intended to separate signals from noise, not to encourage continuous intervention.

An Applied Case: Capability Improvement in a Precision Machining Process

Background

The following is an illustrative case. A manufacturer produces a shaft whose critical diameter has specifications of 24.950 mm to 25.050 mm. The characteristic is measured using a calibrated gauge whose repeatability and reproducibility have been judged acceptable for the application. Five consecutive shafts are measured every hour, producing rational subgroups.

Situation and Problem

Initial analysis shows that the X̄ chart contains a sustained upward shift during several hours, and the R chart shows a concurrent increase in dispersion. The quality team nevertheless reports a preliminary Cpk of 1.36 based on all observations. Production management proposes accepting the result because it exceeds the internal criterion of 1.33.

Analysis

The chart signals indicate that the process is not stable. The combined data include at least two operating states, so the preliminary Cpk is not a defensible estimate of stable capability. Investigation identifies a tool replacement procedure that changes the cutting condition and a coolant-temperature difference between two shifts.

After the procedure is standardized and coolant temperature is controlled, a new representative study is collected. The stable process has a mean of 25.004 mm and a within-subgroup standard deviation of 0.010 mm. Thus:

Cp = (25.050 − 24.950)/(6 × 0.010) = 1.67

Cpu = (25.050 − 25.004)/(3 × 0.010) ≈ 1.53

Cpl = (25.004 − 24.950)/(3 × 0.010) ≈ 1.80

Cpk ≈ 1.53

The lower of Cpu and Cpl shows that the upper specification is the nearer risk. The process has adequate potential spread and is reasonably centered, though not perfectly centered.

The team also calculates Pp and Ppk using the entire post-improvement study. If the overall standard deviation is 0.013 mm, then Pp is approximately 1.28 and Ppk approximately 1.18. The gap between within and overall indices suggests remaining time-related variation, even though the control chart appears substantially improved. Further stratification by shift and tool age is therefore warranted.

Approach and Outcome

The organization establishes an X̄–R chart for routine monitoring, a tool-life check, and a response plan requiring investigation of any confirmed special cause. It does not use specification limits as control limits. A separate capability review is scheduled after sufficient production cycles, including all relevant shifts and tool conditions. The process is released for production with continued monitoring rather than being declared permanently capable from one study.

Learning Points

The case demonstrates that a numerical capability index cannot compensate for instability. It also shows why Cp and Cpk must be read together, why Pp and Ppk reveal longer-term behavior, and why capability is a continuing management responsibility rather than a one-time certificate.

Practical and Managerial Applications

In manufacturing, capability and SPC support machine qualification, preventive maintenance, supplier development, control plans, and release of critical dimensions. In healthcare, charts can monitor waiting time, medication errors, infection indicators, or laboratory measurements, provided definitions and denominators are carefully controlled. In financial and administrative services, individuals may chart processing time, rework, complaint rates, or transaction errors. In software and digital operations, deployment failure rate, response latency, and incident counts can be monitored with suitable charts.

Managers should ask not only whether a metric is above a target but also whether its behavior is stable, whether data are comparable across segments, and whether intervention rules are clear. A dashboard showing averages without variation and time-order information can hide special causes and create false confidence.

Limitations, Risks, and Professional Governance

Several mistakes recur in capability and SPC projects:

• calculating capability before checking stability;
• using Cp when centering is poor;
• confusing Cpk with Ppk or mixing within and overall estimates;
• treating control limits as customer specifications;
• selecting a chart without considering subgroup structure;
• ignoring non-normality, autocorrelation, or mixtures;
• using too few observations to support a high-stakes conclusion;
• deleting inconvenient points without documented causal evidence;
• changing limits repeatedly to suppress signals; and
• measuring a process outcome that is not linked to customer or business risk.

Ethically, analysts must report the data period, exclusions, assumptions, specification source, measurement-system status, and uncertainty. Selective reporting of a favorable subgroup can mislead customers, regulators, or senior management. In regulated industries, record integrity, traceability, validation, and change control may be mandatory. Automated monitoring systems should preserve an audit trail and distinguish an alert from a confirmed cause.

Capability analysis also has a governance dimension. Specifications should be technically justified and periodically reviewed; arbitrarily widening specifications to improve an index transfers risk to customers. Similarly, a stable but poor process should trigger improvement, not normalization of unacceptable performance.

Digital SPC and Emerging Practice

Modern quality systems increasingly combine shop-floor data, laboratory systems, manufacturing execution systems, enterprise resource planning, and cloud analytics. Automated charts can provide near-real-time alerts, stratify data by machine or product family, and link signals to maintenance or corrective-action records. Statistical process monitoring can also be integrated with predictive maintenance and digital twins.

Machine-learning methods may detect nonlinear patterns, interactions, or unusual multivariate states that conventional charts miss. However, predictive algorithms do not eliminate the need for sound sampling, stable definitions, measurement validation, and human interpretation. A black-box anomaly score is not automatically a control chart, and an alert without an operational response can create alarm fatigue. Governance should address model drift, cybersecurity, access control, explainability, data quality, and responsibility for decisions.

Best-Practice Framework for Application

A defensible capability and SPC study normally includes the following practices:

• connect the characteristic to a customer, safety, regulatory, or process requirement;
• verify operational definitions and measurement-system adequacy;
• preserve time order and define rational subgroups;
• use control charts to establish and maintain stability;
• select indices and distributions consistent with the data;
• report both point estimates and, when appropriate, confidence intervals;
• separate short-term capability from long-term performance;
• investigate the limiting specification side and meaningful sources of variation;
• define reaction plans before routine monitoring begins; and
• review the analysis when equipment, material, method, product, or environment changes.

Chapter Summary

Process capability and statistical process control answer related but different questions. Capability indices compare process spread and centering with specifications, while control charts evaluate whether process behavior is stable and predictable over time. Cp measures potential capability from within-subgroup variation; Cpk also considers centering. Pp and Ppk use overall variation and therefore reflect longer-term performance. Sigma-level interpretations require clear assumptions about within or overall variation, distribution, specifications, and opportunities.

A stable process should be established before capability is calculated. Variable charts such as X̄–R, X̄–S, and I–MR are selected according to subgroup structure; p, np, c, and u charts address different attribute-data definitions. EWMA, CUSUM, rare-event, multivariate, and overdispersion-adjusted charts serve specialized conditions. Non-normality, autocorrelation, mixtures, measurement error, and poor sampling can invalidate conventional conclusions. Sound professional practice combines statistical evidence, process knowledge, ethical reporting, documented reaction plans, and appropriate digital governance.

Key Terms

Capability: The degree to which a stable process can meet specified requirements.

Cp: Potential capability index based on specification width and within-subgroup variation.

Cpk: Capability index incorporating process centering and within-subgroup variation.

Pp: Performance index based on specification width and overall variation.

Ppk: Performance index incorporating centering and overall variation.

Common-cause variation: Natural variation produced by the ordinary system of process operation.

Special-cause variation: Variation arising from a specific, identifiable change or unusual condition.

Specification limit: A customer, design, regulatory, or internal boundary defining acceptable output.

Control limit: A statistically calculated boundary used to assess process stability.

Rational subgroup: A group of observations collected under sufficiently similar conditions for a defined analytical purpose.

Sigma level: The standardized distance from the process mean to the nearest specification limit.

Within-subgroup standard deviation: An estimate of short-term process variation.

Overall standard deviation: An estimate of total variation across the study period.

Nonconforming unit: A unit that fails to meet one or more requirements.

Defect: A specific nonconformity; one unit may contain multiple defects.

Shewhart chart: A control chart designed primarily to detect relatively large process shifts.

EWMA chart: A chart using exponentially weighted historical observations to detect small sustained shifts.

CUSUM chart: A chart that accumulates deviations from a target to identify persistent change.

Tampering: Unnecessary adjustment of a stable process in response to routine random variation.

Autocorrelation: Statistical dependence between observations separated in time.

Review Questions

Short Answer Questions

1. Why should process stability normally be assessed before capability?

2. What is the principal difference between Cp and Cpk?

3. How do Pp and Ppk differ conceptually from Cp and Cpk?

4. When is a u chart preferred to a c chart?

5. What is meant by tampering in SPC?

Long Answer Questions

1. Explain the formulas, assumptions, and managerial interpretation of Cp, Cpk, Pp, and Ppk.

2. Discuss how rational subgrouping influences the usefulness of variables control charts.

3. Compare X̄–R, X̄–S, I–MR, p, np, c, and u charts, including the data conditions for their use.

4. Explain why normal capability analysis may fail for non-normal data and evaluate alternative approaches.

5. Describe an integrated workflow for establishing, improving, and monitoring process capability.

Analytical / Application Questions

1. A process has USL = 52, LSL = 48, mean = 50.8, and within-subgroup standard deviation = 0.5. Calculate Cp, Cpu, Cpl, and Cpk, and identify the principal risk.

2. A process has Cp = 1.60, Cpk = 0.95, Pp = 1.20, and Ppk = 0.72. Interpret the pattern and recommend an investigation sequence.

3. A service process produces one transaction result at a time, shows gradual drift, and has strongly skewed completion times. Recommend a monitoring and capability strategy, explaining the limitations of a conventional I–MR chart and normal capability indices.

22. Multiple Choice Questions with Answers

Question 1

A process has Cp = 1.50 and Cpk = 0.80. What is the most appropriate interpretation?

A. The process is stable and perfectly centered

B. The process has adequate spread but is poorly centered

C. The process has excessive within-subgroup variation but ideal centering

D. The process must have a non-normal distribution

Correct Answer: B. The process has adequate spread but is poorly centered

Question 2

Which condition most directly invalidates a conventional capability calculation based on pooled process data?

A. The process has a documented customer specification

B. The sample contains measurements in time order

C. The process contains unresolved special-cause shifts

D. The measurement unit is expressed in millimetres

Correct Answer: C. The process contains unresolved special-cause shifts

Question 3

Which chart is generally appropriate for five-measurement rational subgroups when both subgroup means and ranges are to be monitored?

A. X̄–R chart

B. p chart

C. u chart

D. Individuals chart only

Correct Answer: A. X̄–R chart

Question 4

A quality team counts the number of defects per unit while the number of units inspected varies from sample to sample. Which chart is most appropriate?

A. np chart

B. c chart

C. u chart

D. X̄–S chart

Correct Answer: C. u chart

Question 5

What does a materially lower Ppk than Cpk generally suggest, assuming comparable and appropriate calculations?

A. Long-term or between-subgroup variation is increasing performance risk

B. The process mean is exactly centered

C. The specification limits are necessarily incorrect

D. The measurement system has zero error

Correct Answer: A. Long-term or between-subgroup variation is increasing performance risk

Question 6

A stable process has a control limit above the customer’s USL. What does this comparison mean?

A. The control limit should automatically replace the USL

B. The process is necessarily out of statistical control

C. The process’s natural variation may exceed the required specification range

D. Every observation above the center line is defective

Correct Answer: C. The process’s natural variation may exceed the required specification range

Question 7

Which method is particularly designed to detect small, sustained shifts in a process mean?

A. EWMA chart

B. np chart

C. Simple defect checklist

D. Histogram without time order

Correct Answer: A. EWMA chart

Question 8

A cycle-time distribution is strongly right-skewed. What is the best initial capability practice?

A. Apply normal Cp immediately because all continuous data are normal

B. Ignore the skewness if the average meets its target

C. Investigate an appropriate transformation or non-normal distributional method

D. Replace all values above the mean with the mean

Correct Answer: C. Investigate an appropriate transformation or non-normal distributional method

Question 9

Why can repeated adjustment of a stable process increase defects?

A. It converts specification limits into control limits

B. It responds to common-cause noise and adds unnecessary variation

C. It always reduces the process mean

D. It makes the measurement system more precise

Correct Answer: B. It responds to common-cause noise and adds unnecessary variation

Question 10

For a two-sided specification under the conventional normal interpretation, what is the relationship between Cpk and the nearest within-process sigma distance?

A. The distance is approximately Cpk/3

B. The distance is approximately 3Cpk

C. The distance is approximately 6Cpk

D. There is no mathematical relationship

Correct Answer: B. The distance is approximately 3Cpk