Process Performance and Capability Analysis

Introduction

Six Sigma decisions depend on more than the average level of a process. A process may produce an acceptable average while having excessive variation, may have excellent short-term capability but poor long-term performance, or may meet customer specifications while operating inefficiently. Process performance and capability analysis provides a disciplined way to evaluate these conditions.

The central distinction is between stability, performance, capability, and specification compliance. Stability concerns whether a process behaves predictably over time. Performance describes what the process has actually delivered, usually including all sources of long-term variation. Capability estimates what a stable process can deliver relative to specification limits. Compliance is a judgment about whether individual units or lots satisfy stated requirements. These ideas are related but are not interchangeable.

Capability analysis should therefore be conducted only after the measurement system has been shown to be adequate and the process has been examined for statistical control. Otherwise, a capability index can conceal assignable causes, changes in product mix, measurement error, or inappropriate distributional assumptions.

Learning objectives

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

1. Distinguish stability, yield, throughput, performance, capability, and specification compliance.
2. Select appropriate short-term and long-term variation estimates.
3. Calculate and interpret C_p, C_pk, P_p, P_pk, and Z-bench.
4. Explain the relationship between capability indices, defect rates, and process centering.
5. Evaluate non-normal and attribute-data processes using suitable alternatives.
6. Apply capability analysis to operational and managerial decisions while recognizing its limitations.

1. The logic of process evaluation

A process converts inputs into outputs. Let Y denote a quality characteristic, such as diameter, delivery time, fill volume, or cycle time. A specification is normally expressed by a lower specification limit (LSL), an upper specification limit (USL), or both. The customer requirement is satisfied when the output lies within the applicable limits.

Four questions should be asked in sequence:

1. Is the measurement system adequate? If observed variation is largely measurement error, process indices are misleading.
2. Is the process stable? A control chart or equivalent time-ordered analysis is used to identify special causes.
3. What is the observed performance? This includes the actual spread and centering over a defined period.
4. What is the capability under controlled conditions? This estimates performance relative to specifications after exceptional causes have been removed or separately understood.

A process can be stable but incapable, capable but not stable, or compliant during a sample period while having poor future capability. Stability is a prerequisite for meaningful prediction; capability is not a synonym for current compliance.

Important terminology

• Common-cause variation: Natural variation arising from the routine process system.
• Special-cause variation: Variation attributable to a specific, non-routine source.
• Short-term variation: Within-subgroup variation, intended to represent routine variation over a relatively narrow time interval.
• Long-term variation: Overall variation across time, shifts, operators, materials, and environmental conditions.
• Specification limits: Customer, design, regulatory, or contractual boundaries.
• Control limits: Statistically derived limits used to assess process stability; they are not specification limits.
• Yield: The proportion of units satisfying a defined acceptance criterion.
• Throughput: The rate at which acceptable or total units are completed, depending on the stated definition.
• Capability: The potential or estimated ability of a stable process to meet specifications.
• Performance: Actual observed conformance over a specified period.

2. Stability, yield, and throughput

Stability

A stable process has a consistent probability distribution over the period of interest. In practice, stability is assessed using time-ordered data. For continuous data, individuals–moving range, X̄–R, or X̄–S charts may be appropriate. For attribute data, p, np, c, or u charts may be used according to the sampling structure.

Control limits are calculated from process data and estimate expected common-cause behavior. They should not be confused with engineering tolerances. A process may be statistically in control but have every observation outside the customer specification; conversely, it may meet specifications while showing special-cause signals.

Yield

For a binary acceptance rule, observed yield is:

Yield is empirical and depends on the sample, inspection rule, and definition of conformity. First-pass yield (FPY) counts units that meet requirements without rework or repair. Rolled throughput yield (RTY) measures the probability that a unit passes every step in a sequence without rework. If stage yields are independent, an approximate RTY is:

where Y_i is the first-pass yield at stage i and k is the number of stages. The independence assumption may be unrealistic when the same upstream cause affects several stages; in that case, direct unit-level tracking is preferable.

For example, five sequential stages have first-pass yields of 0.98, 0.97, 0.99, 0.96, and 0.98. Under the independence approximation:

Thus, approximately 88.5% of units pass all stages without rework, even though each individual stage appears highly effective.

Throughput

Throughput is a rate, not a proportion. A simple completed-unit rate is:

Units must be stated, such as units per hour. Effective throughput may be defined as conforming units per hour:

This distinction matters because high output with substantial scrap may produce lower effective throughput than a slower but more capable process. Little’s Law also links work-in-process, throughput, and flow time under relatively stable conditions:

Here, WIP is average work-in-process, throughput is average completion rate, and flow time is average time in the system. The time units must be consistent. Little’s Law is a flow relationship, not a capability equation; it does not by itself diagnose quality variation.

3. Capability indices for approximately normal continuous data

Assume that a quality characteristic is approximately normally distributed, observations are representative, and the process is stable. Let μ be the process mean and σ the relevant standard deviation.

Potential capability: Cp

The potential capability index compares the specification width with six standard deviations:

Here, σ_within estimates short-term, within-subgroup variation. The factor six represents the interval from μ − 3σ to μ + 3σ for a normal process. C_p does not account for process centering. A high value can therefore coexist with substantial nonconformance if the mean is displaced.

Centered capability: Cpk

The capability index that incorporates centering is:

The two terms represent the upper and lower capability margins. If the process is exactly centered between symmetric limits, C_pk = C_p. Otherwise, C_pk < C_p. A negative value indicates that the estimated mean lies outside one specification limit.

The process centering location can be described by:

For symmetric two-sided specifications and the same standard deviation estimate, C_pk = C_p(1 − k). This relationship illustrates why centering improvement may increase C_pk without reducing variation.

Performance indices: Pp and Ppk

Performance indices use overall standard deviation rather than within-subgroup standard deviation:

σ_overall includes variation across the complete observation period. Consequently, P_pk is usually less than or equal to C_pk when overall variation exceeds within-subgroup variation. The comparison is diagnostically useful:

• C_pk close to P_pk: relatively little additional between-subgroup or time-based variation.
• C_pk much greater than P_pk: drift, shift-to-shift differences, autocorrelation, changing inputs, or subgrouping issues may exist.

These indices are not probabilities and should not be interpreted without considering sample size, stability, distribution shape, and the specification context.

Worked example: centered and off-center processes

A shaft diameter has LSL = 9.80 mm and USL = 10.20 mm. The estimated mean is 10.08 mm, within-subgroup standard deviation is 0.04 mm, and overall standard deviation is 0.05 mm.

Potential short-term capability is:

The upper and lower short-term margins are:

Therefore, C_pk = 1.00. The spread is potentially adequate, but centering toward the upper limit makes the upper side critical.

For long-term performance:

The gap between C_pk = 1.00 and P_pk = 0.80 indicates that time-related variation deserves investigation. Treating this process as simply “capable” would be an incomplete conclusion.

4. Z-bench and defect probability

Z-bench expresses the distance from the process distribution to the nearest specification boundary in standard-deviation units. For a two-sided normal process using overall variation:

For two-sided data, the relationship is:

when P_pk uses the same overall standard deviation and the nearest specification side governs. Similarly, a short-term Z value based on within variation is 3C_pk.

A Z value is not automatically a defect rate. Converting it to a probability requires a distributional model. If Z is the lower-tail distance for a standard normal distribution, the lower-tail nonconformance probability is Φ(−Z), where Φ is the standard normal cumulative distribution function. For a centered two-sided process, the total predicted nonconformance is approximately 2Φ(−Z). For an off-center process, the two tail probabilities must be calculated separately and added.

The commonly discussed “1.5 sigma shift” is a convention used in some Six Sigma benchmarking practices, not a statistical law. It should not be silently applied to a measured Z-bench value. Analysts must state whether a reported sigma level is short-term, long-term, shifted, or unshifted.

5. One-sided specifications and target performance

Some characteristics have only one meaningful limit. For an upper limit:

and, using overall variation:

For a lower limit, use C_pl or P_pl with the corresponding lower margin. A one-sided index should not be manufactured into a two-sided index by inventing an irrelevant specification limit.

When a nominal target matters, capability indices alone may be insufficient. A process can meet both limits but consistently miss the desired target. Taguchi-style loss thinking emphasizes that quality loss may increase as output moves away from target, even before a specification boundary is crossed. A target-centered analysis should therefore include the mean-to-target distance and the economic consequences of deviation.

6. Non-normal data

Normal capability indices can be misleading when data are skewed, bounded, multimodal, heavy-tailed, or generated by a mixture of populations. Common examples include cycle time, waiting time, chemical concentration, reliability life, and contamination counts.

The first response should not be automatic transformation. The analyst should investigate whether non-normality arises from:

• a genuine physical distribution;
• mixing shifts, products, machines, or suppliers;
• censoring or detection limits;
• rounding or measurement resolution;
• special causes; or
• an inappropriate sampling plan.

Approaches include:

1. Distribution-based analysis: Fit a defensible distribution, such as lognormal, gamma, Weibull, or another model, and estimate the probability outside specifications. Goodness-of-fit, residual behavior, parameter uncertainty, and physical plausibility must be checked.
2. Transformation: Apply a justified transformation, such as logarithms for multiplicative variation, then interpret results on the original scale. Transformation does not correct instability or mixture structure.
3. Percentile or quantile methods: Estimate the fraction below LSL or above USL directly, often using empirical or bootstrap confidence intervals.
4. Separate populations: Analyze machines, products, shifts, or operating regimes separately when pooling hides meaningful differences.
5. Nonparametric methods: Use distribution-free estimates when sample size and data structure support them, while acknowledging that tail estimates may be imprecise.

For non-normal processes, reporting a conventional C_pk without qualification is poor practice. A stronger report states the model, parameter-estimation method, specification-tail probabilities, confidence intervals, and sensitivity to alternative models.

7. Attribute data and discrete outcomes

Attribute data classify units as conforming/nonconforming or count events. Capability indices based on a continuous normal model are inappropriate unless a defensible continuous latent measurement exists.

For binary data, the observed nonconforming fraction is:

where d is the number of nonconforming units and n is the number inspected. Yield is 1 − p̂. Confidence intervals should be selected for binomial data; a simple normal approximation can be unreliable for small samples or probabilities near zero or one.

For multiple opportunities per unit, define:

and:

A defect is not necessarily a defective unit. One unit may contain several defects, while acceptance may classify it as a single nonconforming unit. The opportunity definition must be explicit and governed to prevent denominator manipulation.

For counts over a constant area, length, time, or unit, a Poisson model may be considered. The Poisson assumption requires that events occur independently with a constant average rate and that multiple events in an infinitesimal interval are negligible. Overdispersion, clustering, and changing opportunity exposure require alternatives such as negative binomial models or stratified analysis.

For attribute processes, useful tools include p, np, c, and u control charts; binomial or Poisson confidence intervals; defect concentration analysis; and reliability or survival models where time-to-failure is recorded. “Six Sigma capability” should be communicated through clearly defined defect probabilities rather than by forcing C_pk onto categorical data.

8. Confidence, sample size, and uncertainty

A capability estimate is not an exact property of the process. It is calculated from finite data. Small samples can produce unstable estimates of the mean and standard deviation, especially in the tails. Confidence intervals or bootstrap intervals should accompany important decisions.

A capability study should document:

• the observation period;
• sampling frequency and subgrouping logic;
• measurement-system results;
• exclusions and treatment of missing data;
• stability evidence;
• distributional assumptions;
• specification source and revision status;
• index type and standard-deviation estimator; and
• uncertainty and decision threshold.

A large C_pk based on a short, favorable window may be less credible than a moderately lower index based on representative operating conditions. Data integrity is therefore part of capability analysis, not an administrative afterthought.

9. Illustrative case: pharmaceutical filling process

Background. A packaging line fills bottles with a labeled target of 500 mL. The regulatory and customer specifications are LSL = 495 mL and USL = 505 mL. The line operates across three shifts and uses two filling heads.

Problem. The reported first-pass yield is 99.2%, but customer complaints suggest occasional underfilled bottles. Management proposes declaring the process capable based on the high yield.

Analysis. A measurement-system study finds that the fill-volume gauge has acceptable repeatability and reproducibility. Time-ordered data show no major special-cause signals, but subgroup means differ by filling head. The combined mean is 500.8 mL, within-subgroup standard deviation is 1.0 mL, and overall standard deviation is 1.45 mL.

The short-term indices are:

The long-term indices are:

Thus, the line appears potentially capable within short operating intervals, but its actual long-term performance is marginal. The difference indicates between-head or between-shift variation. The 99.2% yield also does not contradict the low P_pk: yield is an observed sample result, whereas performance indices estimate distributional behavior and may reveal future risk.

Approach and solution. The team stratifies data by filling head and shift, adjusts head calibration, introduces a start-of-shift verification, and uses a head-specific control chart. It also defines RTY separately from final inspected yield so that reworked bottles are not counted as first-pass successes.

Outcome. After improvement, overall standard deviation falls to 1.05 mL and the mean moves to 500.2 mL. The updated long-term performance is approximately:

The improvement is not merely a higher index. It is a reduction in operational risk supported by a more stable process, better centering, and transparent yield definitions.

Learning points. Yield alone cannot establish capability. Subgroup structure can conceal systematic differences. Separating short-term capability from long-term performance identifies whether the priority is variation reduction, centering, or control of changing conditions.

10. Managerial, ethical, and technological implications

Capability analysis influences release decisions, supplier qualification, maintenance intervals, staffing, and investment. Managers should avoid imposing a universal index threshold without considering safety, cost of failure, measurement uncertainty, and the criticality of the characteristic. A C_pk threshold suitable for a cosmetic dimension may be unacceptable for a safety-critical parameter.

Ethically, analysts must not remove inconvenient observations merely because they are outside specifications. Exclusions require documented, process-based reasons, such as a verified instrument failure or a separately identifiable special cause, and should be accompanied by sensitivity analysis. Specification limits must originate from authorized engineering, customer, or regulatory requirements; they should not be widened to make a process appear capable.

Digital manufacturing systems and AI can support automated data collection, anomaly detection, hierarchical modelling, and real-time capability dashboards. However, AI does not remove the need for stable definitions, representative sampling, measurement validation, and human review. Automated systems may amplify bias when data omit downtime, rework, rejected lots, or difficult shifts. Governance should include version-controlled specifications, audit trails, access controls, model validation, and monitoring for drift.

Best-practice checklist

1. Validate the measurement system before estimating capability.
2. Define the characteristic, unit, opportunity, and specification source.
3. Establish statistical stability using time-ordered data.
4. Choose subgrouping to reflect the physical process, not to manufacture a favorable result.
5. Report both short-term capability and long-term performance when relevant.
6. Inspect centering separately from spread.
7. Use distribution-appropriate methods for skewed, bounded, mixed, or attribute data.
8. Report sample size, assumptions, uncertainty, and observation period.
9. Link capability results to yield, throughput, risk, and economic consequences.
10. Reassess capability after changes to equipment, materials, software, staffing, or specifications.

Chapter Summary

Process performance and capability analysis evaluates whether a process is stable, productive, conforming, and predictable. Stability is assessed through time-ordered evidence and control charts; it is distinct from specification compliance. Yield measures the proportion of conforming output, while throughput measures output rate. Rolled throughput yield shows how losses accumulate across sequential operations.

For approximately normal continuous data, C_p measures potential short-term capability, C_pk incorporates centering, P_p measures long-term spread, and P_pk incorporates long-term centering. Z-bench expresses the nearest specification distance in standard-deviation units, but conversion to defect probability requires explicit distributional assumptions. Non-normal and attribute data require distribution-based, quantile, binomial, Poisson, reliability, or other suitable methods. Capability indices are estimates, not guarantees, and must be supported by measurement-system adequacy, representative sampling, stability analysis, and transparent governance.

Key Terms

Stability: Predictable behavior over time under defined operating conditions.

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

Special-cause variation: Non-routine variation attributable to a specific source.

Specification limit: An authorized boundary defining acceptable output.

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

Yield: The proportion of output meeting acceptance requirements.

First-pass yield: The proportion conforming without rework or repair.

Rolled throughput yield: The proportion passing every process stage without rework.

Throughput: Output completed per unit of operating time.

Capability: The estimated ability of a stable process to meet specifications.

Performance: Actual observed process behavior over a defined period.

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

Cpk: Short-term capability adjusted for process centering.

Pp: Long-term performance based on overall variation.

Ppk: Long-term performance adjusted for centering.

Z-bench: Standard-deviation distance from the process to the nearest specification boundary.

DPO: Defects per opportunity.

DPMO: Defects per million opportunities.

Non-normal data: Data whose distribution does not adequately follow a normal model.

Measurement-system analysis: Evaluation of measurement bias, repeatability, reproducibility, and adequacy.

Review Questions

Short Answer Questions

1. Why must process stability be considered before capability is estimated?
2. Distinguish control limits from specification limits.
3. What is the main difference between C_pk and P_pk?
4. Define rolled throughput yield.
5. Why is C_pk generally inappropriate for purely attribute data?

Long Answer Questions

1. Explain how variation, centering, stability, and specification compliance interact in capability analysis.
2. Derive and interpret C_p, C_pk, P_p, and P_pk, including their assumptions.
3. Discuss how a Six Sigma team should analyze a skewed cycle-time distribution.
4. Explain why observed yield and estimated capability may lead to different managerial conclusions.
5. Discuss ethical risks associated with specification changes, data exclusions, and selective sampling.

Analytical / Application Questions

1. A process has LSL = 98, USL = 102, mean = 100.5, and within-subgroup standard deviation = 0.5. Calculate C_p and C_pk, and identify the limiting specification side.
2. Three sequential operations have first-pass yields of 0.95, 0.98, and 0.97. Calculate the approximate RTY and explain its operational meaning.
3. A stable process has C_pk = 1.20 and P_pk = 0.80. Propose two technically plausible causes of the gap and design a follow-up investigation.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes capability from specification compliance?

A. Capability concerns only the percentage of units currently inspected.

B. Capability estimates process ability relative to specifications, while compliance identifies whether observed units meet requirements.

C. Capability and compliance are identical whenever a control chart is used.

D. Compliance measures variation, whereas capability measures production rate.

Correct Answer: B. Capability estimates process ability relative to specifications, while compliance identifies whether observed units meet requirements.

Question 2

A process has USL = 110, LSL = 90, and within-subgroup standard deviation = 2. What is C_p?

A. 0.83

B. 1.00

C. 1.67

D. 3.33

Correct Answer: C. 1.67

Question 3

If C_p is high but C_pk is much lower, the most likely interpretation is that the process is:

A. Well centered with excessive measurement resolution

B. Off center relative to at least one specification limit

C. Attribute data with no specifications

D. Guaranteed to be unstable

Correct Answer: B. Off center relative to at least one specification limit

Question 4

Which standard deviation is normally used in P_pk?

A. Overall, long-term standard deviation

B. Only the measurement-system standard deviation

C. The specification width divided by two

D. A fixed value of one

Correct Answer: A. Overall, long-term standard deviation

Question 5

A process has first-pass yields of 0.98 and 0.95 at two sequential stages. Assuming independence, the approximate RTY is:

A. 0.030

B. 0.465

C. 0.931

D. 1.930

Correct Answer: C. 0.931

Question 6

For a two-sided normal process using overall variation, Z-bench is related to P_pk by:

A. Z_bench = P_pk ÷ 3

B. Z_bench = 3P_pk

C. Z_bench = 6P_pk

D. Z_bench = P_pk²

Correct Answer: B. Z_bench = 3P_pk

Question 7

Which approach is most appropriate for a strongly skewed, stable cycle-time process?

A. Automatically report normal C_pk without qualification

B. Ignore the upper specification limit

C. Use a justified distributional or quantile-based analysis and report assumptions

D. Replace every value above the mean with the mean

Correct Answer: C. Use a justified distributional or quantile-based analysis and report assumptions

Question 8

DPMO is calculated as:

A. Defects ÷ units

B. Units ÷ opportunities

C. DPO × 1,000,000

D. Yield × 100

Correct Answer: C. DPO × 1,000,000

Question 9

Which condition is required for a conventional normal capability analysis to be credible?

A. The process must have no specification limits

B. The process should be reasonably stable and the distributional assumption defensible

C. The sample must contain only conforming units

D. The mean must equal zero

Correct Answer: B. The process should be reasonably stable and the distributional assumption defensible

Question 10

A large difference between C_pk and P_pk most directly suggests:

A. Additional between-subgroup or time-related variation

B. That specifications are necessarily incorrect

C. That yield is exactly 50%

D. That the measurement system is automatically perfect

Correct Answer: A. Additional between-subgroup or time-related variation