6. Six Sigma Statistics and Variation Management

Introduction and Context

Six Sigma is fundamentally a management system for making decisions under variation. Processes do not produce identical outcomes: dimensions fluctuate, service times differ, transactions contain occasional errors, and customers interpret quality requirements differently. The central statistical challenge is therefore not merely to calculate an average, but to understand how a process behaves, how much it varies, and how confidently its performance can be judged.

In Lean and Six Sigma practice, statistics provides the language for separating signal from noise. Descriptive statistics summarize observed data; probability distributions model possible outcomes; sampling allows conclusions about a population; confidence intervals express uncertainty; and capability indices compare process behaviour with customer or engineering requirements. Measures such as yield, defects per million opportunities (DPMO), and sigma level translate operational performance into forms that support prioritization and improvement.

A statistic is never meaningful in isolation. A high average may conceal excessive variation, a favourable capability index may depend on an unrealistic normality assumption, and a reported defect rate may be distorted by poor sampling or inconsistent definitions. Effective practitioners therefore combine calculation with judgement about measurement systems, data quality, process stability, and the consequences of errors.

Learning Objectives

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

1. Distinguish populations, samples, parameters, and statistics.

2. Explain common probability distributions and identify situations in which they are useful.

3. Calculate and interpret mean, median, range, variance, standard deviation, and percentiles.

4. Construct and interpret confidence intervals while recognizing sampling uncertainty.

5. Explain common-cause and special-cause variation.

6. Calculate yield, rolled throughput yield, defects per unit, DPMO, and approximate sigma level.

7. Assess process capability and performance using Cp, Cpk, Pp, and Ppk.

8. Identify assumptions, limitations, ethical issues, and common errors in Six Sigma statistical analysis.

9. Use statistical evidence to support managerial and improvement decisions.

The Statistical View of Process Performance

A population is the complete set of units, transactions, or events relevant to a question. The output of every invoice processed by a company during a year could constitute a population. A sample is a subset observed to learn about that population. A population characteristic, such as its true mean or standard deviation, is a parameter; the corresponding value calculated from a sample is a statistic.

The distinction matters because most improvement teams cannot inspect every possible outcome. They estimate population behaviour from samples and must therefore consider sampling error. A sample average is not necessarily the population average, and a small or biased sample may produce a misleading impression of performance.

Six Sigma analysis begins with an operational definition. “Defect” must specify what constitutes nonconformance, when it is measured, and whether multiple defects on one unit are counted separately. A unit is the item or transaction being evaluated; a defect is a failure to meet a requirement; and a defect opportunity is a specified chance for a defect to occur. If 500 forms each contain four critical fields, there are 2,000 opportunities. If 12 field errors are found, DPMO is:

DPMO = (Defects ÷ Opportunities) × 1,000,000

= (12 ÷ 2,000) × 1,000,000 = 6,000.

A unit with two defects remains one defective unit but contributes two defects when the objective is to count defect opportunities. Confusing units, defects, and opportunities is one of the most common sources of invalid reporting.

Probability Distributions Used in Six Sigma

Probability describes the likelihood of possible outcomes. A random variable assigns numerical values to outcomes. Discrete random variables take countable values, such as the number of defects in a batch. Continuous random variables can take values within an interval, such as processing time or shaft diameter.

The normal distribution

The normal distribution is continuous, symmetric, and defined by its mean and standard deviation. Approximately 68.27% of observations lie within one standard deviation of the mean, 95.45% within two, and 99.73% within three, provided the data are normally distributed. The standardized value is:

z = (x − μ) ÷ σ

where x is an observation, μ is the mean, and σ is the standard deviation.

The normal model is useful for many measurements and for sample means under suitable conditions. It is not a universal description of process data. Waiting times, repair costs, contamination counts, and financial losses may be skewed, bounded, or multimodal. Applying normal capability formulas to non-normal data without investigation can create false confidence.

The binomial distribution

The binomial distribution models the number of successes or failures in a fixed number of independent trials when each trial has two outcomes and a constant probability of success. It is useful for questions such as the number of defective applications in a sample of 100, provided the definition of defective is consistent.

Its mean is np and variance is np(1 − p), where n is the number of trials and p is the probability of the event of interest. Independence and a stable probability are important assumptions. If defects cluster because of equipment conditions or operator shifts, a binomial model may underestimate variation.

The Poisson distribution

The Poisson distribution is commonly used for counts of events in a defined unit of time, area, length, or volume, such as customer complaints per week or surface flaws per square metre. Its defining property is that the mean and variance are equal. In practice, overdispersion—variance greater than the mean—may indicate clustering, changing conditions, or an unsuitable model.

The exponential distribution

The exponential distribution models waiting time between events in a Poisson process. It is sometimes used for time between failures or arrivals. Its application requires assumptions about a constant event rate and memorylessness. Many real equipment-failure processes do not satisfy these conditions because failure risk changes with age, maintenance, or operating environment.

Describing Data: Centre, Spread, and Shape

The mean is the arithmetic average and is sensitive to extreme values. The median is the middle observation after ordering the data and is more robust when distributions are skewed. The mode is the most frequent value and may be useful for categorical or discrete data. Choosing among them depends on the decision. Median delivery time may better represent a skewed service process, while the mean may be required for a cost calculation.

The range, maximum minus minimum, is easy to calculate but depends heavily on sample size and extreme observations. The variance is the average squared deviation from the mean. For a sample, it is commonly calculated as:

s² = Σ(xᵢ − x̄)² ÷ (n − 1)

The sample standard deviation is s, the square root of s². The denominator n − 1 provides an unbiased estimate of population variance under conventional assumptions. Standard deviation is expressed in the same units as the original measurement, making it easier to interpret.

The interquartile range (IQR) is Q3 − Q1 and describes the middle 50% of observations. It is relatively resistant to outliers. Percentiles indicate the value below which a specified proportion of observations falls. For example, the 95th percentile of response time is a more informative customer-service measure than the average when customers are particularly affected by long delays.

Shape also matters. A symmetric, bell-shaped distribution differs from a right-skewed distribution even when both have the same mean and standard deviation. Histograms, box plots, run charts, and probability plots should therefore accompany numerical summaries. Averages can conceal mixtures of shifts, products, machines, or customer segments.

Sampling and Confidence in Conclusions

Sampling should reflect the question being asked. A simple random sample gives each population element a known chance of selection. A systematic sample selects every kth item after a random start. Stratified sampling deliberately samples important subgroups, such as plants or customer segments. Rational subgrouping collects observations close together in time or under similar conditions to study within-group and between-group variation.

Convenience sampling is easy but often biased. For example, examining only transactions processed during quiet periods may understate normal workload variation. Survivorship bias occurs when failed or abandoned cases are excluded. Time-dependent processes require attention to autocorrelation: observations close together may not be independent.

A confidence interval expresses a range of plausible values for a population parameter under a specified confidence procedure. For a sample mean with known population standard deviation, a 95% interval is:

x̄ ± z₀.₉₇₅(σ ÷ √n).

When σ is unknown, the t distribution is generally used:

x̄ ± t₀.₉₇₅,n−1(s ÷ √n).

A 95% confidence interval should not be interpreted as a 95% probability that the already calculated parameter lies in the interval. In frequentist terms, the method would capture the true parameter in approximately 95% of repeated samples under the assumptions of the method. Wider intervals result from greater variation, smaller samples, or higher confidence requirements. Increasing sample size reduces the standard error, but it does not correct biased measurement or poor sampling.

For a proportion, an approximate interval is p̂ ± z√[p̂(1−p̂)/n], although exact or improved methods may be preferable for small samples or proportions near zero or one. Statistical significance also differs from practical significance. A tiny improvement may be statistically detectable in a very large sample but operationally unimportant; conversely, a meaningful improvement may fail to reach significance in a small sample.

Variation Management: Stability Before Capability

Variation is the dispersion of process outputs. Common-cause variation arises from the ordinary system: methods, materials, equipment, environment, and management policies. Special-cause variation results from identifiable, unusual influences such as a broken tool, incorrect software configuration, or an inexperienced operator performing an unfamiliar task.

A stable process has predictable variation over time, not necessarily good performance. A process can be stable but consistently outside customer requirements. Conversely, an unstable process may occasionally appear capable because the sample happened to exclude special causes.

Control charts support this distinction by displaying data in time order with a centre line and statistically derived control limits. Control limits describe expected process variation; specification limits describe customer or engineering requirements. These are not interchangeable. A process may have all observations within specifications while still showing a non-random trend, or it may be statistically stable while producing many defects because its centre is poorly positioned.

Typical evidence of special cause includes a point beyond a control limit, a sustained shift, a trend, or a non-random pattern. Rules should be applied cautiously and documented before analysis. Excessive rule switching creates false alarms. When a special cause is identified, the team should investigate and document it rather than simply delete the observation. Removing data can be justified only when there is a defensible measurement or recording error, and the decision must be transparent.

Yield, Defects, and Sigma-Level Communication

First-pass yield (FPY) is the proportion of units that meet requirements without rework. If 940 of 1,000 units pass initially, FPY is 94%.

When several sequential steps are involved, rolled throughput yield (RTY) estimates the probability that a unit passes every step without defect or rework. If three process steps have yields of 0.98, 0.95, and 0.97, then:

RTY = 0.98 × 0.95 × 0.97 = 0.90307, or approximately 90.3%.

RTY exposes hidden factory costs. A final yield may look satisfactory because defects are detected and repaired downstream, whereas RTY reveals the probability of right-first-time performance.

Defects per unit (DPU) is total defects divided by total units. Defects per opportunity (DPO) is total defects divided by total opportunities. DPMO is DPO multiplied by one million. Defective units per million opportunities must not be confused with DPMO when units can contain multiple defects.

A sigma level expresses the distance between process performance and a relevant specification in standard-deviation units. For a centred normal process, a yield of approximately 99.73% corresponds to about three short-term standard deviations within the nearest specification limit. Six Sigma terminology often uses a conventional 1.5-sigma long-term shift, producing the familiar benchmark of 3.4 DPMO. This convention is a communication device, not a physical law. The shift may not apply to every process, and sigma conversion tables can obscure distributional assumptions and the distinction between short-term and long-term data.

Capability and Performance Indices

Capability compares the natural spread of a stable process with specification limits. Let USL and LSL denote upper and lower specification limits, and let μ and σ represent the process mean and standard deviation.

Cp = (USL − LSL) ÷ 6σ

Cp measures potential capability assuming the process is centred. It does not account for the location of the mean. A process can have a high Cp but still produce defects if its mean is close to one specification limit.

Cpk = minimum[(USL − μ) ÷ 3σ, (μ − LSL) ÷ 3σ]

Cpk incorporates centring and therefore cannot exceed Cp for the same data and assumptions. A large difference between Cp and Cpk indicates that centring improvement may be valuable. If specifications are one-sided, a one-sided capability measure should be used rather than inventing an irrelevant opposite limit.

Performance indices use overall, or long-term, variation:

Pp = (USL − LSL) ÷ 6sₒᵥₑᵣₐₗₗ

Ppk = minimum[(USL − μ) ÷ 3sₒᵥₑᵣₐₗₗ, (μ − LSL) ÷ 3sₒᵥₑᵣₐₗₗ].

Cp and Cpk commonly use within-subgroup variation, whereas Pp and Ppk use overall variation. If a process is stable and variation estimates are comparable, the indices may be similar. If Ppk is substantially lower than Cpk, between-subgroup or temporal variation may be important.

Capability analysis requires meaningful specifications, a reliable measurement system, a reasonably stable process, and an appropriate distributional model. A capability index calculated from unstable data is a snapshot, not a dependable forecast. For non-normal data, analysts may transform the data, use a suitable distribution, model percentiles directly, or employ a nonparametric approach. The method should be explained rather than selected merely to obtain a favourable index.

An Applied Case: Reducing Errors in a Claims-Processing Centre

Background

The following is an illustrative case. A health-insurance claims centre wanted to reduce payment delays. Each claim contained five critical data fields, and an error in any field could trigger rework. Management initially reported a “97% quality rate” based on the proportion of claims eventually released without complaint.

Situation and analysis

A Six Sigma team defined a unit as one claim, an opportunity as one critical field, and a defect as an incorrect or missing field detected before payment. During a representative two-week period, 8,000 claims generated 40,000 opportunities and 520 defects.

DPO = 520 ÷ 40,000 = 0.013.

DPMO = 13,000.

The team also found that 7,420 claims passed without rework, giving an FPY of 92.75%. The difference between the reported 97% final quality rate and FPY showed that rework had concealed the cost of poor quality.

A time plot revealed that Monday morning batches had a higher error rate. Stratified analysis showed that the problem was concentrated in claims imported from one external format. A Pareto analysis identified missing diagnosis codes and invalid provider identifiers as the leading defect categories. The measurement definition was reviewed because two auditors had classified the same field differently; a short attribute agreement study was conducted before further decisions were made.

Approach and solution

The centre standardized the import template, added validation rules at data entry, and introduced a controlled exception queue rather than allowing processors to bypass error messages. Supervisors monitored first-pass yield by shift and reviewed a weekly control chart. The team did not remove Monday observations as “unusual”; instead, it treated the pattern as evidence of a systematic input condition.

After implementation, a new sample of 10,000 claims contained 250 field defects across 50,000 opportunities. DPMO fell to 5,000, and 9,650 claims passed without rework, producing an FPY of 96.5%. A confidence interval for the post-change defect proportion would help management judge the precision of this estimate rather than treating the sample rate as an exact population value.

Outcome and learning points

The case illustrates why operational definitions, opportunity counts, stratification, and time order matter. Final yield alone failed to reveal rework. The apparent defect reduction was more credible because the measurement system was clarified and the input source was identified. The improvement also required controls to prevent regression; a one-time before-and-after comparison would not establish continued stability.

Managerial Applications and Professional Judgement

Managers use statistical measures to allocate improvement resources, set realistic targets, evaluate suppliers, and determine whether a process is ready for transfer or scale-up. Capability analysis can support supplier discussions, but specifications must not be tightened merely to create an unfavourable supplier result. Conversely, suppliers should not report only favourable subgroups while concealing poor performance in particular lots or shifts.

In healthcare, finance, aviation, and public services, the cost of a defect may be highly asymmetric. A simple defect count may therefore need to be supplemented by severity, risk, and detectability. A low-frequency safety event can deserve more attention than a common, low-consequence administrative error. Statistical optimization must not override legal, safety, privacy, or professional obligations.

Data governance is also part of variation management. Analysts should preserve traceability, protect personal information, document transformations, and distinguish observed data from imputed or modelled values. Automated dashboards and machine-learning systems can detect shifts, classify defects, and forecast failures, but they may amplify biased data or generate false alarms. Human review, model validation, access control, and change management remain necessary.

Common Errors and Limitations

Several mistakes recur in Six Sigma statistics:

• Treating specification limits as control limits.
• Calculating capability before confirming stability.
• Assuming normality because a formula is convenient.
• Reporting DPMO without defining opportunities.
• Using a non-random sample and generalizing to the whole population.
• Confusing correlation with causation.
• Deleting inconvenient observations without investigation.
• Relying on p-values without considering effect size or operational cost.
• Mixing short-term and long-term standard deviations.
• Comparing indices computed from different definitions, periods, or measurement systems.

Statistical methods do not eliminate uncertainty. They make uncertainty visible and provide disciplined ways to act despite it. The strongest analysis combines numerical evidence with process knowledge, measurement-system assessment, graphical analysis, and confirmation over time.

Current Developments and Best Practice

Digital manufacturing systems, event logs, sensors, and process-mining platforms have increased the volume and frequency of quality data. Real-time control charts and anomaly-detection models can identify changes earlier than periodic manual review. Predictive maintenance combines failure histories with sensor data to estimate changing risk. These tools are valuable when data lineage, sampling frequency, alert thresholds, and escalation responsibilities are clearly defined.

Contemporary practice increasingly emphasizes distribution-free methods, Bayesian updating, hierarchical models for multi-site processes, and risk-based monitoring. These approaches are especially useful when data are sparse, nested, non-normal, or continuously updated. They do not replace fundamentals: a sophisticated model cannot compensate for ambiguous definitions, unstable measurement, or poor sampling.

Good practice is to define the unit and opportunity, verify the measurement system, display data over time, assess stability, select a suitable model, report assumptions, quantify uncertainty, and connect every metric to a decision. Statistical capability should be treated as evidence for action, not as a trophy number.

Chapter Summary

Six Sigma statistics provides a disciplined basis for understanding process performance and variation. Populations and samples must be distinguished, and operational definitions must clarify units, defects, and opportunities. Probability distributions such as the normal, binomial, Poisson, and exponential models are useful only when their assumptions fit the process. Descriptive statistics summarize centre, spread, and shape, while sampling and confidence intervals communicate uncertainty.

Variation management requires separation of common and special causes and recognition that control limits differ from specification limits. Yield, RTY, DPU, DPMO, and sigma level translate quality performance into decision-oriented measures, but sigma conversion conventions must be interpreted carefully. Cp and Cpk describe potential capability and centring using within-process variation; Pp and Ppk describe longer-term performance using overall variation. Capability analysis is meaningful only when the measurement system is adequate, the process is sufficiently stable, and the distributional assumptions are defensible. Ethical reporting, data governance, and professional judgement are essential when statistics influence safety, customers, suppliers, or resources.

Key Terms

Population: The complete set of units or outcomes relevant to a study.

Sample: A subset of a population used to draw conclusions.

Parameter: A numerical characteristic of a population.

Statistic: A numerical characteristic calculated from a sample.

Defect: A failure to meet a specified requirement.

Defect opportunity: A defined chance for a defect to occur on a unit.

DPMO: Defects per million opportunities.

First-pass yield: The proportion of units passing without rework.

Rolled throughput yield: The probability of passing all process steps without rework.

Common-cause variation: Routine variation produced by the process system.

Special-cause variation: Non-routine variation arising from an identifiable influence.

Control limit: A statistically calculated boundary describing expected process behaviour.

Specification limit: A customer or engineering requirement boundary.

Standard deviation: A measure of dispersion expressed in the original data units.

Confidence interval: An interval produced by a method designed to capture a population parameter at a stated long-run rate.

Cp: A capability index comparing specification width with within-process spread.

Cpk: A capability index that also accounts for process centring.

Pp: A performance index using overall process variation.

Ppk: A performance index using overall variation and process centring.

Operational definition: A precise statement of how a variable or defect is measured and classified.

Review Questions

Short Answer Questions

1. Distinguish a parameter from a statistic and give one example of each.

2. Why must defects and defect opportunities be defined separately?

3. State the difference between control limits and specification limits.

4. What does Cpk measure that Cp does not?

5. Why is rolled throughput yield often lower than final yield?

Long Answer Questions

1. Explain how descriptive statistics, graphical analysis, and probability distributions work together in Six Sigma analysis.

2. Discuss the assumptions and practical uses of the normal, binomial, and Poisson distributions.

3. Explain common-cause and special-cause variation and describe how control charts assist in distinguishing them.

4. Compare Cp, Cpk, Pp, and Ppk, including the role of within-subgroup and overall variation.

5. Discuss sampling error, confidence intervals, statistical significance, and practical significance in improvement decisions.

Analytical / Application Questions

1. A process produces 2,400 units containing three defect opportunities each. Inspectors find 36 defects, and 2,290 units pass without rework. Calculate DPMO and first-pass yield. Explain why the two measures differ.

2. A stable process has LSL = 40, USL = 60, mean = 52, and within-process standard deviation = 2. Calculate Cp and Cpk and interpret the difference.

3. Three sequential steps have yields of 0.96, 0.98, and 0.95. Calculate RTY. Recommend one additional piece of information needed before deciding which step to improve first.

22. Multiple Choice Questions with Answers

Question 1

A process produces 500 forms with four defect opportunities per form. If 12 defects are found, what is the DPMO?

A. 600

B. 6,000

C. 12,000

D. 24,000

Correct Answer: B. 6,000

Question 2

Which measure best describes the probability that a unit passes every sequential process step without rework?

A. First-pass yield

B. Cp

C. Rolled throughput yield

D. Ppk

Correct Answer: C. Rolled throughput yield

Question 3

A large difference between Cp and Cpk most directly suggests that the process is:

A. Not centred within its specifications

B. Free from common-cause variation

C. Measured with a biased instrument

D. Necessarily non-normal

Correct Answer: A. Not centred within its specifications

Question 4

Which statement correctly distinguishes control limits from specification limits?

A. Both are set only by customer requirements

B. Control limits describe process behaviour, while specification limits describe requirements

C. Specification limits are calculated from sample standard deviation

D. Control limits and specification limits are interchangeable

Correct Answer: B. Control limits describe process behaviour, while specification limits describe requirements

Question 5

Which distribution is commonly appropriate for counting events in a fixed time or area when events occur at a stable average rate?

A. Normal

B. Binomial

C. Poisson

D. Exponential

Correct Answer: C. Poisson

Question 6

If a process is unstable, a capability index calculated from a short sample should be interpreted as:

A. A guaranteed long-term prediction

B. A complete substitute for a control chart

C. A limited snapshot with questionable predictive value

D. Proof that specifications are incorrect

Correct Answer: C. A limited snapshot with questionable predictive value

Question 7

What is the principal effect of increasing sample size, all else equal, on a confidence interval for a mean?

A. It increases sampling error

B. It narrows the interval

C. It removes measurement bias

D. It guarantees statistical significance

Correct Answer: B. It narrows the interval

Question 8

A process has step yields of 0.98, 0.95, and 0.97. Its rolled throughput yield is closest to:

A. 90.3%

B. 92.0%

C. 95.0%

D. 97.0%

Correct Answer: A. 90.3%

Question 9

Which is an example of special-cause variation?

A. Routine variation in material thickness across stable production

B. Normal differences among trained operators

C. A broken sensor causing an abrupt shift in readings

D. The long-term average variation of the process

Correct Answer: C. A broken sensor causing an abrupt shift in readings

Question 10

Why can final yield conceal poor process performance?

A. Final yield always excludes customer requirements

B. Rework can remove evidence of defects that occurred earlier

C. Final yield is calculated only from standard deviation

D. Final yield counts opportunities instead of units

Correct Answer: B. Rework can remove evidence of defects that occurred earlier