Process Stability, Capability and Performance

Introduction

Statistical process control (SPC) requires two related but distinct judgments. First, is the process stable and predictable over time? Second, can that process meet customer, engineering, or regulatory requirements? Stability concerns the behaviour of the process distribution through time; capability concerns the relationship between that distribution and specification limits. A process may be stable but incapable, capable for a short period but unstable, or apparently capable only because special causes have been overlooked.

This distinction is central to quality management. Capability indices such as Cp and Cpk summarize potential or short-term capability, while Pp and Ppk summarize overall or long-term performance. Cpm additionally considers deviation from a target value. None of these indices replaces process knowledge, control charts, measurement-system analysis, or engineering judgment.

Learning Objectives

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

1. Distinguish statistical stability, capability, and performance.
2. Explain the assumptions behind normal-theory capability indices.
3. Calculate and interpret Cp, Cpk, Pp, Ppk, and Cpm.
4. Relate capability indices to yield, nonconformance, and sigma performance.
5. Construct and interpret confidence intervals for capability estimates.
6. Evaluate non-normal data and transformation-based methods.
7. Apply capability analysis responsibly in industrial and service settings.

1. Stability, Capability and Performance

A stable process is governed by a consistent system of common causes. Its probability distribution does not exhibit unexplained shifts, trends, cycles, or changes in variation. Stability is assessed primarily with appropriate control charts and process knowledge. A stable process is predictable in a probabilistic sense, although it may still produce unacceptable output.

A capable process produces output within specification limits with an acceptable probability. Capability therefore requires both a stable process and meaningful specification limits. Specifications are imposed externally by design, customer needs, safety requirements, or regulation; control limits are calculated internally from process data. They must not be confused.

A useful sequence is:

1. Confirm that the measurement system is adequate.
2. Examine process stability using time-ordered data and suitable control charts.
3. Estimate the process distribution only after stability is reasonably established.
4. Compare the distribution with specification limits.
5. Investigate causes and improve the process rather than merely reporting an index.

Important terminology

• USL: upper specification limit.
• LSL: lower specification limit.
• T: nominal target value.
• μ: process mean.
• σ: process standard deviation.
• Within-subgroup variation: short-term variation estimated from variation inside rational subgroups.
• Overall variation: variation estimated from all observations over the period of interest.
• Potential capability: capability implied by process spread if centered, commonly measured by Cp.
• Actual capability: capability accounting for both spread and centering, commonly measured by Cpk.
• Performance: an estimate based on overall observed variation, commonly measured by Pp and Ppk.

2. Measurement and Data Requirements

Capability analysis is only as credible as the data. The measurement system should be assessed for bias, repeatability, reproducibility, resolution, and stability. Excessive measurement error inflates estimated variation and can obscure centering. A gauge repeatability and reproducibility study, calibration evidence, or an equivalent measurement-system evaluation is normally required.

The sampling plan must represent the process and preserve time order. Rational subgroups should contain observations produced under essentially the same short-term conditions. For example, consecutive parts from one machine cycle may form a subgroup, whereas parts collected across several shifts may be better treated as a long-term sample. Mixing unlike sources can conceal important changes.

Specification limits must be technically justified. A control limit is not a specification limit, and a specification limit must not be calculated from the same process data merely to make the process appear capable. The period of analysis must also be stated: a process can have different short-term and long-term capability.

3. Normal-Theory Capability Indices

The following indices assume a continuous characteristic, stable observations, and—unless a validated alternative is used—an approximately normal process distribution. For a two-sided specification, the specification width is USL − LSL and the process spread is measured in standard-deviation units.

3.1 Cp: potential capability

Here, Cp is dimensionless; USL, LSL, and σ must use the same measurement units. The factor 6 represents the width from μ − 3σ to μ + 3σ for a normal distribution. Cp measures spread only and assumes that the process mean is centered between the specifications. It does not detect off-centering.

A Cp of 1.00 indicates that the six-sigma process spread equals the specification width. It does not imply that 99.73% of output is conforming unless the process is centered and stable. Values such as 1.33 or 1.67 are common organizational thresholds, but the required threshold depends on risk, cost, and industry expectations.

3.2 Cpk: capability adjusted for centering

The two terms are the upper and lower capability components:

Cpk is limited by the specification that is closer to the process mean. If the process is exactly centered, Cp = Cpk. If Cpk is substantially smaller than Cp, centering is an important improvement opportunity.

For a stable normal process, the approximate distance from the mean to the nearest specification, in sigma units, is:

This is a benchmark sigma distance, not automatically a “six-sigma performance” claim. It does not include arbitrary 1.5-sigma shifts sometimes used in corporate reporting.

3.3 Pp and Ppk: overall performance

Pp and Ppk use the overall standard deviation, s_overall, calculated from the full set of observations over the selected period:

Here x̄ is the sample mean. Pp and Ppk describe observed long-term performance, including shifts, drifts, and between-subgroup variation present during the study. They are not necessarily estimates of a stable underlying capability.

The relationship between indices is diagnostic. If Cp is materially greater than Pp, short-term variation is smaller than overall variation, suggesting time-dependent changes. Similarly, a large difference between Cpk and Ppk may indicate instability, subgroup-to-subgroup movement, or a change in centering.

3.4 Cpm: capability relative to a target

Cpm penalizes both variation and departure from the target T. It is useful when being near the nominal target matters even if the output remains inside the specification limits. If T equals μ, Cpm equals Cp. Cpm is not the same as Cpk: Cpk focuses on the nearest specification boundary, whereas Cpm incorporates squared target deviation.

Some organizations use a target-centered index, Cpmk, which combines target deviation and one-sided proximity; however, formulas and naming conventions vary across standards and software. The selected definition should always be stated explicitly rather than assumed.

4. Worked Example: Short-Term and Overall Analysis

An injection-moulding process has LSL = 49.50 mm, USL = 50.50 mm, and target T = 50.00 mm. A stable subgroup analysis gives μ = 50.20 mm and within-subgroup σ = 0.10 mm. The complete three-week data set gives x̄ = 50.18 mm and s_overall = 0.14 mm.

Short-term indices

Therefore Cpk = min(1.00, 2.33) = 1.00. The process has good potential spread but is shifted toward the upper specification.

The target-based index is:

Thus, target performance is weak despite a Cp of 1.67.

Overall indices

The overall results show both centering weakness and additional long-term variation. Management should not report “capable” solely on the basis of Cp. Corrective action should first address centering and then investigate why overall variation exceeds within-subgroup variation.

5. Yield, Defect Probability and Sigma Performance

For a normally distributed characteristic with mean μ and standard deviation σ, the probability of falling below a value a is:

where Φ is the standard normal cumulative distribution function. For a two-sided specification, the estimated conforming yield is:

The estimated nonconforming fraction is 1 − Y. For multiple independent characteristics or opportunities, a simple approximation is:

where Y_j is the yield for characteristic j. Independence must be justified; correlated defects cannot be combined this way without adjustment.

For a centered normal process, Cp = 1.00 corresponds to limits at ±3σ and approximately 99.73% conforming output. If Cpk = 1.00, the nearest limit is approximately 3σ from the mean, but the opposite tail may be much smaller. When Cpk = 1.33, the nearest limit is approximately 3.99σ away. These conversions are approximations and depend on normality, stability, and the chosen standard-deviation estimate.

“Defects per million opportunities” is calculated as:

A defect is not necessarily a defective unit: one unit can contain multiple defects. Sigma labels must identify whether they are short-term Z values, long-term Z values, or values adjusted by a stated convention.

6. Confidence Intervals and Uncertainty

Capability indices are estimates, not exact properties. Their uncertainty arises from finite sample size, estimation of variation, non-normality, autocorrelation, subgroup design, and measurement error. A point estimate without an uncertainty statement can encourage overconfident decisions.

For a normal sample of n independent observations, the classical variance result is:

where χ² denotes a chi-square random variable and n − 1 is the degrees of freedom. A two-sided confidence interval for the population variance is:

The square roots provide a confidence interval for σ. Since Cp is proportional to 1/σ when specification limits are fixed, the interval endpoints for Cp are obtained by applying the appropriate inverse relationship to the variance limits. Exact intervals for Cpk and Ppk are more complicated because they involve both mean and standard-deviation uncertainty; software methods may use noncentral distributions, bootstrap resampling, or approximations.

A practical report should state the point estimate, confidence level, sample size, estimation method, and whether within or overall variation was used. A lower confidence bound is often more useful for release decisions than the point estimate alone. Bootstrap intervals can be useful for non-normal or complicated estimators, but resampling must respect serial dependence and subgroup structure.

7. Non-Normal Capability Analysis

Normal capability indices can mislead when the characteristic is skewed, bounded, multimodal, heavy-tailed, or a mixture of populations. A histogram that looks non-normal is not, by itself, sufficient evidence for a particular remedy. Subject-matter knowledge and graphical diagnostics—such as probability plots, empirical distribution functions, and time plots—should be combined with formal assessment.

Common approaches include:

7.1 Distribution-specific analysis

Fit a defensible distribution, such as lognormal, Weibull, gamma, or a suitably bounded distribution, then calculate the probability below LSL and above USL directly from its cumulative distribution function F:

The estimated yield is the probability between the limits. Parameters should be estimated transparently, and goodness-of-fit should not be judged only by a single p-value. Fit, residual behaviour, process mechanism, and validation data matter.

7.2 Transformations

A monotonic transformation g(X) may make the distribution approximately normal. The capability analysis is then performed on Z = g(X), with transformed specification limits g(LSL) and g(USL). For example, a positive right-skewed measurement may use Z = ln(X). Because a monotonic transformation preserves order, the probability of being within corresponding transformed limits is preserved. However, the transformed-scale indices describe spread on that scale and should not be interpreted as ordinary-unit standard deviations.

The transformation must be selected before inspecting results too aggressively, justified by process knowledge or objective diagnostics, and reported so that users can reproduce the analysis. Back-transformation is necessary for communication with engineers and customers.

7.3 Percentile-based and empirical methods

When no reliable parametric distribution is available, quantiles or empirical probabilities can be used. With limited data, tail estimates are unstable, particularly when no observations lie near a specification limit. Confidence intervals should then be emphasized. A capability ratio based only on observed minimum and maximum is not a sound substitute for a statistical model.

8. Autocorrelation, Mixtures and Special Causes

Time-ordered observations may be autocorrelated. Standard deviation estimates and confidence intervals assuming independence can then be misleading. A process with slow drift may appear to have a narrow short-term spread but produce poor long-term performance. Analysts should inspect run charts, consider time-series models where appropriate, and avoid applying capability indices to residuals without explaining what physical output is being assessed.

Mixtures often arise when data combine machines, tools, materials, operators, products, or environmental regimes. The aggregate distribution may be non-normal even though each source is stable. Stratification can reveal that one machine is capable and another is not, or that the combined process is incapable because of between-source differences. If sources are genuinely interchangeable, the aggregate performance may still be relevant; the decision must match the customer experience.

Special causes should not be removed merely because they reduce the index. Removal is defensible only when the cause is verified, its occurrence is not part of normal future operation, and the exclusion rule is documented. Otherwise, deleting observations conceals risk.

9. Applied Case: Precision Shaft Diameter

The following is an illustrative case based on a hypothetical manufacturing operation.

Background

A plant produces shafts specified at 20.000 ± 0.050 mm. Three CNC machines produce the same part. The quality team wants to establish whether the process meets a customer requirement of at least 1.33 for short-term Cpk and at least 1.00 for long-term Ppk.

Problem

An initial pooled analysis reports Cp = 1.58 and Cpk = 1.42. However, customer complaints suggest occasional oversize shafts. The team must determine whether the pooled result reflects real capability.

Analysis

A measurement-system study shows acceptable repeatability and reproducibility relative to the tolerance. Control charts for machine-specific rational subgroups reveal that Machine 2 has a gradual upward shift, while Machines 1 and 3 are stable. The pooled distribution is mildly bimodal. The team therefore stratifies the data by machine and recalculates indices.

Machine 1 has Cpk = 1.48 and Ppk = 1.31. Machine 3 has Cpk = 1.55 and Ppk = 1.38. Machine 2 has Cpk = 1.02 and Ppk = 0.71. The pooled Cp concealed the problem because it summarized a broad specification width relative to a within-machine spread and did not identify the time-dependent shift.

Approach and solution

The plant verifies that a coolant-temperature controller on Machine 2 intermittently fails. It repairs the controller, introduces a start-up verification, and adds a control-chart reaction plan. After collecting a new representative period, the team confirms stability before estimating capability. It reports both within-machine Cpk and overall Ppk, with confidence bounds and the data period clearly stated.

Outcome

Machine 2’s revised Cpk rises to 1.46 and Ppk to 1.29. The customer requirement is met, but the team continues monitoring because the lower confidence bound for Ppk remains close to the threshold. The organization avoids claiming that the entire process was capable merely because the original pooled Cp was high.

Learning points

1. Capability analysis must follow stability analysis.
2. Stratification can reveal hidden mixtures.
3. Cpk and Ppk answer different operational questions.
4. Root-cause correction is preferable to data deletion.
5. Confidence bounds support risk-based decisions.

10. Managerial, Governance and Technology Implications

Managers should specify which index, time horizon, and variation estimate govern acceptance decisions. Supplier contracts and regulatory submissions should not use ambiguous phrases such as “six sigma capable” without defining the calculation convention. Data retention, audit trails, access control, and versioned analysis scripts are important when capability results support product release or safety decisions.

Modern manufacturing execution systems and statistical software can automate data collection, control-chart alerts, distribution fitting, and capability dashboards. Artificial intelligence may assist anomaly detection or select candidate models, but it should not independently redefine specifications, remove outliers, or certify capability. Human review is required for causal interpretation, model validation, and decisions affecting safety or customer conformity. Automated systems should preserve raw data, transformation history, exclusions, and analyst approvals.

11. Best-Practice Decision Framework

A defensible capability study should answer the following questions:

• Is the measurement system fit for purpose?
• Are specifications and target values technically valid?
• Were data collected in time order from the intended process?
• Is the process stable, or have special causes been investigated?
• Are within and overall variation clearly distinguished?
• Is normality plausible, and if not, is an alternative method justified?
• Are subgroups, machines, shifts, and products appropriately stratified?
• Are sample size and confidence intervals adequate for the decision?
• Are exclusions, transformations, and software settings documented?
• Does the action address process causes rather than only the reported index?

Chapter Summary

Stability is the condition of predictable process behaviour; capability is the relationship between a stable process distribution and specification limits; performance describes observed behaviour over a selected period. Cp measures potential spread, Cpk includes centering, Pp uses overall variation, Ppk combines overall variation and centering, and Cpm additionally penalizes departure from a target. These indices require compatible units, meaningful specifications, adequate measurement systems, and appropriate assumptions.

Capability indices are estimates subject to sampling and modelling uncertainty. Yield and sigma interpretations depend on distributional assumptions and clearly stated conventions. Non-normal data require distribution-specific, transformation-based, percentile, or empirical methods rather than automatic use of normal formulas. Stratification, time-order analysis, confidence intervals, and causal investigation are essential for responsible quality decisions.

Key Terms

Stability: Predictable process behaviour governed primarily by common causes.

Capability: Ability of a stable process to meet specification requirements.

Performance: Observed long-term behaviour, usually including between-subgroup variation.

Specification limit: Customer, design, or regulatory boundary for acceptable output.

Control limit: Statistically calculated boundary used to assess process stability.

Cp: Potential capability based on specification width and short-term spread.

Cpk: Short-term capability adjusted for process centering.

Pp: Overall performance based on total observed spread.

Ppk: Overall performance adjusted for centering.

Cpm: Capability index that penalizes deviation from a target.

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

Overall standard deviation: Estimate of variation across the full study period.

Target: Nominal value that the process is intended to achieve.

Yield: Probability or proportion of output within specifications.

DPMO: Defects per million opportunities.

Non-normality: Departure from the distributional form assumed by a method.

Transformation: Mathematical re-expression of a variable for analysis.

Stratification: Separation of data by relevant process source or condition.

Confidence interval: Range expressing uncertainty around an estimated quantity.

Measurement-system analysis: Evaluation of error and adequacy in measurement.

Review Questions

Short Answer Questions

1. Why must stability generally be assessed before capability?
2. What aspect of process behaviour does Cp fail to reflect?
3. How do Pp and Cp differ in their treatment of variation?
4. What additional consideration is included in Cpm?
5. Why is a measurement-system study important before capability analysis?

Long Answer Questions

1. Explain the conceptual and mathematical differences among Cp, Cpk, Pp, and Ppk.
2. Discuss how specification limits, control limits, and target values differ and why confusing them is dangerous.
3. Explain three defensible approaches to non-normal capability analysis.
4. Discuss how confidence intervals improve capability-based management decisions.
5. Develop a complete procedure for conducting a capability study in a multi-machine process.

Analytical / Application Questions

1. A stable process has LSL = 98, USL = 102, μ = 101, and σ = 0.50. Calculate Cp and Cpk, identify the limiting specification, and interpret the result.
2. A process has Cp = 1.60, Cpk = 0.95, Pp = 1.20, and Ppk = 0.78. Diagnose the likely centering and time-dependent variation issues and recommend an investigation plan.
3. A positive, strongly right-skewed measurement has a lower specification limit and no upper limit. Propose a capability-analysis method, state the assumptions that must be checked, and explain how you would communicate the result to non-statistical managers.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes stability from capability?

A. Stability concerns predictability over time; capability concerns compliance with specifications.

B. Stability concerns customer satisfaction; capability concerns measurement resolution.

C. Stability and capability are identical terms.

D. Stability is assessed only after capability is calculated.

Correct Answer: A. Stability concerns predictability over time; capability concerns compliance with specifications.

Question 2

A process has USL = 10, LSL = 4, and σ = 0.50. What is Cp?

A. 0.50

B. 1.00

C. 2.00

D. 4.00

Correct Answer: C. 2.00

Question 3

If Cp is much larger than Cpk, the most likely interpretation is that the process is:

A. Highly autocorrelated but perfectly centered

B. Potentially narrow but off-center

C. Non-measurable

D. Certain to be non-normal

Correct Answer: B. Potentially narrow but off-center

Question 4

Which index uses overall standard deviation?

A. Cp only

B. Cpk only

C. Pp and Ppk

D. Cpm only

Correct Answer: C. Pp and Ppk

Question 5

What does Cpm specifically penalize?

A. Sample size

B. Distance from the target

C. Number of control limits

D. Measurement resolution only

Correct Answer: B. Distance from the target

Question 6

For a stable normal process, the approximate benchmark sigma distance to the nearest specification is:

A. Cpk ÷ 3

B. 3Cpk

C. 6Cpk

D. Cpk²

Correct Answer: B. 3Cpk

Question 7

Why can pooled capability analysis be misleading in a multi-machine process?

A. Pooling always reduces the specification width

B. Pooling can conceal mixtures and machine-specific shifts

C. Pooling eliminates all measurement error

D. Pooling guarantees normality

Correct Answer: B. Pooling can conceal mixtures and machine-specific shifts

Question 8

Which action is most appropriate for a strongly non-normal characteristic?

A. Automatically use Cp without checking assumptions

B. Delete extreme observations

C. Consider a justified distribution-specific method or transformation

D. Replace specifications with control limits

Correct Answer: C. Consider a justified distribution-specific method or transformation

Question 9

For n independent normal observations, which distribution supports a classical confidence interval for variance?

A. Binomial

B. Poisson

C. Chi-square

D. Uniform

Correct Answer: C. Chi-square

Question 10

A DPMO calculation requires the number of defects and the number of:

A. Control limits

B. Opportunities

C. Subgroups only

D. Specification targets only

Correct Answer: B. Opportunities